CHAPMAN & HALL/CRC COMPUTER and INFORMATION SCIENCE SERIES

Handbook of
Bioinspired Algorithms
and Applications

© 2006 by Taylor & Francis Group, LLC

 CHAPMAN & HALL/CRC
COMPUTER and INFORMATION SCIENCE SERIES
Series Editor: Sartaj Sahni

PUBLISHED TITLES
HANDBOOK OF SCHEDULING: ALGORITHMS, MODELS, AND PERFORMANCE ANALYSIS
Joseph Y.-T. Leung
THE PRACTICAL HANDBOOK OF INTERNET COMPUTING
Munindar P. Singh
HANDBOOK OF DATA STRUCTURES AND APPLICATIONS
Dinesh P. Mehta and Sartaj Sahni
DISTRIBUTED SENSOR NETWORKS
S. Sitharama Iyengar and Richard R. Brooks
SPECULATIVE EXECUTION IN HIGH PERFORMANCE COMPUTER ARCHITECTURES
David Kaeli and Pen-Chung Yew
SCALABLE AND SECURE INTERNET SERVICES AND ARCHITECTURE
Cheng-Zhong Xu

HANDBOOK OF BIOINSPIRED ALGORITHMS AND APPLICATIONS

Stephan Olariu and Albert Y. Zomaya

© 2006 by Taylor & Francis Group, LLC

 CHAPMAN & HALL/CRC COMPUTER and INFORMATION SCIENCE SERIES

Handbook of
Bioinspired Algorithms
and Applications

Edited by
Stephan Olariu
Old Dominion University
Norfolk, Virginia, U.S.A.

Albert Y. Zomaya
University of Sydney
NSW, Australia

Boca Raton London New York

© 2006 by Taylor & Francis Group, LLC

C4754_Discl.fm Page 1 Tuesday, August 2, 2005 1:20 PM

Published in 2006 by
Chapman & Hall/CRC
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2006 by Taylor & Francis Group, LLC

Chapman & Hall/CRC is an imprint of Taylor & Francis Group

No claim to original U.S. Government works

Printed in the United States of America on acid-free paper
10 9 8 7 6 5 4 3 2 1

International Standard Book Number-10: 1-58488-475-4 (Hardcover)

International Standard Book Number-13: 978-1-58488-475-0 (Hardcover)

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with
permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish
reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials
or for the consequences of their use.

No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or
other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive, Danvers, MA
01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For
organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Catalog record is available from the Library of Congress

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
Taylor & Francis Group and the CRC Press Web site at
is the Academic Division of T&F Informa plc. http://www.crcpress.com

© 2006 by Taylor & Francis Group, LLC

 Preface

The Handbook of Bioinspired Algorithms and Applications seeks to provide an opportunity for researchers
to explore the connection between biologically inspired (or bioinspired) techniques and the development
of solutions to problems that arise in a variety of problem domains. The power of bioinspired paradigms
lies in their capability in dealing with complex problems with little or no knowledge about the search space,
and thus is particularly well suited to deal with a wide range of computationally intractable optimizations
and decision-making applications.
Vast literature exists on bioinspired approaches for solving an impressive array of problems and there
is a great need to develop repositories of “how to apply” bioinspired paradigms to difficult problems. The
material of the handbook is by no means exhaustive and it focuses on paradigms that are “bioinspired,”
and therefore, chapters on fuzzy logic or simulated annealing were not included in the organization. There
was a decision to limit the number of chapters so that the handbook remains manageable within a single
volume.
The handbook endeavors to strike a balance between theoretical and practical coverage of a range of
bioinspired paradigms and applications. The handbook is organized into two main sections: Models and
Paradigms and Application Domains, and the titles of the various chapters are self-explanatory and a
good indication to what is covered. The theoretical chapters are intended to provide the fundamentals of
each of the paradigms in such a way that allows the readers to utilize these techniques in their own fields.
The application chapters show detailed examples and case studies of how to actually develop a solution
to a problem based on a bioinspired technique. The handbook should serve as a repository of significant
reference material, as the list of references that each chapter provides will become a useful source of further
study.
Stephan Olariu
Albert Y. Zomaya

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page v — #5

 Acknowledgments

First and foremost we would like to thank and acknowledge the contributors of this book for their support
and patience, and the reviewers for their useful comments and suggestions that helped in improving
the earlier outline of the handbook and presentation of the material. Professor Zomaya would like to
acknowledge the support from CISCO Systems and members of the Advanced Networks Research Group
at Sydney University. We also extend our deepest thanks to Jessica Vakili and Bob Stern from CRC Press
for their collaboration, guidance, and, most importantly, patience in finalizing this handbook. Finally,
we thank Mr. Mohan Kumar for leading the production process of this handbook in a very professional
manner.

Stephan Olariu
Albert Y. Zomaya

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page vii — #7

 Editors

Stephan Olariu received his M.Sc. and Ph.D. degrees in computer science from McGill University,
Montreal, in 1983 and 1986, respectively. In 1986 he joined the Old Dominion University where he is a
professor of computer science. Dr. Olariu has published extensively in various journals, book chapters,
and conference proceedings. His research interests include image processing and machine vision, parallel
architectures, design and analysis of parallel algorithms, computational graph theory, computational geo-
metry, and mobile computing. Dr. Olariu serves on the Editorial Board of IEEE Transactions on Parallel
and Distributed Systems, Journal of Parallel and Distributed Computing, VLSI Design, Parallel Algorithms
and Applications, International Journal of Computer Mathematics, and International Journal of Foundations
of Computer Science.
Albert Y. Zomaya is currently the CISCO Systems chair professor of internetworking in the School of
Information Technologies, The University of Sydney. Prior to that he was a full professor in the Electrical
and Electronic Engineering Department at the University of Western Australia, where he also led the
Parallel Computing Research Laboratory from 1990 to 2002. He served as associate, deputy, and acting
head in the same department, and held visiting positions at Waterloo University and the University of
Missouri–Rolla. He is the author/co-author of 6 books and 200 publications in technical journals and
conferences, and the editor of 6 books and 7 conference volumes. He is currently an associate editor
for 14 journals, the founding editor of the Wiley Book Series on Parallel and Distributed Computing, and
the editor-in-chief of the Parallel and Distributed Computing Handbook (McGraw-Hill 1996). Professor
Zomaya was the chair of the IEEE Technical Committee on Parallel Processing (1999–2003) and currently
serves on its executive committee. He has been actively involved in the organization of national and
international conferences. He received the 1997 Edgeworth David Medal from the Royal Society of New
South Wales for outstanding contributions to Australian science. In September 2000 he was awarded the
IEEE Computer Society’s Meritorious Service Award. Professor Zomaya is a chartered engineer (CEng), a
fellow of the IEEE, a fellow of the Institution of Electrical Engineers (U.K.), and member of the ACM. He also
serves on the boards of two startup companies. His research interests are in the areas of high performance
computing, parallel algorithms, networking, mobile computing, and bioinformatics.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page ix — #9

 Contributors

Enrique Alba Jürgen Branke Sajal K. Das

Department of Languages and Institute AIFB Center for Research in Wireless
Computer Science University of Karlsruhe Mobility and Networking
University of Málaga Karlsruhe, Germany Department of Computer
Campus de Teatinos Science & Engineering
Málaga, Spain Forbes Burkowski The University of Texas at
School of Computer Science Arlington
Abdullah Almojel University of Waterloo Arlington, Texas
Ministry of Higher Education Waterloo, Ontario, Canada
Riyadh, Saudi Arabia Tiago Ferra de Sousa
Escola Superior de Tecnologia
S. Cahon
Sanghamitra Bandyopadhyay Instituto Politecnico de Castelo
Laboratoire d’Informatique
Machine Intelligence Unit Branco
Fondamentale de Lille
Indian Statistical Institute Castelo Branco, Portugal
Lille, France
Kolkata, India
Francisco Fernández de Vega
J. Francisco Chicano Grupo de Evolución Artificial
Nilanjan Banerjee
Department of Languages and Centro Universitario de Mérida
Center for Research in Wireless
Computer Science Universidad de Extremadura
Mobility and Networking
University of Málaga Mérida, Spain
Department of Computer
Málaga, Spain
Science & Engineering
The University of Texas at C. Dhaenens
Arlington Ernesto Costa Laboratoire d’Informatique
Arlington, Texas Evolutionary and Complex Fondamentale de Lille
Systems Group Lille, France
Mohamed Belal Centro de Informática e Sistemas
Helwan University da Universidade de Coimbra Bernabe Dorronsoro
Cairo, Egypt Pinhal de Marrocos Central Computing Services
Coimbra, Portugal University of Málaga
Utpal Biswas Campus de Teatinos
Department of Computer Science Carlos Cotta Málaga, Spain
and Engineering Department of Languages and
University of Kalyani Computer Science Hoda El-Sayed
Kalyani, India University of Málaga Bowie State University
Campus de Teatinos Bowie, Maryland
Azzedine Boukerche Málaga, Spain
SITE Mohamed Eltoweissy
University of Ottawa Kris Crnomarkovic Virginia Tech
Ottawa, Canada Advanced Networks Research Falls Church, Virginia
Group
Anthony Brabazon School of Information Muddassar Farooq
Faculty of Commerce Technologies Informatik III
University College Dublin The University of Sydney University of Dortmund
Dublin, Ireland Sydney, Australia Dortmund, Germany

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xi — #11

xii Contributors

Marcos Fernández Kathia Regina Lemos Jucá Gabriel Luque

Instituto de Robótica Federal University of Santa Department of Languages and
Universidad de Valencia Catarina Computer Science
Polígono de la Coma Florianopolis, Brazil ETS Ingeniería
Paterna (Valencia), Spain Informática
M. Khabzaoui University of Málaga
Gianluigi Folino Laboratoire d’Informatique Málaga, Spain
Institute of High Performance Fondamentale de Lille
Computing and Networks Lille, France Ujjwal Maulik
Rende (CS), Italy
Department of Computer Science
Bithika Khargaria and Engineering
Agostino Forestiero High Performance Distributed Jadavpur University
Institute of High Performance Computing Laboratory Kolkata, India
Computing and Networks
The University of Arizona
Rende (CS), Italy
Tuscon, Arizona
N. Melab
Jafaar Gaber Laboratoire d’Informatique
Peter Korošec Fondamentale de Lille
UTMB
Computer Systems Department Lille, France
France
Jožef Stefan Institute
Mario Giacobini Ljubljana, Slovenia
M. Mezmaz
Information Systems Department Laboratoire d’Informatique
University of Lausanne Barbara Koroušić-Seljak
Fondamentale de Lille
Lausanne, Switzerland Computer Systems Department
Lille, France
Jožef Stefan Institute
Michael Guntsch Ljubljana, Slovenia
Institute AIFB Michelle Moore
University of Karlsruhe Zhen Li Department of Computing and
Karlsruhe, Germany The Applied Software Systems Mathematical Sciences
Laboratory Texas A&M University-Corpus
Christi
Salim Hariri Rutgers, The State University of
High Performance Distributed New Jersey Corpus Christi, Texas
Computing Laboratory Camden, New Jersey
The University of Arizona Pedro Morillo
Tuscon, Arizona Kenneth N. Lodding Instituto de Robótica
NASA Langley Research Center Universidad de Valencia
Piotr Jedrzejowicz Hampton, Virginia Polígono de la Coma
Department of Information Paterna (Valencia), Spain
Systems Mi Lu
Faculty of Business
Department of Electrical Anirban Mukhopadhyay
Administration
Engineering Department of Computer Science
Gdynid Maritime University
Texas A&M University and Engineering
Gdynia, Poland
College Station, Texas University of Kalyani
Kalyani, India
Kennie H. Jones
NASA Langley Research Center Francisco Luna
Hampton, Virginia Department of Languages and Mrinal Kanti Naskar
Computer Science Department of Electronics and
L. Jourdan ETS Ingeniería Telecommunication
Laboratoire d’Informatique Informática Engineering
Fondamentale de Lille University of Málaga Jadavpur University
Lille, France Málaga, Spain Kolkata, India

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xii — #12

Contributors xiii

Antonio J. Nebro Manish Parashar Arlindo Silva

Department of Languages and The Applied Software Systems Escola Superior de Tecnologia
Computer Science Laboratory Instituto Politécnico de Castelo
ETS Ingeniería Rutgers, The State University of Branco
Informática New Jersey Castelo Branco, Portugal
University of Málaga Camden, New Jersey and
Málaga, Spain Centro de Informatica e Sistemas
da Universidade de Coimbra
Zhiquan Frank Qiu Pinhal de Marrocos, Portugal
Ana Neves
Intel Corporation
Escola Superior de Tecnologia
Chandler, Arizona
Instituto Politécnico de Castelo João Bosco Mangueira Sobral
Branco Federal University of Santa
Castelo Branco, Portugal Catarina
Borut Robič
and Florianopolis, Brazil
Faculty of Computer and
Evolutionary and Complex and
Information Science
Systems Group
University of Ljubljana Evolutionary and Complex
Centro de Informática e Sistemas Systems Group
Ljubljana, Slovenia
da Universidade de Coimbra
Centro de Informática e Sistemas
Pinhal de Marrocos, Portugal
da Universidade de Coimbra
Abhishek Roy Pinhal de Marrocos, Portugal
Alioune Ngom Center for Research in Wireless
Computer Science Department Mobility and Networking
Tiago Sousa
University of Windsor Department of Computer
Escola Superior de Tecnologia
Windsor, Ontario, Canada Science & Engineering
Instituto Politécnico de Castelo
The University of Texas at
Branco
Arlington
Mirela Sechi Moretti Annoni Castelo Branco, Portugal
Notare Arlington, Texas
and
Barddal University Evolutionary and Complex
Florianopolis, Brazil Systems Group
Hartmut Schmeck
Institute AIFB Centro de Informática e Sistemas
Stephan Olariu University of Karlsruhe da Universidade de Coimbra
Old Dominion University Karlsruhe, Germany Pinhal de Marrocos, Portugal
Norfolk, Virginia
Giandomenico Spezzano
Michael O’Neill Franciszek Seredynski Institute of High Performance
Department of Computer Polish-Japanese Institute of Computing and Networks
Science & Information Systems Information Technologies Rende (CS), Italy
University of Limerick Koszykowa
Limerick, Ireland Warsaw, Poland
and
Michael Stein
Institute AIFB
Institute of Computer Science
Juan Manuel Orduña Polish Academy of Sciences
University of Karlsruhe
Departamento de Informática Germany
Ordona
Universidad de Valencia
Warsaw, Poland
Burjassot (Valencia), Spain
Ivan Stojmenović
Department of Computer Science
Gregor Papa Jurij Šilc School of Information Technology
Computer Systems Department Computer Systems Department and Engineering
Jožef Stefan Institute Jožef Stefan Institute University of Ottawa
Ljubljana, Slovenia Ljubljana, Slovenia Ottawa, Ontario, Canada

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xiii — #13

xiv Contributors

Anna Świȩcicka Marco Tomassini Xin-She Yang

Department of Computer Science Information Systems Department Department of Engineering
Bialystok University of University of Lausanne University of Cambridge
Technology Lausanne, Switzerland
Cambridge, United Kingdom
Bialystok, Poland
Ashraf Wadaa
Old Dominion University Y. Young
Javid Taheri Norfolk, Virginia Civil and Computational
Engineering Centre
Advanced Networks Research
Group Horst F. Wedde School of Engineering
School of Information Informatik III University of Wales Swansea
Technologies University of Dortmund Swansea, United Kingdom
Dortmund, Germany
The University of Sydney
Sydney, Australia Albert Y. Zomaya
B. Wei
Laboratoire d’Informatique Advanced Networks Research
Fondamentale de Lille Group
El-Ghazali Talbi
Cité scientifique, France School of Information
Université des Sciences et Technologies
Technologies de Lille
Benjamin Weinberg The University of Sydney
Cité Scientifique, France Université des Sciences et Sydney, Australia
Technologies de Lille
Domenico Talia Cité Scientifique, France Joviša Žunić
University of Calabria Computer Science Department
Larry Wilson
DEIS Old Dominion University Cardiff University
Rende, Italy Norfolk, Virginia Cardiff, Wales, United Kingdom

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xiv — #14

 Contents

SECTION I Models and Paradigms

1 Evolutionary Algorithms Enrique Alba and Carlos Cotta . . . 1-3

2 An Overview of Neural Networks Models Javid Taheri and

Albert Y. Zomaya . . . . . . . . . . . . . . . . . . . . . 2-21

3 Ant Colony Optimization Michael Guntsch and Jürgen Branke 3-41

4 Swarm Intelligence Mohamed Belal, Jafaar Gaber,

Hoda El-Sayed, and Abdullah Almojel . . . . . . . . . . . . 4-55

5 Parallel Genetic Programming: Methodology, History, and

Application to Real-Life Problems
Francisco Fernández de Vega . . . . . . . . . . . . . . . . 5-65

6 Parallel Cellular Algorithms and Programs Domenico Talia . . 6-85

7 Decentralized Cellular Evolutionary Algorithms Enrique Alba,

Bernabe Dorronsoro, Mario Giacobini, and Marco Tomassini . . 7-103

8 Optimization via Gene Expression Algorithms

Forbes Burkowski . . . . . . . . . . . . . . . . . . . . . 8-121

9 Dynamic Updating DNA Computing Algorithms

Zhiquan Frank Qiu and Mi Lu . . . . . . . . . . . . . . . 9-135

10 A Unified View on Metaheuristics and Their Hybridization

Jürgen Branke, Michael Stein, and Hartmut Schmeck . . . . . 10-147

11 The Foundations of Autonomic Computing Salim Hariri,

Bithika Khargaria, Manish Parashar, and Zhen Li . . . . . . . 11-157

SECTION II Application Domains

12 Setting Parameter Values for Parallel Genetic Algorithms:

Scheduling Tasks on a Cluster Michelle Moore . . . . . . . . 12-179

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xv — #15

xvi Contents

13 Genetic Algorithms for Scheduling in Grid Computing

Environments: A Case Study Kris Crnomarkovic and
Albert Y. Zomaya . . . . . . . . . . . . . . . . . . . . . 13-193

14 Minimization of SADMs in Unidirectional SONET/WDM Rings

Using Genetic Algorithms Anirban Mukhopadhyay,
Utpal Biswas, Mrinal Kanti Naskar, Ujjwal Maulik, and
Sanghamitra Bandyopadhyay . . . . . . . . . . . . . . . . 14-209

15 Solving Optimization Problems in Wireless Networks Using

Genetic Algorithms Sajal K. Das, Nilanjan Banerjee, and
Abhishek Roy . . . . . . . . . . . . . . . . . . . . . . 15-219

16 Medical Imaging and Diagnosis Using Genetic Algorithms

Ujjwal Maulik, Sanghamitra Bandyopadhyay, and
Sajal K. Das . . . . . . . . . . . . . . . . . . . . . . . 16-235

17 Scheduling and Rescheduling with Use of Cellular Automata

Franciszek Seredynski, Anna Świȩcicka, and
Albert Y. Zomaya . . . . . . . . . . . . . . . . . . . . . 17-253

18 Cellular Automata, PDEs, and Pattern Formation

Xin-She Yang and Y. Young . . . . . . . . . . . . . . . . . 18-273

19 Ant Colonies and the Mesh-Partitioning Problem

Borut Robič, Peter Korošec, and Jurij Šilc . . . . . . . . . . . 19-285

20 Simulating the Strategic Adaptation of Organizations Using

OrgSwarm Anthony Brabazon, Arlindo Silva,
Ernesto Costa, Tiago Ferra de Sousa, and Michael O’Neill . . . 20-305

21 BeeHive: New Ideas for Developing Routing Algorithms

Inspired by Honey Bee Behavior Horst F. Wedde and
Muddassar Farooq . . . . . . . . . . . . . . . . . . . . 21-321

22 Swarming Agents for Decentralized Clustering in Spatial Data

Gianluigi Folino, Agostino Forestiero, and Giandomenico
Spezzano . . . . . . . . . . . . . . . . . . . . . . . . . 22-341

23 Biological Inspired Based Intrusion Detection Models for Mobile

Telecommunication Systems Azzedine Boukerche,
Kathia Regina Lemos Jucá, João Bosco Mangueira Sobral, and
Mirela Sechi Moretti Annoni Notare . . . . . . . . . . . . . 23-359

24 Synthesis of Multiple-Valued Circuits by Neural Networks

Alioune Ngom and Ivan Stojmenović . . . . . . . . . . . . 24-373

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xvi — #16

Contents xvii

25 On the Computing Capacity of Multiple-Valued Multiple-Threshold

Perceptrons Alioune Ngom,
Ivan Stojmenović, and Joviša Žunić . . . . . . . . . . . . . 25-427

26 Advanced Evolutionary Algorithms for Training Neural Networks

Enrique Alba, J. Francisco Chicano, Francisco Luna,
Gabriel Luque, and Antonio J. Nebro . . . . . . . . . . . . 26-453

27 Bio-Inspired Data Mining Tiago Sousa, Arlindo Silva,

Ana Neves, and Ernesto Costa . . . . . . . . . . . . . . . 27-469

28 A Hybrid Evolutionary Algorithm for Knowledge Discovery in

Microarray Experiments L. Jourdan, M. Khabzaoui,
C. Dhaenens, and El-Ghazali Talbi . . . . . . . . . . . . . 28-491

29 An Evolutionary Approach to Problems in Electrical Engineering

Design Gregor Papa, Jurij Šilc, and Barbara Koroušić-Seljak . . 29-509

30 Solving the Partitioning Problem in Distributed Virtual

Environment Systems Using Evolutive Algorithms
Pedro Morillo, Marcos Fernández, and Juan Manuel Orduña . . 30-531

31 Population Learning Algorithm and Its Applications

Piotr Jedrzejowicz . . . . . . . . . . . . . . . . . . . . . 31-555

32 Biology-Derived Algorithms in Engineering Optimization

Xin-She Yang . . . . . . . . . . . . . . . . . . . . . . . 32-589

33 Biomimetic Models for Wireless Sensor Networks

Kennie H. Jones, Kenneth N. Lodding, Stephan Olariu,
Ashraf Wadaa, Larry Wilson, and Mohamed Eltoweissy . . . . 33-601

34 A Cooperative Parallel Metaheuristic Applied to the Graph

Coloring Problem Benjamin Weinberg and
El-Ghazali Talbi . . . . . . . . . . . . . . . . . . . . . 34-625

35 Frameworks for the Design of Reusable Parallel and Distributed

Metaheuristics N. Melab, El-Ghazali Talbi, and S. Cahon . . . 35-639

36 Parallel Hybrid Multiobjective Metaheuristics on P2P Systems

N. Melab, El-Ghazali Talbi, M. Mezmaz, B. Wei . . . . . . . . 36-649

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C000” — 2005/8/17 — 18:11 — page xvii — #17

 I
Models and Paradigms

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 1 — #1

 1
Evolutionary
Algorithms

1.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

1.2 Learning from Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4
1.3 Nature’s Way for Optimizing. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
Algorithm Meets Evolution • The Flavors of Evolutionary
Algorithms
1.4 Dissecting an Evolutionary Algorithm . . . . . . . . . . . . . . . . . 1-8
The Fitness Function • Initialization • Selection •
Recombination • Mutation • Replacement
1.5 Fields of Application of EAs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-13
1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-14
Enrique Alba Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-14
Carlos Cotta References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-14

1.1 Introduction
One of the most striking features of Nature is the existence of living organisms adapted for surviving in
almost any ecosystem, even the most inhospitable: from abyssal depths to mountain heights, from volcanic
vents to polar regions. The magnificence of this fact becomes more evident when we consider that the
life environment is continuously changing. This motivates certain life forms to become extinct whereas
other beings evolve and preponderate due to their adaptation to the new scenario. It is very remarkable
that living beings do not exert a conscious effort for evolving (actually, it would be rather awkward to talk
about consciousness in amoebas or earthworms); much on the contrary, the driving force for change is
controlled by supraorganic mechanisms such as natural evolution.
Can we learn — and use for our own profit — the lessons that Nature is teaching us? The answer is a big
YES, as the optimization community has repeatedly shown in the last decades. “Evolutionary algorithm”
is the key word here. The term evolutionary algorithm (EA henceforth) is used to designate a collection
of optimization techniques whose functioning is loosely based on metaphors of biological processes.
This rough definition is rather broad and tries to encompass the numerous approaches currently
existing in the field of evolutionary computation [1]. Quite appropriately, this field itself is continuously
evolving; a quick inspection of the proceedings of the relevant conferences and symposia suffices to
demonstrate the impetus of the field, and the great diversity of the techniques that can be considered
“evolutionary.”

1-3

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 3 — #3

1-4 Handbook of Bioinspired Algorithms and Applications

This variety notwithstanding, it is possible to find a number of common features of all (or at least
most of) EAs. The following quote from Reference 2 illustrates such common points:

The algorithm maintains a collection of potential solutions to a problem. Some of these possible
solutions are used to create new potential solutions through the use of operators. Operators act on
and produce collections of potential solutions. The potential solutions that an operator acts on are
selected on the basis of their quality as solutions to the problem at hand. The algorithm uses this
process repeatedly to generate new collections of potential solutions until some stopping criterion
is met.

This definition can be usually found in the literature expressed in a technical language that uses terms
such as genes, chromosomes, population, etc. This jargon is a reminiscence of the biological inspiration
mentioned before, and has deeply permeated the field. We will return to the connection with biology
later on.
The objective of this work is to present a gentle overview of these techniques comprising both the
classical “canonical” models of EAs as well as some modern directions for the development of the field,
namely, the use of parallel computing, and the introduction of problem-dependent knowledge.

1.2 Learning from Biology

Evolution is a complex fascinating process. Along history, scientists have attempted to explain its
functioning using different theories. After the development of disciplines such as comparative anatomy
in the middle of the 19th century, the basic principles that condition our current vision of evolution
were postulated. Such principles rest upon Darwin’s Natural Selection Theory [3], and Mendel’s work on
genetic inheritance [4]. They can be summarized in the following points (see Reference 5):

• Evolution is a process that does not operate on organisms directly, but on chromosomes. These
are the organic tools by means of which the structure of a certain living being is encoded, that is,
the features of a living being are defined by the decoding of a collection of chromosomes. These
chromosomes (more precisely, the information they contain) pass from one generation to another
through reproduction.
• The evolutionary process takes place precisely during reproduction. Nature exhibits a plethora
of reproductive strategies. The most essential ones are mutation (that introduces variability in
the gene pool) and recombination (that introduces the exchange of genetic information among
individuals).
• Natural selection is the mechanism that relates chromosomes with the adequacy of the entities they
represent, favoring the proliferation of effective, environment-adapted organisms, and conversely
causing the extinction of lesser effective, nonadapted organisms.

These principles are comprised within the most orthodox theory of evolution, the Synthetic
Theory [6]. Although alternate scenarios that introduce some variety in this description have been
proposed — for example, the Neutral Theory [7], and very remarkably the Theory of Punctuated
Equilibria [8] — it is worth considering the former basic model. It is amazing to see that despite the
apparent simplicity of the principles upon which it rests, Nature exhibits unparalleled power in developing
and expanding new life forms.
Not surprisingly, this power has attracted the interest of many researchers, who have tried to translate the
principles of evolution to the realm of algorithmics, pursuing the construction of computer systems with
analogous features. An important point must be stressed here: evolution is an undirected process, that is,
there exists no scientific evidence that evolution is headed to a certain final goal. On the contrary, it can
be regarded as a reactive process that makes organisms change in response to environmental variations.
However, it is a fact that human-designed systems do pursue a definite final goal. Furthermore, whatever

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 4 — #4

Evolutionary Algorithms 1-5

this goal might be, it is in principle, desirable to reach it quickly and efficiently. This leads to the distinction
between two approaches to the construction of natureinspired systems:
1. Trying to reproduce Nature principles with the highest possible accuracy, that is, simulate Nature.
2. Using these principles as inspiration, adapting them in whatever required way so as to obtain
efficient systems for performing the desired task.
Both approaches concentrate nowadays on the efforts of researchers. The first one has given rise to
the field of Artificial Life (e.g., see Reference 9), and it is interesting because it allows re-creating and
studying numerous natural phenomena such as parasitism, predator/prey relationships, etc. The second
approach can be considered more practical, and constitutes the source of EAs. Notice anyway that these
two approaches are not hermetic containers, and have frequently interacted with certainly successful
results.

1.3 Nature’s Way for Optimizing

As mentioned above, the standpoint of EAs is essentially practical — using ideas from natural evolution
in order to solve a certain problem. Let us focus on optimization and see how this goal can be achieved.

1.3.1 Algorithm Meets Evolution

An EA is a stochastic iterative procedure for generating tentative solutions for a certain problem P . The
algorithm manipulates a collection P of individuals (the population), each of which comprises one or more
chromosomes. These chromosomes allow each individual represent a potential solution for the problem
under consideration. An encoding/decoding process is responsible for performing this mapping between
chromosomes and solutions. Chromosomes are divided into smaller units termed genes. The different
values a certain gene can take are called the alleles for that gene.
Initially, the population is generated at random or by means of some heuristic seeding procedure. Each
individual in P receives a fitness value: a measure of how good the solution it represents is for the problem
being considered. Subsequently, this value is used within the algorithm for guiding the search. The whole
process is sketched in Figure 1.1.
As it can be seen, the existence of a set F (also known as phenotype space) comprising the solutions for
the problem at hand is assumed. Associated with F , there also exists a set G (known as genotype space).
These sets G and F respectively constitute the domain and codomain of a function g known as the growth
(or expression) function. It could be the case that F and G were actually equivalent, being g a trivial
identity function. However, this is not the general situation. As a matter of fact, the only requirement
posed on g is subjectivity. Furthermore, g could be undefined for some elements in G .
After having defined these two sets G and F , notice the existence of a function ι selecting some
elements from G . This function is called the initialization function, and these selected solutions (also
known as individuals) constitute the so-called initial population. This initial population is in fact a pool

P
t

c s

vx P⬘

g P⬙ vm

FIGURE 1.1 Illustration of the evolutionary approach to optimization.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 5 — #5

1-6 Handbook of Bioinspired Algorithms and Applications

Evolutionary-Algorithm:

1. P ← apply i on to obtain m individuals (the initial population);

2. while Termination Criterion is not met do
(a) P ⬘ ← apply  on P ; /* selection */
(b) P ⬙ ← apply 1, … , k on P ⬘; /* reproduction */
(c) P ← apply  on P and P ⬙; /* replacement */
endwhile

FIGURE 1.2 Pseudocode of an evolutionary algorithm.

of solutions onto which the EA will subsequently work, iteratively applying some evolutionary operators
to modify its contents. More precisely, the process comprises three major stages: selection (promising
solutions are picked from the population by using a selection function σ ), reproduction (new solutions
are created by modifying selected solutions using some reproductive operators ωi ), and replacement (the
population is updated by replacing some existing solutions by the newly created ones, using a replacement
function ψ). This process is repeated until a certain termination criterion (usually reaching a maximum
number of iterations) is satisfied. Each iteration of this process is commonly termed a generation.
According to this description, it is possible to express the pseudocode of an EA as shown in Figure 1.2.
Every possible instantiation of this algorithmic template1 will give rise to a different EA. More precisely,
it is possible to distinguish different EA families, by considering some guidelines on how to perform this
instantiation.

1.3.2 The Flavors of Evolutionary Algorithms

EAs, as we know them now, began their existence during the late 1960s and early 1970s (some earlier
references to the topic exist, though; see Reference 10). In these years — and almost simultaneouly —
scientists from different places in the world began the task of putting Nature at work in algorithmics, and
more precisely in search of problem-solving duties. The existence of these different primordial sources
originated the rise of three different EA models. These classical families are:

• Evolutionary Programming (EP): This EA family originated in the work of Fogel et al. [11].
EP focuses on the adaption of individuals rather than on the evolution of their genetic informa-
tion. This implies a much more abstract view of the evolutionary process, in which the behavior of
individuals is directly modified (as opposed to manipulating its genes). This behavior is typically
modeled by using complex data structures such as finite automata or as graphs (see Figure 1.3[a]).
Traditionally, EP uses asexual reproduction — also known as mutation, that is, introducing slight
changes in an existing solution — and selection techniques based on direct competition among
individuals.
• Evolution Strategies (ESs): These techniques were initially developed in Germany by Rechenberg
[12] and Schwefel [13]. Their original goal was serving as a tool for solving engineering problems.
With this goal in mind, these techniques are characterized by manipulating arrays of floating-point
numbers (there exist versions of ES for discrete problems, but they are much more popular for
continuous optimization). As EP, mutation is sometimes the unique reproductive operator used
in ES; it is not rare to also consider recombination (i.e., the construction of new solutions by
combining portions of some individuals) though. A very important feature of ES is the utilization
of self-adaptive mechanisms for controlling the application of mutation. These mechanisms are
aimed at optimizing the progress of the search by evolving not only the solutions for the problem
being considered, but also some parameters for mutating these solutions (in a typical situation,

1 Themere fact that this high-level heuristic template can host a low-level heuristic, justifies using the term
metaheuristic, as it will be seen later.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 6 — #6

Evolutionary Algorithms 1-7

(a) Input encoding

(a–n)...(m–n)(a–o)...(m–o)(a–p)..(m–p)(n–q)(o–q)(p–q)...(n–z)(o–z)(p–z)

Input layer a b c ••• m

Hidden layer n o p

Output layer q r s ••• z

(b) IF

AND IS

IS IS FOR PL

POS NL VEL NL

FIGURE 1.3 Two examples of complex representations. (a) A graph representing a neural network. (b) A tree
representing a fuzzy rule.

an ES individual is a pair (x , σ ), where σ is a vector of standard deviations used to control the
Gaussian mutation exerted on the actual solution x ).
• Genetic Algorithms (GAs): GAs are possibly the most widespread variant of EAs. They were con-
ceived by Holland [14]. His work has had a great influence in the development of the field, to the
point that some portions — arguably extrapolated — of it were taken almost like dogmas (i.e., the
ubiquitous use of binary strings as chromosomes). The main feature of GAs is the use of a recom-
bination (or crossover) operator as the primary search tool. The rationale is the assumption that
different parts of the optimal solution can be independently discovered, and be later combined to
create better solutions. Additionally, mutation is also used, but it was usually considered a second-
ary background operator whose purpose is merely “keeping the pot boiling” by introducing new
information in the population (this classical interpretation is no longer considered valid though).

These families have not grown in complete isolation from each other. On the contrary, numerous
researchers built bridges among them. As a result of this interaction, the borders of these classical families
tend to be fuzzy (the reader may check [15] for a unified presentation of EA families), and new variants
have emerged. We can cite the following:

• Evolution Programs (EPs): This term is due to Michalewicz [5], and comprises those techniques
that, while using the principles of functioning of GAs, evolve complex data structures, as in EP.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 7 — #7

1-8 Handbook of Bioinspired Algorithms and Applications

Nowadays, it is customary to use the acronym GA — or more generally EA — to refer to such an

algorithm, leaving the term “traditional GA” to denote classical bit-string based GAs.
• Genetic Programming (GP): The roots of GP can be traced back to the work of Cramer [16], but
it is undisputable that it has been Koza [17] the researcher who promoted GP to its current status.
Essentially, GP could be viewed as an evolution program in which the structures evolved represent
computer programs. Such programs are typically encoded by trees (see Figure 1.3[b]). The final
goal of GP is the automatic design of a program for solving a certain task, formulated as a collection
of (input, output) examples.
• Memetic Algorithms (MAs): These techniques owe their name to Moscato [18]. Some widespread
misconception equates MAs to EAs augmented with local search; although such an augmented EA
could be indeed considered a MA, other possibilities exist for defining MAs. In general, a MA is
problem-aware EA [19]. This problem awareness is typically acquired by combining the EA with
existing algorithms such as hill climbing, branch and bound, etc.

In addition to the different EA variants mentioned above, there exist several other techniques that could
also fall within the scope of EAs, such as Ant Colony Optimization [20], Distribution Estimation Algorithms
[21], or Scatter Search [22] among others. All of them rely on achieving some kind of balance between
the exploration of new regions of the search space, and the exploitation of regions known to be promising
[23], so as to minimize the computational effort for finding the desired solution. Nevertheless, these
techniques exhibit very distinctive features that make them depart from the general pseudocode depicted
in Figure 1.2. The broader term metaheuristic (e.g., see Reference 24) is used to encompass this larger set
of modern optimization techniques, including EAs.

1.4 Dissecting an Evolutionary Algorithm

Once the general structure of an EA has been presented, we will get into more detail on the different
components of the algorithm.

1.4.1 The Fitness Function

This is an essential component of the EA, to the point that some early (and nowadays discredited)
views of EAs considered it as the unique point of interaction with the problem that is intended to be
solved. This way, the fitness function measured how good a certain tentative solution is for the problem
of interest. This interpretation has given rise to several misconceptions, the most important being the
equation “fitness = quality of a solution.” There are many examples in which this is simple not true [19],
for example, tackling the satisfiability problem with EAs (i.e., finding the truth assignment that makes
a logic formula in conjunctive normal form be satisfied). If quality is used as fitness function, then the
search space is divided into solutions with fitness 1 (those satisfying the target formula), and solutions
with fitness 0 (those that do not satisfy it). Hence, the EA would be essentially looking for a needle in
a haystack (actually, there may be more than one needle in that haystack, but that does not change the
situation). A much more reasonable choice is making fitness equal to the number of satisfied clauses in
the formula by a certain solution. This introduces a gradation that allows the EA “climbing” in search of
near-optimal solutions.
The existence of this gradation is thus a central feature of the fitness function, and its actual implemen-
tation is not that important as long this goal is achieved. Of course, implementation issues are important
from a computational point of view, since the cost of the EA is typically assumed to be that of evaluating
solutions. In this sense, it must be taken into account that fitness can be measured by means of a
simple mathematical expression, or may involve performing a complex simulation of a physical system.
Furthermore, this fitness function may incorporate some level of noise, or even vary dynamically. The
remaining components of the EA must be defined accordingly so as to deal with these features of the fitness
function, for example, using a nonhaploid representation [25] (i.e., having more than one chromosome)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 8 — #8

Evolutionary Algorithms 1-9

so as to have a genetic reservoir of worthwhile information in the past, and thus be capable of tackling
dynamic changes in the fitness function.
Notice that there may even exist more than one criterion for guiding the search (e.g., we would like to
evolve the shape of a set of pillars, so that their strength is maximal, but so that their cost is also minimal).
These criteria will be typically partially conflicting. In this case, a multiobjective problem is being faced.
This can be tackled in different ways, such as performing an aggregation of these multiple criteria into a
single value, or using the notion of Pareto dominance (i.e., solution x dominates solution y if, and only
if, fi (x) yields a better or equal value than fi (y) for all i, where the fi ’s represent the multiple criteria being
optimized). See References 26 and 27 for details.

1.4.2 Initialization
In order to have the EA started, it is necessary to create the initial population of solutions. This is
typically addressed by randomly generating the desired number of solutions. When the alphabet used
for representing solutions has low cardinality, this random initialization provides a more or less uniform
sample of the solution space. The EA can subsequently start exploring the wide area covered by the initial
population, in search of the most promising regions.
In some cases, there exists the risk of not having the initial population adequately scattered all over the
search space (e.g., when using small populations and/or large alphabets for representing solutions). It is
then necessary to resort to systematic initialization procedures [28], so as to ensure that all symbols are
uniformly present in the initial population.
This random initialization can be complemented with the inclusion of heuristic solutions in the initial
population. The EA can thus benefit from the existence of other algorithms, using the solutions they
provide. This is termed seeding, and it is known to be very beneficial in terms of convergence speed, and
quality of the solutions achieved [29,30]. The potential drawback of this technique is having the injected
solutions taking over the whole population in a few iterations, provoking the stagnation of the algorithm.
This problem can be remedied by tuning the selection intensity by some means (e.g., by making an
adequate choice of the selection operator, as it will be shown below).

1.4.3 Selection
In combination with replacement, selection is responsible for the competition aspects of individuals in
the population. In fact, replacement can be intuitively regarded as the complementary application of the
selection operation.
Using the information provided by the fitness function, a sample of individuals from the population is
selected for breeding. This sample is obviously biased towards better individuals, that is good — according
to the fitness function — solutions should be more likely in the sample than bad solutions.2
The most popular techniques are fitness-proportionate methods. In these methods, the probability of
selecting an individual for breeding is proportional to its fitness, that is,

fi
pi =
, (1.1)
j∈P fj

where fi is the fitness3 of individual i, and pi is the probability of i getting into the reproduction stage. This
proportional selection can be implemented in a number of ways. For example, roulette-wheel selection rolls

2 At least, this is customary in genetic algorithms. In other EC families, selection is less important for biasing
evolution, and it is done at random (a typical option in evolution strategies), or exhaustively, that is, all individuals
undergo reproduction (as it is typical in evolutionary programming).
3 Maximization is assumed here. In case we were dealing with a minimization problem, fitness should be transformed

so as to obtain an appropriate value for this purpose, for example, subtracting it from the highest possible value of the
guiding function, or taking the inverse of it.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 9 — #9

1-10 Handbook of Bioinspired Algorithms and Applications

a dice with |P| sides, such that the ith side has probability pi . This is repeated as many times as individuals
are required in the sample. A drawback of this procedure is that the actual number of instances of
individual i in the sample can largely deviate from the expected |P| · pi . Stochastic Universal Sampling [31]
(SUS) does not have this problem, and produces a sample with minimal deviation from expected values.
Fitness-proportionate selection faces problems when the fitness values of individuals are very similar
among them. In this case, pi would be approximately |P|−1 for all i ∈ P, and hence selection would be
essentially random. This can be remedied by using fitness scaling. Typical options are (see Reference 5):

• Linear scaling: fi = a · fi + b, for some real numbers a, b.

• Exponential scaling: fi = ( fi )k , for some real number k.
• Sigma truncation: fi = max(0, fi − ( f¯ − c · σ )), where f¯ is the mean fitness of individuals, σ is the
fitness standard deviation, and c is a real number.

Another problem is the appearance of an individual whose fitness is much better than the remaining
individuals. Such super-individuals can quickly take over the population. To avoid this, the best option is
using a nonfitness-proportionate mechanism. A first possibility is ranking selection [32]: individuals are
ranked according to fitness (best first, worst last), and later selected — for example, by means of SUS —
using the following probabilities:

 
1 − + − i−1
pi = η + (η − η ) , (1.2)
|P| |P| − 1

where pi is the probability of selecting the ith best individual, and η− + η+ = 2.

Another possibility is using tournament selection [33]. In this case, a direct competition is performed
whenever an individual needs to be selected. To be precise, α individuals are sampled at random, and
the best of them is selected for reproduction. This is repeated as many times as needed. The param-
eter α is termed the tournament size; the higher this value, the stronger the selective pressure. These
unproportionate selection methods have the advantage of being insensitive to fitness scaling problems
and to the sense of optimization (maximization or minimization). The reader is referred to, for example,
References 34 and 35 for a theoretical analysis of the properties of different selection operators.
Regardless of the selection operator used, it was implicity assumed in the previous discussion that any
two individuals in the population can mate, that is, all individuals belong to an unstructured centralized
population. However, this is not necessarily the case. There exists a long tradition in using structured
populations in EC, especially associated to parallel implementations. Among the most widely known
types of structured EAs, distributed (dEA) and cellular (cEA) algorithms are very popular optimization
procedures [36].
Decentralizing a single population can be achieved by partitioning it into several subpopulations,
where component EAs are run performing sparse exchanges of individuals (dEAs), or in the form of
neighborhoods (cEAs). The main difference is that a dEA has a large subpopulation, usually much
larger than the single individual that a cEA has typically in every component algorithm. In a dEA, the
subpopulations are loosely coupled, while for a cEA they are tightly coupled. Additionally, in a dEA, there
exist only a few subpopulations, while in a cEA there is a large number of them.
The use of decentralized populations has a great influence in the selection intensity, since not all
individuals have to compete among them. As a consequence, diversity is often better preserved.

1.4.4 Recombination
Recombination is a process that models information exchange among several individuals (typically two
of them, but a higher number is possible [37]). This is done by constructing new solutions using the
information contained in a number of selected parents. If it is the case that the resulting individuals (the
offspring ) are entirely composed of information taken from the parents, then the recombination is said to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 10 — #10

Evolutionary Algorithms 1-11

Cut point
Binary mask 00110100011001
01001101011010 01001101011010
Parents
11011010010011 11011010010011

01001110010011 Descendant 01011001010011

FIGURE 1.4 Two examples of recombination on bitstrings: single-point crossover (left) and uniform crossover
(right).

Cutting
Father 1 2 3 4 5 6 7 8 9

1 3 84 5 6 9 2 7
Mappings Child

(1) (6)

Mother 4 3 8 1 7 5 9 2 6
(5)
(2) (4)

(3)

FIGURE 1.5 PMX at work. The numbers in brackets indicate the order in which elements are copied to the
descendant.

be transmitting [38,39]. This is the case of classical recombination operators for bitstrings such as single-
point crossover, or uniform crossover [40], among others. Figure 1.4 shows an example of the application
of these operators.
This property captures the a priori role of recombination: combining good parts of solutions that have
been independently discovered. It can be difficult to achieve for certain problem domains though (the
Traveling Salesman Problem (TSP) is a typical example). In those situations, it is possible to consider other
properties of interest such as respect or assortment. The former refers to the fact that the recombination
operator generates descendants carrying all features common to all parents; thus, this property can be seen
as a part of the exploitative side of the search. On the other hand, assortment represents the exploratory side
of recombination. A recombination operator is said to be properly assorting if, and only if, it can generate
descendants carrying any combination of compatible features taken from the parents. The assortment is
said to be weak if it is necessary to perform several recombinations within the offspring to achieve this
effect.
The recombination operator must match the particulars of the representation of solutions chosen.
In the GA context, the representation was typically binary, and hence operators such as those depicted
in Figure 1.4 were used. The situation is different in other EA families (and indeed in modern GAs too).
Without leaving GAs, another very typical representation is that of permutations. Many ad hoc operators
have been defined for this purpose, for example, order crossover (OX) [41], partially mapped crossover
(PMX; see Figure 1.5) [42], and uniform cycle crossover (UCX) [43] among others. The reader may check
[43] for a survey of these different operators.
When used in continuous parameter optimization, recombination can exploit the richness of the
representation, and utilize a variety of alternate strategies to create the offspring. Let (x1 , . . . , xn ) and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 11 — #11

1-12 Handbook of Bioinspired Algorithms and Applications

+ / + /


+ X + – + + – + X +

X Y Y X X Y X Y Y X X Y

FIGURE 1.6 An example of branch-swapping recombination, as it is typically used in GP.

(y1 , . . . , yn ) be two arrays of real valued elements to be recombined, and let (z1 , . . . , zn ) be the resulting
array. Some possibilities for performing recombination are the following:

• Arithmetic recombination: zi = (xi + yi )/2, 1 ≤ i ≤ n.

√
• Geometric recombination: zi = xi yi , 1 ≤ i ≤ n.
• Flat recombination: zi = αxi + (1 − α)yi , 1 ≤ i ≤ n, where α is a random value in [0, 1].
• BLX-α recombination [44]: zi = ri + β(si − ri ), 1 ≤ i ≤ n, where ri = min(xi , yi ) − α|xi − yi |,
si = max(xi , yi ) + α|xi − yi |, and β is a random value in [0, 1].
• Fuzzy recombination: zi = Q(xi , yi ), 1 ≤ i ≤ n, where Q is a fuzzy connective [45].

In the case of self-adaptive schemes as those typically used in ES, the parameters undergoing self-
adaption would be recombined as well, using some of these operators. More details on self-adaption will
follow in next subsection.
Solutions can be also represented by means of some complex data structure, and the recombination
operator must be adequately defined to deal with these (e.g., References 46 to 48). In particular, the field
of GP normally uses trees to represent LISP programs [17], rule-bases [49], mathematical expressions,
etc. Recombination is usually performed here by swapping branches of the trees involved, as exemplified
in Figure 1.6.

1.4.5 Mutation
From a classical point of view (atleast in the GA arena [50]), this was a secondary operator whose mission is
to keep the pot boiling, continuously injecting new material in the population, but at a low rate (otherwise,
the search would degrade to a random walk in the solution space). EP practitioners [11] would disagree
with this characterization, claiming a central role for mutation. Actually, it is considered the crucial part
of the search engine in this context. This later vision has nowadays propagated to most EC researchers
(atleast in the sense of considering mutation as important as recombination).
As it was the case for recombination, the choice of a mutation operator depends on the representation
used. In bitstrings (and in general, in linear strings spanning  n , where  is arbitrary alphabet) mutation
is done by randomly substituting the symbol contained at a certain position by a different symbol. If a
permutation representation is used, such a procedure cannot be used for it would not produce a valid
permutation. Typical strategies in this case are swapping two randomly chosen positions, or inverting a
segment of the permutation. The interested reader may check [51] or [5] for an overview of different
options.
If solutions are represented by complex data structures, mutation has to be implemented accordingly.
In particular, this is the case of EP, in which, for example, finite automata [52], layered graphs [53],
directed acyclic graphs [54], etc., are often evolved. In this domain, it is customary to use more than one
mutation operator, making for each individual a choice of which operators will be deployed on it.
In the case of ES applied to continuous optimization, mutation is typically done using Gaussian
perturbations, that is,

zi = xi + Ni (0, σi ), (1.3)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 12 — #12

Evolutionary Algorithms 1-13

where σi is a parameter controlling the amplitude of the mutation, and N (a, b) is a random number
drawn from a normal distribution with mean a and standard deviation b. The parameters σi usually
undergo self-adaption. In this case, they are mutated prior to mutating the xi ’s as follows:
 )+N (0,τ )
σi = σi · eN (0,τ i , (1.4)

where τ and τ  are two parameters termed the local and global learning rate, respectively. Advanced schemes
have been also defined in which a covariance matrix is used rather than independent σi ’s. However, these
schemes tend to be unpractical if solutions are highly dimensional. For a better understanding of ES
mutation see Reference 55.

1.4.6 Replacement
The role of replacement is keeping the population size constant.4 To do so, some individuals from the
population have to be substituted by some of the individuals created during reproduction. This can be
done in several ways:
• Replacement-of-the-worst : the population is sorted according to fitness, and the new individuals
replace the worst ones from the population.
• Random replacement : the individuals to be replaced are selected at random.
• Tournament replacement : a subset of α individuals is selected at random, and the worst one is
selected for replacement. Notice that if α = 1 we have random replacement.
• Direct replacement : the offspring replace their parents.
Some variants of these strategies are possible. For example, it is possible to consider the elitist versions
of these, and only perform replacement if the new individual is better than the individual it has to replace.
Two replacement strategies (comma and plus) are also typically considered in the context of ES and
EP. Comma replacement is analogous to replacement of the worst, with the addition that the number of
new individuals |P  | (also denoted by λ) can be larger than the population size |P| (also denoted by µ).
In this case, the population is constructed using the best µ out of the λ new individuals. As to the plus
strategy, it would be the elitist counterpart of the former, that is, pick the best µ individuals out of the µ
old individuals plus the λ new ones. The notation (µ, λ) — EA and (µ + λ) — EA is used to denote these
two strategies.
It must be noted that the term “elitism” is often used as well to denote replacement-of-the-worst
strategies in which |P  | < |P|. This strategy is very commonly used, and ensures that the best individual
found so far is never lost. An extreme situation takes place when |P  | = 1, that is, just a single individual is
generated in each iteration of the algorithm. This is known as steady-state reproduction, and it is usually
associated with faster convergence of the algorithm. The term generational is used to designate the classical
situation in which |P  | = |P|.

1.5 Fields of Application of EAs

Evolutionary algorithms have been thoroughly used in many domains. One of the most conspicuous
fields in which these techniques have been utilized is combinatorial optimization (CO). This way, EAs
have been used to solve classical NP — hard problems such as the Travelling Salesman Problem [57–59],
the Multiple Knapsack Problem [60,61], Number Partitioning [62,63], Max Independent Set [64,65], and
Graph Coloring [66,67], among others.
Other nonclassical — yet important — CO problems to which EAs have been applied are scheduling
(in many variants [43, 68–71]), timetabling [72,73], lot-sizing [74], vehicle routing [75,76], quadratic
assignment [77,78], placement problems [79,80], and transportation problems [81].

4 Although it is not mandatory to do so [56], it is common practice to use populations of fixed size.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 13 — #13

1-14 Handbook of Bioinspired Algorithms and Applications

Telecommunications is another field that has witnessed the successful application of EAs. For example,
EAs have been applied to the placement of antennas and converters [82,83], frequency assignment
[84–86], digital data network design [87], predicting bandwidth demands in ATM networks [88], error
code design [89,90], etc. See also Reference 91.
Evolutionary algorithms have been actively used in electronics and engineering as well. For example,
work has been done in structure optimization [92], aeronautic design [93], power planning [94], circuit
design [95] computer-aided design [96], analogue-network synthesis [97], and service restoration [98]
among other areas.
Besides the precise application areas mentioned before, EAs have been also utilized in many other
fields such as, for example, medicine [99,100], economics [101,102], mathematics [103,104], biology
[105–107], etc. The reader may try querying any bibliographical database or web search engine for
“evolutionary algorithm application” to get an idea of the vast number of problems that have been tackled
with EAs.

1.6 Conclusions
EC is a fascinating field. Its optimization philosophy is appealing, and its practical power is striking.
Whenever the user is faced with a hard search/optimization task that she cannot solve by classical means,
trying EAs is a must. The extremely brief overview of EA applications presented before can convince the
reader that a “killer approach” is in her hands.
EC is also a very active research field. One of the main weaknesses of the field is the absence of
a conclusive general theoretical basis, although great advances are being made in this direction, and
in-depth knowledge is available about certain idealized EA models.
Regarding the more practical aspects of the paradigm, two main streamlines can be identified:
parallelizing and hybridizing. The use of decentralized EAs in the context of multiprocessors or net-
worked systems can result in enormous performance improvement [108], and constitutes an ideal option
for exploiting the availability of distributed computing resources. As to hybridization, it has become
evident in the last years that it constitutes a crucial factor for the successful use of EAs in real-world
endeavors. This can be achieved by hard-wiring problem-knowledge within the EA, or by combining it
with other techniques. In this sense, the reader is encouraged to read other essays in this volume to get
valuable ideas on suitable candidates for this hybridization.

Acknowledgments
This work has been partially funded by the Ministry of Science and Technology (MCYT) and Regional
Development European Found (FEDER) under contract TIC2002-04498-C05-02 (the TRACER project)
http://tracer.lcc.uma.es.

References
[1] T. Bäck, D.B. Fogel, and Z. Michalewicz. Handbook of Evolutionary Computation. Oxford
University Press, New York, 1997.
[2] T.C. Jones. Evolutionary Algorithms, Fitness Landscapes and Search. Ph.D. thesis, University of
New Mexico, 1995.
[3] C. Darwin. On the Origin of Species by Means of Natural Selection. John Murray, London, 1859.
[4] G. Mendel. Versuche über pflanzen-hybriden. Verhandlungen des Naturforschendes Vereines in
Brünn, 4: 3–47, 1865.
[5] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs. Springer-Verlag,
Berlin, 1992.
[6] J. Huxley. Evolution, the Modern Synthesis. Harper, New York, 1942.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 14 — #14

Evolutionary Algorithms 1-15

[7] M. Kimura. Evolutionary rate at the molecular level. Nature, 217: 624–626, 1968.
[8] S.J. Gould and N. Elredge. Punctuated equilibria: The tempo and mode of evolution reconsidered.
Paleobiology, 32: 115–151, 1977.
[9] C.G. Langton. Artificial life. In C.G. Langton, Ed., Artificial Life 1. Addison-Wesley, Santa Fe, NM,
1989, pp. 1–47.
[10] D.B. Fogel. Evolutionary Computation: The Fossil Record. Wiley-IEEE Press, Piscataway, NJ, 1998.
[11] L.J. Fogel, A.J. Owens, and M.J. Walsh. Artificial Intelligence Through Simulated Evolution. John
Wiley & Sons, New York, 1966.
[12] I. Rechenberg. Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologis-
chen Evolution. Frommann-Holzboog Verlag, Stuttgart, 1973.
[13] H.P. Schwefel. Numerische Optimierung von Computer–Modellen mittels der Evolutionsstrategie,
Vol. 26 of Interdisciplinary Systems Research. Birkhäuser, Basel, 1977.
[14] J.H. Holland. Adaptation in Natural and Artificial Systems. University of Michigan Press,
Ann Harbor, MI, 1975.
[15] T. Bäck. Evolutionary Algorithms in Theory and Practice. Oxford University Press, New York, 1996.
[16] M.L. Cramer. A representation for the adaptive generation of simple sequential programs.
In J.J. Grefenstette, Ed., Proceedings of the First International Conference on Genetic Algorithms.
Lawrence Erlbaum Associates, Hillsdale, NJ, 1985.
[17] J.R. Koza. Genetic Programming. MIT Press, Cambridge, MA, 1992.
[18] P. Moscato. On Evolution, Search, Optimization, Genetic Algorithms and Martial Arts: Towards
Memetic Algorithms. Technical report Caltech Concurrent Computation Program, Report 826,
California Institute of Technology, Pasadena, CA, USA, 1989.
[19] P. Moscato and C. Cotta. A gentle introduction to memetic algorithms. In F. Glover and
G. Kochenberger, Eds., Handbook of Metaheuristics. Kluwer Academic Publishers, Boston, MA,
2003, pp. 105–144.
[20] M. Dorigo and G. Di Caro. The ant colony optimization meta-heuristic. In D. Corne, M. Dorigo,
and F. Glover, Eds., New Ideas in Optimization. Maiden head, UK, 1999, pp. 11–32.
[21] P. Larrañaga and J.A. Lozano. Estimation of Distribution Algorithms. A New Tool for Evolutionary
Computation. Kluwer Academic Publishers, Boston, MA, 2001.
[22] M. Laguna and R. Martí. Scatter Search. Methodology and Implementations in C. Kluwer Academic
Publishers, Boston, MA, 2003.
[23] C. Blum and A. Roli. Metaheuristics in combinatorial optimization: Overview and conceptual
comparison. ACM Computing Surveys, 35: 268–308, 2003.
[24] F. Glover and G. Kochenberger. Handbook of Metaheuristics. Kluwer Academic Publishers, Boston,
MA, 2003.
[25] R.E. Smith. Diploid genetic algorithms for search in time varying environments. In Annual
Southeast Regional Conference of the ACM. ACM Press, New York, 1987, pp. 175–179.
[26] C.A. Coello. A comprehensive survey of evolutionary-based multiobjective optimization
techniques. Knowledge and Information Systems, 1: 269–308, 1999.
[27] C.A. Coello and A.D. Christiansen. An approach to multiobjective optimization using genetic
algorithms. In C.H. Dagli, M. Akay, C.L.P. Chen, B.R. Fernández, and J. Ghosh, Eds., Intelligent
Engineering Systems Through Artificial Neural Networks, Vol. 5. ASME Press, St. Louis, MO, 1995,
pp. 411–416.
[28] C.R. Reeves. Using genetic algorithms with small populations. In S. Forrest, Ed., Proceedings of the
Fifth International Conference on Genetic Algorithms. Morgan Kaufmann, San Mateo, CA, 1993,
pp. 92–99.
[29] C. Cotta. On the evolutionary inference of temporal Boolean networks. In J. Mira and
J.R. Álvarez, Eds., Computational Methods in Neural Modeling, Vol. 2686 of Lecture Notes in
Computer Science. Springer-Verlag, Berlin, Heidelberg, 2003, pp. 494–501.
[30] C. Ramsey and J.J. Grefensttete. Case-based initialization of genetic algorithms. In S. Forrest,
Ed., Proceedings of the Fifth International Conference on Genetic Algorithms. Morgan Kaufmann,
San Mateo, CA, 1993, pp. 84–91.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 15 — #15

1-16 Handbook of Bioinspired Algorithms and Applications

[31] J.E. Baker. Reducing bias and inefficiency in the selection algorithm. In J.J. Grefenstette, Ed.,
Proceedings of the Second International Conference on Genetic Algorithms. Lawrence Erlbaum
Associates, Hillsdale, NJ, 1987, pp. 14–21.
[32] D.L. Whitley. Using reproductive evaluation to improve genetic search and heuristic discovery.
In J.J. Grefenstette, Ed., Proceedings of the Second International Conference on Genetic Algorithms.
Lawrence Erlbaum Associates, Hillsdale, NJ, 1987, pp. 116–121.
[33] T. Bickle and L. Thiele. A mathematical analysis of tournament selection. In L.J. Eshelman,
Ed., Proceedings of the Sixth International Conference on Genetic Algorithms. Morgan Kaufmann,
San Francisco, CA, 1995, pp. 9-16.
[34] E. Cantú-Paz. Order statistics and selection methods of evolutionary algorithms. Information
Processing Letters, 82: 15–22, 2002.
[35] K. Deb and D. Goldberg. A comparative analysis of selection schemes used in genetic algorithms.
In G.J. Rawlins, Ed., Foundations of Genetic Algorithms. San Mateo, CA, 1991, pp. 69–93.
[36] E. Alba and J.M. Troya. A survey of parallel distributed genetic algorithms. Complexity, 4: 31–52,
1999.
[37] A.E. Eiben, P.-E. Raue, and Zs. Ruttkay. Genetic algorithms with multi-parent recombination.
In Y. Davidor, H.-P. Schwefel, and R. Männer, Eds., Parallel Problem Solving from Nature
III, Vol. 866 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, Heidelberg, 1994,
pp. 78–87.
[38] C. Cotta and J.M. Troya. Information processing in transmitting recombination. Applied
Mathematics Letters, 16: 945–948, 2003.
[39] N.J. Radcliffe. The algebra of genetic algorithms. Annals of Mathematics and Artificial Intelligence,
10: 339–384, 1994.
[40] G. Syswerda. Uniform crossover in genetic algorithms. In J.D. Schaffer, Ed., Proceedings of the
Third International Conference on Genetic Algorithms. Morgan Kaufmann, San Mateo, CA, 1989,
pp. 2–9.
[41] L. Davis. Handbook of Genetic Algorithms. Van Nostrand Reinhold Computer Library, New York,
1991.
[42] D.E. Goldberg and R. Lingle, Jr. Alleles, loci and the traveling salesman problem.
In J.J. Grefenstette, Ed., Proceedings of an International Conference on Genetic Algorithms.
Lawrence Erlbaum Associates, Hillsdale, NJ, 1985.
[43] C. Cotta and J.M. Troya. Genetic forma recombination in permutation flowshop problems.
Evolutionary Computation, 6: 25–44, 1998.
[44] L.J. Eshelman and J.D. Schaffer. Real-coded genetic algorithms and interval-schemata. In
D. Whitley, Ed., Foundations of Genetic Algorithms 2. Morgan Kaufmann Publishers, San Mateo,
CA, 1993, pp. 187–202.
[45] F. Herrera, M. Lozano, and J.L. Verdegay. Dynamic and heuristic fuzzy connectives-based cros-
sover operators for controlling the diversity and convengence of real coded genetic algorithms.
Journal of Intelligent Systems, 11: 1013–1041, 1996.
[46] E. Alba, J.F. Aldana, and J.M. Troya. Full automatic ann design: A genetic approach. In J. Cabestany,
J. Mira, and A. Prieto, Eds., New Trends in Neural Computation, Vol. 686 of Lecture Notes in
Computer Science. Springer-Verlag, Heidelberg, 1993, pp. 399–404.
[47] E. Alba and J.M. Troya. Genetic algorithms for protocol validation. In H.M. Voigt, W. Ebeling,
I. Rechenberg, and H.-P. Schwefel, Eds., Parallel Problem Solving from Nature IV. Springer-Verlag,
Berlin, Heidelberg, 1996, pp. 870–879.
[48] C. Cotta and J.M. Troya. Analyzing directed acyclic graph recombination. In B. Reusch, Ed.,
Computational Intelligence: Theory and Applications, Vol. 2206 of Lecture Notes in Computer
Science. Springer-Verlag, Berlin, Heidelberg, 2001, pp. 739–748.
[49] E. Alba, C. Cotta, and J.M. Troya. Evolutionary design of fuzzy logic controllers using strongly-
typed GP. Mathware & Soft Computing, 6: 109–124, 1999.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 16 — #16

Evolutionary Algorithms 1-17

[50] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-
Wesley, Reading, MA, 1989.
[51] A.E. Eiben and J.E. Smith. Introduction to Evolutionary Computing. Springer-Verlag, Berlin,
Heidelberg, 2003.
[52] C.H. Clelland and D.A. Newlands. PFSA modelling of behavioural sequences by evolutionary
programming. In R.J. Stonier and X.H. Yu, Eds., Complex Systems: Mechanism for Adaptation.
IOS Press, Rockhampton, Queensland, Australia, 1994, pp. 165–172.
[53] X. Yao and Y. Liu. A new evolutionary system for evolving artificial neural networks. IEEE
Transactions on Neural Networks, 8: 694–713, 1997.
[54] M.L. Wong, W. Lam, and K.S. Leung. Using evolutionary programming and minimum descrip-
tion length principle for data mining of bayesian networks. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 21: 174–178, 1999.
[55] H.-G. Beyer. The Theory of Evolution Strategies. Springer-Verlag, Berlin, Heidelberg, 2001.
[56] F. Fernandez, L. Vanneschi, and M. Tomassini. The effect of plagues in genetic programming:
A study of variable-size populations. In C. Ryan et al., Eds., Genetic Programming, Proceedings of
EuroGP’2003, Vol. 2610 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, Heidelberg,
2003, pp. 320–329.
[57] S. Chatterjee, C. Carrera, and L. Lynch. Genetic algorithms and traveling salesman problems.
European Journal of Operational Research, 93: 490–510, 1996.
[58] D.B. Fogel. An evolutionary approach to the traveling salesman problem. Biological Cybernetics,
60: 139–144, 1988.
[59] P. Merz and B. Freisleben. Genetic local search for the TSP: New Results. In Proceedings of the
1997 IEEE International Conference on Evolutionary Computation. IEEE Press, Indianapolis, USA,
1997, pp. 159–164.
[60] C. Cotta and J.M. Troya. A hybrid genetic algorithm for the 0–1 multiple knapsack problem.
In G.D. Smith, N.C. Steele, and R.F. Albrecht, Eds., Artificial Neural Nets and Genetic Algorithms
3. Springer-Verlag, Wien New York, 1998, pp. 251–255.
[61] S. Khuri, T. Bäck, and J. Heitkötter. The zero/one multiple knapsack problem and genetic
algorithms. In E. Deaton, D. Oppenheim, J. Urban, and H. Berghel, Eds., Proceedings of the 1994
ACM Symposium of Applied Computation proceedings. ACM Press, New York, 1994, pp. 188–193.
[62] R. Berretta, C. Cotta, and P. Moscato. Enhancing the performance of memetic algorithms by
using a matching-based recombination algorithm: Results on the number partitioning problem.
In M. Resende and J. Pinho de Sousa, Eds., Metaheuristics: Computer-Decision Making. Kluwer
Academic Publishers, Boston, MA, 2003, pp. 65–90.
[63] D.R. Jones and M.A. Beltramo. Solving partitioning problems with genetic algorithms. In
R.K. Belew and L.B. Booker, Eds., In Proceedings of the Fourth International Conference on Genetic
Algorithms. Morgan Kaufmann, San Mateo, CA, 1991, pp. 442–449.
[64] C.C. Aggarwal, J.B. Orlin, and R.P. Tai. Optimized crossover for the independent set problem.
Operations Research, 45: 226–234, 1997.
[65] M. Hifi. A genetic algorithm-based heuristic for solving the weighted maximum independent set
and some equivalent problems. Journal of the Operational Research Society, 48: 612–622, 1997.
[66] D. Costa, N. Dubuis, and A. Hertz. Embedding of a sequential procedure within an evolutionary
algorithm for coloring problems in graphs. Journal of Heuristics, 1: 105–128, 1995.
[67] C. Fleurent and J.A. Ferland. Genetic and hybrid algorithms for graph coloring. Annals of
Operations Research, 63: 437–461, 1997.
[68] S. Cavalieri and P. Gaiardelli. Hybrid genetic algorithms for a multiple-objective scheduling
problem. Journal of Intelligent Manufacturing, 9: 361–367, 1998.
[69] D. Costa. An evolutionary tabu search algorithm and the NHL scheduling problem. INFOR, 33:
161–178, 1995.
[70] C.F. Liaw. A hybrid genetic algorithm for the open shop scheduling problem. European Journal
of Operational Research, 124: 28–42, 2000.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 17 — #17

1-18 Handbook of Bioinspired Algorithms and Applications

[71] L. Ozdamar. A genetic algorithm approach to a general category project scheduling problem.
IEEE Transactions on Systems, Man and Cybernetics, Part C (Applications and Reviews), 29: 44–59,
1999.
[72] E.K. Burke, J.P. Newall, and R.F. Weare. Initialisation strategies and diversity in evolutionary
timetabling. Evolutionary Computation, 6: 81–103, 1998.
[73] B. Paechter, R.C. Rankin, and A. Cumming. Improving a lecture timetabling system for university
wide use. In E.K. Burke and M. Carter, Eds., The Practice and Theory of Automated Timetabling
II, Vol. 1408 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, 1998, pp. 156–165.
[74] K. Haase and U. Kohlmorgen. Parallel genetic algorithm for the capacitated lot-sizing prob-
lem. In Kleinschmidt et al., Eds., Operations Research Proceedings. Springer-Verlag, Berlin, 1996,
pp. 370–375.
[75] J. Berger and M. Barkaoui. A hybrid genetic algorithm for the capacitated vehicle routing prob-
lem. In E. Cantú-Paz, Ed., Proceedings of the Genetic and Evolutionary Computation Conference
2003, Vol. 2723 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, Heidelberg, 2003,
pp. 646–656.
[76] J. Berger, M. Salois, and R. Begin. A hybrid genetic algorithm for the vehicle routing problem with
time windows. In R.E. Mercer and E. Neufeld, Eds., Advances in Artificial Intelligence. 12th Biennial
Conference of the Canadian Society for Computational Studies of Intelligence. Springer-Verlag,
Berlin, 1998, pp. 114-127.
[77] P. Merz and B. Freisleben. Genetic algorithms for binary quadratic programming. In W. Banzhaf
et al., Eds., Proceedings of the 1999 Genetic and Evolutionary Computation Conference,
Morgan Kaufmann, San Francisco, CA, 1999, pp. 417–424.
[78] P. Merz and B. Freisleben. Fitness landscape analysis and memetic algorithms for the quadratic
assignment problem. IEEE Transactions on Evolutionary Computation, 4: 337–352, 2000.
[79] E. Hopper and B. Turton. A genetic algorithm for a 2d industrial packing problem. Computers &
Industrial Engineering, 37: 375–378, 1999.
[80] R.M. Krzanowski and J. Raper. Hybrid genetic algorithm for transmitter location in wireless
networks. Computers, Environment and Urban Systems, 23: 359–382, 1999.
[81] M. Gen, K. Ida, and L. Yinzhen. Bicriteria transportation problem by hybrid genetic algorithm.
Computers & Industrial Engineering, 35: 363–366, 1998.
[82] P. Calegar, F. Guidec, P. Kuonen, and D. Wagner. Parallel island-based genetic algorithm for radio
network design. Journal of Parallel and Distributed Computing, 47: 86–90, 1997.
[83] C. Vijayanand, M.S. Kumar, K.R. Venugopal, and P.S. Kumar. Converter placement in all-optical
networks using genetic algorithms. Computer Communications, 23: 1223–1234, 2000.
[84] C. Cotta and J.M. Troya. A comparison of several evolutionary heuristics for the frequency
assignment problem. In J. Mira and A. Prieto, Eds., Connectionist Models of Neurons, Learning
Processes, and Artificial Intelligence, Vol. 2084 of Lecture Notes in Computer Science. Springer-
Verlag, Berlin, Heidelberg, 2001, pp. 709–716.
[85] R. Dorne and J.K. Hao. An evolutionary approach for frequency assignment in cellular radio
networks. In 1995 IEEE International Conference on Evolutionary Computation. IEEE Press, Perth,
Australia, 1995, pp. 539–544.
[86] A. Kapsalis, V.J. Rayward-Smith, and G.D. Smith. Using genetic algorithms to solve the radio link
frequency assignment problem. In D.W. Pearson, N.C. Steele, and R.F. Albretch, Eds., Artificial
Neural Nets and Genetic Algorithms. Springer-Verlag, Wien New York, 1995, pp. 37–40.
[87] C.H. Chu, G. Premkumar, and H. Chou. Digital data networks design using genetic algorithms.
European Journal of Operational Research, 127: 140–158, 2000.
[88] N. Swaminathan, J. Srinivasan, and S.V. Raghavan. Bandwidth-demand prediction in virtual path
in atm networks using genetic algorithms. Computer Communications, 22: 1127–1135, 1999.
[89] H. Chen, N.S. Flann, and D.W. Watson. Parallel genetic simulated annealing: A massively parallel
SIMD algorithm. IEEE Transactions on Parallel and Distributed Systems, 9: 126–136, 1998.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 18 — #18

Evolutionary Algorithms 1-19

[90] K. Dontas and K. De Jong. Discovery of maximal distance codes using genetic algorithms.
In Proceedings of the Second International IEEE Conference on Tools for Artificial Intelligence. IEEE
Press, Herndon, VA, 1990, pp. 805–811.
[91] D.W. Corne, M.J. Oates, and G.D. Smith. Telecommunications Optimization: Heuristic and
Adaptive Techniques. John Wiley, New York, 2000.
[92] I.C. Yeh. Hybrid genetic algorithms for optimization of truss structures. Computer Aided Civil
and Infrastructure Engineering, 14: 199–206, 1999.
[93] D. Quagliarella and A. Vicini. Hybrid genetic algorithms as tools for complex optimisation prob-
lems. In P. Blonda, M. Castellano, and A. Petrosino, Eds., New Trends in Fuzzy Logic II. Proceedings
of the Second Italian Workshop on Fuzzy Logic. World Scientific, Singapore, 1998, pp. 300–307.
[94] A.J. Urdaneta, J.F. Gómez, E. Sorrentino, L. Flores, and R. Díaz. A hybrid genetic algorithm for
optimal reactive power planning based upon successive linear programming. IEEE Transactions
on Power Systems, 14: 1292–1298, 1999.
[95] M. Guotian and L. Changhong. Optimal design of the broadband stepped impedance transformer
based on the hybrid genetic algorithm. Journal of Xidian University, 26: 8–12, 1999.
[96] B. Becker and R. Drechsler. Ofdd based minimization of fixed polarity Reed-Muller expressions
using hybrid genetic algorithms. In Proceedings of the IEEE International Conference on Computer
Design: VLSI in Computers and Processor. IEEE, Los Alamitos, CA, 1994, pp. 106–110.
[97] J.B. Grimbleby. Hybrid genetic algorithms for analogue network synthesis. In Proceedings of the
1999 Congress on Evolutionary Computation. IEEE, Washington D.C., 1999, pp. 1781–1787.
[98] A. Augugliaro, L. Dusonchet, and E. Riva-Sanseverino. Service restoration in compensated dis-
tribution networks using a hybrid genetic algorithm. Electric Power Systems Research, 46: 59–66,
1998.
[99] M. Sipper and C.A. Peña Reyes. Evolutionary computation in medicine: An overview. Artificial
Intelligence in Medicine, 19: 1–23, 2000.
[100] R. Wehrens, C. Lucasius, L. Buydens, and G. Kateman. HIPS, A hybrid self-adapting expert system
for nuclear magnetic resonance spectrum interpretation using genetic algorithms. Analytica
Chimica ACTA, 277: 313–324, 1993.
[101] J. Alander. Indexed Bibliography of Genetic Algorithms in Economics. Technical report
94-1-ECO, University of Vaasa, Department of Information Technology and Production
Economics, 1995.
[102] F. Li, R. Morgan, and D. Williams. Economic environmental dispatch made easy with hybrid
genetic algorithms. In Proceedings of the International Conference on Electrical Engineering, Vol.
2. International Academic Publishers, Beijing, China, 1996, pp. 965–969.
[103] C. Reich. Simulation if imprecise ordinary differential equations using evolutionary algorithms.
In J. Carroll, E. Damiani, H. Haddad, and D. Oppenheim, Eds., ACM Symposium on Applied
Computing 2000. ACM Press, New York, 2000, pp. 428–432.
[104] X. Wei and F. Kangling. A hybrid genetic algorithm for global solution of nondifferentiable
nonlinear function. Control Theory & Applications, 17: 180–183, 2000.
[105] C. Cotta and P. Moscato. Inferring phylogenetic trees using evolutionary algorithms.
In J.J. Merelo, P. Adamidis, H.-G. Beyer, J.-L. Fernández-Villacañas, and H.-P. Schwefel, Eds.,
Parallel Problem Solving from Nature VII, Vol. 2439 of Lecture Notes in Computer Science.
Springer-Verlag, Berlin, 2002, pp. 720–729.
[106] G.B. Fogel and D.W. Corne. Evolutionary Computation in Bioinformatics. Morgan Kaufmann,
San Francisco, CA, 2003.
[107] R. Thomsen, G.B. Fogel, and T. Krink. A clustal alignment improver using evolution-
ary algorithms. In David B. Fogel, Xin Yao, Garry Greenwood, Hitoshi Iba, Paul Marrow,
and Mark Shackleton, Eds., Proceedings of the Fourth Congress on Evolutionary Computation
(CEC-2002) Vol. 1. 2002, pp. 121–126.
[108] E. Alba. Parallel evolutionary algorithms can achieve super-linear performance. Information
Processing Letters, 82: 7–13, 2002.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C001” — 2005/8/6 — 13:20 — page 19 — #19

 2
An Overview of
Neural Networks
Models

2.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21

2.2 General Structure of a Neural Network . . . . . . . . . . . . . . . . 2-22
Single- and Multi-Layer Perceptrons • Function Representation
2.3 Learning in Single-Layer Models . . . . . . . . . . . . . . . . . . . . . . . 2-26
Supervised Learning
2.4 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-30
K-Means Clustering • Kohonen Clustering • ART1 • ART2
2.5 Learning in Multiple Layer Models . . . . . . . . . . . . . . . . . . . . . 2-32
The Back Propagation Algorithm • Radial Basis Functions
2.6 A Sample of Neural Network Applications . . . . . . . . . . . . 2-35
Expert Systems • Neural Controllers • Decision Makers •
Robot Path Planning • Adaptive Noise Cancellation
Javid Taheri 2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-36
Albert Y. Zomaya References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-36

2.1 Introduction
Artificial Neural Networks have been one of the most active areas of research in computer science over
the last 50 years with periods of intense activity interrupted by episodes of hiatus [1]. The premise for
the evolution of the theory of artificial Neural Networks stems from the basic neurological structure of
living organisms. A cell is the most important constituent of these life forms. These cells are connected
by “synapses,” that are the links that carry messages between cells. In fact, by using synapses to carry the
pulses, cells can activate each other with different threshold values to form a decision or memorize an
event. Inspired by this simplistic vision of how messages are transferred between cells, scientists invented
a new computational approach, which became popularly known as Artificial Neural Networks (or Neural
Networks for short) and used it extensively to target a wide range of problems in many application
areas.
Although the shape or configurations of different Neural Networks may look different at the first
glance, they are almost similar in structure. Every neural network consists of “cells” and “links.” Cells are
the computational part of the network that perform reasoning and generate activation signals for other

2-21

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 21 — #1

2-22 Handbook of Bioinspired Algorithms and Applications

cells, while links connect the different cells and enable messages to flow between cells. Each link is usually
a one directional connection with a weight which affects the carried message in a certain way. This means,
that a link receives a value (message) from an input cell, multiplies it by a given weight, and then passes it
to the output cell. In its simplest form, a cell can have three states (of activation): +1 (TRUE), 0, and −1
(FALSE) [1].

2.2 General Structure of a Neural Network

Cells (or neurons) can have more sophisticated structure that can handle complex problems. These
neurons can basically be linear or nonlinear functions with or without biases. Figure 2.1 shows two simple
neurons, unbiased and biased.

2.2.1 Single- and Multi-Layer Perceptrons

The single-layer perceptron is one of the simplest classes of Neural Networks [1]. The general overview of
this network is shown in Figure 2.2 while the network has n inputs and generates only one output. The
input of the function f (·) is actually a linear combination of the network’s inputs. In this case, W is a
vector of neuron weights, X is the input vector, and y is the only output of the network defined as follows:

y = f (W · X + b),

W = (w1 w2 ... wn ),

X = (x1 x2 ... xn )T .

The above-mentioned basic structure can be extended to produce networks with more than one output.
In this case, each output has its own weights and is completely uncorrelated to the other outputs. Figure 2.3

(a)
w
x f (.) y

Input layer Neuron Output layer

y = f (wx)

(b)
w
x Σ f (.) y

Input layer Neuron Output layer

y = f (wx + b)

FIGURE 2.1 (a) Unbiased and (b) biased structure of a neural network.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 22 — #2

An Overview of Neural Networks Models 2-23

x1
w1
x2 w2
Σ f ( .) y

wn
xn
b

Input Layer Neuron Output Layer

FIGURE 2.2 A single output single-layer perceptron.

x1
w1, 1

w1, 2
Σ f1 (.) y1
x2

b1

Σ f2 (.) y2

b2

wm–1,n Σ fm (.) ym

bm
wm,n
xn
1

Input layer Layer -1 Output layer

FIGURE 2.3 A multi output single-layer perceptron.

shows an instant of such a network with the following formulas:

Y = F (W · X + B),
 
w1,1 w1,2 . . . w1,n
 w2,1 
 
W = . ,
 .. 
wm,1 . . . wm,n
X = (x1 x2 ... xn ) T ,
Y = ( y1 y2 ... ym ) T ,
B = (b1 b2 ... b m )T ,
F (·) = ( f1 (·) f2 (·) ... fm (·))T ,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 23 — #3

2-24 Handbook of Bioinspired Algorithms and Applications

x1 y1
w11,1 z p1
z11 w 21,1 z 21 w p1,1
Σ f 11(.) Σ f 21(.) Σ f p1(.)
x1
b11 b12 p f P1(.)
b1
1 1 1
z p2 y2
z12 z 22
Σ .)
f 12( Σ f 22 ( .) Σ f P2( .)

b12 b 22 b p2
1 1 1

zm1 z 2m2
Σ f 1m1( .) Σ f 2m (.) Σ f 2m (.)
2
w 2m2,n2 w pmp,np 2

b1m1 b2m2 b pmp z pmp

xn w1m1,n1 ym
1 1 1

Input layer Layer-1 Layer-2 Layer-p Output layer

FIGURE 2.4 The basic structure of a multi-layer neural network.

where n is the number of inputs, m the number of outputs, W the weighing matrix, X the input vector,
Y the output vector, and F (·) the array of output functions.
A multi-layer perceptron can simply be constructed by concatenating several single-layer perceptron
networks. Figure 2.4 shows the basic structure of such network with the following parameters [1]: X is
the input vector, Y the output vector, n the number of inputs, m the number of outputs, p the total number
of layers in the network, mi the number of outputs for the ith layer and, ni the number of inputs for the
ith layer.
Note that in this network, every internal layer of the network can have its own number of inputs and
outputs only by considering the concatenation rule, that is, ni = mi−1 . The output of the first layer is
calculated as follows:

Z 1 = F 1 (W 1 · X + B 1 ),
 
1
w1,1 1
w1,2 ... 1
w1,n
 1 
 w2,1 
 
W1 = 
 ..
,

 . 
 
wm1
1 ,1
. . . wm
1
1 ,n

X = (x1 x2 ... x n )T ,

B 1 = (b11 b21 ... 1 T

bm1
) ,

Z 1 = (z11 z21 ... Zm1 1 )T ,

F 1 (·) = ( f11 (·) f21 (·) ... fm11 (·))T .

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 24 — #4

An Overview of Neural Networks Models 2-25

Consequently the output of the second layer would be:

Z 2 = F 2 (W 2 · Z 1 + B 2 ),
 2 
w1,1 w1,2 2 ... 2
w1,n
 2 
 w2,1 
 
W2 =  .
,

 .. 
 
2
wm 2 ,1
. . . wm
2
2 ,m1

B 2 = (b12 b22 ... 2 T

bm2
) ,

Z 2 = (z12 z22 ... 2 T

zm2
) ,

F 2 (·) = ( f12 (·) f22 (·) ... fm22 (·))T ,

and finally the last layer formulation can be presented as follows:

Y = Z p = F p (W p · Z p−1 + B p ),
 p p p 
w1,1 w1,2 . . . w1,n
 p 
w 
 2,1 
W =
p
 ..
,

 . 
 
p p
wm1 ,1 . . . wmp ,mp−1
p p p
B p = (b1 b2 ... bmp )T ,
p p p
Z p = (z1 z2 ... zmp )T ,
p p p
F p (·) = ( f1 (·) f2 (·) ... fmp (·))T .

Notice that in such networks, the complexity of the network raises in a fast race based on the number of
layers. Practically experienced, each multi-layer perceptron can be evaluated by a single-layer perceptron
with comparatively huge number of nodes.

2.2.2 Function Representation

Two of the most popular uses of Neural Networks are to represent (or approximate) functions and model
systems. Basically, a neural network would be used to imitate the behavior of a function by generating
relatively similar outputs in comparison to the real system (or function) over the same range of inputs.
2.2.2.1 Boolean Functions
Neural networks were first used to model simple Boolean functions. For example, Figure 2.5 shows how
a neural network can be used to model an AND operator, while Figure 2.6 gives the truth table. Note
that, “1” stands for “TRUE” while “−1” represents a “FALSE” value. The network in Figure 2.5 actually
simulates a linear (function) separator, which simply divides the decision space into two parts.
2.2.2.2 Real Valued Functions
In this case, the network weights must be set so that it can generate continuous outputs of a real system.
The generated network is also intended to act as an extrapolator that can generate output data for inputs
that are different from the training set.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 25 — #5

2-26 Handbook of Bioinspired Algorithms and Applications

x1 1.4

Σ sgn( .) y

1.4
x2 –0.7

Input layer Neuron Output layer

FIGURE 2.5 A neural network to implement the logical AND.

x2

+1

–1 +1 x1

–1

x1
–1 +1
x2 –1 –1 –1
+1 –1 +1

FIGURE 2.6 Implementation of the logical AND of Figure 2.5.

2.3 Learning in Single-Layer Models

The main, and most important, application of all Neural Networks is their ability to model a process or
learn the behavior of a system. Toward this end, several algorithms were proposed to train the adjustable
parameters of a network (i.e., W ). Basically, training a neural network to adjust the W ’s is categorized
into two different classes: supervised and unsupervised [2–6].

2.3.1 Supervised Learning

The main purpose of this kind of training is to “teach” a network to copy the behavior of a system or a
function. In this case, there is always a need to have a “training” data set. The network topology and the
algorithm that the network is trained with are highly inter-related. In general, a topology of the network
is chosen first and then an appropriate training algorithm is used to tune the weights (W ) [7,8].

2.3.1.1 Perceptron Learning

As mentioned earlier, the perceptron is the most basic form of Neural Networks. Essentially, this network
tries to classify input data by mapping it onto a plane. In this approach, to simplify the algorithm,
suppose that the network’s input is restricted to {+1, 0, −1}, while the output can be {+1, −1}. The aim
of the algorithm is to find an appropriate set of weights, W , by sampling a training set, T , that will capture

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 26 — #6

An Overview of Neural Networks Models 2-27

the mapping that associates each input to an output, that is,

W = (w0 w1 ... wn ),

T = {(R 1 , S 1 ), (R 2 , S 2 ), . . . , (R L , S L )},

where n is the number of inputs, R i is the ith input data, S i represents the appropriate output for the ith
pattern, and, L is the size of the training set. Note that, for the above vector W , wn is used to adjust the
bias in the values of the weights. The Perceptron Learning can be summarized as follows:

Step 1: Set all elements of the weighting vector to zero, that is, W = (0 0 · · · 0).
Step 2: Select training pattern at random, namely kth datum.
Step 3: IF the current W has not been classified correctly, that is, W · R k = S k , then, modify the
weighing vector as follows: W ← W + R k S k .
Step 4: Repeat steps 1 to 3 until all data are classified correctly.

2.3.1.2 Linear Auto-Associators Learning

An auto-associate network is another type of network which has some type of memory. In this network,
the input and output nodes are basically the same. Hence, when a datum enters the network, it passes
through the nodes and converges to the closest memorized data, which was previously stored in the
network during the training process [1].
Figure 2.7 shows an instance of such network with five nodes. It is worthwhile mentioning that the
weighing matrix of such network is not symmetrical. That is, wi,j which relate the node “i” to node “j”
may have different value than wj,i . The main key of designing such network is the training data. In this
case, the assumption is to have orthogonal training data or at least approximately orthogonal, that is,


0 i = j,

Ti , Tj  ≈
1 i = j,

where Ti is the ith training data and

· is the inner product of two vectors. Based on the above assumption
the weight matrix for this network is calculated as follows where ⊗ stands for outer product of two vectors:


N
W = Ti ⊗ T i .
i=1

5 2

4 3

FIGURE 2.7 A sample linear auto-associate network with five nodes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 27 — #7

2-28 Handbook of Bioinspired Algorithms and Applications

As it can be seen, the main advantage of this network is in its one-shot learning process, by considering
orthogonal data. Note that, even if the input data are not orthogonal in the first place, they can be
transferred to a new space by a simple transfer function.
2.3.1.3 Iterative Learning
Iterative learning is another approach that can be used to train a network. In this case, the network’s weights
are modified smoothly, in contrast to the one-shot learning algorithms. In general, network weights are set
to some arbitrary values first, then, trained data are fed to the network. In this case, in each training cycle,
network weights are modified smoothly. Then, the training process proceeds until achieving an acceptable
level of acceptance for the network. However, the training data could be selected either sequentially or
randomly in each training cycle [9–11].
2.3.1.4 Hopfield’s Model
A Hopfield neural network is another example of an auto-associative network [1,12–14]. There are two
main differences between this network and the previously described auto-associate network. In this
network, self-connection is not allowed, that is, wi,i = 0 for all nodes. Also, inputs and outputs are either
0 or 1. This means that the node activation is recomputed after each cycle of convergence as follows:


N
Si = wi,j · uj (t ), (2.1)
j=1

1 if Si ≥ 0,
uj = (2.2)
0 if Si < 0.

After feeding a datum into the network, in each convergence cycle, the nodes are selected by a uniform
random function, the input are used to calculate Equation (2.1) and then followed by Equation (2.2) to
generate the output. This procedure is continued until the network converges.
The proof of convergence for this network uses the notion of “energy.” This means that an energy value
is assigned to each state of the network and through the different iterations of the algorithm, the overall
energy is decreased until it reaches a steady state.
2.3.1.5 Mean Square Error Algorithms
These techniques emerged as an answer to the deficiencies experienced by using Preceptrons and other
simple networks [1,15]. One of the most important reasons is the inseparability of training data. If the data
used to train the network are naturally inseparable, the training algorithm never terminates (Figure 2.8).
The other reason for using this technique is to converge to a better solution. In Perceptron learning,
the training process terminates right after finding the first answer regardless of its quality (i.e., sensitivity
of the answer). Figure 2.9 shows an example of such a case. Note that, although the answer found by
the Perceptron algorithm is correct (Figure 2.9[a]), the answer in Figure 2.9(b) is more robust. Finally,
another reason for using Mean Square Error (MSE) algorithms, which is crucial for most neural network
algorithms, is that of speed of convergence.
The MSE algorithm attempts to modify the network weights based on the overall error of all data. In this
case, assume that network input and output data are represented by Ti , Ri for i = 1, . . . , N , respectively.
Now the MSE error is defined as follows:

1 
N
E= (W · Ti − Ri )2 .
N
i=1

Note that, the stated error is the summation of all individual errors for the all the training data. Inspite
of all advantages gained by this training technique, there are several disadvantages, for example, the
network might not be able to correctly classify the data if it is widely spread apart (Figure 2.10). The other

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 28 — #8

An Overview of Neural Networks Models 2-29

–1

–2

–3
–3 –2 –1 0 1 2 3

FIGURE 2.8 Inseparable training data set.

(a) 3 (b) 3

2 2

1 1

0 0
0

–1 –1

–2 –2

–3 –3
–3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3

FIGURE 2.9 Two classifications for sample data.

disadvantage is that of the speed of convergence which may completely vary from one set of data to
another.
2.3.1.6 The Widow–Hoff Rule or LMS Algorithm
In this technique, the network weight is modified after each iteration [1,16]. A training datum is selected
randomly, then, the network weights are modified based on the corresponding error. This procedure
continues until converging to the answer. For a randomly selected kth entry in the training data, the error
is calculated as follows:

ε = (W · Tk − Rk )2 .

The gradient vector of this error would be:

∂ε ∂ε ∂ε
∇ε = ··· .
∂W0 ∂W1 ∂WN

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 29 — #9

2-30 Handbook of Bioinspired Algorithms and Applications

–1

–2

–3
–3 –2 –1 0 1 2 3

FIGURE 2.10 A data set with far apart solutions.

Hence,

∂ε
= 2(W · Tk − Rk ) · Tk .
∂Wj

Based on the Widow–Hoff algorithm, the weights should be modified opposite the direction of the
gradient. As a result, the final update formula for the weighting matrix W would be:

W  = W − ρ · (W · Tk − Rk ) · Tk .

Note that, ρ is known as the learning rate and it absorbs the multiplier of value “2.”

2.4 Unsupervised Learning

This class of networks attempts to cluster input data without the need for the traditional“learn by example”
technique that is commonly used for Neural Networks. Note that, clustering applications tend to be the
most popular type of applications that these networks are normally used for. The most popular networks
in this class are: K-means, Kohonen, ART1, and ART2 [17–21].

2.4.1 K-Means Clustering

This is the simplest technique used for classifying data. In this technique, a network with a predefined
number of clusters is considered, then, each datum is assigned to one of these clusters. This process
continues until all data are checked and classified properly. The following algorithm shows how this
algorithm is implemented:
Step 1: Consider a network with K clusters.
Step 2: Assign all data to one of the above clusters, with respect to the distance between the center of
the cluster and the datum.
Step 3: Modify the center of the assigned cluster.
Step 4: Check all data in the network to ensure proper classification.
Step 5: If a datum has to be moved from one cluster to another, then, update the center of both clusters.
Step 6: Repeat steps 4 and 5 until no datum is wrongly classified.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 30 — #10

An Overview of Neural Networks Models 2-31

(a) (b)

+ + +

+ + +

FIGURE 2.11 Results for a K-means clustering with (a) correct (b) incorrect number of clusters.

FIGURE 2.12 Output topology of a Kohonen network.

Figure 2.11 shows an instance of applying such network for data classification with the correct and
incorrect number of clusters.

2.4.2 Kohonen Clustering

This classification method clusters input data based on how the topological representation of the data.
The outputs of the network are arranged so that each output has some neighbors. Thus, during the
learning process, not only one output, but a group of close outputs are modified to classify the data. To
clarify the situation, assume that a network is supposed to learn how a set of data is to be distributed in a
two-dimensional representation (Figure 2.12).
In this case, each point is a potential output with a predefined neighborhood margin. For example,
the cell marked as “X” and eight of its neighbors are given. Therefore, whenever this cell gets selected
for update, all its neighbors are included in the process too. The main idea behind this approach for
classifying the input data is analogus to some biological facts. In a mammalian brain, all vision, auditory,
tactile sensors are mapped into a number of “cell sheets.” Therefore, if one of the cells is activated all cells
close to it will be affected, but at different levels of intensity.

2.4.3 ART1
This neural classifier, known as“Adaptive Resonance Theory”or ART, deals with digital inputs (Ti ∈ {0, 1}).
In this network, each “1” in the input vector represents information while a “0” entry is considered noise or
unwanted information. In ART, there is no predefined number of classes before the start of classification;
in fact, the classes are generated during the classification process.
Moreover, each class prototype may include the characteristics of more than a training datum. The
basic idea of such network relies on the similarity factor for data classification. In summary, every time

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 31 — #11

2-32 Handbook of Bioinspired Algorithms and Applications

a datum is assigned to a cluster, firstly, the nearest class with this datum is found, then, if the similarity of
this datum and the class prototype is more than a predefined value, known as a vigilance factor, then, the
datum is assigned to this class and the class prototype is modified to have more similarity with the a new
data entry [1,22,23].
The following procedure shows how this algorithm is implemented. However, the following needs to
be noted before outlining the algorithm:

1. X  is the number of 1’s in the vector X .

2. X · Y is the number of common 1’s between these vectors X and Y .
3. X ∩ Y is the bitwise AND operator applied on vectors X and Y .

Step 1: Let β be a small number, n be the dimension of the input data; and ρ be the vigilance factor
(0 ≤ ρ < 1).
Step 2: Start with no class prototype.
Step 3: Select a training datum by random, Tk .
Step 4: Find the nearest unchecked class prototype, Ci , to this datum by minimizing (Ci ·Tk )/(β +Ci ).
Step 5: Test if Ci is sufficiently close to Tk by verifying if (Ci · Tk )/(β + Ci ) > (Tk /(β + ρ)).
Step 6: If it is not similar enough, then, make a new class prototype and go to step 3.
Step 7: If it is sufficiently similar check the vigilance factor: (Ci · Tk /Tk ) ≥ ρ.
Step 8: If vigilance factor is exceeded, then, modify the class prototype by Ci = Ci ∩ Tk and go to step 3.
Step 9: If vigilance factor is not exceeded, then, try to find another unchecked class prototype in step 4.
Step 10: Repeat steps 3 to 9 until none of the training data causes any change in class prototypes.

2.4.4 ART2
This is a variation to ART1 with the following differences:

1. Data are considered continuous and not binary.

2. The input data is processed before passing it to the network. Actually, the input data is normalized,
then, all elements of the result vector that are below a predefined value are set to zero and the vector
normalized again. The process is used for noise cancellation.
3. When a class prototype is found for a datum, the class prototype vector is moved fractionally toward
the selected datum. As a result, contrary to the operation of ART1, the weights are moved smoothly
toward a new datum. The main reason for such a modification is to ‘memorize’ previously learnt
rules.

2.5 Learning in Multiple Layer Models

As mentioned earlier, multi-layer Neural Networks consist of several concatenated single-layer networks
[1,24–26]. The inner layers, known as hidden layers, may have different number of inputs and outputs.
Because of the added complexity the training process becomes more involved. This section presents two
of the most popular multi-layer neural network are presented.

2.5.1 The Back Propagation Algorithm

Back propagation algorithm is one of the most powerful and reliable techniques that can be used to adjust
the network weights. The main idea of this approach is to use gradient information of a cost function to
modify the network’s weights.
However, using this approach to train multi-layer networks is a little different from single-layer
networks. In general, multi-layer networks are much harder to train than single-layer ones. In fact,
convergence of such networks is much slower and very error sensitive.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 32 — #12

An Overview of Neural Networks Models 2-33

x1 w11,1 y1
z1 w 21,1
Σ f11(.) Σ f12(.)

x2
z2 y2
Σ f21(.) Σ f22(.)

zs
Σ f 1m1(.) Σ fm2(.)
w 2s,m
w 1n,s
xn ym

Input layer Layer-1 Layer-2 Output layer

FIGURE 2.13 A single hidden layer network.

In this approach, an input is presented to the network and allowed to “forward” propagate through
the network and the output is calculated. Then, the output will be compared to a“desired”output (from the
training set) and an error calculated. This error is then propagated “backward” into the network and the
different weights updated accordingly. To simplify describing this algorithm, consider a network with
a single hidden layer (and two layers of weights) given in Figure 2.13.
In relation to the above network, the following definitions apply. Of course, the same definitions can
be easily extended to larger networks.

Ti , Ri for i = 1, . . . , L: The training set of input and outputs, respectively.

N , S, M : The size of the input, hidden, and output layers, respectively.
W 1 : Network weights from the input layer to the hidden layer.
W 2 : Network weights from the hidden layer to the output layer.
X , Z , Y : Input and output of the hidden layer, and the network output, respectively.
F 1 (·): Array of network functions for the hidden layer.
F 2 (·): Array of network functions for the output layer.

It is important to note that, in such network, different combinations of weights might produce the
same input/output relationship. However, this is not crucial as long as the network is able to “learn” this
association. As a result, the network weights may converge to different sets of values based on the order of
the training data and the algorithm used for training although their stability may differ.

2.5.2 Radial Basis Functions

Radial Basis Function (RBF) Neural Network is another popular multi-layer neural network [27–31]. The
RBF network consists of two layers, one hidden layer and one output layer. In this network, the hidden
layer is implemented by radial activation functions while the output layer is simply a weighted sum of the
hidden layer outputs.
The RBF neural network is able to model complex mappings, which Perceptron Neural Networks
can only accomplish by means of multiple hidden layers. The outstanding characteristics of such network
makes it applicable for variety of applications, such as, function interpolation [32,33], chaotic time serious
modeling [34,35], system identification [36–38], control systems [39,40], channel equalization [41–43],
speech recognition [44,45], image restoration [46,47], motion estimation [48], pattern classification [49],
and data fusion [50].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 33 — #13

2-34 Handbook of Bioinspired Algorithms and Applications

x1 Σ y1
l1,1

Φ1 (.) l1,2

x2 Σ y2
Φ2(.)

Φp(.)
lp,m
xn Σ ym

Input layer Hidden layer Output layer

FIGURE 2.14 The basic structure of a RBF network.

The main topology of this network as is shown in Figure 2.14. Many functions were introduced for
possible use in the hidden layer; however, radial functions (Gaussian) remain the most effective to use for
data or pattern classification. The Gaussian functions are defined as follows:

i (X ) = exp−(X − µj )T j−1 (X − µj ),

where j = 1, 2, . . . , L and L represents the number of nodes in the hidden layer, X is the input vector,
µj and j are the mean vector and covariance matrix of the jth Gaussian function, respectively. In some
approaches, a polynomial term is appended to the above expression while in others the functions are
normalized to the sum of all Gaussian components as in the Gaussian mixture estimation. Geometrically,
a radial basis function in this network represents a bump in the N -dimensional space where N is the
number of entries (input vector size). In this case, the µj represents the location of this bump in the space
and j models its shape.
Because of the nonlinear behavior of this network, training procedure of the RBF network (as in
multi-layer networks) is approached in a completely different manner to that of single-layer networks.
In this network, the aim is to find the center and variance factor of all hidden layer Gaussian functions as
well as the optimal weights for the linear output layer. In this case, the following cost function is usually
considered as the main network objective:
N 

T
Min ([Y (Ti ) − Ri ] · [Y (Ti ) − Ri ]) ,
i=0

where N is the number of inputs in the training data set, Y (X ) is the output of the network for input X
and,
Tk , Rk  is the kth training data pair. So, the actual output of the network is a combination of a
nonlinear computation followed by a linear operation. Therefore, finding an optimal set of weights for
hidden layers and output layer parameters is hardly achievable.
In this case, several approaches were used to find the optimal set of weights, however, none of these can
provide any guarantees that optimality can be achieved. For example, many approaches suggest that the
hidden layer parameters are set randomly and the training procedure is just carried on for the output linear
components. In contrast, in some other cases, the radial basis functions are homogenously distributed

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 34 — #14

An Overview of Neural Networks Models 2-35

over the sample space before finding the output linear weights. However, the back propagation algorithm
seems to be the most suitable approach for training such a network.

2.6 A Sample of Neural Network Applications

This section briefly reviews a number of application areas in which Neural Networks were used effectively,
and this is by no means an exhaustive list of applications.

2.6.1 Expert Systems

One popular application is the use of Neural Networks as expert systems. Several definitions were presented
to clearly distinguish this kind of systems from other approaches [55–57]. Generally, an expert system
is defined as a system than can imitate the action of a human being for a given process. This definition
does not restrict the design of such systems by traditional Artificial Intelligence approaches. Therefore,
a variety of such systems can be built by using Fuzzy Logic, Neural Networks, and Neuro-Fuzzy techniques.
In most of these systems there is always a knowledge-based component that holds information about the
behavior of the system as simple rules followed by operators (usually in Fuzzy Systems) or a large database
collected from the system performance that a neural network can be trained to emulate.

2.6.2 Neural Controllers

Neural controllers are a specific class of the expert systems that deal with the process of regulating a linear
or nonlinear system. There are two methods to train such system, supervised and unsupervised. In the
supervised approach, another controller usually exists and the neural controller is trained to imitate its
behavior. In such case, the neural controller is connected in parallel to the other controller and during the
process, by sampling inputs and outputs, the network is trained to generate similar outputs for similar
inputs of the real controller. This process is known as on-line training. In contrast, in the case of off-line
training a database of the real controller inputs and outputs can be employed to train the network [58–60].

2.6.3 Decision Makers

In this specific class, which can also be viewed as an expert system, a neural network is used to make critical
decisions in unexpected situations. One such popular application is in financial markets such as stock
market applications. One of main characteristics of such systems that distinguish them from simple expert
systems is their stability. In fact, these systems must be able to produce acceptable output for untrained
situations. Therefore, a sufficiently “rich” data set must be used for the training process [61–63].

2.6.4 Robot Path Planning

Another complex scenario in which Neural Networks were used with some promise is that of robot
path planning. In this case, the robot tries to navigate its way to reach a target location. The situation
can be made more complicated by adding obstacles in the environment or even other mobile robots.
Normally, this situation is modeled as an optimization problem in which some cost function is minimized
(e.g., minimize the distance that the robot needs to travel) while satisfying certain constraints (e.g., no
collisions) [64–66].

2.6.5 Adaptive Noise Cancellation

Neural networks have been used very effectively to filter noise. In this case, the target signal (in the training
set) is the non-noisy signal that the input should be generating. The network must learn how to make
imitate the noise and in the process manage to neutralize it. Many approaches were introduced in the
literature over the years and some of these were deployed in real environments [67–69].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 35 — #15

2-36 Handbook of Bioinspired Algorithms and Applications

2.7 Conclusion
In this chapter, a general overview of the artificial Neural Networks was presented. These networks vary
in their sophistication from the very simple to the more complex. As a result, their training techniques
vary as well as their capabilities and suitability for certain applications. Neural networks have attracted a
lot of interest over the last few decades and it is expected they will be an active area of research for years
to come. Undoubtedly, more robust neural techniques will be introduced in the future that could benefit
a wide range of complex application.

References
[1] Gallant, S.I. Neural Network Learning and Expert Systems, MIT Press, Cambridge, MA, 1993.
[2] Karayiannis, N.B. and Venetsanopoulos, A.N. Efficient learning algorithms for Neural Networks
(ELEANNE). IEEE Transactions on Systems, Man, and Cybernetics, 23, 1993, 1372–1383.
[3] Hassoun, M.H. and Clark, D.W. An adaptive attentive learning algorithm for single-layer Neural
Networks. In Proceedings of the IEEE International Conference on Neural Networks, Vol. 1,
July 24–27, 1988, pp. 431–440.
[4] Ulug, M.E. A single layer fast learning fuzzy controller/filter, Neural Networks. In Proceedings of
the IEEE World Congress on Computational Intelligence, Vol. 3, June 27–July 2, 1994, pp. 1662–1667.
[5] Karayiannis, N.B. and Venetsanopoulos, A.N. Fast learning algorithms for Neural Networks.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 39, 1992, 453–474.
[6] Hrycej, T. Back to single-layer learning principles. In Proceedings of the International Joint
Conference on Neural Networks, Seattle, Vol. 2, July 8–14, 1991, p. 945.
[7] Healy, M.J. A logical architecture for supervised learning, Neural Networks. In Proceedings of
the IEEE International Joint Conference on Neural Networks, Vol. 1, November 18–21, 1991,
pp. 190–195.
[8] Brandt, R.D. and Feng, L. Supervised learning in Neural Networks without feedback network. In
Proceedings of the IEEE International Symposium on Intelligent Control, September 15–18, 1996,
pp. 86–90.
[9] Gong, Y. and Yan, P. Neural network based iterative learning controller for robot manipulators. In
Proceedings of the IEEE International Conference on Robotics and Automation, Vol. 1, May 21–27,
1995, pp. 569–574.
[10] Park, S. and Han, T. Iterative inversion of fuzzified Neural Networks. IEEE Transactions on Fuzzy
Systems, 8, 2000, 266–280.
[11] Zhan, X., Zhao, K., Wu, S., Wang, M., and Hu, H. Iterative learning control for nonlinear systems
based on Neural Networks. In Proceedings of the IEEE International Conference on Intelligent
Processing Systems, Vol. 1, October 28–31, 1997, pp. 517–520.
[12] Chen, C.J., Haque, A.L., and Cheung, J.Y. An efficient simulation model of the Hopfield Neural
Networks. In Proceedings of the International Joint Conference on Neural Networks, Vol. 1, June 7–11,
1992, pp. 471–475.
[13] Galan-Marin, G. and Munoz-Perez, J. Design and analysis of maximum Hopfield networks. IEEE
Transactions on Neural Networks, 12, 2001, 329–339.
[14] Nasrabadi, N.M. and Li, W. Object recognition by a Hopfield neural network. IEEE Transactions
on Systems, Man and Cybernetics, 21, 1991, 1523–1535.
[15] Xu, J., Zhang, X., and Li, Y. Kernel MSE algorithm: a unified framework for KFD, LS-SVM and
KRR. In Proceedings of the International Joint Conference on Neural Networks, Vol. 2, July 15–19,
2001, pp. 1486–1491.
[16] Hayasaka, T., Toda, N., Usui, S., and Hagiwara, K. On the least square error and prediction square
error of function representation with discrete variable basis. In Proceedings of the Workshop on
Neural Networks for Signal Processing, Vol. VI. IEEE Signal Processing Society, September 4–6,
1996, pp. 72–81.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 36 — #16

An Overview of Neural Networks Models 2-37

[17] Park, D.-C. Centroid neural network for unsupervised competitive learning. IEEE Transactions on
Neural Networks, 11, 2000, 520–528.
[18] Pedrycz, W. and Waletzky, J. Neural-network front ends in unsupervised learning. IEEE
Transactions on Neural Networks, 8, 1997, 390–401.
[19] Park, D.-C. Development of a neural network algorithm for unsupervised competitive learning.
In Proceedings of the International Conference on Neural Networks, Vol. 3, June 9–12, 1997,
pp. 1989–1993.
[20] Hsieh, K.-R. and Chen, W.-T. A neural network model which combines unsupervised and
supervised learning. IEEE Transactions on Neural Networks, 4, 1993, 357–360.
[21] Dajani, A.L., Kamel, M., and Elmastry, M.I. Single layer potential function neural network for
unsupervised learning. In Proceedings of the International Joint Conference on Neural Networks,
Vol. 2, June 17–21, 1990, pp. 273–278.
[22] Georgiopoulos, M., Heileman, G.L., and Huang, J. Properties of learning in ART1. In Proceedings
of the IEEE International Joint Conference on Neural Networks, Vol. 3, November 18–21, 1991,
pp. 2671–2676.
[23] Heileman, G.L., Georgiopoulos, M., and Hwang, J. A survey of learning results for ART1 networks.
In Proceedings of the IEEE International Conference on Neural Networks, IEEE World Congress on
Computational Intelligence, Vol. 2, June 27–July 2, 1994, pp. 1222–1225.
[24] Song, J. and Hassoun, M.H. Learning with hidden targets. In Proceedings of the International Joint
Conference on Neural Networks, Vol. 3, June 17–21, 1990, pp. 93–98.
[25] Kwan, H.K. Multilayer feedbackward Neural Networks. In Proceedings of the International
Conference on Acoustics, Speech, and Signal Processing, Vol. 2, April 14–17, 1991, pp. 1145–1148.
[26] Shepanski, J.F. Fast learning in artificial neural systems: multilayer perceptron training using
optimal estimation. In Proceedings of the IEEE International Conference on Neural Networks, Vol. 1,
July 24–27, 1988, pp. 465–472.
[27] Karayiannis, N.B. and Randolph-Gips, M.M. On the construction and training of reformulated
radial basis function Neural Networks. IEEE Transactions on Neural Networks, 14, 2003, 835–846.
[28] Leonard, J.A. and Kramer, M.A. Radial basis function networks for classifying process faults. IEEE
Control Systems Magazine, 11, 1991, 31–38.
[29] Li, R., Lebby, G., and Baghavan, S. Performance evaluation of Gaussian radial basis function
network classifiers. In Proceedings of the IEEE, SoutheastCon, April 5–7, 2002, pp. 355–358.
[30] Heimes, F. and van Heuveln, B. The normalized radial basis function neural network. In Proceedings
of the IEEE International Conference on Systems, Man, and Cybernetics, Vol. 2, October 11–14, 1998,
pp. 1609–1614.
[31] Craddock, R.J. and Warwick, K. Multi-layer radial basis function networks. An extension to the
radial basis function. In Proceedings of the IEEE International Conference on Neural Networks, Vol. 2,
June 3–6, 1996, pp. 700–705.
[32] Carr, J.C., Fright, W.R., and Beatson, R.K. Surface interpolation with radial basis functions for
medical imaging. IEEE Transactions on Medical Imaging, 16, 1997, 96–107.
[33] Romyaldy, M.A., Jr. Observations and guidelines on interpolation with radial basis function
network for one dimensional approximation problem. In Proceedings of the 26th Annual Conference
of the IEEE Industrial Electronics Society, Vol. 3, October 22–28, 2000, pp. 2129–2134.
[34] Leung, H., Lo, T., and Wang, S. Prediction of noisy chaotic time series using an optimal radial
basis function neural network. IEEE Transactions on Neural Networks, 12, 2001, 1163–1172.
[35] Katayama, R., Kajitani, Y., Kuwata, K., and Nishida, Y. Self generating radial basis function as neuro-
fuzzy model and its application to nonlinear prediction of chaotic time series. In Proceedings
of the Second IEEE International Conference on Fuzzy Systems, Vol. 1, March 28–April 1, 1993,
pp. 407–414.
[36] Warwick, K. and Craddock, R. An introduction to radial basis functions for system identification.
A comparison with other neural network methods. In Proceedings of the 35th IEEE Decision and
Control Conference, Vol. 1, December 11–13, 1996, pp. 464–469.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 37 — #17

2-38 Handbook of Bioinspired Algorithms and Applications

[37] Lu, Y., Sundararajan, N., and Saratchandran, P. Adaptive nonlinear system identification using
minimal radial basis function Neural Networks. In Proceedings of the IEEE International Conference
on Acoustics, Speech, and Signal Processing, Vol. 6, May 7–10, 1996, pp. 3521–3524.
[38] Tan, S., Hao, J., and Vandewalle, J. A new learning algorithm for RBF Neural Networks with
applications to nonlinear system identification. In Proceedings of the IEEE International Symposium
on Circuits and Systems, Vol. 3, April 28–May 3, 1995, pp. 1708–1711.
[39] Ibayashi, T., Hoya, T., and Ishida, Y. A model-following adaptive controller using radial basis
function networks. In Proceedings of the International Conference on Control Applications, Vol. 2,
September 18–20, 2002, pp. 820–824.
[40] Dash, P.K., Mishra, S., and Panda, G. A radial basis function neural network controller for UPFC.
IEEE Transactions on Power Systems, 15, 2000, pp. 1293–1299.
[41] Deng, J., Narasimhan, S., and Saratchandran, P. Communication channel equalization using
complex-valued minimal radial basis function Neural Networks. IEEE Transactions on Neural
Networks, 13, 2002, 687–696.
[42] Lee, J., Beach, C.D., and Tepedelenlioglu, N. Channel equalization using radial basis function
network. In Proceedings of the IEEE International Conference on Neural Networks, Vol. 4, June 3–6,
1996, pp. 1924–1928.
[43] Lee, J., Beach, C.D., and Tepedelenlioglu, N. Channel equalization using radial basis function
network. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal
Processing, Vol. 3, May 7–10, 1996, pp. 1719–1722.
[44] Sankar, R. and Sethi, N.S. Robust speech recognition techniques using a radial basis function
neural network for mobile applications. In Proceedings of IEEE Southeastcon, April 12–14, 1997,
pp. 87–91.
[45] Ney, H. Speech recognition in a neural network framework: discriminative training of Gaussian
models and mixture densities as radial basis functions. In Proceedings of the IEEE International
Conference on Acoustics, Speech, and Signal Processing, Vol. 1, April 14–17, 1991, pp. 573–576.
[46] Cha, I. and Kassam, S.A. Nonlinear image restoration by radial basis function networks.
In Proceedings of the IEEE International Conference on Image Processing, Vol. 2, November 13–16,
1994, pp. 580–584.
[47] Cha, I. and Kassam, S.A. Nonlinear color image restoration using extended radial basis function
networks. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal
Processing, Vol. 6, May 7–10, 1996, pp. 3402–3405.
[48] Bors, A.G. and Pitas, I. Optical flow estimation and moving object segmentation based on median
radial basis function network. IEEE Transactions on Image Processing, 7, 1998, 693–702.
[49] Gao, D. and Yang, G. Adaptive RBF Neural Networks for pattern classifications. In Proceedings of
the International Joint Conference on Neural Networks, Vol. 1, May 12–17, 2002, pp. 846–851.
[50] Fan, C., Jin, Z., Zhang, J., and Tian, W. Application of multisensor data fusion based on RBF
Neural Networks for fault diagnosis of SAMS. In Proceedings of the 7th International Conference on
Control, Automation, Robotics and Vision, Vol. 3, December 2–5, 2002, pp. 1557–1562.
[51] Tou, J.T. and Gonzalez, R.C. Pattern Recognition, Addison-Wesley, Reading, MA, 1974.
[52] Lo, Z.-P., Yu, Y., and Bavarian, B. Derivation of learning vector quantization algorithms.
In Proceedings of the International Joint Conference on Neural Networks, Vol. 3, June 7–11, 1992,
pp. 561–566.
[53] Burrascano, P. Learning vector quantization for the probabilistic neural network. IEEE Transactions
on Neural Networks, 2, 1991, 458–461.
[54] Karayiannis, N.B. and Randolph-Gips, M.M. Soft learning vector quantization and clustering
algorithms based on non-Euclidean norms: multinorm algorithms. IEEE Transactions on Neural
Networks, 14, 2003, 89–102.
[55] Medsker, L. Design and development of hybrid neural network and expert systems. In Proceedings
of the IEEE International Conference on Neural Networks, IEEE World Congress on Computational
Intelligence, Vol. 3, June 27–July 2, 1994, pp. 1470–1474.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 38 — #18

An Overview of Neural Networks Models 2-39

[56] Kurzyn, M.S. Expert systems and Neural Networks: a comparison, artificial Neural Networks and
expert systems. In Proceedings of the First International Two-Stream Conference on Neural Networks,
New Zealand, November 24–26, 1993, pp. 222–223.
[57] Hudli, A.V., Palakal, M.J., and Zoran, M.J. A neural network based expert system
model. In Proceedings of the Third International Conference on Tools for Artificial Intelligence,
November 10–13, 1991, pp. 145–149.
[58] Wang, W.-Y., Cheng, C.-Y., and Leu, Y.-G. An online GA-based output-feedback direct adaptive
fuzzy-neural controller for uncertain nonlinear systems. IEEE Transactions on Systems, Man and
Cybernetics, Part B, 34, 2004, 334–345.
[59] Zhang, Y., Peng, P.-Y., and Jiang, Z.-P. Stable neural controller design for unknown nonlinear
systems using backstepping. IEEE Transactions on Neural Networks, 11, 2000, 1347–1360.
[60] Nelson, A.L., Grant, E., and Lee, G. Developing evolutionary neural controllers for teams of mobile
robots playing a complex game. In Proceedings of the IEEE International Conference on Information
Reuse and Integration, October 27–29, 2003, pp. 212–218.
[61] Rothrock, L. Modeling human perceptual decision-making using an artificial neural network.
In Proceedings of the International Joint Conference on Neural Networks, Vol. 2, June 7–11, 1992,
pp. 448–452.
[62] Mukhopadhyay, S. and Wang, H. Distributed decomposition architectures for neural
decision-makers. In Proceedings of the 38th IEEE Conference on Decision and Control, Vol. 3,
December 7–10, 1999, pp. 2635–2640.
[63] Rogova, G., Scott, P., and Lolett, C. Distributed reinforcement learning for sequential decision
making. In Proceedings of the Fifth International Conference on Information Fusion, Vol. 2, July 8–11,
2002, pp. 1263–1268.
[64] Taheri, J. and Sadati, N. Fully modular online controller for robot navigation in static and
dynamic environments. In Proceedings of the 2003 IEEE International Symposium on Computational
Intelligence in Robotics and Automation, Vol. 1, July 16–20, 2003, pp. 163–168.
[65] Sadati, N. and Taheri, J. Genetic algorithm in robot path planning problem in crisp and fuzzified
environments. In Proceedings of the IEEE International Conference on Industrial Technology, Vol. 1,
December 11–14, 2002, pp. 175–180.
[66] Sadati, N. and Taheri, J. Solving robot motion planning problem using Hopfield neural network in
a fuzzified environment. In Proceedings of IEEE International Conference on Fuzzy Systems, Vol. 2,
May 12–17, 2002, pp. 1144–1149.
[67] Bambang, R. Active noise cancellation using recurrent radial basis function Neural Networks.
In Proceedings of the Asia-Pacific Conference on Circuits and Systems, Vol. 2, October 28–31, 2002,
pp. 231–236.
[68] Chen, C.K. and Chiueh, T.-D. Multilayer perceptron Neural Networks for active noise cancellation.
In Proceedings of the IEEE International Symposium on Circuits and Systems, Vol. 3, May 12–15,
1996, pp. 523–526.
[69] Tao, L. and Kwan, H.K. A neural network method for adaptive noise cancellation, Circuits and
Systems. In Proceedings of the IEEE International Symposium on Circuits and Systems, Vol. 5,
May 30–June 2, 1999, pp. 567–570.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C002” — 2005/8/17 — 18:12 — page 39 — #19

 3
Ant Colony
Optimization

3.1 Biological Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-41

3.2 An ACO Algorithm for the TSP . . . . . . . . . . . . . . . . . . . . . . . . . 3-43
3.3 Design Decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-45
Matrix Interpretation • Solution Construction • Pheromone
Update • Pheromone Initialization • Integration of
Local Search
3.4 Population Based Ant Colony Optimization. . . . . . . . . . . 3-49
Michael Guntsch 3.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-50
Jürgen Branke References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-51

3.1 Biological Background

Although only 2% of all insect species are social, they comprise more than 50% of the total insect biomass
globally [1], and more than 75% in some areas like the Amazon rain forest [2]. By social, we mean
that these insects, including all ants and termites and some subspecies of bees and wasps, live in colonies
composed of many interacting individuals. Insect colonies are capable of solving a number of optimization
problems that none of the individual insects would be able to solve by itself. Some examples are finding
short paths when foraging for food, task allocation when assigning labor to workers, and clustering when
organizing brood chambers, all of which are problems that have counterparts in real world optimization
problems.
In order for a swarm of insects to cooperate in problem solving, some form of communication is
necessary. This communication between the individuals of a colony can be more or less direct, depending
on the exact species. When a social bee has found a food source, it communicates the direction and
distance of the location where it found the food to the other bees by performing a characteristic dance [3].
This is a very direct communication, as the other bees must perceive the dance the one bee is performing
in order to be able to locate the food source. Other forms of direct communication include stimulation
by physical contact or the exchange of food or liquid.
Indirect communication between the individuals of a colony is more subtle and requires one individual
to modify the environment in such a way that it will alter the behavior of individuals passing through
this modified environment at a later time. One scenario where this type of environmentally induced
action exists in nature is when termites construct a nest that has a very complicated structure and exhibits
properties like temperature control [4]. Whenever a construction phase has ended, the surroundings of

3-41

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 41 — #1

3-42 Handbook of Bioinspired Algorithms and Applications

the worker have changed, and the next phase of working is encouraged, which in turn results in new
surroundings, and so forth. Another example for indirect communication is the laying of pheromone
trails performed by certain species of ants. An ant foraging for food will mark its path by distributing an
amount of pheromone on the trail it is taking, encouraging (but not forcing) ants who are also foraging
for food to follow its path. The principle of modifying the environment in order to induce a change in
behavior as a means of communication is called stigmergy and was first proposed in Reference 5.
Stigmergy is the basis for the organization in many ant colonies. Although the ants have a queen, this
is a specialized ant which is only responsible for laying eggs and does not have any governing function.
Instead, the ants of a colony are self-organized. The term self-organization (SO) is used to describe the
complex behavior which emerges from the interaction of comparatively simple agents. Its origins lie in
the fields of physics and chemistry, where SO is used to describe microscopic operations resulting in
macroscopic patterns, see Reference 6. Through SO, the ants are able to solve the complex problems which
they encounter on a daily basis. The benefits of SO as a basis for problem solving are especially apparent
in its distributed and robust character. Effectively, an ant colony can maintain a meaningful behavior even
if a large number of ants are incapable of contributing for some amount of time.
To better understand the mechanism behind an ant colony’s ability to converge to good solutions when
looking for a short path from the nest to a food source, some experiments were undertaken in References 7
and 8. In Reference 8 a nest of the Argentine ant Linepithema humile was given two paths of identical
length that it could take to reach a food source, and after some time had passed, it was observed that the
ants had converged to one of the paths, following it practically to the exclusion of the alternative. To test
whether this type of ant would converge to the shorter of two alternate paths, an experimental setup
similar to the one depicted in Figure 3.1 was evaluated in Reference 7.
The Argentine ant is practically blind, so it has no means of directly identifying a shorter path. However,
despite this deficiency, a swarm of these ants is capable of finding the shorter path connecting the nest to
the foraging area containing the food, as the experiment shows. Initially, all ants are located at the nest
site. A number of ants start out from the nest in search of food, each ant laying pheromone on its path,
and reach the first fork at point A. Since the ants have no information which way to go, that is, no ant has
walked before them and left a pheromone trail, each ant will choose to go either right or left with equal
probability. As a consequence, about one half of the foraging ants will take the shorter route, the others the
longer route to intersection B. The ants which were on the shorter track will reach this intersection first,
and have to decide which way to turn. Again, there is no information for the ants to use as orientation,
so half of the ants reaching intersection B will turn back toward the nest, while the rest continues toward
the foraging area containing the food. The ants on the longer branch between intersections A and B,
unaffected by the other ants they met head-on, arrive at intersection B and will also split up; however,

Foraging
area

B
12.5 cm

Nest

FIGURE 3.1 Single bridge experimental setup.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 42 — #2

Ant Colony Optimization 3-43

since the intensity of the pheromone trail heading back toward the nest is roughly twice as high as that of
the pheromone trail heading for the foraging area, the majority will turn back toward the nest, arriving
there at the same time as the other ants which took the long way back. Interestingly, since more ants have
now walked on the short branch between intersections A and B in comparison to the long one, future ants
leaving the nest will now already be more inclined to take the short branch, which is a first success in the
search for the shortest path.
The ants which continued toward the foraging area pick up some food to carry back to the nest. Arriving
at intersection B, the ants will prefer the short branch by the same argument as used above for ants starting
out fresh from the nest. Since the amount of pheromone at intersection A on the path back to the nest is
(roughly) equal to the sum of the pheromone amounts on the two branches leading away from the nest,
the shortest complete path from the foraging area back to the nest is also the most likely to be chosen by
the returning ants. Since the ants are continually distributing pheromone as they walk, the short path is
continually reinforced by more and more ants, until the amount of pheromone placed thereon in relation
to the alternative routes is so high that practically all ants use the shortest path, that is, the system converges
to the shortest path through self-reinforcement.
One point that we have neglected to mention so far is that the pheromone used by ants to mark their
trails slowly evaporates over time. This does not render the arguments used to explain the double bridge
experiment any less valid, it simply makes some of the math used for explanation less rigorous than
implied. Indeed, due to the evaporation of pheromone, a path that has not been chosen for some time,
invariably the long one, will contain almost no traces of pheromone after a sufficient amount of time,
further increasing the likelihood of ants taking the short path identified by the continually updated
pheromone.
In the rest of this chapter, we explain how the concept of marking paths with pheromones can be
exploited to construct algorithms capable of solving highly complex combinatorial optimization problems,
which was first proposed in Reference 9. The following section outlines the structure of Ant Colony
Optimization (ACO) algorithms using the Traveling Salesman Problem as an example. Afterwards, the
design decisions which must be made when implementing an ACO algorithm for a particular problem
are discussed. Finally, a number of applications where ACO algorithms lead to good results are surveyed.

3.2 An ACO Algorithm for the TSP

Consider the problem of finding the shortest tour connecting a number of points v ∈ V in a complete,
weighted graph G = (V , V × V , d) with d : V × V → IR + indicating the distance between nodes. This
problem, also known as the Traveling Salesman Problem (TSP), is a generalization of the problem solved
by the real ants looking for the shortest path between the food source and the nest.
An ACO algorithm for the TSP proceeds iteratively. In every iteration, a number of ants independently
construct tours by moving progressively from the current city to one that has not been visited yet,
completing the tour after all cities have been visited. The decision of which city to move to from the
current location is made using pheromone information τij , which denotes whether going from city i to
city j led to good solutions in the past, and heuristic information ηij , which signifies an immediate benefit
of moving from city i to city j. For both types of information, larger values are better. In case of the TSP, we
set ηij = 1/dij to indicate that short distances are preferable to long ones. Unlike biological ants, artificial
ants perform an a posteriori update on the global pheromone information depending on how good the
respective constructed tour is in comparison to the competing ants of the same iteration. Algorithm 3.1
shows a standard ACO algorithm for the TSP.
As can be seen in Algorithm 3.1, the artificial ants used to solve this problem are augmented in a
number of ways in comparison to their biological inspiration. The algorithm starts by initializing all
pheromone values to a level τ0 > 0, which is necessary since the ants make decisions according to relative
pheromone values and not absolute ones. We will discuss the merits of the different values for τ0 in
Section 3.3. For now, we use τ0 = 1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 43 — #3

3-44 Handbook of Bioinspired Algorithms and Applications

Algorithm 3.1 Basic ACO Algorithm for the TSP

1: initialize pheromone values ∀i, j ∈ [1, n] : τij

→ τ0
2: repeat
3: for each ant l ∈ {1, . . . , m} do
4: initialize selection set S
→ {1, . . . , n}
5: randomly choose starting city i0 ∈ S for ant l
6: move to starting city i
→ i0
7: while S = ∅ do
8: remove current city from selection set S
→ S \ {i}
β
τijα · ηij
9: choose next city j in tour with probability pij =
β
τihα · ηih
h∈S
10: update solution vector πl (i)
→ j
11: move to new city i
→ j
12: end while
13: finalize solution vector πl (i)
→ i0
14: end for
15: for each solution πl , l ∈ {1, . . . , m} do

n
16: calculate tourlength f (πl )
→ diπl (i)
i=1
17: end for
18: for all (i, j) do
19: evaporate pheromone τij
→ (1 − ρ) · τij
20: end for
21: determine best solution of iteration π + = arg min f (πl )
l∈[1,m]
22: if π + better than current best π  , i.e. f (π + ) < f (π  ) then
23: set π 
→ π +
24: end if
25: for all (i, j) ∈ π + do
26: reinforce τij
→ τij + ∆/2
27: end for
28: for all (i, j) ∈ π  do
29: reinforce τij
→ τij + ∆/2
30: end for
31: until condition for termination met

After the initialization of the pheromone matrix, m ants per iteration each independently construct
a solution for the TSP instance. The solution of ant l is stored in a permutation vector πl , which contains
the edges that were traversed during tour construction, that is, ant l moved from city i to city πl (i) in
the tour. In order to ensure that each solution constructed is feasible, each ant maintains a selection set S
which, while the ant constructs a solution, contains all cities that still need to be visited to complete the
tour. The starting city for the ants is chosen randomly, since a tour is a circle visiting all cities and therefore
the starting position is arbitrary. Afterwards, as long as there are still cities that need to be visited, the ant
chooses the next city in the tour in a probabilistic fashion according to Equation (3.1), which is called the
random proportional transition rule:
β
τijα · ηij
pij =  β
. (3.1)
α
h∈S τih · ηih

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 44 — #4

Ant Colony Optimization 3-45

The artificial ant tends to move to the city which has the best combination of pheromone information,
signifying a promising overall solution quality, and heuristic information, which equates to a short distance
to the immediate successor in the tour. The pheromone and heuristic information can be weighted with
the parameters α, β ∈ IR + in order to gauge the influence of the respective information on the decision
process. For the TSP, using α = 1 and β = 5 yields good results, see Section 3.3 for more details.
Once the artificial ant has visited all the cities initially in S, it returns to the starting city, thereby
completing the tour. After m ants have completed their individual tours in this fashion, the solutions are
evaluated and the pheromone information is updated. The update consists of two parts: evaporation and
intensification. The purpose of evaporation is to diminish all pheromone values by a relative amount,
and is accomplished by multiplying all pheromone values with a factor (1 − ρ), where ρ is called the
evaporation rate. For the intensification, the best ant of the iteration (with solution π + ) and the best
ant found by the algorithm over all iterations (with solution π  ) are used, each updating the pheromone
values corresponding to the edges traversed on the tour with ∆/2. In conjunction with τ0 = 1, setting
∆ = ρ usually works well.
After a number of iterations, the exact number correlating to how low the pheromone evaporation is set,
the ants will cease finding new solutions because the pheromone matrix has converged to one solution.
It is customary to stop the ACO algorithm after a number of iterations chosen in accordance with the
evaporation rate ρ and problem size n in order to facilitate the aggregation and comparison of results.

3.3 Design Decisions

In Section 3.2, we introduced a basic ACO algorithm for the TSP. All ACO algorithms follow a basic layout
which consists of ants repeatedly constructing solutions, which is modeled as a walk through a decision
graph, and updating the pheromone values in order to influence the behavior of future ants [10]. Thus,
the TSP is a very intuitive application for ACO algorithms. In this section, we discuss the design decisions
made for Algorithm 3.1 in more detail and show alternatives used for other problem classes.

3.3.1 Matrix Interpretation

In Section 3.2, we stated that the pheromone value τij denotes how opportune it was in the past to move
from city i to city j in the tour. When pheromone values denote a relationship between two items of
a problem instance, for example, two cities in the TSP, we say that the pheromone matrix is encoded in an
item × item fashion. The reason for using an item × item encoding for the TSP is that the tourlength is a
function of the edges traversed in a tour, and hence the pheromone values guide the ants toward solutions
with low overall tourlength. This type of encoding is also used for many other problems, for example,
scheduling problems with setup costs, where the solution quality is derived from a predecessor–successor
relationship. However, for problems like the Quadratic Assignment Problem (QAP), where facilities are
allocated to sites, the goal being to minimize the sum of distance-weighted flows between the facilities
(see Reference 11), it makes no difference whether a facility i was positioned before or after another
facility j. The only information which is relevant for the evaluation function is the location of the site that
facility i is assigned to. Hence, for problems where items need to be positioned optimally, whether in a
geometric sense as for the QAP or in a schedule, for example, when minimizing tardiness, an item × place
encoding is employed, with τij denoting how good it was in the past to put item i on place j.
Not all problems allow for the pheromone values to be arranged sensibly in a matrix fashion.
In Reference 12, ACO algorithms are used for constructing a shortest super-sequence to a given number
of strings, that is, the shortest string which contains all given strings as a sequence. The characters of each
string are marked with pheromone information to increase the likelihood of being chosen as the next
character when the ants construct a supersequence. Marking individual nodes (or items) with pheromone
values is proposed in Reference 13 when using ACO algorithms to solve subset problems like the Multiple
Knapsack Problem (MKP), where the optimal subset (in the form of a binary vector stating which items

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 45 — #5

3-46 Handbook of Bioinspired Algorithms and Applications

are in the subset) to a given number of items is sought which satisfies a number of linear constraints and
maximizes a linear function. The higher the pheromone value assigned to an item is, the more likely its
inclusion in the knapsack becomes.

3.3.2 Solution Construction

As we mentioned in Section 3.2, ants construct solutions by making a number of consecutive local
decisions which result in a global solution. For the TSP and most other item × item problems, the exact
sequence of these decisions is the result of the individual decisions made, as the ant continually moves
to the successor to make the next decision. However, for item × place problems, the decision sequence
is arbitrary, unless there are precedence constraints. In the absence of such constraints, it is usually best
to randomly permute the decision sequence every time an ant constructs a solution in order to avoid
a bias in the construction process, see Reference 14. Although it is generally preferable to proceed in a
random sequence, for some problems, for example, when scheduling jobs to minimize total tardiness, it
is necessary to start at place 1 and proceed in order so that heuristic information is available (see below);
however, even in these cases, a combination with a random decision sequence yields better results [15].
For the QAP, where so far no effective heuristic guidance for the ants has been discovered, a random
decision sequence should always be employed.
The basic rule by which ants make decisions is the random proportional transition rule defined
by Equation (3.1), in which the existence of heuristic information is already implied. If no heuristic
information is used, the formula is reduced to

τijα
pij =  α. (3.2)
h∈S τih

In References 16 and 17, a modification to the random proportional transition rule is proposed which
allows for greater control over the balance between exploration and exploitation. A parameter q0 ∈ [0, 1)
denotes the probability for choosing the best combination of pheromone and heuristic information
perceived by the ant, that is, for choosing j with
β
j = arg max τihα · ηih , (3.3)
h∈S

instead of proceeding probabilistically according to Equation (3.1), which is done with probability (1−q0 ).
This pseudo-random proportional rule allows for a calibration of the amount of exploitation (q0 -case)
versus biased exploration ((1−q0 )-case) performed by the ants, similar to the temperature setting in Simu-
lated Annealing or to reinforcement learning, specifically Q-learning [18]. In most cases, however, it is suf-
ficient to use only the random proportional transition rule defined by Equation (3.1), that is, to set q0 = 0.
For an instance of size n, constructing a solution takes O(n 2 ) steps for an ant, since it must make
n decisions and has O(n) candidates for each decision. A common method for reducing this computational
burden for very large problem instances is using so called candidate lists. For example, when using an
ACO algorithm for the TSP, it could make sense to define a candidate list Li for each city i which contains
the L nearest neighbors of city i. When the ant makes a decision, it uses Si = S ∩ Li instead of S as the
selection set, unless Si = ∅, in which case the ant reverts to using S. On average, this will greatly reduce the
computational expense of constructing a solution while still maintaining good solution quality, see also
Reference 19.
Equations (3.1), (3.2), or (3.3) are usually applied in permutation problems. For some problem classes,
however, extensions or alternatives exist. For scheduling problems which minimize total tardiness, using
the sum over all pheromone values up to and including i, that is,

i
τij = τlj , (3.4)
l=1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 46 — #6

Ant Colony Optimization 3-47

instead of using only the pheromone value τij , is studied in Reference 20. This type of pheromone
evaluation more accurately reflects that a job which was not chosen for a particular place in the schedule
despite a high pheromone value should be scheduled soon afterwards.
In Reference 21, an ACO algorithm for the Maximum Clique Problem (MCP) is proposed, where new
nodes for the clique C being constructed are chosen by evaluating how strong the pheromone-based con-
nection between nodes already in C and the candidate nodes in S are, that is, the probability to add node i is
α
τiC
pi =  α (3.5)
h∈S τhC

with τiC = j∈C τij and S being updated in such a way that C  = C ∪ {i} is still a clique for all i ∈ S.
Correspondingly, a pheromone update is undertaken on all edges of the largest final clique of the iteration.
In addition to the pheromone information, ants have the possibility of exploiting heuristic information,
if available. Although not a strict necessity, heuristic guidance usually plays an important role for attaining
a good solution quality. Two known exceptions to this rule are the QAP, where no beneficial heuristic
guidance seems to exist, and the MCP mentioned above, for which results in Reference 21 showed that
employing a straightforward heuristic guidance scheme ultimately led to a worse solution quality.
For the problem classes for which beneficial heuristic guidance does exist, we distinguish between
constant and dynamic heuristic values. For the TSP, the heuristic values ηij = 1/dij are computed once
and remain constant. For other problems like the Probabilistic TSP (see References 22 and 23) or the single
machine total weighted tardiness problem (SMTWTP), the ηij values are functions of past decisions which
the ant has made, that is, they need to be recalculated during every solution construction. For example,
when considering which job j to schedule at place i when minimizing the total tardiness of all jobs, the
heuristic values
1
ηij = (3.6)
max(T + pj , dj ) − T

are used in Reference 20, where dj is the due date and pj the processing time of job j, and T = i−1
h=1 pπ(h)
is the sum of processing times of all jobs that have already been scheduled. Even when heuristic values
are dynamic, the effort for their computation should be restricted to constant time if possible, since
otherwise the amount of time necessary for constructing a solution becomes prohibitively large. Another
disadvantage of the heuristic information for the SMTWTP is that it cannot be combined with a random
decision sequence.
Finally, it is possible to influence the ant’s decisions by tuning the pheromone and heuristic weights
α and β. For most applications, setting α = β = 1 is sufficient. However, tuning these parameters can
lead to better performance for some problem classes. For the TSP, choosing β > 1 has been shown to yield
good results, for example, using β = 2 in References 17 and 24 or β = 5 in References 25 and 26. Using
a steadily decreasing value of β has also been applied successfully for the Resource Constrained Project
Scheduling Problem (RCPSP) in Reference 27. Using values of α > 1 has been shown in Reference 21 to
achieve quicker convergence at the cost of a lower solution quality in the long run.

3.3.3 Pheromone Update

The purpose of the pheromone update is to focus the search process of the ants on a promising portion
of the solution space, which is then explored more extensively in the hope of finding the optimal solution.
The field of update strategies is perhaps the most studied part of ACO algorithms, and a variety of methods
to update the pheromone values have been proposed. In this section, we briefly explain the conceptually
different approaches which have been suggested to update pheromone values, and also explain in detail
some of the more successful strategies.
At the most abstract level, the pheromone update should accomplish a positive reinforcement of those
search space regions which seem promising and a negative reinforcement of all others. The principal

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 47 — #7

3-48 Handbook of Bioinspired Algorithms and Applications

mechanisms used for this are pheromone evaporation (for the negative reinforcement), which diminishes
all pheromone values by a relative amount each time it is applied, and pheromone intensification (for
the positive reinforcement), achieved by adding an update term to selected pheromone values. Formally,
an update takes the form

∀i, j ∈ [1, n] : τij

→ (1 − ρ) · τij + ∆ij , (3.7)

where ρ ∈ (0, 1] is a parameter of the algorithm denoting how much of the pheromone information is lost
with every application of evaporation. A high evaporation rate will cause a more rapid convergence which
is coupled with less exploration than a low evaporation rate. Thus, the evaporation rate should be tuned
in accordance with the number of iterations that the ACO algorithm is allowed to run. ∆ij is an update
value, which is 0 if the edge (i, j) was not traversed by the ant and some value greater than 0 if it was.
The exact value of ∆ij and especially the strategy when an update is performed is the key difference
between most types of ACO algorithm. There are two aspects to consider when characterizing pheromone
intensification: which solutions update, and how much is updated by these solutions.
Generally, updates of the pheromone values take place after an iteration of m ants has constructed
solutions. In the Ant System (AS), which was introduced in Reference 9 for solving the TSP, every ant of
the iteration contributes to the pheromone update. For each ant l ∈ [1, m], the update value ∆ij (l), is

calculated, and the update is performed with the sum of update values ∆ij = m l=1 ∆ij (l). Three different
methods for determining the individual ∆ij were tested: assigning a constant, using the inverse to the
distance dij between customers i and j, and, performing best and used subsequently, inverse to the length
of the entire tour, that is, the solution quality. In addition to the m ants of an iteration being allowed
to perform an update, it was also proposed to let a number of so called elitist ants, which represent the
best solution found by all ants so far, update the pheromone trail. Using a small number of these elitist
ants, inspired by the elitist strategy in Reference 28, intensifies the search near the currently best solution,
leading to better results overall.
Further research resulted in the introduction of the Ant Colony System (ACS) [16,17]. Here, an online
update of the pheromone values was proposed in order to enforce exploration: each time an ant traversed
an edge (i, j), it would reduce the corresponding pheromone value according to

τij
→ (1 − ρ) · τij + ρ · τ0 , (3.8)

thus encouraging subsequent ants to choose different edges (note that this holds only for τij ≥ τ0 ). Also,
the global update by all ants at the end of an iteration was replaced by one update performed along the
best tour found so far, that is, by one elitist ant.
Originating from AS and ACS, many other update schemes have been proposed. In References 24
and 29, the MAX–MIN Ant System (MMAS) is introduced, which uses only the best ant of the iteration
and an elitist ant for the positive pheromone update and avoids stagnation in the search process by limiting
pheromone values to the predetermined interval [τmin , τmax ]. Limiting the pheromone values also bounds
the minimum and maximum probability with which an edge is selected according to Equation (3.1), if the
heuristic values are bounded as well. In Reference 25, a modification of the AS called AS-rank is proposed,
where the m ants of an iteration are ranked by their solution quality and, together with a number of elitist
ants of maximum rank, update the pheromone trail in proportion to their rank.
Some methods also exist that operate without using evaporation. In Reference 30, pheromone update is
accomplished by comparing the solution quality of the ant to the average quality of the m previous ants. If it
is better, a positive update is performed along the path; if it is worse, the update is negative. Thus, the update
is accomplished in O(m · n) time for m ants, compared to O(n 2 ) for Equation (3.7). In Reference 31,
a population of solutions is maintained from which the pheromone matrix is derived. Updates are
performed on the population, with the insertion of a solution being equivalent to a positive update of the
pheromone matrix, and a deletion to a negative update (nullifying the previous positive update which was
undertaken upon insertion). This update mechanism will be studied in more detail in Section 3.4.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 48 — #8

Ant Colony Optimization 3-49

The number of ants m also plays a role in the exact effect of the pheromone updates. The more solutions
are constructed before an update is undertaken, the higher the expected quality of the updating solutions.
However, the number of ants is a linear factor in the runtime of the algorithm, which is why a trade-off
between runtime and quality of the updating solutions must be found. In Reference 17, a method for
finding the optimal number of ants is discussed which relies on knowledge of the average size of the
pheromone values before and after a change, both being a function of the problem size n. Modeling the
local pheromone update as a first-order linear recurrence relation allows m to be expressed as a function
of the average pheromone levels before and after an update, and of the initial pheromone values. Although
the authors cannot provide these pheromone levels, they argue that experimental results show m = 10 to
work well, and this is also the case in our own experiments.

3.3.4 Pheromone Initialization

Having discussed a number of different possibilities for updating pheromone information, we now illus-
trate some possibilities for initializing the pheromone values. Pheromone initialization is usually uniform,
that is, all pheromone values τij are assigned an initial value τ0 . A notable exception to this rule is studied in
Reference 32, where the initialization is performed after a number of preprocessing steps used to identify
local optima.
For uniform initialization, the value of τ0 must be seen in relation to the value for ∆ in order to
understand its impact on the behavior of the algorithm. Setting τ0  ∆ will cause the updates performed
during the initial phase of the ACO algorithm to have practically no effect. Instead, the ants are guided
almost exclusively by the heuristic information until all pheromone values have dropped to lower values.
Conversely, for τ0  ∆, there is a high risk of the ACO algorithm converging on the first good solution
found, since the first solution to receive some updates will be highly favored in future decisions due to
Equation (3.1). Setting

∆
τ0 = , (3.9)
ρ

which is the maximum value attainable in the long run for any τij via Equation (3.7), usually represents
a good trade-off between runtime and exploration.

3.3.5 Integration of Local Search

Local search is a possible augmentation for the ACO algorithm which can be used to further improve
the solutions found by the ants of an iteration (either all or only those used for pheromone intens-
ification) before the pheromone update is performed. Some common local search methods, which
have been especially popular with the TSP and other routing problems, are 2-opt, 3-opt [33], and
Lin–Kernighan (LK) [34]. When applicable, using local search represents a trade-off between quickly
improving solution quality by exploring the neighborhood of a solution and premature convergence to a
local optima. Ideally, the local search performed should be possible in O(n 2 ) steps for instances of size n,
since otherwise it becomes the dominant part of the search process instead of the ants constructing new
solutions.

3.4 Population Based Ant Colony Optimization

Population based Ant Colony Optimization (PACO), introduced in Reference 31, represents a con-
ceptually new and different update scheme to the ACO algorithm. Instead of explicitly maintaining a
pheromone matrix and using evaporation to gradually reduce old or unnecessary pheromone informa-
tion, the PACO algorithm maintains a population of solutions P = {πi , . . . , πk }, which is updated instead

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 49 — #9

3-50 Handbook of Bioinspired Algorithms and Applications

of the pheromone matrix, and from which the pheromone values τij can be derived via

τij = τ0 + ∆ · |{π ∈ P|(i, j) ∈ π }|, (3.10)

where ∆ is the update value used.

After m ants have constructed a solution, the best ant is taken as a candidate to update the popu-
lation. For the first k iterations, the candidate solution is added to P in all cases. Once |P| = k, it
becomes necessary to remove a solution from the population in order to maintain a population size
of k. Managing the Population as a FIFO queue, that is, removing the oldest solution and adding
the best of the iteration, has proven to be a very good update method [31] for population sizes of
k ≤ 5, outperforming MMAS and the standard ACO algorithm on the TSP and the QAP. Other
update methods are introduced in Reference 35. It is also possible to define one of the solutions
in the population as the elitist solution, which can only be removed if a better solution is added in
its place.
Since old information is removed via the deletion of solutions from the population instead of pher-
omone evaporation, the PACO algorithm can move to a completely new portion of the search space in
k iterations. Coupled with the explicit storage of solutions which are available for repair, this makes the
algorithm well suited for dynamic problems, see Reference 35. When a number of alternative solutions are
sought, for example, for multi-criteria optimization problems, PACO uses the population to store non-
dominated solutions, achieving a good approximation of the Pareto front [36,37]. Finally, the population
of solutions can be used as an interface with other population driven algorithms like Genetic Algorithms
[38]. See Reference 37 for a more detailed study of the PACO algorithm.

3.5 Applications
In this section, we present a survey of some of the noteworthy applications of ACO. Of course, this survey
cannot hope to present a complete overview, and the interested reader is referred to References 26, 39,
and 40 for additional surveys.
One of the earliest and most intuitive applications of ACO was the TSP [9]. Since all ACO algorithms
depend in some fashion on the metaphor of an ant moving through a graph [10], using the TSP to illustrate
the basic principles of Ant Algorithms is a logical choice, and it is also used as the introductory example in
Reference 39. ACO has delivered good results on many TSP instances, especially when combined with local
search [24]. However, due to the existence of very good heuristics like Lin–Kernighan [34] and polynomial
time approximation schemes [41] for the Euclidean TSP, ACO algorithms are not the best choice for this
problem class. The situation is better for the Sequential Ordering Problem (SOP), an extension of the
TSP, where the goal is to find a minimum weight Hamiltonian Path with precedence constraints among
the nodes. Here, a form of Ant Algorithm called Hybrid Ant System (HAS-ACO) [42] is currently one of
the best algorithms available. Other variations of the standard TSP, like the Probabilistic TSP (PTSP) and
Dynamic TSP (DTSP), are also handled well by ACO, using proper heuristic guidance [23] for the PTSP
and pheromone resetting strategies [43,44] or the PACO algorithm [35] for the DTSP.
Another problem related to the TSP is the Vehicle Routing Problem (VRP), in which a number of
customers must be serviced exactly once, and all vehicles begin and end their respective tours at a depot.
The goal is to minimize the number of vehicles while meeting constraints such as capacity per vehicle,
maximum tourlength per vehicle, and time windows. Solving this problem with Ant Systems was first
proposed in Reference 45, and further research has lead to a unified approach for VRPs [46], where the
Ant System is combined with an insertion heuristic from Reference 47.
The QAP, defined in Reference 11 and shown to be NP-hard in Reference 48, is a conceptually different
optimization problem compared to the TSP and its derivates in the sense that the pheromone matrix
is not interpreted in an item × item fashion, but rather as item × place. Applying Ant System to the
QAP was first undertaken in Reference 49, including a heuristic guidance scheme for the ants when

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 50 — #10

Ant Colony Optimization 3-51

constructing a solution. Adding local search to the AS algorithm was shown to be beneficial in Reference 50.
In Reference 51, the Hybrid Ant System (HAS) was introduced and applied to the QAP with good results.
The HAS algorithm uses ants to modify solutions instead of building them, and the pheromone values
are used to remember beneficial changes.
Another class of problems in which ACO algorithms have seen wide and successful application is in
scheduling problems. For scheduling with due dates, for example, the Single Machine Total Weighted
Tardiness Problem (SMTWTP), the pheromone matrix is also interpreted in a item × place fashion.
However, in contrast to the QAP, “place” in this case refers to the place in the schedule and not a physical
location. An ACO algorithm for the SMTWTP was applied in Reference 52, where ACO found the optimal
solution to 125 benchmark problems more often than the other heuristics evaluated. Ant Algorithms have
also been applied to somewhat more complex scheduling problems, for example, job shop scheduling
[53], flow shop scheduling [54], and, most notably, the Resource Constrained Project Scheduling Problem
(RCPSP) [27], where ACO was state of the art at the time of publishing.
Lately, the ACO algorithm has been extended to be able to deal with multi-criteria optimization
problems, in particular the Single Machine Total Tardiness with Setup Costs Problem. Here, two criteria
exist which must be optimized simultaneously, yet cannot be aggregated into a single optimization func-
tion. Rather, the algorithm needs to find a number of solutions which represent different trade-offs
between the two (or more) criteria. The PACO algorithm was modified for optimizing multi-criteria prob-
lems in Reference 36 with further improvements in Reference 28 yielding an algorithm which performs
very well and can deal with an arbitrary number of criteria.
So far, all the problems discussed have been permutation problems, which can be handled quite
well by ACO. However, some efforts have been undertaken to apply ACO to areas where solutions are
not permutations. As mentioned above, in Reference 12, ACO is successfully applied to the shortest
supersequence problem. Also, some partitioning problems, for example, graph coloring [55] and data
clustering [56], have been solved with ACO, with varying degrees of success. In Reference 57, ACO is used
as a generic algorithm for solving Constraint Satisfaction Problems (CSPs) with promising results.
As a final note, although not being an application, in the recent past it has been shown that under certain
conditions, some versions of ACO can provably find the optimal solution to the instance of a problem
with a probability arbitrarily close to 1 [58,59]. Although these results have no immediate impact on
the applicability of ACO algorithms, they put ACO on the same level as Simulated Annealing or Genetic
Algorithms in terms of solution finding capability. Note that with a lower bound greater than 0 on the
probability to find the solution or move closer to the solution in a given iteration, any method will find
the optimum with a probability arbitrarily close to 1, given enough time.

References
[1] R.H. Arnett. American Insects: A Handbook of the Insects of America North of Mexico. Van Nostrand
Rehinhold, New York, 1985.
[2] E.J. Fittkau and H. Klinge. On biomass and trophic structure of the central amazonian rain forest
ecosystem. Biotropica, 5: 2–14, 1973.
[3] K. von Frisch. The Dance Language and Orientation of Bees. Harvard University Press, 1967.
[4] M. Lüscher. Air-conditioned termite nests. Scientific American, 205: 138–145, 1961.
[5] P. Grassé. La reconstruction du nid et les coordinations interindividuelles chez bellicositermes
natalensis et cubitermes sp. la theorie de la stigmergie: essai d’interpretation du comportement
des termites constructeurs. Insectes Sociaux, 6: 41–81, 1959.
[6] G. Nicolis and I. Prigogine. Self-Organization in Non-Equilibrium Systems. John Wiley & Sons,
New York, 1977.
[7] S. Goss, S. Aron, J.-L. Deneubourg, and J. Pasteels. Self-organized shortcuts in the argentine ant.
Naturwissenschaften, 76: 579–581, 1989.
[8] J.-L. Deneubourg, S. Aron, S. Goss, and J. Pasteels. The self-organizing exploratory pattern of the
argentine ant. Journal of Insect Behavior, 3: 159–168, 1990.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 51 — #11

3-52 Handbook of Bioinspired Algorithms and Applications

[9] M. Dorigo. Ottimizzazione, apprendimento automatico, ed algoritmi basati su metafora naturale

(Optimization, learning and natural algorithms) (Italian). Ph.D. thesis. Dipartimento di Elettronica,
Politecnico di Milano, Italy, 1992.
[10] M. Dorigo and G.D. Caro. Ant colony optimization: A new meta-heuristic. In P.J. Angeline,
Z. Michalewicz, M. Schoenauer, X. Yao, and A. Zalzala, Eds., Congress on Evolutionary Computation
(CEG), Vol. 2. IEEE Press, Washington, 1999, pp. 1470–1477.
[11] T.C. Koopmans and M.J. Beckman. Assignment problems and the location of economic activities.
Econometrica, 25(1): 53–76, 1957.
[12] R. Michel and M. Middendorf. An ACO algorithm for the shortest common supersequence prob-
lem. In D. Corne, M. Dorigo, and F. Glover, Eds., New Ideas in Optimization, McGraw-Hill,
New York, 1999, pp. 51–62.
[13] G. Leguizamón and Z. Michalewicz. A new version of ant system for subset problems.
In P.J. Angeline, Z. Michalewicz, M. Schoenauer, X. Yao, and A. Zalzala, Eds., Congress of
Evolutionary Computation, Vol. 2. IEEE Press, Washington, 1999, pp. 1459–1464.
[14] D. Merkel and M. Middendorf. On the behaviour of ant algorithms: Studies on simple problems.
In Metaheuristics International Conference (MIC), July 2001, pp. 573–577.
[15] D. Merkle and M. Middendorf. A new approach to solve permutation scheduling problems with
ant colony optimization. In Applications of Evolutionary Computing — Evo Workshops, Vol. 2037
of Lecture Notes on Computer Science. Springer-Verlag, Heidelberg, 2001, pp. 484–493.
[16] M. Dorigo and L.M. Gambardella. Ant colonies for the traveling salesman problem. BioSystems,
43: 73–81, 1997.
[17] M. Dorigo and L.M. Gambardella. Ant colony system: A cooperative learning approach to the
traveling salesman problem. IEEE Transactions on Evolutionary Computation, 1: 53–66, 1997.
[18] C.J.C.H. Watkins and P. Dayan. Technical note: Q-learning. Machine Learning, 8: 279–292,
1992.
[19] T. Stützle and M. Dorigo. ACO algorithms for the traveling salesman problem. In K. Miettinen,
M. Makela, P. Neittaanmaki, and J. Periaux, Eds., Evolutionary Algorithms in Engineering and
Computer Science, Wiley, New Jersey, 1999, pp. 163–183.
[20] D. Merkle and M. Middendorf. An ant algorithm with a new pheromone evaluation rule for total
tardiness problems. In Applications of Evolutionary Computing — Evo Workshops, Vol. 1803 of
Lecture Notes on Computer Science. Springer-Verlag, Heidelberg, 2000, pp. 287–296.
[21] S. Fenet and C. Solnon. Searching for maximum cliques with ant colony optimization. In G. Raidl
et al., Eds., Applications of Evolutionary Computing — Evo Workshops, Vol. 2611 of Lecture Notes
on Computer Science, Springer, 2003, pp. 236–245.
[22] J. Branke and M. Guntsch. New ideas for applying ant colony optimization to the probabilistic
TSP. In Applications of Evolutionary Computing — Evo Workshops, Vol. 2611 of Lecture Notes in
Computer Science. Springer-Verlag, Heidelberg, 2003, pp. 165–175.
[23] J. Branke and M. Guntsch. Solving the probabilistic tsp with ant colony optimization. Journal of
Mathematical Modelling and Algorithms (JMMA), 3(4): 403–425, 2004.
[24] T. Stützle and H. Hoos. Max–min ant system. Future Generation Computer Systems, 16: 889–914,
2000.
[25] B. Bullnheimer, R. Hartl, and C. Strauss. A new rank based version of the ant system — a com-
putational study. Central European Journal for Operations Research and Economics, 7: 25–38,
1999.
[26] D. Corne, M. Dorigo, and F. Glover. New Ideas in Optimization. McGraw-Hill, New York, 1999.
[27] D. Merkel, M. Middendorf, and H. Schmeck. Ant colony optimization for resource-constrained
project scheduling. IEEE Transactions on Evolutionary Computation, 6: 333–346, 2002.
[28] J. Holland. Adapdation in Natural and Articial Systems. MIT Press, Ann Arbor, MI, 1975.
[29] T. Stützle and H.H. Hoos. The max–min ant system and local search for the traveling salesman
problem. In IEEE International Conference on Evolutionary Computation. IEEE Press, Piscataway,
NJ, 1997, pp. 309–314.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 52 — #12

Ant Colony Optimization 3-53

[30] V. Maniezzo. Exact and approximate nondeterministic tree-search procedures for the quadratic
assignment problem. INFORMS Journal on Computing, 11(4): 358–369, 1999.
[31] M. Guntsch and M. Middendorf. A population based approach for ACO. In S. Cagnoni et al., Eds.,
Applications of Evolutionary Computing — Evo Workshops, Vol. 2279 of Lecture Notes on Computer
Science. Springer, 2002, pp. 72–81.
[32] C. Solnon. Boosting ACO with a preprocessing step. In S. Cagnoni, J. Gottlieb, E. Hart,
M. Middendorf, and G. Raidl, Eds., Applications of Evolutionary Computing — Evo Workshops,
Vol. 2279. Springer-Verlag, Kinsale, Ireland, 2002, pp. 161–170.
[33] S. Lin. Computer solutions for the traveling salesman problem. Bell Systems Technical Journal,
44(10): 2245–2269, 1965.
[34] S. Lin and B. Kernighan. An effective heuristic algorithm for the traveling salesman problem.
Operations Research, 21: 498–516, 1973.
[35] M. Guntsch and M. Middendorf. Applying population based ACO to dynamic optimization
problems. In International Workshop on Ant Algorithms ANTS, Vol. 2463 of Lecture Notes on
Computer Science. Springer-Verlag, Heidelberg, 2002, pp. 111–122.
[36] M. Guntsch and M. Middendorf. Solving multi-criteria optimization problems with population-
based aco. In C. Fonseca, P. Fleming, E. Zitzler, K. Deb, and L. Thiele, Eds., Evolutionary Multi-
Criterion Optimization (EMO), Vol. 2632 of Lecture Notes on Computer Science. Springer, Berlin,
Heidelberg, 2003, pp. 464–478.
[37] M. Guntsch. Ant Algorithms in Stochastic and Multi-Criteria Environments, Ph.D. thesis. Insti-
tute AIFB, University of Karlsruhe, January 2004. http://www.ubka.uni-karlsruhe.de/cgibin/
psview?document=2004/wiwi/3.
[38] J. Branke, C. Barz, and I. Behrens. Ant-based crossover for permutation problems. In E. Cantu-Paz,
Ed., Genetic and Evolutionary Computation Conference, Vol. 2723 of Lecture Notes in Computer
Science. Springer, 2003, pp. 754–765.
[39] E. Bonabeau, M. Dorigo, and G. Théraulaz. Swarm Intelligence. Oxford University Press, Oxford,
1999.
[40] T. Stützle and M. Dorigo. The ant colony optimization metaheuristic: Algorithms, applica-
tions, and advances. In F. Glover and G. Kochenberger, Eds., Handbook of Metaheuristics. Kluwer
Academic Publishers, Norwell, MA, 2002.
[41] S. Arora. Polynomial time approximation schemes for Euclidean traveling salesman and other
geometric problems. Journal of the ACM, 45: 753–782, 1998.
[42] L.M. Gambardella and M. Dorigo. An ant colony system hybridized with a new local search for
the sequential ordering problem. IN-FORMS Journal on Computing, 12: 237–255, 2000.
[43] M. Guntsch, M. Middendorf, and H. Schmeck. An ant colony optimization approach to dynamic
TSP. In L. Spector et al., Eds., Genetic and Evolutionary Computation Conference (GECCO).
Morgan Kaufmann Publishers, San Francisco, CA, 2001, pp. 860–867.
[44] M. Guntsch and M. Middendorf. Pheromone modification strategies for ant algorithms applied
to dynamic TSP. In E. Boers et al., Eds., Applications of Evolutionary Computing — Evo Work-
shops, Vol. 2037 of Lecture Notes in Computer Science. Springer-Verlag, Heidelberg, 2000,
pp. 213–222.
[45] B. Bullnheimer, R. Hartl, and C. Strauss. An improved ant system algorithm for the vehicle routing
problem. Technical report, POM Working Paper No. 10/97, University of Vienna, 1997.
[46] M. Reimann, K. Doerner, and R. Hartl. Analyzing a unified ant system for the vrp and
some of its variants. In G. Raidl et al., Eds., Applications of Evolutionary Computing —
Evo Workshops, Vol. 2611 of Lecture Notes on Computer Science. Springer, Heidelberg, 2003,
pp. 300–310.
[47] M.M. Solomon. Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research, 35(2): 254–265, 1987.
[48] S. Sahni and T. Gonzales. P-complete approximation problems. Journal of ACM, 23(3): 555–565,
1976.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 53 — #13

3-54 Handbook of Bioinspired Algorithms and Applications

[49] V. Maniezzo and A. Colorni. The ant system applied to the quadratic assignment problem. IEEE
Transactions on Knowledge and Data Engineering, 5: 769–778, 1998.
[50] T. Stützle and H. Hoos. Max–min ant system and local search for combinatorial optimization
problems. In Metaheuristics International Conference (MIC), Kluwer Academic, Norwell, MA,
1997.
[51] L.M. Gambardella, E.D. Taillard, and M. Dorigo. Ant colonies for the QAP. Journal of Operations
Research Society, 2: 167–176, 1999.
[52] A. Bauer, B. Bullnheimer, R. Hartl, and C. Strauss. An ant colony optimization approach for the
single machine total tardiness problem. In Congress on Evolutionary Computation (CEC), IEEE
Press Piscataway, NJ, 1999, pp. 1445–1450.
[53] A. Colorni, M. Dorigo, V. Maniezzo, and M. Trubian. Ant system for job-shop scheduling.
JORBEL — Belgian Journal of Operations Research, Statistics and Computer Science, 34: 39–53,
1994.
[54] T. Stützle. An ant approach for the flow shop problem. In 6th European Congress on Intelligent
Techniques & Soft Computing (EUFIT), Vol. 3. Verlag Mainz, Aachen, 1998, pp. 1560–1564.
[55] A. Vesel and J. Zerovnik. How good can ants color graphs? Journal of Computing and Information
Technology — CIT, 8: 131–136, 2000.
[56] N. Monmarche. On data clustering with artificial ants. In A.A. Freitas, Ed., Data Mining with
Evolutionary Algorithms: Research Directions. AAAI Press, Orlando, FL, 1999, pp. 23–26.
[57] C. Solnon. Ants can solve constraint satisfaction problems. IEEE Transactions on Evolutionary
Computation, 6: 347–357, 2002.
[58] W. Gutjahr. ACO algorithms with guaranteed convergence to the optimal solution. Information
Processing Letters, 82: 145–153, 2002.
[59] T. Stützle and M. Dorigo. A short convergence proof for a class of ACO algorithms. IEEE
Transactions on Evolutionary Computation, 6: 358–365, 2002.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C003” — 2005/8/6 — 13:21 — page 54 — #14

 4
Swarm Intelligence

4.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-55

4.2 Swarm Intelligence Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-55
Emergent Strategy-Based Paradigm • Highly Distributed
Control Paradigm • Organizing Principles • Swarm
Intelligence Communication Forms • The Limitations of
Swarm Intelligence
Mohamed Belal 4.3 The Main Applications of Swarm Intelligence . . . . . . . . . 4-58
Jafaar Gaber Ant Colony Optimization • Particle Swarm Optimization
Hoda El-Sayed 4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-62
Abdullah Almojel References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-62

4.1 Introduction
Swarm Intelligence (SI) is a computational and behavioral metaphor for solving distributed problems
inspired from biological examples provided by social insects such as ants, termites, bees, and wasps and
by swarm, herd, flock, and shoal phenomena in vertebrates such as fish shoals and bird flocks.
In other words, SI is based on the principles underlying the behavior of natural systems consisting
of many agents, and exploiting local communication forms and highly distributed control. Thus, the
SI approach constitutes a very practical and powerful model that greatly simplifies the design of distributed
solutions to different kind of problems. In the last few years, SI principles have been successfully applied
to a series of applications including optimization algorithms, communications networks, and robotics.

4.2 Swarm Intelligence Overview

The SI approach emphasizes two important paradigms: the highly distributed control paradigm and the
emergent strategy-based paradigm.

4.2.1 Emergent Strategy-Based Paradigm

Collective behavior demonstrated by social insects (ants, bees, termites, etc.) often emerges from a small
set of simple low-level interactions between individuals, and between individuals and the environment.
The following example illustrates the concept of emergence. To solve a given task, for example, to sort
elements scattered on the ground, one can write an algorithm wherein a centralized part distributes the
task to achieve between a set of distributed agents. The centralized program, based on the global goal

4-55

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 55 — #1

4-56 Handbook of Bioinspired Algorithms and Applications

and plans, the current input, and the current state, collects agent results, analyzes them, and decides the
actions to be executed next.
One way of achieving the required task without a centralized part is by the addition of the individual
efforts of a multitude of agents who do not have any idea of the global objective to be reached; that is, there
is the emergence of collective behavior. Deneubourg et al. [1] introduced a model of sorting behavior
in ants. They found that simple model ants were able to sort into piles objects initially strewn randomly
across a plane. To be precise, near an anthill, one can observe that ants run in all directions to gather
corpses, to clean up their nests, or transport their eggs to order them by size etc. One can only imagine
that something, such as a special chemical marker, indicates to individual ants where to place their chips,
and allows them to distinguish an object already arranged from an object to be arranged. But how are
these markers placed and on what criteria? In fact, such interesting collective behavior can be mediated by
nothing more than similar, simple individual behavior. For example, F(1) each ant wanders a bit, (2) if an
ant meets an object and if it does not carry one, it takes it, and (3) if an ant transports an object and there
is a similar object in the same way in front of it, it deposits its load. By following these local strategic rules
with only local perceptual capacities, ants display the ability to perform global sorting and clustering of
objects.
Swarm Intelligence is a new way to control multiple agent systems. The swarm-type approach to
emergent strategy deals with large numbers of homogeneous agents, each of which has fairly limited
capabilities on its own. However, when many such simple agents are brought together, globally interesting
behavior can emerge as a result of the local interactions of the agents and the interactions between the
agents and the environment. A key research issue in such a scenario is determining the proper design of
the local control laws that will allow the collection of agents to solve a given problem.

4.2.2 Highly Distributed Control Paradigm

Swarm Intelligence is, intrinsically, a bottom-up approach. Bottom-up approaches are carried out by
programming large numbers of independent entities with relatively simple sets of rules. Brought together,
constructive behavior emerges, as it does in insects that create complex social behavior and structures
from the combined efforts of individuals with extremely limited intelligence. In contrast, the top-down
approach is based on the classic centralized method (e.g., the Client/Server approach), wherein central
coordination should take place. SI can be applied to fully distributed systems that consist of several
autonomous agents working together with local communication and minimal perception capabilities to
complete one or more tasks.

4.2.3 Organizing Principles

A study of the SI approach reveals a useful set of organizing principles that can guide the design of efficient
distributed applications for different kinds of problems. SI has the following notable features:

Autonomy: The system does not require outside management or maintenance. Individuals are
autonomous, controlling their own behavior both at the detector and effector levels in a
self-organized way.
Adaptability: Interactions between individuals can arise through direct or indirect communication
via the local environment; two individuals interact indirectly when one of them modifies the
environment and the other responds to the new environment at a later time. By exploiting such
local communication forms, individuals have the ability to detect changes in the environment
dynamically. They can then autonomously adapt their own behavior to these new changes. Thus,
swarm systems emphasize auto-configuration capabilities.
Scalability: SI abilities can be performed using groups consisting of a few, up to thousands of individuals
with the same control architecture.
Flexibility: No single individual of the swarm is essential, that is, any individual can be dynamically
added, removed, or replaced.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 56 — #2

Swarm Intelligence 4-57

Robustness: SI provides a good example of a highly distributed architecture that greatly enhances
robustness; no central coordination takes place, which means that there is no single point of failure.
Moreover, like most biological and social systems, and by combining scalability and flexibility
capabilities, the swarm system enables redundancy, which is essential for robustness.
Massively parallel: The swarm system is massively parallel and its functioning is truly distributed.
Tasks performed by each individual within its group are the same. If we view each individual
as a processing unit, SI architecture can be thought of as single instruction stream–multiple data
stream (SIMD) architecture or systolic networks.
Self-organization: Swarm systems emphasize self-organization capabilities. The intelligence exhibited is
not present in the individuals, but rather emerges somehow out of the entire swarm. In other words,
if we view every individual as a processing unit, solutions to problems obtained are not predefined
or preprogrammed but are determined collectively as a result of the running program.
Cost effectiveness: The swarm-type system consists of a finite collection of homogeneous agents, each
of which has fairly limited capabilities on its own. Also, each agent has the same capabilities and
control algorithm. It is clear that the autonomy and the highly distributed control afforded by the
swarm model greatly simplify the task of designing the implementation of parallel algorithms and
hardware. For example, for swarm-type multi-robotic systems, robots are relatively simple and
their design process effort can be kept minimal in terms of sensors, actuators, and resources for
computation and communication.

4.2.4 Swarm Intelligence Communication Forms

SI exploits local communication forms. Interactions between individuals can arise through direct or
indirect communication.
4.2.4.1 Indirect Communication
Indirect communication is implicit communication that takes place between individuals via the environ-
ment. This is known as Stigmergy communication. The Stigmergy concept describes a class of mechanisms
mediating animal–animal interactions through stimuli. When an animal does not explicitly distinguish
between its own activity and the activities of others, its actions include modification of its local envi-
ronment. By sensing its environment, an animal will perform an appropriate action as a response to the
new environment at a later time. Thus, interaction takes place in stages through changes in the local
environment. Note that the behavior of each insect can then be described as a series of stimulus–response
sequences.
There are two forms of Stigmergy. In the Stigmergy Sematectonic communication form, information
is communicated through physical modification of the environment. For example, opening a hole in
the body of a termitary causes a disruption of the termitary’s carefully maintained internal atmosphere
(intense gradients in temperature, humidity, carbon dioxide, and oxygen). Sensing some problem in
the body of the termitary, termites perform the rebuilding function and attack intruders while repairing
the breach in order to restore the termitary’s equilibrium.
In the second form of Stigmergy, some signal substance is deposited in the environment that makes
no direct contribution to the task being undertaken but is used to influence the subsequent behavior
that is task related [2]. For example, for building their nests, termites use highly volatile chemicals
called pheromones. Termites place tiny balls of mud near other balls of mud that have high pheromone
concentrations and, as a consequence, mounds develop. As the mounds grow, pheromones at the bases
evaporate and the termites bring the mud to the top, driving the height of some mounds upward of 30 ft
and causing adjacent mounds to meet in arches.
Pheromone-based Stigmergy is well developed in ants. Ants are capable of finding the shortest path
from a food source to the nest. Also, they are capable of adapting to changes in the environment, and find
a new shortest path once the old one is no longer feasible due to an obstacle [3]. Ants deposit a certain
amount of pheromone while walking, and each ant probabilistically prefers to follow a direction rich in
pheromone rather than a poorer one. Hence, the shorter path will receive a higher amount of pheromone

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 57 — #3

4-58 Handbook of Bioinspired Algorithms and Applications

and this will in turn cause a higher number of ants to choose the shorter path. This elementary behavior
of real ants explains how they can find the shortest path. The collective behavior that emerges is a form of
autocatalytic behavior (or positive feedback), whereby the more the ants follow the trail the more likely
they are to do so.

4.2.4.2 Direct Communication

Direct communication is explicit communication that can also take place between individuals. Examples
of such interactions are the waggle dance of the honeybee, using antennas, trophallaxis (food or liquid
exchange, e.g., mouth-to-mouth food exchange in honeybees), mandibular contact, visual contact,
chemical contact (the odor of nearby nest mates), etc.
Direct communication can be implemented by mobile wireless ad hoc networks. Individuals have
a very limited memory with the added feature that they are mobile; therefore, they can be considered
mobile agents. Indirect interactions through the environment can be thought of as distributed short-
term memory. Indeed, agents communicate through pheromone trails. When walking toward the colony
or food sources, ants will simply walk toward a high concentration of pheromone. The accumulated
pheromone then serves as a distributed shared memory. Note also that we need an analog for indirect
interaction through the local environment to implement the autoadaptive mechanism. Such a system
can adapt to changes in user behavior and system software through the pheromones. In other words,
pheromones will monitor the state of the machines and the network.

4.2.5 The Limitations of Swarm Intelligence

The swarm approach provides a rich source of inspiration and its principles are directly applicable to
computer systems. However, although it emphasizes auto-configuration, auto-organization, and adapt-
ability capabilities, the swarm-type approach remains useful for non-time-critical applications involving
numerous repetitions of the same activity over a relatively large area, such as finding the shortest path or
collecting rock samples on Mars [4]. Indeed, the swarm-type approach deals with the cooperation of large
numbers of homogeneous agents. Such approaches usually rely on mathematical convergence results
(such as the random walk) that reach the desired outcome over a sufficiently long period of time [4].
Notice that, in addition, the agents involved are homogeneous.

4.3 The Main Applications of Swarm Intelligence

Swarm Intelligence principles have been successfully applied in a variety of problem domains and applic-
ations. An example of successful research direction in SI is ant colony optimization (ACO), which focuses
on discrete optimization problems. Particle swarm optimization (PSO) is also an efficient and general
approach to solve nonlinear optimization problems with constraints. Another example of interesting
research direction is swarm robotics, where the focus is on applying SI techniques to the control of large
groups of cooperating autonomous robots.

4.3.1 Ant Colony Optimization

Ant colony optimization has been applied successfully to a large number of difficult, discrete optimization
problems including the traveling salesman, the quadratic assignment, scheduling, vehicle routing, etc.,
as well as to routing in telecommunication networks. Ant algorithms are a subset of SI. In other words,
ant algorithms can be viewed as multi-agent systems (ant colony), where agents (individual ants) solve
required tasks through cooperation in the same way that ants create complex social behavior from the
combined efforts of individuals.

4.3.1.1 Basic Ant Algorithm

The basic concept underlying the ant algorithm is inspired by the foraging behavior of real ants. When ants
search for food, they start from their nest and move at random toward the food. Ants use highly volatile

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 58 — #4

Swarm Intelligence 4-59

chemicals called pheromones to provide a sophisticated signaling system. While walking, ants deposit
quantities of pheromone marking the selected routes that they follow with a trail of the substance. When
an ant encounters an intersection, it has to decide which path to follow next. The concentration of
pheromone on a certain path is an indication of its usage. An ant chooses a path with a high probability
to follow and thereby reinforces it with a further quantity of pheromone. Over time, the concentration of
pheromone decreases due to diffusion. This foraging process is an autocatalytic process characterized by
a positive feedback loop, where the probability that an ant chooses any given path increases according to
the number of ants choosing the path on previous occasions. Ants that take the shortest path will reach
the food source first. On their way back to the nest, the ants again have to select a path. After a sufficiently
long period of time, the pheromone concentration on the shorter path will be higher than on other longer
paths. Thus, all the ants will finally choose the shorter path.
This ant foraging process can be used to find the shortest path in networks. Also, ants are capable of
adapting to changes in the environment, and find a new shortest path once the old one is no longer feasible
due to some obstacle. Thus, this process is appropriate to mobile ad hoc networks wherein link changes
occur frequently [5].
Let G = (V , E) be a connected graph with N = |V | nodes. The simple ant colony optimization meta-
heuristic can be used to find the shortest path between a source node vs and a destination node vd on the
graph G. The path length is defined by the number of nodes on the path. A variable ϕi,j corresponding to
the artificial pheromone concentration is associated with each edge (i, j). An ant located in node vi uses
pheromone ϕi,j to compute the probability of node vj being the next hop. This transition probability pi,j
is defined as:
 ϕi,j

 if j ∈ Vi ,
ϕi,j 
pi,j = j∈Vi with pi,j = 1 for 1 ≤ i ≤ N .


0 if j ∈
/ Vi , j∈Vi

During the process, ants deposit pheromone on the edges. In the simplest version of the algorithm, the
ants deposit a constant amount of pheromone, that is, the amount of pheromone of the edge (i, j) when
an ant moves from node vi to node vj is updated from the formula:

ϕi,j = ϕi,j + ϕ.

Moreover, like real ant pheromone, the artificial pheromone concentration should decrease over time.
In the simple ant algorithm this is shown by:

ϕi,j (t + τ ) = (1 − q)ϕi,j (t ), where 0 < q ≤ 1.

4.3.1.2 The Traveling Salesman Problem

The traveling salesman problem (TSP) is one of the most studied NP-hard problems in combinatorial
optimization. In the following, we show how the basic ant algorithm is adapted to solve this problem.
Consider a graph G = (N , E), where N is a set of nodes representing cities and E is a set of arcs fully
connecting the nodes. The distance between cities i and j is denoted dij . The TSP consists of finding a
minimal length Hamiltonian circuit on the graph G = (N , E). A Hamiltonian circuit of graph G is a
closed tour visiting once and only once all the n = |N | nodes of G, and its length is given by the sum of
the lengths of all the arcs of which it is composed.
A swarm of m ants build tours by executing n steps (one step by node). If all the iterations are done in
parallel, the m tours will be built in n iterations. The number of ants m at each iteration is kept constant.
The addition of new pheromone and pheromone evaporation are executed after all ants have completed
their tour, that is, after they have built a complete tour. Each ant has a memory that contains the list of
already visited cities. This list is used to define the set of cities that the ant located on city i still has to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 59 — #5

4-60 Handbook of Bioinspired Algorithms and Applications

visit. Recall that a feasible tour visits a city exactly once. Additionally, this allows the ant to cover the same
tour (i.e., path) to deposit delayed pheromones on the visited arcs. The probability with which an ant k
chooses to go from city i to city j while building its tour at the algorithm iteration t is:

 a (t ) (k)
  i,j if j ∈ Vi ,
(k)
pi,j (t ) = (k) ai,l (t )
l∈Vi


0 otherwise,

(k)
where Vi denotes the set of the neighborhood of node i that ant k has not visited yet.
The ant decision value ai,j (t ) is obtained by the composition of the local pheromone trail value with a
local heuristic value that supports the nearest neighbors as follows:

[ϕi,j (t )]α [dij ]−β

ai,j (t ) =  α −β
for j ∈ Ni ,
l∈Ni [ϕi,l (t )] [dil ]

where Ni is the set of neighbors of node i, and α and β are two parameters that control the relative weight
of the pheromone trail and the heuristic value. A heuristic value should measure or estimate the relevance
of adding an arc (i, j). A reasonable heuristic for TSP is 1/dij , the inverse of the distance between cities
i and j.
After all the ants have completed their tour, pheromone evaporation on arcs is executed. Each ant k
deposits a quantity of pheromone

 1
 if arc (i, j) ∈ T (k) (t ),
(k)
ϕij (t ) = L (k) (t )


0 otherwise,

where T (k) (t ) is the tour by ant k at iteration t and L (k) (t ) is its length. Note that the shorter the tour of
ant k, the greater is the amount of pheromone deposited.
The addition of new pheromone and pheromone evaporation are set by the following formula:


m
(k)
ϕij (t ) = (1 − q)ϕij (t − n) + ϕij (t ),
k=1

where q is the pheromone decay trail, 0 < q ≤ 1. The initial amount of pheromone ϕij (0) is set to the
same small positive value on all arcs. The suitable value for q is 0.5, which ensures a tradeoff between
sufficient positive feedback and the exploration of new cycles. For α and β, optimal values are α ≈ 1 and
1 ≤ β ≤ 5. Note that with α = 0, the algorithm corresponds to the classical greedy algorithm, and with
α > 2, all the agents converge to the same cycle, which is not necessarily optimal.
The comparison of this algorithm with other heuristics such as tabu search and simulated annealing on
small TSP problems (n = 30 cities), emphasizes its efficiency. The same technique presented here for TSP
was applied to solve other optimization problems such as job scheduling, QAP, and routing in networks
[2,5–8].
4.3.1.3 Comparison with Other Nature-Inspired Algorithms
A number of modern optimization techniques are inspired by nature. In simulated annealing modeled
from the thermodynamic behavior of solids, particles in solution space move under the control of
a randomized scheme, with probabilities according to some typically Boltzmann-type distribution.
Genetic algorithms (GAs) start with a randomly generated population and use crossover and mutation
operators to update it together with fitness function to evaluate the individuals. Neural networks (NNs)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 60 — #6

Swarm Intelligence 4-61

are a distributed learning technique in which the knowledge associated with a trained neural network is
not stored in any specific location but encoded in a distributed way across its weight matrix.
Ant colony optimization shares many common points with these nature-inspired approaches. ACO,
SA, and GA share the same update mechanism with random techniques. Randomness is present in the
fuzzy behavior of ants [8]. ACO shares with GA some organizing principles of social population such as
interaction and self-organization. ACO shares with NN trained networks the property that knowledge is
distributed throughout the network. Moreover, ACO, like NN, exhibits emergence capabilities [8].

4.3.2 Particle Swarm Optimization

Particle swarm optimization algorithms are also a subset of SI. The basic concept of PSO is inspired by the
social behavior of bird flocking and fish schooling. More precisely, PSO is a parallel evolutionary computa-
tion technique that provides a collaborative population-based search model. Individuals of the population
called particles fly around in a multidimensional search space. During flight, each particle adjusts its pos-
ition according to its own experience and according to the experience of a neighboring particle, moving
toward the best position encountered by itself or its neighbors. Thus, the PSO system combines local
search methods (through self-experience) with global search methods (through neighboring experience),
attempting to balance exploration and exploitation [9,10].
In practice, a PSO algorithm is initialized with a population of random candidate solutions or particles.
Two factors characterize a particle status on the search space: its position and its velocity. Additionally, the
performance of each particle is measured according to a problem-dependent fitness function (i.e., cost
function). Each particle is assigned a randomized velocity and is iteratively moved through the problem
space. It is attracted towards the location of the best fitness achieved so far by the particle itself and by the
location of the best fitness achieved so far across its neighborhood. Two versions of PSO exist depending
on the neighborhood topology used to exchange experience among particles. In the global version of the
algorithm, the neighborhood of the particle is the entire population (i.e., the entire swarm). In the local
version, the swarm is divided into overlapping neighborhoods of particles.

4.3.2.1 The Standard PSO Algorithm

The basic PSO algorithm can be described by the following equations:

(1) (2)
vi (t + 1) = avi (t ) + b1 r1 (pi − xi (t )) + b2 r2 (pi − xi (t )),
xi (t + 1) = xi (t ) + vi (k + 1),

where vi (t ) denotes the velocity of particle i, which represents the distance to be traveled by this particle
from its current position, that is, the difference between two successive particle positions; xi (t ) represents
(1) (2)
the particle position; pi represents its own previous best position; and pi is the best value obtained so
(2)
far by any particle among the neighbors. In the global version of the algorithm, pi represents the globally
best position among the whole swarm. Particles change their position (or state) in the following manner.
At iteration t , the velocity vi (t ) is updated based on its current value affected by a tuning parameter a,
and on a term that attracts the particle towards previously found best positions. The strength of attraction
is given by the coefficients b1 and b2 . The particle position xi (t ) is updated using its current value and
the newly computed velocity vi (t + 1). The three tuning parameters, a, b1 , and b2 , influence greatly the
algorithm performance. The inertia weight a is a user specified parameter, a large inertia weight pressures
towards global exploration in a new search area while a small inertia pressures towards fine tuning the
current search area. Positive constant acceleration coefficients (or learning factors) b1 and b2 control the
maximum step size of the particle, usually b1 = b2 = 2. Suitable selection of the tuning factors a, b1 ,
and b2 can provide a balance between the global (i.e., state space exploration), and local search (i.e., state
space exploitation). Random numbers r1 and r2 are selected in the range [0, 1] and they introduce useful
randomness for state space exploitation.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 61 — #7

4-62 Handbook of Bioinspired Algorithms and Applications

In the following, we show how this basic PSO algorithm can be adapted to solve an optimization
problem. The task-mapping problem (TMP) is one of the most studied NP-hard problems in distributed
computing.

4.3.2.2 PSO for Task Assignment Problems

An important issue in distributed computing is the efficient assignment of computations into different
processing elements. Given the graph of clustered tasks and the graph of the target distributed architecture,
one should find a mapping by placing the highly communicative tasks on adjacent nodes of the processor
network. It is well known that TMP is NP-hard. Salman et al. [9] use the PSO approach to address this
problem.
For each problem, particles should be designed such that potential solutions can be represented. This
is the key issue in designing a PSO algorithm. For a TMP, each particle is represented by a vector of length
equal to M , which corresponds to the number of vertices in the task-graph. The value of each element in
the particle is an integer in the range [1, N ], where N is the number of processors in the target architecture.
For example, for a task-graph with eight vertices and three processors, the particle representation (1, 2,
1, 2, 2, 3, 1, 3) means that tasks 1, 3, and 7 are assigned to processor 1; tasks 2, 4, and 5 are assigned
to processor 2; and tasks 6 and 8 are assigned to processor 3. Thus, a particle position corresponds to
an M -coordinate position in an M -dimensional search space. According to the PSO equations presented
here, each particle is iteratively moved (i.e., flies) through an M -dimensional search space [9].

4.3.2.3 Comparison with Other Nature-Inspired Algorithms

Particle swarm optimization and genetic algorithm have some common features. Both algorithms start
with a group of a randomly generated population and use fitness values to evaluate the individuals.
Moreover, both update the population and search for the optimum with random techniques. However,
PSO does not use genetic operators like crossover and mutation. Particles update themselves with internal
velocities (i.e., the difference between two successive particle positions). Note also that a PSO algorithm
uses a tracking memory that accelerates its convergence to the best solution, even in the local version in
most cases [11].
The performance of a PSO algorithm presented in Reference 9 is evaluated in comparison with a GA on
randomly generated mapping problem instances. The results showed that the quality of the PSO algorithm
solution is better than that of the GA’s in most of the test cases and, in addition, the PSO algorithm runs
faster than that of the GA’s.
Further, the PSO approach seems to be a promising method to train ANN. Indeed, PSO can be used to
replace the back-propagation learning algorithm in ANN. For example, Reference 12, showed that PSO is
faster and gets better results in most cases than the evolutionary approach.

4.4 Conclusion
Swarm Intelligence is a rich source of inspiration for our computer systems. Specifically, SI has many fea-
tures that are desirable for distributed computing. These include auto-configuration, auto-organization,
autonomy, scalability, flexibility, robustness, emergent behavior, and adaptability. These capabilities sug-
gest a wide variety of applications that can be solved by SI principles. We believe that the emergence
paradigm and the highly distributed control paradigm will be fruitful to new technologies, such as
nanotechnology, massively parallel supercomputers, embedded systems, and scalable systems for deep
space applications.

References
[1] J.L. Deneubourg, S. Goss, N. Franks, A. Sendova-Franks, C. Detrain, and L. Chrétien. The
dynamics of collective sorting, robot-like ants and ant-like robots. In Simulation of Animal

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 62 — #8

Swarm Intelligence 4-63

Behaviour: From Animal to Animals (J.A. Meyter and S. Wilson, Eds.), MIT Press, Cambridge,
MA, 1991, pp. 356–365.
[2] T. White. Swarm intelligence and problem solving in telecommunications. Canadian Artificial
Intelligence Magazine, spring, 1997.
[3] R. Beckers, J.L. Deneubourg, and S. Goss. Trails and U-turns in the selection of the shortest path
by the ant Lasius niger. Journal of Theoretical Biology, 159, 397–415, 1992.
[4] Lynne E. Parker. ALLIANCE: An architecture for fault tolerant multi-robot cooperation. IEEE
Transactions on Robotics and Automation, 14, 220–240, 1998.
[5] Mesut Günes and Otto Spanio. Ant-routing-algorithm for mobile multi-hop ad-hoc networks. In
Network Control and Engineering for Qos, Security and Mobility II, Kluwer Academic Publishers,
2003, pp. 120–138.
[6] Eric Bonabeau and Guy Theraulaz. Intelligence Collective, Edition HERMES, Paris, 1994.
[7] Marc Dorigo, Eric Bonabeau, and Guy Theraulaz. Ant algorithms and stigmergy. Future Generation
Computer Systems, 16, 851–871, 2000.
[8] B. Denby and S. Le Hégarat-Mascle. Swarm intelligence in optimization problems. Nuclear
Instruments and Methods in Physic Research Section A, 502, 364–368, 2003.
[9] Ayed Salman, Imtiaz Ahmad, and Sabah Al-Madani. Particle swarm optimization for task
assignment problem. Microprocessors and Microsystems, 26, 363–371, 2002.
[10] Ioan Cristian Trelea. The particle swarm optimization algorithm: Convergence analysis and
parameter selection. Information Processing Letters, 85, 317–325, 2003.
[11] X. Hu, R. Eberhart, and Y. Shi. Particle swarm with extended memory for multiobjective
optimization. In Proceedings of the IEEE Swarm Intelligence Symposium, 2003, Indianapolis,
IN, USA.
[12] L. Messerschmidt and A.P. Engelbrecht. Learning to play games using a PSO-based competitive
learning approach. In Proceedings of the 4th Asia-Pacific Conference on Simulated Evolution and
Learning, 2002.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C004” — 2005/8/6 — 13:21 — page 63 — #9

 5
Parallel Genetic
Programming:
Methodology, History,
and Application to
Real-Life Problems
5.1 Introduction to Genetic Programming . . . . . . . . . . . . . . . . 5-65
How GP Works
5.2 Models of Parallel and Distributed GP . . . . . . . . . . . . . . . . . 5-70
Parallelizing at the Fitness Level • Parallelizing at the
Population Level
5.3 The Length of Individuals and Measurement of
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-72
5.4 Parallel GP: The History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-73
5.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-74
Typical GP Problems
5.6 Real-Life Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-75
Placement and Routing in FPGA • Medical Knowledge
Representation by Means of GP
5.7 Concluding Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-82
Francisco Fernández Acknowledgment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-82
de Vega References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-82

5.1 Introduction to Genetic Programming

Software industry costs are mainly influenced by human resources required for developing software,
because software development is still supported by human expertise. Although prices for hardware have
dropped during the last 50 years, costs for hiring computer engineers have steadily increased in the same
period. Hence the advent of new techniques for automatic software development would be welcomed by
the software industry.
Scientists have been interested for long in the search for techniques capable of automatic software
development. The first attempts to endow computers with the ability to learn can be traced back to
the 1950s. The term machine learning, which was coined at that time, embodied the idea of computer
algorithms that can learn by themselves through experience [1].

5-65

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 65 — #1

5-66 Handbook of Bioinspired Algorithms and Applications

The finding of reliable methods for automatic programming would be a revolution for computer
science, and would completely change our perception of the software industry.
Although the proposal described in the 1950s was of interest, until very recently results have not unveiled
the potential that techniques of machine learning can attain. This potential has been clearly shown by a new
technique that can be considered part of “evolutionary algorithms” (EAs), genetic programming (GP) [2].
Genetic programming is aimed at evolving computer programs. It begins with a basic description of
the problem to be solved, after which the initialization process takes place; GP proceeds by automatically
generating a set of candidate solutions for the problem to be solved. Each of the candidate solutions
takes the shape of a computer program. Finally, GP enters a loop that evaluates each of the solutions,
selects the best one — according to a measurement criteria — and produces a new set of candidate
solutions employing the information contained in those selected solutions that acts as parents for the next
generation.
Not only GP but also the set of techniques comprised within the EA field resembles the natural evolution
of species in nature, as described by the Natural Evolution Theory [3].
Among the techniques that arose under the umbrella of natural evolution, genetic algorithms [4],
evolutionary programming [5], and evolution strategies [6,7] have matured and demonstrated their
usefulness.
During the last few years, GP has demonstrated not only its capability of automatically developing
software modules [8], but has also been employed for designing industrial products with outstanding
quality — such as electronic circuits that have been recently patented [9].
Although GP has proved its usefulness, the computational resources required for solving complex
problems may become huge. In such instances, improvement of the technique is sought by using concepts
borrowed from the parallel processing area. Not only GP but also EA have incorporated some degree of
parallelization when difficult problems are to be solved.
This chapter focuses on parallel genetic programming. We first provide some basic ideas about how GP
works, and then how it can be parallelized. Finally, after reviewing the history of the field, we show some
problems — benchmark and real-life problems — that are solved by Parallel GP.

5.1.1 How GP Works

5.1.1.1 The GP Algorithm
The GP algorithm is quite similar to other EAs. The idea behind this group of techniques is to improve a set
of candidate solutions for the problem at hand by means of a series of iterations. Thus, by applying several
genetic operators, partial solutions from different individuals in the population will combine to produce
an optimal or at least a useful solution for the problem to be solved. One of the main features of GP is that
every individual from the population is a computer program — a candidate program that tries to solve
the problem we are facing. At the end of the evolutionary process, the best individual from the population
will be the program that solves the problem in the best way among all the programs evaluated so far.
The main steps in the algorithm are the following:
1. Initialize the population of candidate solutions (individuals) for the problem to be solved.
2. Evaluate all of the individuals in the population and assign them a fitness value.
3. Select individuals in the population who will become parents. A selection algorithm is employed
in this step.
4. Apply genetic operations to the selected individuals in order to create new ones — descendants.
5. Introduce these new individuals into the new population.
6. If the population is not saturated go to step 3.
7. If the termination criterion is reached, then present the best individual as the output. Otherwise,
replace the existing population with the new population and go to step 3.
In the following sections, we will describe the different steps of the algorithm; but before that, an
introduction to program representation is offered.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 66 — #2

Parallel Genetic Programming 5-67

5.1.1.2 Terminal and Function Sets

As stated earlier, GP populations are made up of computer programs. Therefore, each individual is
a program, and usually, these programs are encoded be means of tree-like structures. The reason for
employing tree structures is simply a tradition that has been inherited from the times when Koza described
the technique [2]. Recently, other data structures have been successfully employed [10].
Each individual is made up of functions — internal nodes — and terminals — the leaves of the tree
(see Figure 5.1). Both sets have to be decided according to the problem to be solved. The intuition about
the general shape that a solution for the problem might adopt usually helps GP designer for the selection
of these sets. For instance, if we try to discover the shape of a function by means of GP when solving
a symbolic regression problem, probably, we should include arithmetic functions in the function set.
Nevertheless, no guarantee about the suitability of the set is frequently provided.
The terminal set is usually made up of variables and constant values that are significant for the problem
at hand.
Thus, the first concern for GP designers is to appropriately define the function and terminal sets: even
when the solution to the faced problem is not known, one should take care that a solution can be found
using the functions and terminals selected.
For making things easier, we will now describe a very simple problem, and we will see how terminal
and function sets can be decided, and how the problem is then solved.
For instance, suppose we want to solve this problem: find an algorithm that accepts an integer number
as the input and returns double the value of the input. Although the problem is extremely easy, it will
be illustrative for demonstrating the power of GP. The technique is capable of developing the algorithm
without any help from human programmers.
The first step is to build both terminal and function sets useful for the problem at hand. In this problem,
we know that the function that the algorithm has to implement is simply F (n) = (n + n), so we know that
the function set should include an operator for computing addition, so that F = {+}, and the terminal
set must include a parameter for the input T = {n}. Once the appropriate sets have been established, GP
can develop by means of the evolution of an individual like that depicted in the left part of Figure 5.2.
But of course, a different selection for the function and terminal sets might also lead to a solution for
the problem. Consider the following sets: F = {∗}, and T = {2, n}. This time we have not only included
a parameter in the terminal set, but also a constant value. Again, GP could obtain a solution like that
depicted in the right part of Figure 5.2.
Although one could think that any function and terminal set will be useful for solving the problem, this
is not always the case. If we choose the sets F = {−} and T = {n}, no matter which function we might
build (n, n − n, n − n − n, . . .) using these sets, the solution would never be found. Notice that we use a
binary subtraction operator; if we add the equivalent unary operator, the solution could be built as (n–n).

+
F = {+}
n + T = {n}

n n

FIGURE 5.1 Individuals in GP are usually encoded by means of tree structures.

+
*
n n 2 n
F ={+} F = {*}
T = {n} T = {2,n}

FIGURE 5.2 Two solutions for the problem, employing different terminal and function sets.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 67 — #3

5-68 Handbook of Bioinspired Algorithms and Applications

Even when the designer does not know the optimal solution, he should be able to extract some inform-
ation from the high-level specification of the problem, which helps him to define the appropriate function
and terminal sets.

5.1.1.3 Fitness Function

It is not only the task of defining the terminal and function sets that is crucial for obtaining good results
when using GP, a good fitness function is also required. The idea behind any EA is that good parents will
probably have similar or sometimes better children, and after a number of generations, descendants will
be good enough for solving the problem. But behind this idea, there is a need for a measure that correctly
classifies individuals from the population as good or bad. If the function in charge of measuring and
classifying individuals — who would later become parents — does not take into account good traits for
the problem, and instead assigns high fitness values to individuals with a poor performance, the algorithm
will probably never find a solution for the problem.
So, the fitness function is in charge of evaluating individuals from the population and assigning them
a value according to the performance obtained when solving the problem. High fitness values will thus
favor individuals for being selected when generating descendants. On the other hand, if a minimization
problem is faced, low fitness values are preferred instead.
For the problem described above, a good fitness function would be in charge of evaluating an indi-
vidual — computer program — and comparing the values returned by that program with those values that
should really be obtained, and are computed by using the objective function F (n) = {n + n}. Therefore,
the fitness function will compute an error value. This error value is considered the fitness value for the
individual; low fitness values are thus preferred for this problem.
The selection of a good fitness function is easy for the example shown here, because we know the
objective function. However, this is not a typical case, and the election of a good fitness function is a
difficult task, given that no clue about the shape of the solution for the problem is known.
After the fitness function has been defined, the GP algorithm possesses the capability of choosing
good individuals for breeding purposes. Good individuals have higher probabilities of producing good
and sometimes better descendants than bad individuals. Nevertheless, the GP algorithm choose’s good
individuals with a given probability, so that worse ones also have opportunities for transmitting their
genetic material to descendants.
The process of transmitting genetic material to descendants is not merely performed by copying selected
individuals. A set of genetic operations are applied instead, so that the information transmitted is somehow
altered and descendants will thus differ from parents while at the same time they will inherit their features.
Therefore, in the process of creating new candidate solutions, several genetic operators act as the source
for the variation that is required.

5.1.1.4 Genetic Operations

Genetic operators are the variation mechanisms that generate new candidate solutions, similar to their
parents but including some differences. If parents are good, they are allowed to breed new individuals
that share some features with them but that are not completely identical to them. Possibly, some of these
offspring can have better fitness than their parents. Some of them will also have worse fitness values, but
evolution will discard them as generations are computed. On the other hand, sometimes bad individuals
are selected for breeding, and may produce individuals that will help to solve the problem by introducing
new genetic material in the population.
A couple of genetic operators are usually in charge of this task: crossover and mutation. Crossover takes
two parents and mixes them up with a given probability so that new individuals are generated. Mutation
takes an individual and randomly changes a part of it with a certain probability.
In GP, when tree structures are employed, crossover exchanges two randomly selected subtrees (see
Figure 5.3) of the parents, while mutation randomly modifies a subtree from an individual (see Figure 5.4).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 68 — #4

Parallel Genetic Programming 5-69

Crossover points are

+ randomly selected +

+ n n +

Parent A Parent B
n n n n

+
Subtrees are exchanged +
in decendants

n n n +

Decendant #1
+ n
Decendant #2

Crossover nn n

FIGURE 5.3 Tree-based GP crossover. The two parents exchange one of their subtrees in order to generate two new
offspring.

+ +

+ n n n

Subtree selected New randomly

n n for mutation generated subtree

Mutation

FIGURE 5.4 Tree-based GP mutation. A subtree is randomly selected from the parent and replaced with another
randomly generated subtree.

5.1.1.5 Termination Criterion

The GP algorithm may be stopped when a satisfactory solution has been found, or after a given number
of generations have been computed.
Although the first criteria is the preferred one, given that it means that the solution has been found,
the second one is also employed because computational resources available are always limited. Frequently,
the process has to be stopped when no more time for computing is available. This latter case is a problem
for GP researchers, and also for any research dealing with EAs. The fact is that EAs are useful for difficult
problems, and sometimes the difficulty of the problem is so large that the computational resources required
for solving it are not available.
Researchers have been fighting this problem during the last decade and have tried to improve the
algorithms for tackling more difficult problems. One of the more successful way of improving algorithms
is by adding some degree of parallelization. Parallel EAs thus allow us to employ parallel architectures
and distributed computing resources. In the next section we describe the development of a parallel
version of GP.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 69 — #5

5-70 Handbook of Bioinspired Algorithms and Applications

5.2 Models of Parallel and Distributed GP

Typically, the reason for parallelizing any computer algorithm is to achieve time saving. In EAs there is a
second reason: the implementation of the new parallel EA may improve the convergence process, so that
a smaller number of iterations will be required for finding a solution of similar quality as when using the
classic algorithm. Although this second improvement is not always present, some smart decisions when
developing the parallel algorithm will provide these two advantages.
The necessity for a parallel implementation is frequently crucial, given that a large set of individuals
has to be evaluated for many generations, and large amounts of computational resources are consequently
required. In GP, parallelization is even more important: researchers have demonstrated that the size of
individuals tends to grow when generations are progressively computed and evolved [11]. Therefore,
computational resources required for the GP algorithm will not be constant along generations, but will
increase as they are computed.
When parallelism is added to any EA, different models can be employed, depending upon the operations
that are parallelized. Usually, parallelization can be applied at the level of the population and at the level
of the fitness evaluation. In the following sections, we briefly describe these models, concentrating on GP;
for a more detailed discussion, see Reference 12.

5.2.1 Parallelizing at the Fitness Level

The simplest way of applying parallelism to the GP algorithm, is to distribute the evaluation of individuals
among several processors (see Figure 5.5). The main algorithm — the master — is computed in one
processor, and the individuals are sent to other processors that in turn will evaluate and return the fitness
values to the main process. This model is useful when the evaluation process is the most time-consuming
step of the algorithm, which is usually the case in many real-life problems.
A well-known feature of GP individuals is the differences in size and complexity when compared to other
individuals belonging to the same population. This requires the application of a load balancing policy (see
for instance Reference 13). Load balancing can automatically be obtained if steady-state reproduction is
used instead of generational reproduction.

3 0 5 7 1
3 — Fitness values are returned
to the main processor
1— Individuals are sent
for evaluation

2 — Each processor evaluates one individual

FIGURE 5.5 Parallelizing at the fitness level. Different processors or workstations are in charge of evaluating
individuals, while the main processor runs the GP algorithm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 70 — #6

Parallel Genetic Programming 5-71

We have to bear in mind that the previously described parallel algorithm is basically the same as the
sequential version, the only difference is the parallel evaluation of individuals.
In the next section we describe other ways of developing parallel GP, but with some changes to the basic
algorithm, which will also help to solve the problem at hand more quickly.

5.2.2 Parallelizing at the Population Level

Populations of any species in nature feature a spatial distribution that usually depends on the orography
of the landscape. They are said to be grouped in demes, semi-independent groups of individuals that
are relatively isolated and scarcely interact with other neighboring demes by migrating some of their
individuals.
This idea can be applied to GP, and different models have been proposed during the last few years.
Although a detailed description of the algorithms can be found in Reference 14, we present here the main
features of the new models.

5.2.2.1 The Island Model

The idea of having several semi-independent groups of individuals can also be applied to GP (see
Figure 5.6). We could thus build a parallel algorithm by distributing the whole population of individuals
among the processors available. Each of the processors will thus be in charge of applying evolution to its
deme, and sometimes will exchange good individuals with other neighboring processors. This model is
usually called the Island Model [15,16].
The idea is that each of the subpopulation focuses on a different area of the search space, because the
convergence process within each of the deme takes a different path; and this is an important difference
from the classic algorithm. Researchers have found that this difference helps to solve problems tackled
by GP in a smaller number of steps, which is a new advantage for the parallel model. Researchers have
also found that this model helps to add diversity to populations, which is a good feature for avoiding
premature convergence.
Different connection topologies have been employed during the last few years. The most common
ones are rings, two- and three-dimensional meshes, stars and hypercubes, although recently a random
communication topology has also been defined in which a given subpopulation sends its emigrants to
another randomly chosen subpopulation [17]. The most common replacement policy replaces the worst k
individuals in the receiving population with k immigrants, who are the best k individuals of their original
island.
We may notice that the new parallel GP algorithm requires several new parameters:

1. Number of subpopulations
2. Frequency of exchange
3. Number of exchanged individuals
4. The communication topology
5. Replacement policy

FIGURE 5.6 General island topology.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 71 — #7

5-72 Handbook of Bioinspired Algorithms and Applications

Some of these important parameters have been studied recently [14]. Researchers have found that a good
choice for the previous parameters is to send 10% of individuals from each subpopulation to an adjacent
one every 10 generations. The communication topology does not affect significantly the convergence
process. On the other hand, a compromise between the number of subpopulations and the total number
of individuals employed for solving the problem has to be adopted: if each of the subpopulations is made
up of a small number of individuals, the exploration phase performed by each of the subpopulation will
not be satisfactory.
A difference may be established between the parallel model we use and the parallel architecture we
employ to run that algorithm. When dealing with distributed EAs, such as island GP, one can run the
algorithm on both distributed memory multiprocessor machines and also on the network of workstations
(NOWs).
In these architectures, the address spaces of each processor are separated and communication between
processors must be implemented through some form of message passing. NOWs are widely used because
of their low cost and ubiquity, although their performances are limited by communication latencies and
heterogeneous workload distribution. In the case of parallel EAs, given that the communication step is
not the most time-consuming part of the algorithm, and the migration step is rarely performed, these
kinds of low-cost architecture are useful enough.
The migrations between the different demes can be implemented, for example, using the Message
Passing Interface Standard (MPI) with synchronous communication operations, that is, each island
runs a standard generational GP and individuals are exchanged at fixed synchronization points between
generations. Implementation details can be found in Reference 18.
Researchers have also found that the Island Model may obtain better results than the panmictic model —
classic model — even when it is run on a standard sequential machine [19]. The improvement does not
come from the parallelization of the operations, because only one processor is employed, but from the
change in the model. The Island Model introduces the migration step that helps to conserve diversity
during the run, which is good for the finding of better solutions.
Other spatial distributions are also available for parallel EAs. For instance, one could distribute each of
the individuals from the population on a two-dimensional grid (other dimensions are also possible). The
idea is that each of the individuals interact only with their direct neighbors. Therefore, reproduction and
mating take place locally for each of the individuals.
The model allows the slow diffusion of information from good individuals across the grid, and semi-
isolated niches of individuals arise in different areas of the space.

5.3 The Length of Individuals and Measurement of Results

As described earlier, GP individuals may feature different sizes and complexities because variable-size
individuals are employed. This means that the evaluation of a population of individuals does not neces-
sarily always require the same amount of computing resources, even when the number of individuals is
always the same.
There is a second factor that directly influences the way we should measure results: Researchers have
found that the length of individuals progressively increases in a given experiment as generations are com-
puted [20]. Therefore, even for a particular experiment, different generations require different computing
time to be evaluated. So, if one wants to compare a technique with another one, the comparison has to
be carefully carried out. This is the case for parallel GP: we are not only interested in solving a particular
problem, but also want to know the power of the new technique when compared with the more common
implementation of GP.
Traditionally, researchers present results comparing fitness values — quality of solutions — and the
number of generations required to find that particular solution. In the case of GP, a better way of comparing
results is by computing the total number of nodes evaluated until a given fitness value is found [17]. This
is particularly useful for evaluating parallel GP when compared with GP, because the measure is not

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 72 — #8

Parallel Genetic Programming 5-73

biased by the different speeds of each of the processors or computers in charge of evaluating each of the
subpopulation. This measure will by itself show us the advantage of the model, and this advantage will
be observed even when it is employed in a sequential machine. Of course, in a given architecture and a
real-life problem, the time required for obtaining a particular solution is the right measure: researchers
try to obtain a solution for a problem as soon as possible.
For computing fitness values, a useful figure is the Mean Best Fitness (MBF, the average over a number
of runs) of the best fitness value at the end of the run. Nevertheless, when difficult problems are evaluated,
none knows in advance whether the global optimum has been obtained or not. Therefore, the idea is to
take the measure when a specified amount of computational effort has been spent.
In the comparisons we show below, we employ the measure described here: MBF versus computing
effort (total number of nodes evaluated).

5.4 Parallel GP: The History

A number of researchers have applied ideas borrowed from the parallel processing field to GP. Probably,
the first attempt to develop a parallel version of GP was carried out by Tufts in 1993 [21]. The idea
described was simply to evaluate simultaneously a number of individuals from the GP population, by
using several processors. The problem addressed was a classification over a set of customers, in order to
predict credit card attrition.
A couple of years later, Juillè and Pollack described a parallel implementation of GP on a fine-grained
SIMD architecture [22]. They employed a MasPar MP-2 Computer. Although they first presented a
study on the parallel evaluation of S-expressions on the computer, which basically corresponds to a
parallelization based on the fitness level, they also described a proposal based on the Island Model,
employing a ring topology for communicating subpopulations. They demonstrated the usefulness of the
implementation by using the Tic-Tac-Toe problem.
Koza and Andre [23] described latter an implementation of parallel GP based on the Island Model
employing a network of Transputers — single VLSI devices containing 32-bit on-chip processor, on-chip
memory, and several independent serial bi-direction physical on-chip communication links. The physical
topology was based on a central supervisory process (the boss process) and 64 transputers, each running
the basic GP algorithm, and the migration step for sending and receiving individuals. They employed
a cellular topology, so that every node was physically connected to four neighbors in the network. The
problem selected for the tests was the even-5-parity function. Several migration rates were employed.
The main conclusion was that parallelization delivered more than linear speedup in solving the problem.
Nevertheless, this result was questioned a couple of years later by Punch [24]. The fact is that the new
implementation of the algorithm not only uses a parallel architecture but also modifies the main algorithm,
so that the new one helps to improve the performances obtained when solving a problem.
The same year other implementation based on the bulk synchronous programming (BSP) model was
presented [25], although this time no super-linear speedup was reported. The results showed that the
Island Model achieved better speedups than panmictic GP — the classic version.
Another version of parallel GP was described by Stoffel and Spector [23]. Instead of using tree-based GP,
they employed linear programs that were run on a stack-based virtual machine. By means of a symbolic
regression problem, authors showed that parallel GP can save computing effort.
In 1997, Oussaidène et al. [13] applied a combination of fitness level parallelization and the island-based
approach of parallel GP to trading model induction. They simultaneously tackled the load balancing
problem that arises because of the different shapes and sizes of individuals in GP, which affects the
performance of the parallel system when different individuals are evaluated on different processors. The
proposal was to employ a dynamic scheduling algorithm based on a steady-state version of GP.
In a paper published by Punch [24] in 1998, the super-linear speedup reported by Koza [23] is questioned
by a series of tests. The benchmark problems selected there were the royal tree and the ant problem.
According to Punch, the features of the problem influences the results obtained. Particularly, the number

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 73 — #9

5-74 Handbook of Bioinspired Algorithms and Applications

of solutions in the search space greatly influences the performance of the Island Model. Therefore, the
conclusion was that multiple-solution problems would be more amenable to multiple populations than
single-solution problems. On the other hand, nondeceptive problems would also be more amenable to
multiple populations than deceptive problems.
In the aforementioned papers, different parallel GP models are employed to study the improvement
achieved when different benchmark and real-life problems are tackled. Nevertheless, no indepth study on
specific parameters of the new models is presented until 2000. For instance, Tongchim and Chongstitvatana
[26] studied synchronous and asynchronous versions of parallel GP. Results demonstrated that the parallel
asynchronous algorithms obtains better results.
On the other hand, a whole study on the migration policies, topology, and other important parameters
for the Island Model is presented for GP in 2003 [14]. Similarly, a study is performed, employing a parallel
version of GP based on a cellular approach [27].
Plenty of results dealing with the application of parallel GP to real-life problems, and also papers
focusing on any of the important parameters for the new model are available today. Recently, researchers
have shown that the technique is even capable of solving some problems and obtaining solutions better
than any other techniques previously invented by human being. The results are so impressive that they
have even been patented. For instance, Koza [9] describes several applications of parallel GP that had led
to solutions to problems; these applications are novel and useful enough to be patented. Particularly, Koza
[28] describes some results that can be considered as inventions. He presented several analogue circuits
that have been discovered by means of parallel GP. Some of those circuits — later patented — can probably
be considered as the first inventions developed by computers.
In the following section, we show by means of two examples, how parallel GP can be applied to solving
a real-life problem.

5.5 Applications
In this section, we briefly describe several benchmark problems that have been traditionally employed for
testing GP performances, and also present two different real-life problems that have been addressed by
means of parallel GP.

5.5.1 Typical GP Problems

We briefly describe several benchmark problems that have been traditionally used for experimenting with
GP: the even parity problem, the artificial ant on the Santa Fe trail problem [2,20], the symbolic regression
problem [24] and the royal tree problem [24]. They are useful for analyzing properties of genetic operators
or different implementations of GP, and their description are provided as a reference for readers interested
in GP. Although we don’t show here results obtained with these problems, interested readers may refer to
References 19 and 17. We focus instead on a couple of real-life problems that have been tackled by means
of parallel GP.

Even Parity k Problem. The boolean even parity k function of k boolean arguments returns true if an even
number of its boolean arguments evaluates to true, otherwise it returns false. If k = 4, then 16 fitness cases
must be checked to evaluate the fitness of an individual. The fitness is computed as 16 minus the number of
hits over the 16 cases. Thus a perfect individual has fitness 0, while the worst individual has fitness 16. The
set of functions to be employed for GP individuals might be the following one: F = {AND, OR, NOT }.
The terminal set in this problem is composed of k different boolean variables T = {v1 , . . . , vN }.
Artificial Ant Problem on the Santa Fe Trail. In this problem, an artificial ant is placed on a 32 × 32
toroidal grid. Some of the cells from the grid contain food pellets. The goal is to find a navigation
strategy for the ant that maximizes its food intake. Typically, the function set employed is the following
one: F = {if − food − ahead} while the terminal set is T = {left, right, forward} as described in Ref-
erence 2. As fitness function, we use the total number of food pellets lying on the trail (89) minus the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 74 — #10

Parallel Genetic Programming 5-75

C
B B B B
A A A A A A A A A

X X X X X X X X X

FIGURE 5.7 The Royal Tree Problem. Examples of perfect trees.

amount of food eaten by the ant from the path. This turns the problem into a minimization one, like the
previous one.
Symbolic Regression Problem. The problem aims to find a program that matches a given equation. We
employ the classic polynomial equation f (x) = x 4 + x 3 + x 2 + x, and the input set is composed
of 1000 fitness cases. For this problem, the set of functions used for GP individuals is the following:
F = {∗, //, +, −}, where // is like / but returns 0 instead of error when the divisor is equal to 0, thus
allowing syntactic closure. The fitness computes the sum of the square errors at each test point. Again,
lower fitness means a better solution.
The Royal Tree Problem. This Problem [24] is commonly used as a standard function for testing the
effectiveness of GP. It consists of a single base function that is specialized into as many cases as necessary,
depending on the desired complexity of the resulting problem.
A series of functions, a, b, c, etc., with increasing arity are defined. (An a function has arity 1, a
b function has arity 2, and so on.) A number of terminals x, y, and z are also defined.
A level-a tree is an a root node with a single x child. A level-b tree is a b root node with two level a trees
children. A level-c tree is a c root node with three level-b trees as children. A level-e tree has depth 5 and
326 nodes, while a level-f tree has depth 6 and 1927 nodes. Perfect trees are defined as shown in Figure 5.7.
The raw fitness of a subtree is the score of its root. Each function calculates its score by adding up the
weighted scores of its direct children. If the child is a perfect tree of the appropriate level (for instance, a
complete level-c tree beneath a d node), then the score of that subtree, times a FullBonus weight, is added
to the score of the root. If the child’s root is incorrect, then the weight is Penalty. After scoring the root,
if the function is itself the root of a perfect tree, the final sum is multiplied by CompleteBonus. Typical
values used are: FullBonus = 2, PartialBonus = 1, Penalty = 1/3, and CompleteBonus = 2. The score
base case is a level-a tree, which has a score of 4 (the a–x connection is worth 1 times the FullBonus, times
the CompleteBonus).

5.6 Real-Life Applications

As described in Section 5.4, there is plenty of research dealing with parallel GP. We describe here a couple
of real-life problems that have been tackled by using it. Results show that the technique can successfully
solve the problems addressed.

5.6.1 Placement and Routing in FPGA

Field Programmable Gate Arrays (FPGAs) are integrated devices used for the implementation of digital
circuits by means of a configuration or programming process. Several manufacturers and different kinds
of FPGAs are available.
One of the best known is the island-based FPGA (this island has nothing to do with island-based EAs).
This model includes three main components: configurable logic blocks (CLBs), input–output blocks
(IOBs), and connection blocks (see Figure 5.8). Configurable logic blocks are used to implement all
the logic circuitry — they have different configuration possibilities, and are positioned like matrix on
the FPGA.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 75 — #11

5-76 Handbook of Bioinspired Algorithms and Applications

FIGURE 5.8 Island-based FPGA.

The IOBs allow the connection of the circuit implemented by the CLBs with any external system. Finally,
the connection blocks (switch-boxes and interconnection lines) are employed for the internal routing of
the circuit.
One of the main steps in the FPGA design process is the placement and routing. We present a meth-
odology that is based on parallel GP. The methodology has also been employed for tackling Multi-FPGA
Systems Synthesis [29].
The problem we try to solve begins with a circuit description, and the goal is to place components
and wires in an FPGA. Genetic programming is thus in charge of encoding circuits, so that a graph —
circuit — is described by means of a tree — GP individual. In the following, we describe how graphs are
encoded by means of trees.
Although several authors have implemented GP in hardware [30,31], the idea here is completely
different: we use GP for implementing circuits on hardware.

5.6.1.1 Circuits Encoding Using Trees

When implementing a circuit on an FPGA, each of the circuit components has to be implemented into
a different CLB, and then the CLBs have to be connected according to the circuit topology. Given that
circuits are encoded by means of trees, and that evolution will generate new circuits, a fitness function is
required for analyzing these circuits, deciding if they are correct or not, and their degree of resemblance
with the circuit that is being to implemented.
Any circuit is made up of components and connections. Given that components compute very easy
logic functions, any of them can be implemented into any CLB from the FPGA. Therefore, we can describe
a given circuit in a way similar to the example depicted in Figure 5.9. This means that we only have to
connect CLBs from the FPGA according to the interconnection model that a given circuit implements, and
then we can configure each of the CLBs with the function that each component performs in the circuit.
Circuits have to be encoded by means of trees, because we employ tree-based GP. We can first label
each component from the circuit with a number, and then assign components’ labels to the ends of
wires connected to them (as shown in Figure 5.9). Wires can now be disconnected without losing any
information. All the wires can be included within a tree by connecting each of the wires as a branch of
the tree and keeping them all together in the same tree. The circuit can be easly rebuilt later by using the
labels as a guide.
By labeling both extremes of wires, we will have all the information required for reconstructing the
circuits. This way of representing circuits allows us to go back and construct the real graph. Moreover, any
given tree, randomly generated, will always correspond to a particular graph, regardless of the usefulness
of the associated circuit. In this proposal, each node from the tree is representing a connection, and each
branch is representing a wire.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 76 — #12

Parallel Genetic Programming 5-77

i1
1
i2 5
2 01
7 8
i3
3
i4 6
4

FIGURE 5.9 Representing a circuit with black boxes and labeling connections.

FIGURE 5.10 Mapping an individual into a circuit.

5.6.1.2 GP Sets
The function set for our problem contains only one element: F = {SW}, Similarly, the terminal set
contains only one element T = {CLB}. But SW and CLB may be interpreted differently depending on the
position of the node within a tree. Sometimes a terminal node corresponds to an IOB connection, while
sometimes it corresponds to a CLB connection in the FPGA. Similarly, an internal node — SW node —
sometimes corresponds to a CLB connection (the first node in the branch), while others affect switch
connections in the FPGA (internal node in a branch, see Figure 5.10). Each of the nodes in the tree will
thus contain different information:

1. If we are dealing with a terminal node, it will have information about the position of CLBs, the
number of pins selected, the number of wires to which it is connected, and the direction we are
taking when placing the wire.
2. If we are, instead, in a function node, it will have information about the direction we are taking.
This information enables us to establish the switch connection, or in the case of the first node of
the branch, the number of the pin where the connection ends.

5.6.1.3 Evaluating Individuals

To evaluate an individual we must convert the genotype (tree structure) to the phenotype (circuit in the
FPGA), and then compare it to the circuit provided by the partitioning algorithm. We developed an FPGA
simulator for this task. This software allows us to simulate any circuit and to check its resemblance to
other circuits. Therefore, this software tool is in charge of taking an individual from the population and
evaluating every branch from the tree in a sequential way, establishing the connections that each branch
specifies. Circuits are thus mapped by visiting each of the useful nodes of the trees and making connections
on the virtual FPGA, thus obtaining phenotype.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 77 — #13

5-78 Handbook of Bioinspired Algorithms and Applications

5.6.1.4 Results
Figure 5.9 graphically depicts one of the test circuits that has been used for validating the methodology.
The main parameters employed were the following: maximum number of generations equal to 500,
maximum tree depth equal to 30, steady state, tournament selection of size 10, crossover probability
equal to 98%, mutation probability equal to 2%, ramped half-and-half initialization, elitism (i.e., the best
individual has been added to the new population at each generation). We employed 5 subpopulations of
500 individuals each, with a period of migration of 10 generations. The GP tool we used is described in
Reference 18.
Figure 5.11 shows one of the proposed solutions among those obtained with parallel GP for the circuit.
A very important fact is that each of the solutions that GP found possesses different features, such as area
of the FPGA used and position of the input/output terminals. This means that the methodology could
easily be adapted for managing typical constraints in FPGA placement and routing.
Figure 5.12 presents a comparison of parallel GP and classic GP when applied to the problem of
placement and routing on FPGAs. We can see that parallel GP, employing 5 populations and 500 individuals
per population, achieved better convergence results than GP with 2500 individuals — the same total
amount of individuals: PGP converges more quickly and obtains slightly better results.
The methodology has been successfully applied to Multi-FPGAs System Synthesis [29]. Figure 5.13
shows a picture of the device that was built for testing the methodology.

4 i4

8 o1

6 3 i3

7 1 i1

2 i2

FIGURE 5.11 A solution generated by PGP.

FIGURE 5.12 Comparison between parallel GP — 5 pop., 500 individuals each — and classic GP — 2500 individuals.
50 runs have been performed and averaged for each curve.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 78 — #14

Parallel Genetic Programming 5-79

Parallel GP
10 × 105
GP

6 × 105

Fitness 2 × 105

5 × 105 15 × 105 25 × 105

Effort

FIGURE 5.13 A Multi-FPGA board designed for testing the methodology.

5.6.2 Medical Knowledge Representation by Means of GP

In this section we present another real-life problem that has been solved by means of parallel GP. We apply
GP to the acquisition of medical knowledge. We are focussing on burn diagnosing.
An adaptive system capable of classifying is developed by means of parallel GP. It uses a set of parameters,
obtained by specialist doctors, to predict the evolution of a burn according to its initial stages. The system
is first trained with a set of parameters and the results of evolutions that have been recorded over a set
of clinical cases. Once the system is trained, it is an aid to deciding how new cases will probably evolve.
Thanks to the use of parallel GP, an explicit expression of the input parameter is provided, and this explicit
expression takes the form of a decision tree, which will be incorporated into software tools that help
physicians in their everyday work.
Of course the aim here is not to suppress the task of specialists, but to help them make more accurate
diagnoses. Furthermore, developing software tools for medical diagnosis may help nonspecialist doctors
when immediate treatment must be applied.

5.6.2.1 The Problem of Burn Diagnosing

When somebody suffers a burn it is necessary to find out its degree, in order to apply the best treatment
from the beginning. This classification problem is recognized as being difficult because it is not always
clear how much tissue damage there is and how it will evolve.
If we are interested in developing an automatic classifier system capable of deciding the diagnosis of a
burn, this system must first be able to extract information from a burn by analyzing a picture of it, and
second embody that information in a knowledge system. This knowledge system could take the form of a
decision tree. Bearing in mind the feasibility of GP for representing trees, the link between GP and medical
knowledge systems is straightforward.
Two different problems must thus be solved: image processing and knowledge representation. We
describe here a methodology for representing the knowledge used by a specialist when diagnosing by
means of parallel GP. So, the input for the methodology is the features that have been previously extracted
by an image processing system. The system must be capable of predicting how a burn will evolve, allowing
users to choose the most suitable treatment.
We believe that EAs have an important advantage over other kinds of machine learning techniques such
as neural networks: they have the ability to provide information not only about classification, but also
about the route taken to reach a decision.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 79 — #15

5-80 Handbook of Bioinspired Algorithms and Applications

5.6.2.2 Classifying by Means of Decision Trees

Medical diagnosis is a classification problem, in which the search space is made up of a set of points with
n-coordinates. Each coordinate is allocated a value for a given symptom. The aim is to find the category
to which a point belongs and this data will give us the diagnosis.
Sometimes we are not only interested in obtaining a classifier but also in retrieving important inform-
ation from the problem’s input parameters. Several classifier systems for medical diagnosis have been
designed based on decision trees and decision rules [32]. They take the parameters and classify the data.
The decision tree approach [33] falls into the category of inductive learning methods. Taking a set of
examples, the aim is to construct a tree that is able to classify new samples within the search space.
Decision trees usually classify the members of a set as either positive or negative examples. In this
application of parallel GP, due to the number of possibilities involved in burn diagnosis, we extend the
classic approach to multi-class decision-making, as in other medical diagnosis research [34]. In fact
a decision tree is a chain of if-then-else constructions that can be seen as a computer program. Each
condition is applied to some specific input parameters. The parameters will thus lead us along a branch
of the tree, finally reaching a decision, which is the category to which the input data belongs.
We want to generate the decision tree automatically according to a group of clinic cases that make up
the training set. This will be done by means of parallel GP.
Bearing in mind that in GP each individual adopts the shape of a tree, a decision tree can be considered
to be an individual. If we apply GP to finding a specific decision tree, at the end of the process the best
individual takes the form of that decision tree, that is, the genotype of the individual presents the chain of
decisions that are necessary for a diagnosis to be made.
When using GP for decision tree extraction, the function set will only be composed of the if-then-else
instruction, and some logical operators (see Section 5.3). As a consequence, any tree can be constructed.
The terminal set will contain the parameters we have decided to study in each burn.

5.6.2.3 A Case Study: Burns Unit, Virgen del Rocío Hospital. Seville, Spain
In order to apply the methodology, the collaboration of medical specialists from the area of burns treatment
was required.
The Burns Unit, at Virgen del Rocío Hospital, was in charge of retrieving information about real cases
of burns. A form was provided for several specialists to collect information on cases, which would be
useful for testing the problem. Several photographs were taken in each of the cases, and several features
of burns were noted in the form.
Thirty one different clinical cases were studied. Each completed form was accompanied by two pho-
tographs. No image processing has been done, but photographs are necessary for studying the different
parameters that will be automatically retrieved in the future. In this research, following the specialists’
indications, just three parameters have been studied.
No photographs are shown here, in order to preserve the privacy of the people who took part in the
study, but all of them are stored in the Burns Unit at the Virgen del Rocío Hospital, Seville, Spain.
We are aware that an accurate diagnosis requires the study of additional features, but we believe that
these three parameters are enough to see how the methodology works, and to obtain a first sketch for a
knowledge system based on GP.
Following specialist doctors’ advice, we decided to develop the first version of the knowledge system
taking into account just three different parameters from each picture:

1. Color (C): Several possible values: White, Black, Yellow, Red, Pink, and combinations
2. Dryness (D): Two possibilities: True or False
3. Ampoule (A): Two possibilities: Absent or Not

Although studying a wider set of parameters would be more useful, this simple model confirms the
validity of the methodology employed. The aim was not to create a tool that can classify 100% of the cases

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 80 — #16

Parallel Genetic Programming 5-81

but to demonstrate how easily a doctor’s knowledge can automatically be captured by means of GP and
decision trees.
Four kinds of diagnosis are possible for a burn: first degree, surface second degree, deep second degree,
and third degree.
In order to train the system, a table with parameters, taken by studying the photographs and diagnoses,
was built.

5.6.2.4 Results
A set of 31 clinical cases were provided by doctors, and these cases were used to train the system. Each of
them was allocated its corresponding parameters together with the result of its evolution over a couple of
weeks. This table of values was given to the algorithm in order for it to train.
Due to how GP works, the terminal and function sets were established as F = {ifthenelse, >, =, <,
AND, NOT} and T = {C, D, A}.
We ran the algorithm employing 5 populations of 2000 individuals each, with a period of migration
of 10 generations. We waited for 60 generations before taking the results. At the end, the best decision
tree we obtained was able to classify correctly 27 out of 31 cases. This was due to the presence of several
cases with the same parameters but with different evolution values. It was, consequently, impossible to
categorize these cases accurately. Figure 5.14 shows the decision tree obtained.
The meanings for nodes are the following: A, Ampoule; D, Dryness; C, Color; D2, Deep second degree;
S2, Surface second degree; 3, third degree; P, Pink color; Y, Yellow color.
Bearing in mind that each color has an integer number associated with it, comparisons among them are
meaningful. Of course, many different versions of the same tree can be obtained in different executions
of the algorithm. For the sake of simplicity, we show just one.
Using this methodology, we are able to represent medical knowledge by means of decision trees. This
is just an example about the use of GP for the task proposed, although for obtaining a completely reliable
decision tree a much larger set of examples must be used, and training and test sets have to be built and
employed in the learning process.
Finally, Figure 5.15 presents a comparison between parallel GP and GP when applied to the problem
of medical knowledge representation. We can see that parallel GP employing 5 populations and 2000
individuals per population achieved better convergence results than GP with 10000 individuals (the same
total amount of individuals).
We have to consider that the curve shown for PGP is not taking into account the time savings obtained
when using a parallel architecture, such as a multiprocessor system or a cluster of computers. If we take
both improvements into account — time savings and improvement in the convergence process — parallel
GP is far superior to plain GP.

IF

!A D2 IF

!D S2 IF

<= 3 IF

C Y > D2 S2

C P

FIGURE 5.14 Burn diagnosing decision tree.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 81 — #17

5-82 Handbook of Bioinspired Algorithms and Applications

Parallel GP
1400 GP

1000

Fitness 600

200

2 × 105 6 × 105 10 × 105 14 × 105

Effort

FIGURE 5.15

5.7 Concluding Discussion

In this chapter, we have presented parallel GP and its application to some real-life problems. We have
described the methodology and its history. We have also provided a set of benchmark problems commonly
used in GP. We have finally shown how parallel GP can be applied to the problem of placing and routing
circuits on FPGAs and also to the problem of extracting medical knowledge for burn diagnosing.

Acknowledgment
Part of this research has been possible thanks to Ministerio de Ciencia y Tecnología research projects
number TIC2002-04498-C05-01.

References
[1] T. Mitchell. Machine Learning. McGraw Hill, New York, 1996.
[2] J.R. Koza. Genetic Programming. The MIT Press, Cambridge, MA, 1992.
[3] C. Darwin. On the Origin of Species by Means of Natural Selection. John Murray, London, 1859.
[4] John H. Holland. Adpatation in Natural and Artificial Systems. University of Michigan Press, Ann
Arbor, MI, 1975.
[5] L.J. Fogel, A.J. Owens, and M.J. Walsh. Artificial intelligence through a simulation of evol-
ution. In M. Maxfield, A. Callahan, and L.J. Fogel, Eds., Biophysics and Cybernetic Systems:
Proceedings of the second Cybernetic Sciences Symposium, Spartan Books, Washington, D.C., 1965,
pp. 131–155.
[6] I. Rechenberg. Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen
Evolution. frommann-holzbog, Stuttgart, 1973. German.
[7] H.P. Schwefel. Evolutionsstrategie und numerische Optimierung. Ph.D thesis, Technische
Universitat Berlin, Berlin, 1975.
[8] W.B. Langdon. Data Structures and Genetic Programming: Genetic Programming + Data Structures =
Automatic Programming! Kluwer Academic Publishers, New York, 1998.
[9] J.R. Koza, F.H. Bennett III, and O. Stiffelman. Genetic programming as a Darwinian invention
machine. Genetic Programming: Proceedings of EuroGP ’99, LNCS, Vol. 1598. Springer-Verlag, May
1999.
[10] K. Stoffel and L. Spector. High-performance, parallel, stack-based genetic programming. In
Koza, J.R., et al., Eds., Genetic Programming 1996: Proceedings of the First Annual Conference.
Stanford University, MIT Press, CA, 1996, pp. 224–229.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 82 — #18

Parallel Genetic Programming 5-83

[11] W.B. Langdon and R. Poli. Fitness causes bloat. In R. Roy P.K. Chawdhry and R.K. Pant, Eds., Soft
Computing in Engineering Design and Manufacturing. Springer-Verlag, London, 1997, pp. 13–22.
[12] M. Tomassini. Parallel and distributed evolutionary algorithms: a review. In K. Miettinen,
M. Mäkelä, P. Neittanmäki, and J. Périaux, Eds., Evolutionary Algorithms in Engineering and
Computer Science. John Wiley & Sons, New York, 1999, 113–133.
[13] M. Oussaidéne, B. Chopard, O. Pictet, and M. Tomassini. Parallel genetic programming and its
application to trading model induction. Parallel Computing, 23: 1183–1198, 1997.
[14] F. Fernández, M. Tomassini, and L. Vanneschi. An empirical study of multipopulation genetic
programming. Genetic Programming and Evolvable Machines, 4: 21–52, 2003.
[15] J.P. Cohoon, S.U. Hegde, W.N. Martin, and D. Richards. Punctuated equilibria: A parallel genetic
algorithm. In J.J. Grefenstette, Ed., Proceedings of the Second International Conference on Genetic
Algorithms. Lawrence Erlbaum Associates, Mahwah, NJ, 1987, p. 148.
[16] R. Tanese. Parallel genetic algorithms for a hypercube. In J.J. Grefenstette, Ed., Proceedings of
the Second International Conference on Genetic Algorithms. Lawrence Erlbaum Associates, 1987,
pp. 177–183.
[17] F. Fernández de Vega. Distributed Genetic Programming Models with Application to Logic Syn-
thesis on FPGAs. Ph.D. thesis, Computer Science Department, University of Extremadura, Cáceres,
Spain, 2001.
[18] F. Fernández, M. Tomassini, L. Vanneschi, and L. Bucher. A distributed computing environment
for genetic programming using MPI. In J. Dongarra, P. Kaksuk, and N. Podhorszki, Eds., Recent
Advances in Parallel Virtual Machine and Message Passing Interface, Vol. 1908 of Lecture Notes in
Computer Science. Springer-Verlag, Heidelberg, 2000, pp. 322–329.
[19] F. Fernandez, M. Tomassini, and J.M. Sanchez. Experimental study of isolated multipopulation
genetic programming. In IEEE International Conference on Industrial Electronics, Control and
Instrumentation, Nagoya, Japan, 2000. IEEE Press, Washington, 2000, pp. 2672–2677.
[20] W.B. Langdon and R. Poli. Foundations of Genetic Programming. Springer-Verlag, Berlin, 2002.
[21] P. Tufts. Parallel case evaluation for genetic programming. In 1993 Lectures in Complex Systems,
Vol. VI of Santa Fe Institute Studies in the Science of Complexity, 1993, pp. 591–596.
[22] H. Juille and J.B. Pollack. Parallel genetic programming and fine-grained SIMD architecture. In
E.V. Siegel and J.R. Koza, Eds., Working Notes for the AAAI Symposium on Genetic Programming.
MIT, Cambridge, MA, November 10–12, 1995. AAAI, pp. 31–37.
[23] P.J. Angeline and K.E. Kinnear Jr. (Eds.). Advances in Genetic Programming 2. The MIT Press,
Cambridge, MA, 1996.
[24] W. Punch. How effective are multiple populations in genetic programming. In J.R. Koza,
W. Banzhaf, K. Chellapilla, K. Deb, M. Dorigo, D.B. Fogel, M. Garzon, D. Goldberg, H. Iba,
and R.L. Riolo, Eds., Genetic Programming 1998: Proceedings of the Third Annual Conference.
Morgan Kaufmann, San Francisco, CA, 1998, pp. 308–313.
[25] D. C. Dracopoulos and S. Kent. Bulk synchronous parallelisation of genetic programming. In Jerzy
Waśniewski, Ed., Applied Parallel Computing: Industrial Strength Computation and Optimization;
Proceedings of the third International Workshop, PARA ’96, Springer-Verlag, Berlin, Germany, 1996,
pp. 216–226.
[26] S. Tongchim and P. Chongstitvatana. Nearest neighbor migration in parallel genetic programming
for automatic robot programming. In Proceedings of the Sixth International Conference on Control,
Automation, Robotics and Vision, Singapore, December 2000.
[27] G. Folino, C. Pizzuti, and G. Spezzano. A scalable cellular implementation of parallel genetic
programming. IEEE Transactions on Evolutionary Computation, 7: 37–53, 2003.
[28] J.R. Koza, F.H. Bennett III, D. Andre, and M.A. Keane. Genetic Programming III: Darwinian
Invention and Problem Solving. Morgan Kaufmann, San Francisco, CA, 1999.
[29] F. Fernandez, I. Hidalgo, J. Lanchares, and J.M. Sanchez. A methodology for reconfigurable hard-
ware designed based upon evolutionary computation. Microprocessors and Microsystems. Elsevier,
28: 363–371, 2004.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 83 — #19

5-84 Handbook of Bioinspired Algorithms and Applications

[30] P. Martin. A hardware implementation of a genetic programming system using FPGAs and
Handel-C. Genetic Programming and Evolvable Machines, 2: 317–343, 2001.
[31] M.I. Heywood and A.N. Zincir-Heywood. Register based genetic programming on fpga computing
platforms. In R. Poli, W. Banzhaf, W.B. Langdon, J.F. Miller, P. Nordin, and T.C. Fogarty, Eds.,
Proceedings of the European Conference on Genetic Programming, Vol. 1802 of Lecture Notes in
Computer Science. Springer-Verlag, London, 2000, pp. 44–59,
[32] J. H. Holmes. Discovering risk of disease with a learning classifier system. In T. Bäck, Ed., Proceed-
ings of the Seventh International Conference on Genetic Algorithms (ICGA97). Morgan Kaufmann,
San Francisco, CA, 1997.
[33] J.R. Quinlan. Decision trees and instance-based classifiers. In The Computer Science and
Engineering Handbook, 1997, pp. 521–535.
[34] M. Kurzynski. The application of unified and combined recognition decision rules to the
multistage diagnosis problem. In Proceedings of the 20th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, Vol. 3. Hong-Kong, 1998, pp. 1194–1197.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C005” — 2005/7/20 — 19:53 — page 84 — #20

 6
Parallel Cellular
Algorithms and
Programs

6.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-85

6.2 Cellular Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-86
6.3 Parallel CA Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-87
CAMELot—CARPET • Other Cellular-Based Parallel Systems
6.4 Programming Standard and Nonstandard Parallel
Cellular Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-93
Inhomogeneous CA • Asynchronous and Probabilistic CA •
Partitioned CA
6.5 Programming Emergent Systems as Massively
Parallel CA: Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-95
The Q2R Ising Model • Epidemic Diffusion Simulation
6.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-98
Domenico Talia References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-100

6.1 Introduction
An emergent phenomenon is the large-scale group behavior of a system that does not seem to have any
explanation in terms of the single constituent parts only. In other words, emergence can be defined by
saying that “the whole is greater than the sum of the parts.” In emergent systems, we can consider two
different levels of description: the microscopic level, where all the single components are taken into account;
and the macroscopic level, where emergent behavior occurs as the synthesis of the complex interaction of
the microscopic components. To bring emergent systems out of a speculative horizon it is necessary to
experiment and test them. In particular, emergent system simulation on parallel computers is an essential
practice for an indepth analysis and evaluation of the accuracy of the proposed models of emergent
behavior.
The programming of emergent phenomena and systems using traditional programming models and
tools is very difficult and involves long and complex coding. This is mainly because these approaches
are based on the design of a system as a whole; hence, design and programming do not start from basic
elements. It is better to design emergent systems by means of paradigms that allows for expressing the
behavior of single elements and their interactions. The global behavior of these systems then emerges from

6-85

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 85 — #1

6-86 Handbook of Bioinspired Algorithms and Applications

the evolution and interaction of a massive number of elements; hence, it does not need to be explicitly
coded.
The cellular automata (CA) model is a massively parallel computational model that can be effectively
used for the investigation and simulation of emergent phenomena and systems. Cellular automata are
inherently parallel; therefore, they can be used to model and simulate very large-scale emergent systems on
parallel computers [1,2]. Cellular parallel tools allow for the exploitation of the inherent parallelism of CA
in the implementation of natural solvers that simulate dynamical emergent systems by a massive number
of simple agents (cells) that interact locally. Parallel cellular languages and environments provide useful
design and programming tools for the development of scalable simulations and models of emergent
behavior. This approach is a valid alternative to complex and expensive laboratory experiments and
simulations [3].
We discuss here how the basic CA concepts are related to emergent systems, describe parallel CA
environments and tools for programming emergence in complex systems, and present some significant
programming examples of emergent systems. The remainder of the chapter is organized as follows:
Section 6.2 introduces CA and Section 6.3 outlines the main issues in parallel CA programming, describes
the main features of the CAMELot environment and its programming language CARPET and discusses
some related systems. Section 6.4 shows how different classes of CA can be implemented by using the
CARPET language. Section 6.5 presents two examples of emergent phenomena programmed according to
the parallel CA model in the CARPET cellular language and gives performance figures for them. Finally,
Section 6.6 draws some conclusions.

6.2 Cellular Automata

Cellular automata are discrete dynamical systems in that space, time, and properties can have only a finite
number of states [4]. They are an effective model for exploring systems with no central control. Given that
the rules are executed in parallel on every cell, we can easily explore systems, using a decentralized control,
in which simple local interaction takes place for a large population of cells over some period of time. The
basic idea is to describe a complex system and simulate it by the interaction of a massive number of cells
following simple rules. Thus, a complex system is not described with complex global equations but the
complexity emerges from the interaction of simple local rules.
A CA can be defined as a d-dimensional Euclidean space (where d = 1, 2, 3 is used in practice),
partitioned into cells of uniform size, each one embedding an identical elementary automaton (ea). Input
for each ea is given by the states of the elementary automata in the neighboring cells, where neighborhood
conditions are determined by a pattern invariant in time and constant over the cells. At the time t = 0,
ea are in arbitrary states and the CA evolves changing the state of all ea at discrete times, according to a
local rule. Each cell in the regular spatial lattice can have any one of a finite number of states. As mentioned
before, the states of the cells in the lattice are updated according to a local rule called the state transition
function. That is, the state of a cell at a given time depends only on its own state in the previous time step
and the states of its nearby neighbors at the previous time step [5].
Here, we summarize the basic CA concepts useful for programming emergence.

• Transition function: Set of rules that define how the state of each cell changes on the basis of its
current state and the states of its neighbor cells.
• State: The state of a cellular automaton (global state) is completely specified by the values of the
variables at each cell (local state). The state of a cell is a simple or structured variable (see substate)
that takes values in a finite set. The cell state can be either a numeric value or a property. For
instance, if each cell represents part of a landscape, then the state might contain the altitude or the
type of land.
• Substate: If the cell state is represented as a structured variable, substates are the fields of the
structure that represent the attributes of the cell state. For example, if each cell represents a particle,
its state can be composed of two substates that represent particle mass and speed.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 86 — #2

Parallel Cellular Algorithms and Programs 6-87

C C
C

FIGURE 6.1 CA neighborhoods.

• Neighborhood: The set of cells that a cell interacts with. The neighborhood of a cell is typically
taken to be all immediately adjacent cells. Simple neighborhoods of a cell (C) in a two-dimensional
lattice are shown in Figure 6.1.
Let us define a CA as the 4-tuple (E d , S, N , σ ), where
• E d is a regular lattice (the elements of E d are called cells).
• S is a finite set of states.
• N a finite set (with |N | = n) of neighborhood indices such that for all x in N , all c in E d : c + x
in E d .
• σ : S n → S a transition function.
A configuration Ct : E d → S is a function that associates a state with each cell of the lattice. The effect
of the transition function σ is to change the configuration Ct into the new configuration Ct +1 according
to Ct +1 (c) = σ ({Ct (i) : i in N (c)}), where N (c) denotes the set of neighbors of cell c, N (c) = {i in E d :
c − i in N }. In standard CA, all cells of the automaton are updated synchronously in parallel; whereas
extended CA models define asynchronous updating [6]. In section 6.4, we discuss nonstandard CA models.
The state of the entire automaton advances in discrete time steps. Therefore, in CA the transition function
plays a role analogous to that of the evolution equation in classical dynamical models. The global behavior
of the system is not directly specified but it is determined, in other words, it emerges by the evolution of
the states of all cells as a result of multiple interactions. Cellular automata capture the peculiar features of
systems that may be seen to evolve exclusively according to the local interactions of their constituent parts,
and guarantee computational universality. Furthermore, applied aspects of modeling have been widely
investigated from a theoretical viewpoint [4,5].

6.3 Parallel CA Systems

For the implementation of CA on parallel computers two main approaches can be used. One is to write
programs that encode the CA rules in a general-purpose parallel programming language such as HPF,
HPC++, Linda, or CILK or to use a high-level sequential language such as C, Fortran, or Java with one
of the low-level toolkits/libraries currently used to implement parallel applications such as MPI, RMI,
PVM, or OpenMP. This approach does not require a parallel programmer to learn new language syntax
and programming techniques for cellular programming. However, it is not simple enough to be used by
programmers who are not experts in parallel programming, and the code consists of a large number of
instructions even if simple cellular models must be implemented.
The other possibility is to use a high-level language specifically designed for CA, in which it is possible
to directly express the features and the rules of CA, and then use a compiler to translate the CA code
into a program executable on parallel computers. This second approach has the advantage that it offers a
programming paradigm that is very close to the CA abstract model and that the same CA description could
possibly also be compiled into different codes for various parallel machines. Furthermore, in this approach,
parallelism is transparent from the user, so the programmers can concentrate on the specification of the
model without worrying about architecture-related issues. In summary, it leads to the writing of software

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 87 — #3

6-88 Handbook of Bioinspired Algorithms and Applications

that does express the cellular paradigm in a natural way; these programs are simpler to read, change, and
maintain. On the other hand, the regularity of computation and locality of communication allow CA
programs to get good performance and scalability on parallel architectures.
Recently, several CA environments have been implemented on current desktop computers. Examples of
these systems are CAT, CelLab, CaSim, CDM, Cellsim, DDLab, and Mathematica. A longer list can be found
in Reference 7, in which the main features of these systems are outlined. Sequential CA-based systems can
be used for educational purposes and very simple simulations, but real-world phenomena simulations
generally take a very long time, or in some cases cannot be executed on this class of systems because
of memory or computing power limits. Therefore, massively parallel computers are the appropriate
computing platform for the execution of CA models when real-life problems must be solved. In fact,
for two- and three-dimensional CA of large size the computational load can be enormous. Thus, if CA
are to be used for investigating large complex phenomena, their implementation on high performance
computers composed of several processors is a must.
In particular, general-purpose distributed-memory parallel computers offer a very useful architecture
for a scalable CA machine in terms of speed-up, programmability, and portability. These systems are based
on a large number of interconnected processing elements (PE), which perform a task in parallel. According
to this approach, in recent years several parallel cellular software environments have been developed.
Significant examples of these parallel cellular systems are CAMELot, StarLogo, NEMO, P-CAM [8],
Cellular, ParCel-1, PECANS [9], and DEVS. Some are discussed in Reference 1. Together with these
software systems, parallel CA hardware has been developed for a more efficient execution of CA algorithms.
Two examples of CA hardware are the CAM-8 machine [10] and the CEPRA FPGA machine. These two
systems are special purpose machines that exploit CA parallelism in a very efficient way, although they are
specialized machines that do not support general computation models.
Cellular automata parallel systems allow a user to exploit the inherent parallelism of CA to support the
efficient simulation of complex systems that can be modeled by a very large number of simple elements
(cells) with local interaction only. Cellular automata-based languages share several features such as a
common computational paradigm and some differences such as, for example, different constructs to
specify details of a CA or of mapping and output visualization. Many real-world applications in science
and engineering, such as lava-flow simulations, molecular gas simulation, landslide modeling, freeway
traffic flow, three-dimensional rendering, soil bioremediation, biochemical solution modeling, and forest
fire simulation, have been implemented by using these CA languages. Moreover, parallel CA languages can
be used to implement a more general class of fine-grained applications such as finite elements methods,
partial differential equations, and systolic algorithms.
The main issues that influence the way in which CA languages support the design of applications on
high performance architectures are:
1. The programming approach: The unit of programming is the single cell of the automaton.
2. The cellular lattice declaration: It is based on the definition of the lattice dimension and the lattice
size.
3. The cell state definition and operations: Cell state is defined as a single variable or a record of typed
variables; cell state access and update operations are needed.
4. The neighborhood declaration and use: Neighborhood concept is used to define interaction among
cells in the lattice.
5. The Parallelism exploitation: The unit of parallelism is the cell, and parallelism, like communication,
is implicit.
6. The cellular automata mapping: Data partitioning and process-to-processor mapping is implicit at
the language level.
7. The output visualization: Automaton global state, as the collection of the cell states, is shown as it
evolves.
Many of these issues are taken into account in parallel CA systems and similar or different solutions are
provided by parallel CA languages. By discussing these concepts, we intend to illustrate how this class of

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 88 — #4

Parallel Cellular Algorithms and Programs 6-89

languages can be effectively used to implement high-performance applications in science and engineering
using the massively parallel cellular approach.

Programming Approach. When a programmer starts to design a parallel cellular program, she/he must
define the structure of the lattice that represents the abstract model of a computation in terms of cell-
to-cell interaction patterns. Then it must concentrate on the unit of computation that is a single cell of
the automaton. The computation to be performed must be specified as the evolution rule (transition
function) of the cells that compose the lattice. Thus, as against other approaches, a user does not specify
a global algorithm that contains the program structure in an explicit form. The global algorithm consists
of all the transition functions of all cells that are executed in parallel for a certain number of iterations
(steps).
It is worth to notice that in some CA languages it is possible to define transition functions that change
in time and space to implement inhomogeneous CA computations. Thus, after defining the dimension
(e.g., one-, two-, or three-dimensional) and the size of the CA lattice, she/he needs to specify, by the
conventional and the CA statements, the transition function of the CA that will be executed by all the
cells. Then the global execution of the cellular program is performed as a massively parallel computation
in which implicit communication occurs only among neighbor cells that access each other’s state.
Cellular Lattice Declaration. As mentioned in the previous section, the lattice declaration defines the
lattice dimension and the lattice size. Most languages support two-dimensional rectangular lattices only
(e.g., CANL and CDL). However, some of them, such as CARPET and Cellang, allow the definition of
one-, two-, and three-dimensional lattices. Some languages allow also the explicit definition of boundary
conditions, such as CANL [9], which allows adiabatic boundary conditions where absent neighbor cells
are assumed to have the same state as the center cell. Others implement reflecting conditions that are
based on mirroring the lattice at its borders. Most languages use standard boundary conditions such as
fixed and toroidal conditions.
Cell State. The cell state contains the values of data on which the cellular program works. Thus the
global state of an automaton is defined by the collection of the state values of all the cells. While low-level
implementations of CA allow to define the cell state as a small number of bits (typically 8 or 16 bits),
cellular languages such as CARPET, CANL, DEVS-C++, and CDL allow a user to define cell states as a
record of typed variables as follows:

cell=(direction : int ; speed : float);

where two substates are declared for the cell state. According to this approach, the cell state can be composed
of a set of substates that are of integer, real, char, or Boolean type and in some cases (e.g., CARPET), arrays
of these basic types can also be used. Together with the constructs for cell state definition, CA languages
define statements for state addressing and updating that address the substates by using their identifiers;
for example, cell.direction indicates the direction substate of the current cell.
Neighborhood. An important feature of CA languages that differentiate them from array-based languages
and standard data-parallel languages is that that they do not use explicit array indexing. Thus, cells are
addressed with a name or the name of the cells belonging to the neighborhood. In fact the neighborhood
concept is used in the CA setting to define interaction among cells in the lattice. In CA languages, the
neighborhood defines the set of cells whose state can be used in the evolution rule of the central cell. For
example, if we use a simple neighborhood composed of four cells we can declare it as follows

neigh cross=(up, down, left, right);

and address the neighbor cell states by the ids used in the above declaration (e.g., down.speed,
left.direction). The neighborhood abstraction is used to define the communication pattern among

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 89 — #5

6-90 Handbook of Bioinspired Algorithms and Applications

cells. It means that at each time step, a cell sends to and receives from the neighbor cells the state values.
In this way implicit communication and synchronization are realized in cellular computing.
The neighbor mechanism is a concept similar to the region construct that is used in the ZPL
language [11], in which regions replace explicit array indexing making the programming of vector-
or matrix-based computations simpler and more concise. Furthermore, this way of addressing the lattice
elements (cells) does not require compile-time sophisticated analysis and complex run-time checks to
detect communication patterns among elements.
Parallelism Exploitation. Cellular automata languages do not provide statements to express parallelism
at the language level. It turns out that a user does not need to specify what portion of the code must be
executed in parallel. In fact, in parallel CA languages the unit of parallelism is a single cell, and parallelism,
like communication and synchronization, is implicit. This means that in principle the transaction function
of every cell is executed in parallel with the transaction functions of the other cells. In practice, when
coarse-grained parallel machines are used, the number of cells N is greater than the number of available
processors P, so each processor executes a block of N /P cells that can be assigned to it using a domain
decomposition approach.
CA Mapping. Like parallelism and communication, data partitioning and process-to-processor mapping
is implicit in CA languages. The mapping of cells (or blocks of them) onto the physical processors that
compose a parallel machine is generally done by the run-time system of each particular language and the
user usually intervenes in selecting the number of processors or some other simple parameter.
Some systems that run on MIMD computers use load balancing techniques that assign at run-time the
execution of cell transition functions to processors that are unloaded, or use greedy mapping techniques
that avoid some processor to become unloaded or free during the CA execution for a long period. Examples
of these techniques can be found in References 1, 12, and 13.
Output Visualization and Monitoring. A computational science application is not just an algorithm.
Therefore it is not sufficient to have a programming paradigm for implementing a complete application. It
is also significant to dispose of environments and tools that help a user in all the phases of the application
development and execution. Most of the CA languages we are discussing here provide a development
environment that allows a user not only to edit and compile the CA programs, but also to monitor the
program behavior during its execution on a parallel machine, by visualizing the output as composed of
the states of all cells. This is done by displaying the numerical values or by associating colors to these
values. Examples of these parallel environments are CAMEL for CARPET, PECANS for CANL, and DEVS
for DEVS-C++. Some of these environments provide dynamic visualization of simulations together with
monitoring and tuning facilities. Users can interact with the CA environment to change values of cell
states, simulation parameters, and output visualization features. These facilities are very helpful in the
development of complex scientific applications and make it possible to use these CA environments as real
problem-solving environments (PSEs) [14].
In the rest of this section we outline some of the listed issues by discussing the main features of
CAMELot, a general-purpose system that can be easily used for programming emergent systems using
the CARPET cellular programming language according to a massively parallel paradigm and some related
parallel CA environments and languages.

6.3.1 CAMELot—CARPET
CAMELot (CAMEL open technology) is a parallel software environment designed to support the parallel
execution of cellular algorithms, the visualization of the results, and the monitoring of cellular program
execution [15]. CAMELot is an MPI-based portable version of the CAMEL system based on the CARPET
language. CARPET (CellulAR Programming EnvironmenT ) offers a high-level cellular paradigm that
offers to a user the main CA features to assist her/him in the design of parallel cellular algorithms without
apparent parallelism [6].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 90 — #6

Parallel Cellular Algorithms and Programs 6-91

A CARPET user can develop cellular programs describing the actions of many simple active elements
(implemented by cells) interacting locally. Then, the CAMELot system executes in parallel cell evolution
and allows a user to observe the global complex evolution that arises from all the local interactions.
CARPET uses a C-based grammar with additional constructs to describe the rules of the transition
function of a single cell. In a CARPET program, a user can define the basic rules of the system to be
simulated (by the cell evolution rule), but she/he does not need to specify details about the parallel
execution. The language includes

1. A declaration part (cadef) that allows to specify:

• the dimension of the automaton (dimension);
• the radius of the neighborhood (radius);
• the type of the neighborhood (neighbor);
• the state of a cell as a record of substates (state);
• a set of global parameters to describe the global characteristics of the system (parameter);
• a set of constructs for addressing and updating the cell states (e.g., update, GetX,
GetY, GetZ).

Figure 6.2 shows a simple CA programmed in CARPET that implements the Fredkin’s rule. This is a
simple rule: a cell becomes alive if the number of living cells in its neighborhood is odd; if the number of
living cells in its neighborhood is even a cell becomes dead. Fredkin’s rule is very simple, however it has
the fascinating property that any initial pattern of living cells is replicated several times on a larger scale.
CARPET and CAMELot have been used for implementing high-performance real-world simulations
based on the emergence paradigm such as lava flow, traffic flow, and forest fire simulations [6]. In
Section 6.5 the CARPET language is used to program two significant examples of emergent systems. Its
main linguistic features are outlined by describing how it supports the implementation of real emergent
applications.

#define dead 0
#define alive 1

cadef
{
dimension 2;
radius 1;
state (short value);
neighbor Neumann[4] ([0,-1]North, [-1,0]West,
[0,1]South,[1,0] East);
}
int sum=0;
{ for (i=0; i<4; i++)
sum = sum + Neumann[i]_value;
if (sum%2 == 0)
update(cell_value, dead);
else
update (cell_value, alive);
}

FIGURE 6.2 The Fredkin’s rule written in CARPET.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 91 — #7

6-92 Handbook of Bioinspired Algorithms and Applications

6.3.2 Other Cellular-Based Parallel Systems

As mentioned before, some cellular-based systems have recently been developed. Among these we discuss
some representative examples such as Cellular, DEVS-C++, and ZPL. In particular, ZPL is a language for
parallel programming that has not been specifically designed for CA, but it shares with parallel cellular
languages the same rationale and some basic ideas.
The Cellular system [16] consists of a CA-based programming language, named Cellang, an abstract
virtual CA machine for execution of cellular programs, named avcam, and a viewer named cellview, which
is used to examine the cell values in a graphic format. Cellang combines the classic CA programming
paradigm, like CARPET or CDL, with that of agents. The results of an execution can either be viewed
graphically, as an output stream of cell locations and values, or passed through a custom filter before
being reported. The current implementation of the Cellang compiler generates codes both for sequential
computers and shared-memory multiprocessor systems. However, the system has not been implemented
on distributed-memory parallel machines or PC clusters.
The NEMO (Neighbourhood Modeling) system is an environment based on CA designed by the
Carleton’s PARAllel DIGital Modeling (PARADIGM) group [12]. The main component of the NEMO
systems is the Cell driver. The Cell driver is the parallel kernel of NEMO that provides the parallel
execution of a NEMO application and interacts with the other components. In addition to the Cell driver,
NEMO contains the Neighbourhood Analysis (NAN) driver, which is designed to support the parallel
computation of spatial statistics on very large lattices, and the Propagation driver (PD), which supports
the modeling in parallel of processes in a lattice that spreads along an active boundary. Moreover, NEMO
provides a display facility for the visualization of graphical data from the simulation. As against CAMELot
and Cellular, NEMO does not offer a high-level language to implement cellular programs although some
default functions can be used.
DEVS-C++ is a high-performance environment based on the DEVS (Discrete Event System Specifica-
tion) formalism that supports the analysis, design, and simulation of discrete event dynamical systems [17].
The DEVS formalism, based on CA theory and discrete event simulation, provides a means of specifying
a mathematical object called system defined as a lattice of cells. DEVS-C++ is a DEVS parallel imple-
mentation that uses the C++ language for programming simulation and models designed by the DEVS
formalism. Applications are programmed using a set of specifically designed classes (included in the Con-
tainer and Devs libraries) provided by the DEVS-C++ environment in an object-oriented programming
framework. The Devs class is the basic class to provide methods for the DEVS formalism. Such methods
are implemented as virtual methods to be defined by users. A Java version of the system has been recently
developed.
ZPL is a data-parallel array programming language designed for fast execution on both sequential and
parallel computers [11]. Like CARPET, ZPL is implicitly parallel, that is, the programmer does not express
the parallelism; the ZPL compiler and run-time system will generate parallel codes and map processes on
processors. As cellular languages, ZPL is suitable only for regular parallel computations.
A fundamental concept in ZPL is the notion of a region. Fortran and other array languages refer to
subarrays using the so-called slice or triple notation, whereas ZPL uses the concept of a region. A region
is simply a set of indices, that is,

region X = [1..n, 1..n];

specifies the standard indices of an nxn array. A ZLP program next declares a set of directions. Directions
are used to transform regions, as in the expression north of X . As in cellular programming, array indexing
is avoided in ZPL by referring to adjacent array elements using the @ operator, which is similar to the
neighborhood mechanism of CARPET. An expression A@d, executed in the context of a region X , results
in an array of the same size and shape as X composed of elements of A offset in the direction d. ZPL does
not have parallel directives as in data-parallel languages such as HPF or other forms of explicit parallelism.
This implicit computation can be parceled out to different processors to get parallelism. Thus, parallelism,
as in cellular programming, comes simply from the semantics of the array operations.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 92 — #8

Parallel Cellular Algorithms and Programs 6-93

6.4 Programming Standard and Nonstandard Parallel

Cellular Automata
Nonstandard models of CA can be divided between modifications and generalizations according to ter-
minology introduced by Worsch [7], where several nonstandard automata are described. By modifications
of CA we denote computational models that can simulate CA and can be simulated by CA with a linear
overhead in time and space (number of cells). On the other hand, generalizations of CA are models that
cannot be shown to be simulated by CA in linear time.
In Section 6.4.1, we concentrate on discussing how CARPET can be used for implementing four types
of nonstandard CA: inhomogeneous, synchronous, probabilistic, and partitioned CA.

6.4.1 Inhomogeneous CA
Inhomogeneous CA are a generalization of standard CA. In CA we can have spatial inhomogeneity,
temporal inhomogeneity, or both. In spatially inhomogeneous CA, there is not a single transition function
σ , but there is a set of different transition functions σ1 , σ2 , . . . , σN , associated to different cells or regions
of a CA in which also different neighborhoods can be defined. This class of automata can be implemented
in CARPET using the operations GetX, GetY, GetZ that return the value of the coordinates X, Y, and
Z of a cell in the CA lattice. For example, if we want to use a different transition function in a rectangular
region of a two-dimensional lattice we can write a code as follows

{
. . . .
if ((GetX >= 100 && GetX <= 200) && (GetY >= 80 && GetY <= 400)
{ trans-funct1()}
else
{ trans-funct2()}
. . . .
}

The same approach can be used for different transition functions for cells that belong to the border of
lattice to define boundary conditions. For instance, if two border sides are identified by the coordinates
X = 0 and X = 400, to use a transition function for the cells that are placed on that borders we can write

. . . .
if ((GetX == 0 | | GetX == 400)
{ border-trans-funct()}
else
{ normal-trans-funct()}

In temporal inhomogeneous CA, the transaction function changes during time. This generalization of
standard CA is very useful in modeling and simulation of phenomena that consist of more computational
phases. Thus, for a certain number of iterations a function σta is used, then another function σtb is used
for another time interval, and so on depending on the kind of computation that must be performed.
This class of CA can be easily programmed in CARPET by using the predefined variable step. For
example, Figure 6.3 shows the structure of a CARPET algorithm for a two-dimensional CA composed of a
sequence of three different transition functions. Temporal inhomogeneity can also be implemented using
conditions that include the step variable and more complex logic expressions.

6.4.2 Asynchronous and Probabilistic CA

Asynchronous CA are only a modification of the standard CA model. In an asynchronous CA, each cell at
each time step t can choose nondeterministically between keeping its current state s(t − 1) or changing its

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 93 — #9

6-94 Handbook of Bioinspired Algorithms and Applications

cadef
{
dimension 2;
radius 1;
....
}

{
if (step == 1)
trans_funct1();
else
if (step > 1 && step <= 10 )
trans_funct2();
else
trans_funct2();
}

FIGURE 6.3 CARPET code for a temporal inhomogeneous CA.

state according to the transition function σ . This class of CA is useful in the simulation of asynchronous
systems, where it is not necessary to update the state of its components at the same time.
Asynchronous CA can be programmed in CARPET by using the random function that allows to
express nondeterminism in the local transition function of each cell. The random(n) function returns
a pseudo random integer number between 0 and n. We can use it when the state of a cell must be updated.
For example, the following instructions show the nondeterministic update of a substate of a cell (notice
that the randomize function call creates a different seed for the random number generator, avoiding
that the same random value is generated for every cell):

cadef
{
. . . .

state ( int substate1, float substate2);

. . .
}
. . . . .
randomize();
if(random(n)> n/2)
update(cell_substate2, X)
else
. . ./* the cell state is not updated */

The random function can be also used for the implementation of probabilistic CA that in some
aspects are similar to asynchronous CA. In a probabilistic CA, given a local configuration c, each cell
can enter a new state s with a probability σ (c, s) and s∈S σ (c, s) = 1. This type of CA is useful in the
simulation of probabilistic phenomena that occur in physics and other sciences. As mentioned before,
by the use of the random function it is possible to implement this modification of the standard CA
model in CARPET. In particular, this can be done by a using a switch statement with probability values
(computed on the basis of a random number) assigned to the different case branches that must contain
an update operation whose execution will depend on the probability assigned to its own branch.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 94 — #10

Parallel Cellular Algorithms and Programs 6-95

cadef
{
dimension 2;
radius 1;
state (int sub1, sub2, sub3, sub4);
neighbor VonNeu[4] ([0,-1]North,[-1,0]West,[0,1]South,[1,0] East);
}
int sum=0;
{
sum = VonNeu[1]_sub1+VonNeu[2]_sub2+VonNeu[3]_sub3+VonNeu[4]_sub4;
if (sum <= 2)
{
update(cell_sub1, 0);
update(cell_sub2, 1);
}
else
{
update(cell_sub3, 0);
update(cell_sub4, 1);
}
}

FIGURE 6.4 A simple two-dimensional partitioned CA written in CARPET.

6.4.3 Partitioned CA
Another significant modification of the standard CA model is represented by the partitioned CA. Whereas
in a standard CA, a cell can use the whole state of each neighboring cell to compute its next state, in a
partitioned CA only the component Sni of a neighbor ni is used to determine the new state of a cell. Thus,
each cell reads only the component i of a cell ni that belongs to its neighborhood. The transition function
of a partitioned CA can be expressed as

σ : Sn1 × Sn2 × · · · × Snk → Sn1 × Sn2 × · · · × Snk

where N = {n1 , n2 , . . . , nk }. Partitioned CA is useful in modeling systems where each cell in a neigh-
borhood contributes toward the change of the state of a central cell. Moreover, they can be used to make
possible the implementation of transition functions that otherwise would be infeasible for software or
hardware limits. In fact, in them the domain of σ is of size |S| instead of |S||N | .
The implementation of partitioned CA in CARPET is quite direct because the language allows the
definition of the cell state as a set of substates. Thus we can define the cell state of a partitioned CA as
composed of a number of substates equal to the number of neighbors. Figure 6.4 shows a simple CARPET
program that implements a two-dimensional partitioned CA.
The state of a cell is updated according to the sum of the ith component of the neighbor cell i. Since a
radius 1 von Neumann neighborhood is defined, the cell state is composed of four substates. If we use a
Moore neighborhood, a state composed of eight substates must be used.

6.5 Programming Emergent Systems as Massively

Parallel CA: Examples
Emergent behavior not only occurs in nature and in life sciences such as biology and medicine, but
significant examples of emergent systems can be found in physics, computer science, and engineering.
For instance, network synchronization, distributed mobile sensors, cooperative environments, peer-to-
peer computing, and distributed control systems are emergent systems that today occur in many real-life

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 95 — #11

6-96 Handbook of Bioinspired Algorithms and Applications

problems. We cannot discuss these systems in detail, but we give here two simple but significant examples
of how to program emergent systems on parallel machines using a massively parallel cellular paradigm.
In particular, this section presents two examples of emergent systems programmed by using the
CARPET language that we implemented on a Linux cluster. The first example is the Q2R Ising model and
the second one is an epidemics diffusion model.

6.5.1 The Q2R Ising Model

The Ising model is one of the pillars of statistical mechanics. Each cell represents a spin that can have
two values (1/0, up/down), and neighboring cells have an energetic preference to be the same value. As a
system of 1/0 spins, it is a model for magnetism: like iron, there is a temperature (the Neel point) above
which the magnetization “melts” away. Run at high temperatures to see the melted state, and run at low
temperatures to see the magnetized state. The Q2R model is an Ising model with particular properties.
Q2R cellular automata are a good example for addressing an old question: How is thermodynamic
irreversibility, as seen in entropy increase, compatible with microscopic reversibility, as in Newton’s
equations of motion? The Boltzmann–Zermelo argument says that the time for returning to the low-
entropy initial state of a large system are far longer than the age of the universe. How can we test this
assertion? Molecular dynamics in its usual form is inadequate to address this question, since arithmetic
rounding errors as well as discretization of space and time preclude return to the exact initial configuration.
Monte Carlo simulations of Ising models avoid such errors but require random numbers and are therefore
not reversible in the usual sense. The Q2R update rules of CA for microcanonical Ising models are reversible
and the system returns exactly to the initial configuration after an exponentially long time.
In Q2R, all of the spins are flipped simultaneously, while in the canonical Ising model only one randomly
selected spin is flipped at a time. The spin-flipping decision is deterministic in Q2R while it is probabilistic
in the Metropolis model. Both versions of the Ising model give many comparable results. Computer
simulation of Q2R CA helps us to understand some very old fundamental problems.
Here we show the simple implementation of the Q2R model using the CARPET language and some
preliminary performance results on a parallel machine. Figure 6.5 shows the CARPET code composed
of a very short number of instructions that have been executed on a cluster of PentiumIII connected
by Myrinet. The program uses a two-dimensional automaton with a Moore neighborhood composed of
eight neighbor cells; to flip the cell spin, the neighbor cells’ spins are summed and then, depending on the
current value of the cell, the central cell spin is updated.
Table 6.1 shows the performance results (elapsed time in seconds) up to 12 processors for the execution
of 100 iteration steps considering four different lattice sizes. As shown in Figure 6.6, the algorithm is scalable
despite its simplicity. Relative speedup is very close to ideal. Although the code of each cell is simple, the
application makes efficient use of the 12 parallel processors. In fact efficiency (E = speedup/Procs) ranges
from 0.9 to 1, this means that 90% to 100% of the computing power of each CPU is exploited.

6.5.2 Epidemic Diffusion Simulation

The dominant features of an epidemic, its distribution over areas of land, and its evolution through time,
are the result of the dynamic interactions of the host and pathogen systems, both of which are influenced
by numerous complex biological processes. Hence, models must somehow describe the relationships
among the individuals, the time, and the space to apply possible infection conditions on them depending
on the neighboring infected individuals. Based on these properties, epidemic diffusion is an emergent
phenomenon that can be effectively modeled using CA. In a cellular model of a plant epidemic, the
host population is divided into small-scale units, each of which is located at a cell on a two-dimensional
lattice. Models can be used to develop understanding of quantitative or qualitative behavior, or to make
predictions of the dynamics of biological systems.
Figure 6.7 shows the CARPET code that implements a simulation of an epidemic with three types of
cells: cells representing healthy cells/creatures, other cells representing sick ones, and blank cells. Healthy

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 96 — #12

Parallel Cellular Algorithms and Programs 6-97

#define up 1
#define down 0
cadef
{
dimension 2;
radius 1;
state (short spin);
neighbor Moore[8]([0,-1]N,[-1,-1]NW,
[-1,0]W,[-1,1]SW, [0,1]S,
[1,1]SE,[1,0]E,[1,-1]NE);
}
int i;
short sum = 0;
{
for (i = 0; i < 8; i++)
sum = Moore[i]_spin + sum;
if (sum == 4)
{ if (cell_spin == down)
update(cell_spin, up);
else
update(cell_spin,down);
}
}

FIGURE 6.5 The Q2R Ising model written in the CARPET language.

TABLE 6.1 Performance of the Q2R Ising Program

Automata 1 2 4 8 12
size Proc Procs Procs Procs Procs

1152 × 1152 17.2 8.1 4.4 2.2 1.4

2304 × 2304 67.4 37.8 17.4 9.3 5.8
4608 × 4608 275.7 139.8 69.2 33.9 23.8
9216 × 9216 1035.4 561.8 268.8 144.1 94.9

Note: Elapsed time in sec × 100 steps

cells become sick with a chance of 55% when next to a sick cell. Sick cells recover at any given time step
with a probability of 80%, but have a 10% chance of dying. The simple behavior of a single cell can
be programmed in CARPET as shown in Figure 6.7 and the global behavior of the epidemic implicitly
emerges by the parallel execution of a large number of cells that dynamically interact according to a simple
neighbor pattern.
Such a CA program can be used to investigate spatial clustering effects in the spread of a simple epidemic
or it can be the basic skeleton for a more complex model with many degrees of freedom. Its use makes
it very easy to try and test different scenarios and “what if ” situations such as effects of vaccination,
preventive measures, or the effects of new drugs.
Table 6.2 shows the performance results (elapsed time in seconds) up to 12 processors for the execution
of 100 iteration steps considering four different lattice sizes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 97 — #13

6-98 Handbook of Bioinspired Algorithms and Applications

Ising
1152x1152 2304x2304 4068x4068 9216x9216

14

12

10

8
Speedup

0
0 2 4 6 8 10 12 14
Processors

FIGURE 6.6 The speedup of the Q2R Ising CA program.

Figure 6.8 shows the corresponding speedup. As the number of processors increases there is a cor-
responding increase in the relative speedup, which implies that the computational power of the cluster
processors is exploited in an efficient way. In fact, efficiency of the parallel epidemics program ranges from
0.97 to 1, this means that all the CPUs are fully used during the CA program execution.

6.6 Conclusion
A large number of applications can be naturally expressed by combining the emergent system model and
the massively parallel paradigm. In many cases, designers do not use this approach because the available
tools do not support it. However, parallel implementation of CA-based emergent computation systems
and phenomena represent a valid and effective approach in the study of several classes of problems. These
kinds of simulation are very helpful in vital scientific areas such as biology, physics, chemistry, medicine,
social science, and economy. Cellular automata are a viable abstract model for representing emergent
decentralized systems and phenomena in those and other areas.
In this chapter, we discussed how to use parallel CA for complex system development and described
the main features of parallel CA environments for developing scalable emergent systems and phenomena.
Parallel cellular languages and tools provide a high-level framework for emergent computation and, at the
same time, they offer a scalable setting for getting high performance using parallel architectures. While
efforts in traditional sequential computer languages and systems design focused on how to express and
implement imperative operations and data, the main goal of the cellular paradigm is to offer a massively
parallel computational model based on a large number of simple cellular objects and operations that are
suitable for defining emergent complex systems.
We showed, through CARPET example programs, how the combination of the CA model with parallel
computing techniques and systems could be exploited to efficiently implement emergent computation
structures. Finally, modeling and simulation work through parallel cellular methods helps researchers by

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 98 — #14

Parallel Cellular Algorithms and Programs 6-99

#define blank 0
#define healthy 1
#define sick 2
cadef
{
dimension 2;
radius 1;
state (short status);
neighbor Moore[8]([0,-1]N,[-1,-1]NW,[-1,0]W,[-1,1]SW, [0,1]S,
[1,1]SE,[1,0]E,[1,-1]NE);
parameter (probS 0.55, probR 0.8, probD 0.1);
}
int i; float probX; short cond = 0;
{
for (i = 0; i < 8 && cond == 0; i++)
if (cell_status == healthy && Moore[i]_status == sick)
{ probX = random(1);
if (probX <= probS)
update(cell_status, sick);
cond = 1;
}
if (cell_status == sick)
{ probX = random(1);
if (probX <= probR)
update(cell_status, healthy);
else
{ probX = random(1);
if(probX <= probD)
update(cell_status,blank);
}
} }

FIGURE 6.7 The simple epidemics model written in the CARPET language.

TABLE 6.2 Performance of the Epidemics Program

Automata 1 2 4 8 12
size Proc Procs Procs Procs Procs

1152 × 1152 19.3 9.5 4.9 2.4 1.6

2304 × 2304 77.5 43.4 19.1 10.2 6.7
4608 × 4608 316.4 154.2 80.5 39.5 26.9
9216 × 9216 1220.9 627.2 313.2 166.5 104.3

Note: Elapsed Time in sec × 100 steps

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 99 — #15

6-100 Handbook of Bioinspired Algorithms and Applications

14 Epidemics
1152 × 1152 2304 × 2304
4068 × 4068 9216 × 9216
12

10

Speedup 8

0
0 2 4 6 8 10 12 14
Processors

FIGURE 6.8 Speedup of the epidemics CA program.

setting up a cross-disciplinary framework and assisting the designers during software development from
the design phase to the execution, tuning, and validation phases. Therefore, these environments can be
used as virtual laboratories where scientists may work cooperatively by programming and experimenting
as in a real laboratory, getting data and knowledge on the modeled systems.
A Linux version of the CAMELot system is available online at the webpage www.icar.cnr.it/
spezzano/camelot/carpet.html. Programmers, scientists, students, and professionals may
download and use it to simulate complex systems according to the CA model.

References
[1] D. Talia, “Cellular processing tools for high-performance simulation,” IEEE Computer, 33, 44–52,
2000.
[2] J.R. Weimar, Simulation with Cellular Automata, Logos-Verlag, Berlin, 1997.
[3] M. Sipper, “The Emergence of Cellular Computing,” IEEE Computer, 32, 18–26, 1999.
[4] J. von Neumann, Theory of Self Reproducing Automata, University of Illinois Press, IL, 1966.
[5] S. Wolfram,“Computation theory of cellular automata,” Communicates in Math. Physics, 96, 15–57,
1984.
[6] D. Talia, “Implementing Standard and Nonstandard Parallel Cellular Automata in CARPET,”
In Proceedings of the 9th Euromicro Workshop on Parallel and Distributed Processing (PDP 2001),
Mantova, Italy, IEEE Computer Society Press, Washington, 2001, pp. 243–249.
[7] T. Worsch, “Simulation of cellular automata,” Future Generation Computer Systems, 16, 157–170,
1999.
[8] A. Schoneveld and J.F. de Ronde, “P-CAM: a framework for parallel complex systems simulations,”
Future Generation Computer Systems, 16, 217–234, 1999.
[9] L. Carotenuto, F. Mele, M. Mango Furnari, and R. Napolitano, “PECANS: A parallel environment
for cellular automata modeling,” Complex Systems, 10, 23–41, 1996.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 100 — #16

Parallel Cellular Algorithms and Programs 6-101

[10] T. Toffoli and N. Margolus, Cellular Automata Machines: A New Environment for Modeling. The
MIT Press, Cambridge, MA, 1986.
[11] B.L. Chamberlain et al., “The Case for high level parallel programming in ZPL,” IEEE
Computational Science & Engineering, 5, 76–86, 1998.
[12] C. Hecker, D. Roytenberg, J.-R. Sack, and Z. Wang, “System development for parallel cellular
automata and its applications,” Future Generation Computer Systems, 16, 235–247, 1999.
[13] M. Cannataro, S. Di Gregorio, R. Rongo, W. Spataro, G. Spezzano, and D. Talia, “A Parallel Cellular
automata environment on multicomputers for computational science,” Parallel Computing, 21
803–824, 1995.
[14] E. Gallopoulos, E.N. Houstis, and J.R. Rice, “Workshop on problem-solving environments:
findings and recommendations,” ACM Computing Surveys, 27, 277–279, 1995 .
[15] G. Spezzano and D. Talia, “Programming cellular automata algorithms on parallel computers,”
Future Generation Computer Systems, 16, 203–216, 1999.
[16] J.D. Eckart, “Cellang 2.0: Reference Manual,” ACM Sigplan Notices, 27, 107–112, 1992.
[17] P.Z. Zeigler et al., “The DEVS environment for high-performance modeling and simulation,” IEEE
Computational Science & Engineering, 4, 61–71, 1997.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C006” — 2005/7/20 — 19:53 — page 101 — #17

 7
Decentralized Cellular
Evolutionary
Algorithms

7.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-103

7.2 Synchronous and Asynchronous cEAs . . . . . . . . . . . . . . . . . 7-104
7.3 New cEA Variants Based on a Modified Ratio . . . . . . . . . 7-105
7.4 Selection Pressure, Grid Shape, and Time. . . . . . . . . . . . . . 7-107
7.5 Experiments in Discrete Optimization . . . . . . . . . . . . . . . . . 7-107
Massively Multimodal Deceptive Problem (MMDP) •
Frequency Modulation Sounds (FMS) • Multimodal Problem
Generator (P-PEAKS) • Error Correcting Code Design
Problem (ECC) • Maximum Cut of a Graph (MAXCUT) •
Minimum Tardy Task Problem (MTTP) • Satisfiability
Problem (SAT) • Experimental Analysis
7.6 Experiments in Continuous Optimization. . . . . . . . . . . . . 7-115
Rastrigin’s Function • Ackley’s Function •
Enrique Alba Fractal Function • Experimental Analysis
Bernabe Dorronsoro 7.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-118
Mario Giacobini Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-119
Marco Tomassini References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-119

7.1 Introduction
This chapter focuses on the class of algorithms called cellular evolutionary algorithms (cEAs). Here,
we present the canonical algorithm and suggest the interesting variants targeted to solve complex problems
accurately with a minimum effort for customization. These techniques, also called diffusion or fine-
grained models, have been popularized, among others, by the early work of Gorges-Schleuter [1] and
Manderick and Spiessens [2]. The basic idea is to add some structure to the population of tentative
solutions. The pursued effect is to improve on the diversity and exploration capabilities of the algorithm
while, still admitting an easy combination with local search and other search techniques to improve on
exploitation.

7-103

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 103 — #1

7-104 Handbook of Bioinspired Algorithms and Applications

These structured models are based on a spatially distributed population in which genetic operations
may only take place in a small neighborhood of each individual. Usually, individuals are arranged on
a regular grid of dimensions, d = 1, 2, or 3. cEAs are a kind of decentralized EA model [3]. They are not
a parallel implementation of an EA; in fact, although parallelism could be used to speed up the search,
we do not address parallel implementations in this work. In addition, it is worth remarking that, although
SIMD (single instruction stream–multiple data stream) machine implementations were popular a decade
ago, this is no longer true, and today the best distributed implementation of a cEA should make use of
domain decomposition on clusters of networked machines.
Although, fundamental theory is still an open research line for cEAs, they have been empirically reported
as being useful in maintaining diversity, and promoting slow diffusion of solutions through the grid. Part
of their behavior is due to a lower selection pressure compared with that of panmictic EAs (here panmictic
means that, any individual may mate with any other individual in the population). The influence of the
selection method [4,5], neighborhood [6], and grid topology on the efficiency of cEAs in comparison with
other EAs [7] have been investigated in detail, and tested on different applications, such as combinatorial
and numerical optimization.
Cellular evolutionary algorithms can be seen as stochastic cellular automata (CA) [8,9] where the
cardinality of the set of states is equal to the number of points in the search space. CAs, as well as cEAs,
usually assume a synchronous or “parallel” update policy, in which all the cells are updated simultaneously.
However, this is not the only option available. Indeed, several works on asynchronous CAs have shown
that sequential update policies have a marked effect on their dynamics [10,11]. Furthermore, the shape of
the structure in which individuals evolve has a deep impact on the performance of the cEA. The algorithm
admits a special, easy modulation of its shape that can sharpen the exploration or the exploitation
capabilities of the canonical technique, as shown in Reference 7. Thus, it is interesting to investigate
asynchronous cEAs and nonsquare shaped cEAs, in order to analyze their problem solving capabilities,
which is the subject of the second part of this chapter.
This work is organized as follows. Section 7.2 contains some background on synchronous and asyn-
chronous cEAs. In Section 7.3, we discuss the ability of cEAs for changing their behavior depending
on the population shape. In Section 7.4, we illustrate, quantitatively, the modifications on the selection
intensity due to asynchronicity and population shape. We deal with a set of discrete and continuous
benchmark problems in Sections 7.5 and 7.6, respectively, with the goal of analyzing the actual computa-
tional power of the algorithms. Finally, Section 7.7 discusses our conclusions, as well as some comments
on future work.

7.2 Synchronous and Asynchronous cEAs

In this section we summarize the canonical behavior of cEAs. A cEA starts with the cells (individuals) in a
random state and proceeds by successively updating them using evolutionary operators, until a termination
condition is met. Updating a cell in a cEA means selecting two parents in the individual’s neighborhood
(including the individual itself), applying genetic operators to them, and finally replacing the individual
if an offspring has a better fitness or using another replacement policy.
Cells can be updated synchronously or asynchronously. In synchronous (parallel) update all the cells
change their states simultaneously, while in asynchronous, or sequential, update cells are updated one at
a time in some order. There exist many ways for sequentially updating the cells of a cEA (an excellent
discussion of asynchronous update in cellular automata, which are essentially the same system as a cEA,
is available in Reference 10). We consider four asynchronous update methods: fixed line sweep (LS), fixed
random sweep (FRS), new random sweep (NRS), and uniform choice (UC).

• In fixed line sweep, the simplest method, the n grid cells are updated sequentially (1, 2, . . . , n) line
after line.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 104 — #2

Decentralized Cellular Evolutionary Algorithms 7-105

• In fixed random sweep, the next cell to be updated is chosen with uniform probability without
j p
replacement; this will produce a certain update sequence (c1 , c2k , . . . , cnm ), where cq means that
cell number p is updated at time q and ( j, k, . . . , m) is a permutation of the n cells. The same
permutation is then used for all update cycles.
• The new random sweep method works like FRS, except that a new random cell permutation is used
for each sweep through the array.
• In uniform choice, the next cell to be updated is chosen at random with uniform prob-
ability and with replacement. This corresponds to a binomial distribution for the update
probability.

The concept of generation that is customarily used in EAs and in synchronous cEAs has to be replaced
by time step in the asynchronous cases. A time step is defined as updating n times sequentially, which
corresponds to updating all the n cells in the grid for LS, FRS, and NRS, and possibly less than n different
cells in the UC method, since some cells might be updated more than once. It should be noted that, with
the exception of LS, the other asynchronous updating policies are stochastic, representing an additional
source of nondeterminism besides that of the genetic operators.

7.3 New cEA Variants Based on a Modified Ratio

After explaining the basic algorithm and the asynchronous variants in Section 7.2, we now pro-
ceed to characterize the population grid itself. For this goal, we use the “radius” definition given in
Reference 7, which is refined from the seminal one appeared in Reference 6 to account for non-
square grids. The grid is considered to have a radius equal to the dispersion of n ∗ points in a circle
centered in (x̄, ȳ) (Equation [7.1]). This definition always assigns different numerical values to different
grids.

  n∗ n∗
(xi − x̄)2 + (yi − ȳ)2 i=1 xi i=1 yi
rad = , x̄ = , ȳ = . (7.1)
n∗ n∗ n∗

Although it is called a “radius,” rad measures the dispersion of n ∗ patterns. Other possible measures
for symmetrical topologies would allocate the same numerical value to different topologies (which is
undesirable). Two examples are, the radius of a circle surrounding a rectangle containing the topology, or
an asymmetry coefficient. The definition not only characterizes the grid shape but also provides a radius
value for the neighborhood. As proposed in Reference 6, the grid-to-neighborhood relationship can be
quantified by the ratio between their radii (Equation [7.2]).

radNeighborhood
ratiocEA = . (7.2)
radTopology

When solving a given problem with a constant number of individuals (n = n ∗ , for making fair
comparisons) the topology radius will increase as the grid gets thinner (Figure 7.1[b]). Since the neighbor-
hood is kept constant in size and shape throughout this chapter (we always use linear5 (L5), Figure 7.1[a]),
the ratio will be smaller as the grid gets thinner.
This ratio value directly influences the behavior of the algorithm. During the search for reducing the
ratio means reducing the global selection intensity on the population (see Section 7.4), thus promoting
exploration, that is, the importance of such a ratio measure. This is expected to allow a higher diversity
in the genotype that could improve the results in difficult problems (such as in multimodal or epistatic
tasks). On the other hand, the search performed inside each neighborhood is guiding the exploitation
of the algorithm. In this chapter we study how the ratio affects the search efficiency over a variety of
domains. Changing the ratio during the search is a unique feature of cEAs that can be used to shift from

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 105 — #3

7-106 Handbook of Bioinspired Algorithms and Applications

(a) (b) rad1

rad2

radlinear 5 = 2 + 2 = 0.8944 rad2 > rad1 then radio2 < ratio1

5

FIGURE 7.1 (a) Radius of neighborhood L5. (b) 5 × 5 = 25 and 3 × 8 ≈ 25 grids; equal number of individuals
with two different ratios.

(a) 0 1 2 3 4 5 6 7 (b)
0
0 1 2 3 4 5 6 7 8 9 10 1112 13 14 15
1
2 Relocation 0
3 1
4 2
5 3
6
7 (2, 4) ([2·8 + 4] div 16, [2·8 + 4] mod 16) = (1, 4)

FIGURE 7.2 Relocation of an individual and its neighbors when the grid changes from (a) 8 × 8 to other (b) with
shape 4 × 16.

exploration to exploitation at a minimum complexity without introducing another new algorithm family
in the literature.
Many techniques for managing the exploration/exploitation trade-off are possible. Among them, it is
worth to mention the heterogeneous EAs [12,13], in which algorithms with different features run in
separate subpopulations and collaborate in order to avoid premature convergence. A different alternative
is using Memetic Algorithms [14], in which local search is combined with the genetic operators in order to
promote local exploitation.
Since, a shift between exploration and exploitation can be made by changing the shape of the population
(and thus its radius), one can think on changing it during the search. Hence, we theoretically consider
the population as a list of length m · n, such that, the first row of the m × n grid is composed by the first
n individuals of the list, the second row is made up with the next n individuals, and so on. Therefore,
when performing a change from a m × n grid to a new m
× n
grid (being m · n = m
· n
) the individual
placed at position (i, j) will be relocated as follows:

(i, j) → ([i ∗ n + j] div n

, [i ∗ n + j] mod n
). (7.3)

We call this redistribution method contiguous, because the new grid is filled up by following the order of
appearance of the individuals in the list. Figure 7.2 contains an example of a grid change from 8×8 to 4×16.
It can be shown, how an individual in position (2, 4) is relocated at (1, 4), changing its neighbors placed
at its north and south positions, and keeping close to those placed at its east and west positions. Hence,
the change in the grid shape can be seen as an actual migration of individuals among neighborhoods,
which will introduce additional diversity into the population for the forthcoming generations. Note that
in this chapter we only use static ratios, that is, grid and neighborhood shapes that do not change during
the run.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 106 — #4

Decentralized Cellular Evolutionary Algorithms 7-107

1.0

0.9
0.8

0.7

Best proportion
0.6
0.5
0.4
0.3
0.2
L5 (32 × 32)
0.1 C21 (64 × 64)
0
0 4 8 12 16 20 24 28 32 36 40
Number of generations

FIGURE 7.3 Growth curves of the best individual for two cEAs with different neighborhood and population shapes,
but similar ratio values. The vertical axis represents the proportion of population consisting of best individual as a
function of time.

7.4 Selection Pressure, Grid Shape, and Time

Selection pressure is related to the concept of takeover time, which is the time taken for a single best
individual to colonize the whole population with copies of itself under the effects of selection only [15].
Shorter takeover time mean a more intense selection.
Algorithms with similar ratio show a similar selection pressure, as stated in Reference 5. In Figure 7.3,
we plot such a similar behavior for two algorithms with different neighborhood and population radii, but
having two similar ratio values. The algorithms plotted are those using a L5 neighborhood with a 32 × 32
population, and a compact21 (C21), neighborhood with a population of 64 × 64 individuals. In the C21
neighborhood a central cell is surrounded by two cells in all directions, including the diagonals, and the
four corner cells are cut out.
Hence, it may be very interesting to see how the shape of the grid influences the search of the algorithm.
Thus, we study the selection pressure for synchronous cGAs with different grid shapes. In Figure 7.4 we
plot the selection pressure for different cGAs using L5 neighborhood and six possible grid shapes for a
population of 1024 individuals. Note that the selection pressure induced in synchronous rectangular grids
falls under the curve for a synchronous square grid (32 × 32 population), which means that thinner grids
favor a more explorative style of search.
If we now keep the shape of the grid constant (say a square), but allow the cell update mode to
change, we observe a similar effect on the selection pressure: the global selection pressure induced by
the various asynchronous policies fall between the low synchronous limit and the high panmictic bound
(see Figure 7.5, [16]). Thus, by varying the update policies it is possible to influence the explorative or
exploitative character of the search.

7.5 Experiments in Discrete Optimization

In this section we present the set of discrete problems that we have chosen, to study the behavior of our
algorithms. The selected benchmark is representative because it contains many different interesting fea-
tures found in optimization, such as epistasis, multimodality, deceptiveness, use of constraints, parameter
identification, and problem generators. These are important ingredients in any work trying to evaluate
algorithmic approaches with the objective of getting reliable results, as stated by Whitley et al. [17].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 107 — #5

7-108 Handbook of Bioinspired Algorithms and Applications

1
Ratio 0.003 (1 × 1024 population)
0.9 Ratio 0.006 (2 × 512 population)
Ratio 0.012 (4 × 256 population)
0.8 Ratio 0.024 (8 × 128 population)
Ratio 0.047 (16 × 256 population)

Best individual proportion

0.7 Ratio 0.069 (32 × 32 population)

0.6

0.5

0.4

0.3

0.2

0.1

100 101 102 103

Number of generations

FIGURE 7.4 Takeover times with tournament selection using a L5 neighborhood in a population of 1024 individuals
with different grid shapes; mean values over 100 runs. The vertical axis represents the proportion of population
consisting of best individual as a function of time. Horizontal axis is in logarithmic scale.

0.9

0.8

0.7
Best individual proportion

0.6

0.5

0.4

0.3 Synchronous
UC
0.2 NRS
LS
0.1 Panmictic

0
0 10 20 30 40 50
Time steps

FIGURE 7.5 Takeover times with tournament selection using a L5 neighborhood in a 32 × 32 grid; mean values over
100 runs. The vertical axis represents the proportion of population consisting of best individual as a function of time.

We experiment with the set of problems studied in Reference 7 which includes the massively multimodal
deceptive problem (MMDP), the frequency modulation sounds (FMSs), and the multimodal problem
generator P-PEAKS; we then extend this basic three-problem benchmark with error correcting code
design (ECC), maximum cut of a graph (MAXCUT), the minimum tardy task problem (MTTP), and the
satisfiability problem (SAT).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 108 — #6

Decentralized Cellular Evolutionary Algorithms 7-109

Unitation Subfunction value Massively Multimodal Deceptive Problem

0 1.000000

Subfunction value
1.0
1 0.000000 0.8
2 0.360384 0.6
3 0.640576 0.4
4 0.360384 0.2
5 0.000000 0.0
6 1.000000 0 1 2 3 4 5 6
Unitation

FIGURE 7.6 Basic deceptive bipolar function (si ) for MMDP.

The choice of this set of problems is justified by both their difficulty and their application domains
(parameter identification, telecommunications, combinatorial optimization, scheduling, etc.). This gives
us a high level of confidence in the results obtained, although the evaluation of conclusions is consequently
more laborious than with a small test suite.
The problems selected for this benchmark are explained in Sections 7.5.1 to 7.5.7. We include the
explanations in the chapter to make it self-contained and to avoid the typical small information lacks that
could preclude other researchers from reproducing the results. Finally, in Section 7.5.8 we present and
analyze the results.

7.5.1 Massively Multimodal Deceptive Problem (MMDP)

The MMDP is a problem that has been specifically designed to be difficult for an EA [18]. It is made up
of k deceptive subproblems (si ) of six bits each, whose value depends on the number of ones (unitation)
that a binary string has (see Figure 7.6). It is easy to see (graphic of Figure 7.6) that these subfunctions
have two global maxima and a deceptive attractor in the mid-point.
In MMDP, each subproblem si contributes to the fitness value according to its unitation (Figure 7.6).
The global optimum has a value of k and it is attained when every subproblem is composed of zero
or six ones. The number of local optima is quite large (22k ), while there are only 2k global solutions.
Therefore, the degree of multimodality is regulated by the k parameter. Here, we use a considerably large
instance of k = 20 subproblems. The instance we try to maximize for solving the problem is shown in
Equation (7.4), and its maximum value is equal to k.


k
fMMDP ( s ) = fitnesssi . (7.4)
i=1

7.5.2 Frequency Modulation Sounds (FMS)

The FMS problem [19] is defined as, determining the six real parameters, x = (a1 , w1 , a2 , w2 , a3 , w3 ),
of the frequency modulated sound model given in Equation (7.5) for approximating it to the sound wave
given in Equation (7.6) (where θ = 2 · π/100). The parameters are defined in the range [−6.4, +6.35],
and we encode each parameter into a 32 bit substring in the individual.

y(t ) = a1 · sin(ω1 · t · θ + a2 · sin(ω2 · t · θ + a3 · sin(ω3 · t · θ))), (7.5)

y0 (t ) = 1.0 · sin(5.0 · t · θ − 1.5 · sin(4.8 · t · θ + 2.0 · sin(4.9 · t · θ))). (7.6)

The goal is to minimize the sum of square errors given by Equation (7.7). This problem is a highly com-
plex multimodal function having strong epistasis, with optimum value 0. Due to the extreme difficulty of
solving this problem with high accuracy, without applying local search or specific operators for continuous

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 109 — #7

7-110 Handbook of Bioinspired Algorithms and Applications

optimization (such as gradual GAs [13]), we stop the algorithm when the error falls below 10−2 . Hence,
our objective for this problem will be to minimize Equation (7.7)


100
fFMS (x ) = (y(t ) − y0 (t ))2 . (7.7)
t =0

7.5.3 Multimodal Problem Generator (P-PEAKS)

The P-PEAKS problem [20] is a multimodal problem generator. A problem generator is an easily param-
eterizable task, which has a tunable degree of epistasis, thus admitting to derive instances with growing
difficulty. In addition, using a problem generator removes the opportunity to hand-tune algorithms to
a particular problem, therefore allowing a larger fairness when comparing algorithms. With a problem
generator, we evaluate our algorithms on a high number of random problem instances, since a different
instance is solved each time the algorithm runs, the predictive power of the results for the problem class
as a whole is increased.
The idea of P-PEAKS is to generate P random N -bit strings that represent the location of P peaks in
the search space. The fitness value of a string is the number of bits the string has in common with the
nearest peak in that space, divided by N (as shown in Equation [7.8]). By using a small/large number of
peaks we can get weakly/strongly epistatic problems. In this chapter we have used an instance of P = 100
peaks, of length N = 100 bits each, which represents a medium/high epistasis level [7]. The maximum
fitness value for this problem is 1.0.

1
fP-PEAKS (x ) = max {N − Hamming D(x , Peak i )}. (7.8)
N 1≤i≤p

7.5.4 Error Correcting Code Design Problem (ECC)

The ECC problem was presented in Reference 21. We will consider a three-tuple (n, M , d), where n is
the length of each codeword (number of bits), M is the number of codewords, and d is the minimum
Hamming distance between any pair of codewords. Our objective will be to find a code that has a value
for d as large as possible (reflecting greater tolerance to noise and errors), given previously fixed values
for n and M . The problem we have studied is a simplified version of that in Reference 21. In our case,
we search half of the codewords (M /2) that will compose the code, and the other half is made up by the
complement of the codewords computed by the algorithm.
The fitness function to be maximized is:

1
fECC = M M , (7.9)
i=1 j=1,i =j dij−2

where dij represents the Hamming distance between codewords i and j in the code C (made up of
M codewords, each of length n). Here, we consider an instance where M = 24 and n = 12. The search
space is of size 4096
24 , which is approximately 10 . The optimum solution for M = 24 and n = 12 has a
87

fitness value of 0.0674 [22].

7.5.5 Maximum Cut of a Graph (MAXCUT)

The MAXCUT problem looks for a partition of the set of vertices (V ) of a weighted graph G = (V , E)
into two disjoint subsets V0 and V1 so that the sum of the weights of the edges with one endpoint in V0
and the other in V1 is maximized. For encoding the problem, we use a binary string, (x1 , x2 , . . . , xn ), of
length n where each digit corresponds to a vertex. If a digit is 1 then the corresponding vertex is in set V1 ;

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 110 — #8

Decentralized Cellular Evolutionary Algorithms 7-111

if it is 0 then the corresponding vertex is in set V0 . The function to be maximized [23] is:


n−1 
n
fMAXCUT (x ) = wij · [xi · (1 − xj ) + xj · (1 − xi )]. (7.10)
i=1 j=i+1

Note that wij contributes to the sum only if nodes i and j are in different partitions. While one can
generate different random graph instances to test the algorithm, here we have used the case “cut20.09,”
with 20 vertices and a probability of 0.9 of having an edge between any two randomly chosen vertices.
The maximum fitness value for this instance is 56.740064.

7.5.6 Minimum Tardy Task Problem (MTTP)

The MTTP [24] is a task-scheduling problem wherein each task i from the set of tasks T = {1, 2, . . . , n}
has a length li (the time it takes for its execution), a deadline di (before which a task must be scheduled,
and its execution completed), and a weight wi . The weight is a penalty that has to be added to the objective
function in the event that the task remains unscheduled. The lengths, weights, and deadlines of the tasks
are all positive integers. Scheduling the tasks of a subset S of T is to find the starting time of each task in S,
such that at most one task at a time is performed and each task finishes before its deadline.
We characterize a one-to-one scheduling function g defined on a subset of tasks S ⊆ T : S  → Z + ∪ {0},
so that for all tasks i, j ∈ S has the following properties:

1. A task cannot be scheduled before the previous one is completed: g (i) < g ( j) ⇒ g (i) + li ≤ g ( j).
2. Every task finishes before its deadline: g (i) + li ≤ di .

The objective function for this problem is to minimize the sum of the weights of the unscheduled tasks.
Therefore, the optimum scheduling minimizes Equation (7.11):

fMTTP (x ) = wi . (7.11)
i∈T −S

The schedule of tasks S can be represented by a vector x = (x1 , x2 , . . . , xn ) containing all the tasks
ordered by its deadline. Each xi ∈ {0, 1}, where if xi = 1 then task i is scheduled in S, while if xi = 0
means that task i is not included in S. The fitness function is the inverse of Equation (7.11), as described
in Reference 23. We have used in this study an instance called “mttp20,” with size 20, and maximum fitness
value of 0.02439.

7.5.7 Satisfiability Problem (SAT)

The SAT problem has received much attention by the scientific community since it plays a central role in
NP-completeness [25]. The SAT problem was the first that was demonstrated to belong to the NP class of
problems.
The SAT problem consists of assigning values to a set of n boolean variables x = (x1 , x2 , . . . , xn ) such
that they satisfy a given set of clauses c1 (x ), . . . , cm (x ), where ci (x ) is a disjunction of literals, and a literal
is a variable or its negation. Hence, we can define SAT as a function f : B n → B, B = {0, 1} like:

fSAT (x ) = c1 (x ) ∧ c2 (x ) ∧ · · · ∧ cm (x ). (7.12)

An instance of SAT, x , is called satisfiable if fSAT (x ) = 1, and unsatisfiable otherwise. A k-SAT instance
consists of clauses with length k. When k ≥ 3 the problem is NP-hard [25]. In this chapter we will consider
an instance of 3-SAT made up of 430 clauses and 100 variables. This instance belongs to the well-known

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 111 — #9

7-112 Handbook of Bioinspired Algorithms and Applications

phase transition of hard SAT instances. The fitness function is a linear function of the number of satisfied
clauses. In this, we use the so-called stepwise adaptation of weights (SAWs) [26]:

fSAW (x ) = w1 · c1 (x ) + · · · + wm · cm (x ). (7.13)

This function weighs the values of the clauses with wi ∈ B in order to give more importance to those
clauses which are not yet satisfied by the current best solution. These weights are adjusted dynamically
according to wi = wi + 1 − ci (x ∗ ), x ∗ being the current fittest individual.

7.5.8 Experimental Analysis

Although a full-length study of the problems presented in Section 7.5.7 is beyond the scope of this work,
we present results comparing synchronous and asynchronous cEAs, and also cEAs having different values
of the ratio, always with a constant neighborhood shape (L5). Note that it is not our aim to compare cEAs
performance with state-of-the-art algorithms and heuristics for combinatorial and numerical optimi-
zation. To end this, we should at least tune the parameters and include local search capabilities in the
algorithm, which is not the case. Thus, the results only pertain to the relative performance of the different
cEA update methods and ratios among themselves.
In this section we present the results of solving some problems using JCell v1.5, our custom simulation
program written in Java, with three different static ratios. The configuration of the algorithm for the
binary encoded problems is shown in Table 7.1, and the static ratios used are shown in Table 7.2.
The following tables show the results for the problems mentioned preciously: MMDP (Table 7.3),
FMS (Table 7.4), P-PEAKS (Table 7.5), ECC (Table 7.6), MAXCUT (Table 7.7), MTTP (Table 7.8), and
SAT (Table 7.9). In these tables we report the average of the final best fitness of the algorithm, the
average number of evaluations to obtain the optimum value (if obtained), and the hit rate (percentage
of successful runs). Therefore, we are analyzing the final distance to the optimum (especially interesting
when the optimum is not found), the effort of the algorithm, and the expected efficacy of the algorithm,
respectively. In order to get reliable results, we have performed 100 independent runs for any algorithm
and for every problem in the test suite.
From the inspection of these tables some conclusions can be clearly drawn. First, the studied asyn-
chronous algorithms tend to need a smaller number of generations than the synchronous ones to locate
an optimum, in general. Moreover, the differences among asynchronous and synchronous algorithms

TABLE 7.1 Parameterization Used in the Algorithm for the

Binary Encoded Problems

Population size 400 individuals

Selection of parents Binary tournament + binary tournament
Recombination DPX, pc = 1.0
Bit mutation Bit-flip, pm = 1/L (10/L for FMS)
Individual length L
Replacement Rep_if_Better

DPX indicates standard double point crossover.

TABLE 7.2 Studied Ratios

Name (shape of population) Value of ratio

Square (20 × 20 individuals) 0.11

Rectangular (10 × 40 individuals) 0.075
Narrow (4 × 100 individuals) 0.031

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 112 — #10

Decentralized Cellular Evolutionary Algorithms 7-113

TABLE 7.3 MMDP with a Maximum of

1000 Generations

Average
solution Average Hit rate
Algorithm (best = 20) generations (%)

Square 19.813 214.2 57

Rectangular 19.824 236.1 58
Narrow 19.842 299.7 61
LS 19.518 343.5 23
FRS 19.601 209.9 31
NRS 19.536 152.9 28
UC 19.615 295.7 36

TABLE 7.4 FMS Problem with a Maximum of

3000 Generations

Average
solution Average Hit rate
Algorithm (best ≥ 100) generations (%)
Square 90.46 437.4 57
Rectangular 85.78 404.3 61
Narrow 80.76 610.9 63
LS 81.44 353.4 58
FRS 73.11 386.2 55
NRS 76.21 401.5 56
UC 83.56 405.2 57

TABLE 7.5 P-PEAKS Problem with a Maximum of

100 Generations

Average
solution Average Hit rate
Algorithm (best = 1) generations (%)
Square 1.0 51.8 100
Rectangular 1.0 50.4 100
Narrow 1.0 53.9 100
LS 1.0 34.8 100
FRS 1.0 38.4 100
NRS 1.0 38.8 100
UC 1.0 40.1 100

TABLE 7.6 ECC Problem with a Maximum of

500 Generations

Average
solution Average Hit rate
Algorithm (best = 0.0674) generations (%)
Square 0.0670 93.9 85
Rectangular 0.0671 93.4 88
Narrow 0.0673 104.2 94
LS 0.0672 79.7 89
FRS 0.0672 82.4 90
NRS 0.0672 79.5 89
UC 0.0671 87.3 86

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 113 — #11

7-114 Handbook of Bioinspired Algorithms and Applications

TABLE 7.7 MAXCUT Problem with a Maximum of

100 Generations

Average
solution Average Hit rate
Algorithm (best = 56.74) generations (%)

Square 56.74 11.3 100

Rectangular 56.74 11.0 100
Narrow 56.74 11.9 100
LS 56.74 9.5 100
FRS 56.74 9.7 100
NRS 56.74 9.6 100
UC 56.74 9.6 100

TABLE 7.8 MTTP with a Maximum of 50 Generations

Average
solution Average Hit rate
Algorithm (best = 0.02439) generations (%)

Square 0.02439 8.4 100

Rectangular 0.02439 8.3 100
Narrow 0.02439 8.9 100
LS 0.02439 5.9 100
FRS 0.02439 6.2 100
NRS 0.02439 6.3 100
UC 0.02439 6.3 100

TABLE 7.9 SAT Problem with a Maximum of

3000 Generations

Average
solution Average Hit rate
Algorithm (best = 430.0) generations (%)

Square 429.54 703.1 79

Rectangular 429.67 706.3 84
Narrow 429.61 763.7 81
LS 429.52 463.2 78
FRS 429.67 497.7 85
NRS 429.49 610.5 75
UC 429.50 725.5 76

are statistically significant (with two exceptions), thus indicating that the asynchronous versions perform
more efficiently with respect to cEAs with a changing ratio.
On the contrary, synchronous algorithms perform like asynchronous or even better in terms of the
percentage of solutions found (hit rate), while the quality of solutions found by the algorithms does not
always have significant differences (the exceptions are probably due to the difference on the hit rate).
Another interesting result is that, we can define two classes of problems: those solved by any method
to optimality (100% hit rate) and those in which no 100% rate is achieved at all. Problems seem to be
amenable for cEAs directly, or to need some (yet unstudied) help, for example, by including local search.
In order to summarize the large set of results and get some useful conclusion, we present a ranking with
the best algorithms by following three different metrics: average best final solution, average number of
generations on success, and hit rate. Table 7.10 shows the three mentioned rankings. These rankings have

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 114 — #12

Decentralized Cellular Evolutionary Algorithms 7-115

TABLE 7.10 Ranking of the Algorithms with Discrete Problems

Average Average
solution generations Hit rate (%)

1 Narrow 10 1 LS 14 1 Narrow 6
1 Rectangular 10 2 NRS 16 2 Rectangular 10
3 Square 14 3 FRS 18 3 FRS 14
4 FRS 15 4 UC 30 4 LS 15
5 LS 18 5 Rectangular 33 5 Square 17
5 UC 18 6 Square 37 6 UC 19
7 NRS 21 7 Narrow 48 7 NRS 21

been computed by adding the position (from better to worse: 1, 2, 3, . . .) that algorithms are allocated for
the previous results presented from Table 7.3 to Table 7.9, according to the three criteria.
As we would expect after the previous comments, according to the average final best fitness and hit rate
criteria, synchronous algorithms with any of the three studied ratios are, in general, more accurate than
all the asynchronous ones for our test problems, with a special leading position for narrow population
grids. On the other hand, asynchronous versions clearly outperform any of the synchronous algorithms,
in terms of the average number of generations (efficiency), with a trend toward LS as being the best ranked
flavor of cEA for our discrete test suite.

7.6 Experiments in Continuous Optimization

We will extend the work of the previous sections by testing all the algorithms with some continuous
functions in order to get a more extensive study. This study may be interesting for analyzing the behavior
of the algorithms in continuous optimization, in contrast to the study performed on discrete optimi-
zation. The functions selected for the study are the three typical multimodal numerical benchmarks:
Rastrigin’s (RASTR), Ackley’s (ACKL), and fractal (FRAC) function. These three problems are real-
coded in the algorithms, while previously the standard binary-coded individuals were used (Section 7.5).
That is the cause for our special interest on experimenting with a more traditional global optimiza-
tion. The codification employed for these three problems has been made by following Michalewicz’s
implementation [27].

7.6.1 Rastrigin’s Function

The generalized RASTR function is a sinusoidally modulated function with a global minimum of zero at
the origin. It is a typical example of nonlinear multimodal function. It was first proposed by Rastrigin
as a two-dimensional function and has been generalized by Mühlenbein and Schlierkamp-Voosen [28].
This function is a fairly difficult problem due to its large search space and large number of local minima,
although the function is separable and the local minima are symmetrical (Equation [7.14]).


n
fRASTR (x) = nA + xi2 − A cos(ωxi ). (7.14)
i=1

The constants are given by A = 10 and ω = 2π . The domain of variables xi , i = 1, . . . , n, is

−5.12 ≤ xi ≤ 5.12 and n = 10. The function has a global minimum at the point f (0) = 0.

7.6.2 Ackley’s Function

Ackley’s function is a multimodal test function obtained by cosine modulation of an exponential func-
tion. Originally proposed by Ackley [29] as a two-dimensional function, it has been generalized by

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 115 — #13

7-116 Handbook of Bioinspired Algorithms and Applications

Bäck et al. [30] (see Equation [7.15]). Contrary to Rastrigin’s, Ackley’s function is not separable, even
though it shows a regular arrangement of the local optima.

  1/2  
1 2 1
n n
fACKL (x) = −a exp −b xi − exp cos(cxi ) + a + e. (7.15)
n n
i=1 i=1

The constants are a = 20, b = 0.2, and c = 2π. The variables xi , i = 1, . . . , n are in the domain −32.768 ≤
xi ≤ 32.768. This function has a global minimum at the point f (0) = 0.

7.6.3 Fractal Function

This function has been taken from Reference 31, were its construction, as well as motivations for introduc-
ing it as a test problem, are given. Indeed, the function allows the degree of ruggedness to be controlled,
and it is likely to capture features of real-world noisy objective functions.


n
fFRAC (x) = (C
(xi ) + xi2 − 1), (7.16)
i=1

where

 C(z)
 if z = 0,

C(1)|z|2−D
C (z) =


1 if z = 0,

and
∞
 1 − cos(b j z)
C(z) = .
b (2−D)j
j=−∞

For the runs, we have chosen the constants D = 1.85 and b = 1.5. The 20 variables xi (xi = 1, . . . , 20)
vary in the range [−5, 5]. The infinite sum in the function C(z) is practically calculated starting with
j = 0 and alternating the signs of the j values. The sum is stopped when the relative difference between
its previous and present value is lower than 10−8 , or when j = 100 is reached.

7.6.4 Experimental Analysis

In this section we will study the results of our experiments with the proposed continuous problems,
as we did in Section 7.5.8 for the discrete case. We maintain, in this case, the ratios used for the synchronous
algorithms with respect to those studied in Section 7.5, as well as the asynchronous update criteria. On
the other hand, we needed a specific configuration for the genetic operators and theirs probabilities in the
case of real-encoded problems. This codification is given in detail in Table 7.11.
The following tables present the results of our experiments with RASTR (Table 7.12), ACKL (Table 7.13),
and FRAC (Table 7.14) problems. Like in the case of discrete problems, these tables contain values for
the average of the final best fitness, the average generations needed for finding it, and the hit rate. These
three values are calculated over 100 independent runs. For the three real-coded problems a run is stopped
successfully as soon as an individual is found with fitness within 0.1 from the optimum.
The results obtained with continuous problems are not as clear as in the discrete case. Regarding the
average number of generations needed to find an optimal solution, asynchronous algorithms do not
always perform a lesser number of generations with respect to synchronous ones; for example, ACKL and
FRAC problems. Differences among synchronous and asynchronous algorithms are usually significant.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 116 — #14

Decentralized Cellular Evolutionary Algorithms 7-117

TABLE 7.11 Parameterization Used in the Algorithm for the

Real-Encoded Problems

Population size 400 individuals

Selection of parents Binary tournament + binary tournament
Recombination AX, pc = 1.0
Bit mutation Uniform, pm = 1/2L
Individual length L
Replacement Rep_if_Better

AX stands for standard arithmetic crossover.

TABLE 7.12 RASTR Problem with a Maximum of

700 Generations

Average
solution Average Hit rate
Algorithm (best ≤ 0.1) generations (%)
Square 0.0900 323.8 100
Rectangular 0.0883 309.8 100
Narrow 0.0855 354.2 100
LS 0.0899 280.9 100
FRS 0.0900 289.6 100
NRS 0.0906 292.2 100
UC 0.0892 292.4 100

TABLE 7.13 ACKL Problem with a Maximum of

500 Generations

Average
solution Average Hit rate
Algorithm (best ≤ 0.1) generations (%)
Square 0.0999 321.7 78
Rectangular 0.0994 293.1 73
Narrow 0.1037 271.9 65
LS 0.0932 302.0 84
FRS 0.0935 350.6 92
NRS 0.0956 335.5 87
UC 0.0968 335.0 85

TABLE 7.14 FRAC Problem with a Maximum of 100

Generations

Average
solution Average Hit rate
Algorithm (best ≤ 0.1) generations (%)

Square 0.0224 75.2 94

Rectangular 0.0359 62.8 78
Narrow 0.1648 14.6 16
LS 0.0168 69.7 98
FRS 0.0151 71.5 100
NRS 0.0163 73.6 98
UC 0.0138 72.8 96

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 117 — #15

7-118 Handbook of Bioinspired Algorithms and Applications

TABLE 7.15 Ranking of the Algorithms with Continuous Problems

Average Average
solution generations Hit rate (%)

1 UC 8 1 LS 7 1 FRS 3
2 LS 9 2 Narrow 9 5 NRS 5
2 FRS 9 2 Rectangular 9 4 LS 7
4 NRS 13 4 FRS 13 6 UC 8
4 Rectangular 13 5 UC 14 7 Square 11
6 Narrow 15 6 NRS 15 3 Rectangular 13
7 Square 16 7 Square 17 2 Narrow 15

This result tells us about a larger efficiency of changing ratio cGAs, which contrast with the findings for
discrete problems. On the other hand, contrary to that observed in the case of discrete problems, the
success rates of asynchronous algorithms are greater than those of synchronous, in general. Contrary to
the results of Section 7.5.8, where either all the algorithms get a 100% hit rate or none finds the solution
in every run for any problem, the FRS cEA is the unique algorithm which is able to find the solution in all
the executions for FRAC problem.
In order to summarize these results, and following the structure of Section 7.5.8, we present in Table 7.15
a ranking with the best algorithms in all the problems by means of the average solution found, the number
of generations needed to find an optimal solution, and the success rate. It can be seen in this table that
there exists a trend of asynchronous algorithms that perform better than synchronous ones in terms of
the average solution found and the success rate, while synchronous changing-ratio algorithms seem to be
more efficient than most asynchronous ones, in general (square and LS are the exceptions).

7.7 Conclusions
In the first part of this chapter we have described several asynchronous update policies for the population
of a cEA, followed by some ratio policies, all of them inducing a different kind of search in the cEA. One
can tune the selection intensity of a cEA by choosing the update policy and grid ratio without having to
deal with additional numerical parameter settings. This is a clear advantage of the algorithms that allows
users to utilize existing knowledge instead of inventing a new class of heuristic.
Our conclusion is that cEAs can be easily induced to promote exploration or exploitation by simply
changing the update policy or the ratio of the population. This opens new research lines to decide efficient
manners of shifting from one given policy/ratio to another in order for the optimum to be reached with
a smaller effort when compared with the basic cEA or other types of EAs.
In a later part of the chapter we have applied our extended cEAs to a set of both discrete and continuous
test problems. Although our goal has not been to compete with state-of-the-art specialized heuristics, the
results clearly show that cEAs are very efficient optimization techniques, that could be further improved
by being hybridized with local search techniques [32]. The results on the test problems largely confirm,
with some small exceptions, that the solving abilities using the various update/ratio modes are directly
linked to their induced selection pressures, showing that exploitation plays an important role. It is clear
that the role of exploration might be more important on even harder problem instances, but this aspect
can be addressed in our algorithms by using more explorative settings, as well as by using different cEA
strategies at different times during the search, dynamically [33].
We conclude that, with respect to discrete problems, asynchronous algorithms are more efficient than
synchronous; with statistically significant values for most problems. On the other hand, if we pay attention
to the success (hit) rate, it can be concluded that the synchronous policies for the different evaluated ratios
outperform the asynchronous algorithms: slightly in terms of the average final fitness, and clearly in terms
of probability of finding a solution (i.e., frequency of optimum location).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 118 — #16

Decentralized Cellular Evolutionary Algorithms 7-119

On the contrary, if we pay attention to the experiments on continuous problems we can get different
(somewhat complementary) conclusions. Asynchronous algorithms outperform synchronous ones in the
cases of average solutions found and hit rate (significant differences in many cases), while in the average
number of generations synchronous algorithms are, in general, more efficient than asynchronous ones.
As future research, we plan to address a single problem of large difficulty and to characterize selection
intensity analytically for all the models.

Acknowledgments
This work has been partially funded by the Ministry of Science and Technology (MCYT) and Regional
Development European Fund (FEDER) under contract TIC2002-04498-C05-02 (the TRACER project)
http://tracer.lcc.uma.es. We thank J. Kaempf for performing a part of the computer simula-
tions for the real-valued problems. M. Giacobini gratefully acknowledges financial support by the Fonds
National Suisse pour la Recherche Scientifique under contract 200021-103732/1.

References
[1] M. Gorges-Schleuter. ASPARAGOS an asynchronous parallel genetic optimisation strategy.
In J.D. Schaffer, Ed., Proceedings of the Third International Conference on Genetic Algorithms,
Morgan Kaufmann, San Francisco, CA, 1989, pp. 422–427.
[2] B. Manderick and P. Spiessens. Fine-grained parallel genetic algorithms. In J.D. Schaffer, Ed.,
Proceedings of the Third International Conference on Genetic Algorithms, Morgan Kaufmann,
San Francisco, CA, 1989, pp. 428–433.
[3] E. Alba and M. Tomassini. Parallelism and evolutionary algorithms. IEEE Transactions on
Evolutionary Computation, 6: 443–462, 2002.
[4] M. Gorges-Schleuter. An analysis of local selection in evolution strategies. In Genetic and
Evolutionary Conference, GECCO99, Vol. 1, Morgan Kaufmann, San Francisco, CA, 1999,
pp. 847–854.
[5] J. Sarma and K.A. De Jong. An analysis of local selection algorithms in a spatially structured
evolutionary algorithm. In T. Bäck, Ed., Proceedings of the Seventh International Conference on
Genetic Algorithms, Morgan Kaufmann, San Francisco, CA, 1997, pp. 181–186.
[6] J. Sarma and K.A. De Jong. An analysis of the effect of the neighborhood size and shape on
local selection algorithms. In H.M. Voigt, W. Ebeling, I. Rechenberg, and H.P. Schwefel, Eds.,
Parallel Problem Solving from Nature (PPSN IV), Vol. 1141 of Lecture Notes in Computer Science,
Springer-Verlag, Heidelberg, 1996, pp. 236–244.
[7] E. Alba and J.M. Troya. Cellular evolutionary algorithms: Evaluating the influence of ratio.
In M. Schoenauer et al., Eds., Parallel Problem Solving from Nature, PPSN VI, Vol. 1917 of Lecture
Notes in Computer Science, Springer-Verlag, Heidelberg, 2000, pp. 29–38.
[8] M. Tomassini. The parallel genetic cellular automata: Application to global function optimization.
In R.F. Albrecht, C.R. Reeves, and N.C. Steele, Eds., Proceedings of the International Conference on
Artificial Neural Networks and Genetic Algorithms, Springer-Verlag, Heidelberg, 1993, pp. 385–391.
[9] D. Whitley. Cellular genetic algorithms. In S. Forrest, Ed., Proceedings of the Fifth International
Conference on Genetic Algorithms, Morgan Kaufmann Publishers, San Mateo, CA, 1993, p. 658.
[10] B. Schönfisch and A. de Roos. Synchronous and asynchronous updating in cellular automata.
BioSystems, 51: 123–143, 1999.
[11] M. Sipper, M. Tomassini, and M.S. Capcarrere. Evolving asynchronous and scalable non-uniform
cellular automata. In Proceedings of International Conference on Artificial Neural Networks and
Genetic Algorithms (ICANNGA97), Springer-Verlag KG, Vienna, 1998, pp. 67–71.
[12] P. Adamidis and V. Petridis. Co-operating populations with different evolution behaviours. In Pro-
ceedings of the Third IEEE Conference on Evolutionary Computation, IEEE Press, Washington, 1996,
pp. 188–191.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 119 — #17

7-120 Handbook of Bioinspired Algorithms and Applications

[13] F. Herrera and M. Lozano. Gradual distributed real-coded genetic algorithms. IEEE-EC, 4: 43–62,
2000.
[14] P. Moscato. Memetic algorithms. In P.M. Pardalos and M.G.C. Resende, Eds., Handbook of Applied
Optimization, Oxford University Press, Oxford, 2000, pp. 157–167.
[15] D.E. Goldberg and K. Deb. A comparative analysis of selection schemes used in genetic algorithms.
In G.J.E. Rawlins, Ed., Foundations of Genetic Algorithms, Morgan Kaufmann, San Francisco, CA,
1991, pp. 69–93.
[16] M. Giacobini, E. Alba, and M. Tomassini. Selection intensity in asynchronous cellular evolutionary
algorithms. In E. Cantú-Paz et al., Eds., Proceedings of the Genetic and Evolutionary Computation
Conference GECCO’03, Springer-Verlag, Berlin, 2003, pp. 955–966.
[17] D. Whitley, S. Rana, J. Dzubera, and K.E. Mathias. Evaluating evolutionary algorithms. Artificial
Intelligence, 85:245–276, 1997.
[18] D.E. Goldberg, K. Deb, and J. Horn. Massively multimodality, deception and genetic algorithms.
In R. Männer and B. Manderick, Eds., Parallel Problem Solving from Nature, PPSN II,
North-Holland, 1992, pp. 37–46.
[19] S. Tsutsui and Y. Fujimoto. Forking genetic algorithm with blocking and shrinking modes.
In S. Forrest, Ed., Proceedings of the fifth International Conference of Genetic Algorithms,
Morgan Kaufmann, San Mateo, CA, 1993, pp. 206–213.
[20] K.A. De Jong, M.A. Potter, and W.M. Spears. Using problem generators to explore the effects of
epistasis. In T. Bäck, Ed., Proceedings of the Seventh ICGA, Morgan Kaufmann, San Francisco, CA,
1997, pp. 338–345.
[21] F.J. MacWilliams and N.J.A. Sloane. The Theory of Error-Correcting Codes. North-Holland,
Amsterdam, 1977.
[22] H. Chen, N.S. Flann, and D.W. Watson. Parallel genetic simulated annealing: A massively parallel
SIMD algorithm. IEEE Transactions on Parallel and Distributed Systems, 9: 126–136, 1998.
[23] S. Khuri, T. Bäck, and J. Heitkötter. An evolutionary approach to combinatorial optimization
problems. In Proceedings of the 22nd ACM Computer Science Conference, ACM Press, Phoenix, AZ,
1994, pp. 66–73.
[24] D.R. Stinson. An Introduction to the Design and Analysis of Algorithms, 2nd ed. (1987). The Charles
Babbage Research Center, Winnipeg, Manitoba, Canada, 1985.
[25] M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness.
Freeman, San Franciso, CA, 1979.
[26] A. Eiben and J. van der Hauw. Solving 3-SAT with adaptive genetic algorithms. In Proceedings of
the Fourth IEEE Conference on Evolutionary Computation, IEEE, Piscataway, NJ, 1997, pp. 81–86.
[27] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs, 3rd ed. Springer-
Verlag, Heidelberg, 1996.
[28] H. Mühlenbein and D. Schlierkamp-Voosen. The science of breeding and its application to the
breeder genetic algorithm (bga). Evolutionary Computation, 1: 335–360, 1993.
[29] D.H. Ackley. A Connectionist Machine for Genetic Hillclimbing. Kluwer, Boston, MA, 1987.
[30] T. Bäck, G. Rudolf, and H.-P. Schwefel. Evolutionary programming and evolution strategies:
Similarities and differences. In D.B. Fogel and W. Atmar, Eds., Proceedings of the Second Conference
on Evolutionary Programming, Evolutionary Programming Society, La Jolla, CA, 1993, pp. 11–22.
[31] T. Bäck. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary
Programming, Genetic Algorithms. Oxford University Press, New York, 1996.
[32] G. Folino, C. Pizzuti, and G. Spezzano. Parallel hybrid method for SAT that couples genetic
algorithms and local search. IEEE Transactions on Evolutionary Computation, 5: 323–334, 2001.
[33] E. Alba and B. Dorronsoro. The exploration/exploitation tradeoff in dynamic cellular evolutionary
algorithms. IEEE Transactions on Evolutionary Computation, 9(2): 126–142, 2005.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C007” — 2005/8/6 — 13:21 — page 120 — #18

 Optimization via
8
Gene Expression
Algorithms

8.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-121

8.2 The Central Dogma of Biology . . . . . . . . . . . . . . . . . . . . . . . . . 8-122
8.3 Borrowing from Nature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-123
8.4 The Gene Expression Algorithm . . . . . . . . . . . . . . . . . . . . . . . 8-123
8.5 Advantages in Using Gene Expression. . . . . . . . . . . . . . . . . . 8-125
8.6 Prior Research in Gene Expression Algorithms. . . . . . . . 8-125
8.7 Gene Expression for the TSP . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-126
Construction Heuristics • Genetic Algorithms and the TSP •
The TSP Gene Expression Algorithm
8.8 A New Gene Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-129
The Inversion Vector • Recovering the Permutation • Using
the Inversion Vector
8.9 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-131
8.10 Final Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-132
Forbes Burkowski References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-133

8.1 Introduction
The extreme diversity of life forms on our planet is a testament to the many ways in which millions of
different species have adapted to the huge number of biological niches that are available ranging, as they
do, over extremes in temperature and pressure in air, land, and water. Viewed in an abstracted formulation,
we may consider evolutionary adaptation of species within a biological niche to be a robust search and
optimization mechanism (Fogel, 1994), the phenotype changing over time so that the species survives
in its competitive environment. Under the Darwinian view, evolutionary processes are involved with
reproduction, mutation, competition, and selection. The bioinspired strategies of genetic algorithms and
evolutionary optimization, in general, strive to capture these biologically robust search mechanisms by
formulating mathematical models that search function domains using algorithmic procedures that mimic
some of the evolutionary processes of the cell, for example, mutation, and crossover. Before discussing
these procedures, let us go over a superficial but quick review of the main biological events in the cell.

8-121

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 121 — #1

8-122 Handbook of Bioinspired Algorithms and Applications

DNA with introns and exons

Transcription:
Primary
mRNA
transcript
RNA
splicing:

Mature mRNA (introns eliminated)

mRNA goes
to a ribosome
for translation:

A protein sequence is produced

FIGURE 8.1 The central dogma.

8.2 The Central Dogma of Biology

During the 1950s Watson and Crick proposed the central dogma of biology. The basic idea was that
information within a cell flowed from DNA to RNA to protein. Protein was not able to influence DNA.
We now know that this is too simple. For example, a virus such as HIV can change the DNA of a host cell
to achieve the replication of the virus.
Nonetheless, as a generalization, the central dogma is a reasonable assumption. Figure 8.1 illustrates the
main events. The interested reader may consult almost any undergraduate text on genetics or biomolecular
chemistry to get the details, for example, Lewin (1997). A single DNA molecule is comprised of a sequence
of nucleotides: Adenine, Guanine, Cytosine, and Thymine typically denoted as A, G, C, and T. The
sequence can be subdivided into intron and exon subsequences. Exons contain nucleotides that specify
the content of protein, while introns (sometimes called“junk”DNA) contain nucleotide subsequences with
a significance that is largely unknown. During the operation of transcription, some particular sequence
of introns and exons making up a gene will be transcribed to form a primary mRNA transcript, which
essentially contains the same code sequence as the DNA except that thymine is replaced with a different
nucleotide called Uracil, that is, U replaces T in the reproduced string. Later a splicing event removes all
intron nucleotides and the mature mRNA is ready for translation by a ribosome, a very large assembly
of molecules that takes the available amino acids in the cell and links them into long chains that make
up a protein. The ordering of the amino acids in the protein sequence is dictated by three letter codons
originally within the exons of the DNA. The now somewhat outmoded “one gene, one protein” hypothesis
states that, a single gene in the DNA is responsible for the specification of one type of protein in the cell.
As seen in this setting, mutation is an event that typically introduces a single nucleotide substitution
in the DNA of a gamete cell that will participate in reproduction (sperm or egg). In crossover, there is an
interchange of material between two different molecules of DNA, each containing the same sequence of
genes but typically with slightly different nucleotide sequences within these genes.
Of course, nature achieves its complexity using strategies that go well beyond simple mutation and
crossover of genes. For example, while codons in the DNA will prescribe the sequence of amino acids in
protein, other mechanisms at the molecular level work on some proteins to modify them. This posttran-
slational modification and alternate splicing strategies that leave out or reorder exons allow the cell to
have many more different proteins than the original “one gene, one protein” hypothesis indicated. Another
example of increased complexity is the huge evolutionary leap that took place when some early primordial
cell became endowed with mitochondria, the energy producing organelles is now present in all cells.
This event is considered to have started as a symbiotic relationship between two cellular structures, one
enveloping the other. In support of this view, we note that mitochondria have their own DNA, which does

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 122 — #2

Optimization via Gene Expression Algorithms 8-123

Modify gene
and generate
phenotype

Is
Yes
phenotype
fit?
No

Replace
phenotype

FIGURE 8.2 The reproductive cycle.

not exactly follow the same encoding as regular DNA. This symbiotic combination illustrates a mechanism
that achieves a remarkable advance in complexity without any reliance on mutation and crossover per se.

8.3 Borrowing from Nature

Ever since the seminal work by John Holland in 1975 (Holland, reprinted 1992), the evolutionary
computation research community has experimented with a wide range of mathematical models for
function optimization. Since the full mechanism of biological evolution is daunting in its formidable
complexity, these mathematical models must necessarily pick some limited feature of the entire process
and this source of “bioinspiration” becomes the central theme of the model. For example, the binary
representation of a gene is formatted to represent a feasible solution of some optimization problem and
a fitness function is applied to the gene to assess how well the feasible solution meets the requirements
of the problem. If the gene has produced a high fitness evaluation then that gene stays in a population
of such genes, otherwise it is replaced. Figure 8.2 provides a summary of the basic idea. Creation of
new genes and selection strategies to chose the surviving genes are basic to the procedures involved with
genetic algorithms. As optimization strategies, genetic algorithms take inspiration from the evolutionary
strategies of the cell and incorporate concepts and mathematical objects that show some relationship with
biological genes, mutation, crossover, etc. We will assume that the reader is familiar with these ideas.
Researchers have borrowed many other features from evolutionary biology. For example, research
developing genetic algorithms inspired by gene duplication (Sawai and Adachi, 1999) and gene signaling
(Goldberg et al., 1989). Although not discussed in Section 8.3, most genes occur in a cell with duplication.
This is known as diploidy and the corresponding mathematical exploration of this within a genetic
algorithm framework has been studied (Bagley, 1967; Smith, 1988). Recalling that introns do not directly
specify amino acid sequences, the presence of noncoding areas in a binary gene for genetic algorithms has
also been considered (Wu and Lindsay, 1995).
The main content of this chapter presents yet another mathematical model that is based on a biological
feature, namely that of gene expression.

8.4 The Gene Expression Algorithm

The many thousands of proteins produced by living cells collectively make up the phenotype of the
individual. The external environment will subject this individual to competitive pressures that effectively
serve to determine whether that individual will reproduce or be eliminated. In general, genes are expressed

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 123 — #3

8-124 Handbook of Bioinspired Algorithms and Applications

Problem
specification

Fitness
Binary gene
value
Evaluate
objective
function

FIGURE 8.3 Fitness evaluation from a binary gene.

Problem
specification

Gene expression algorithm

Binary gene Intermediate Problem Fitness

representation phenotype value
Evaluate
objective
function

FIGURE 8.4 Fitness evaluation from a phenotype.

by being chosen for transcription leading ultimately to some particular protein with possible modification.
The important issue is that, fitness is essentially due to the capabilities of protein making up the phenotype.
More significantly, fitness is not an immediate property of the genes that have been expressed.
In the traditional genetic algorithm (Goldberg, 1989: 60), the phenotype of an individual is the decoded
parameter or parameters of the binary gene. We use Figure 8.3 to describe a somewhat simplified version
of the algorithm, showing how the problem specification will dictate the format of a binary gene. A
subsequent calculation using an objective function evaluates the fitness of that gene. In this scenario, the
fitness function is applied directly to the binary gene. As noted by Mitchell (1996) p. 6:, there is often no
notion of a phenotype in the context of genetic algorithms, although more recently, various works have
presented various models to incorporate them.
When a genetic algorithm involves a gene expression strategy, the fitness function is applied to a
computed phenotype that is more than a simple alternate representation of the binary gene. Typically,
the computation of this phenotype is fairly complicated and is achieved by an algorithm that is called a
gene expression algorithm. As shown in Figure 8.4, the gene expression strategy utilizes an intermediate
stage that produces a final phenotype, which is then subjected to fitness evaluation. To make the biological
inspiration more obvious, we simplify the transformation of cellular information as follows:

DNA gene → mRNA → protein for the phenotype.

In our computational environment, we will have various binary representations that run in parallel with
this natural model:
gene → intermediate representation → phenotype.

The intermediate representation may be absent or may exist as a sequence of intermediate transformations
depending on the requirements of the algorithm. The most significant component of our strategy is
that, we use a gene expression algorithm to produce a binary phenotype representation that is ready for
fitness evaluation. The phenotype meets various feasibility requirements and, in its construction, the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 124 — #4

Optimization via Gene Expression Algorithms 8-125

gene expression algorithm utilizes various parameters that eventually take on values that are determined
through the use of a genetic algorithm. Another description of the gene expression strategy would be the
use of optimization heuristics that are dependent on parameters derived via evolutionary computation.
It is important to understand how an algorithm designer might approach a problem with the intention of
utilizing a gene expression algorithm. In the traditional deployment of a genetic algorithm, the algorithm
designer would have a binary string acting as a gene and then would exercise the genetic algorithm with an
appropriate fitness function applied to such a gene. With a gene expression algorithm, some appropriate
data structure (usually a binary string) acts as a phenotype structure which provides an input for the
fitness function. The “cleverness” of the algorithm designer is then challenged by two goals:
1. Specify the format of a binary gene (suitable for mutation and crossover operations) that can then
be transformed into the phenotype structure.
2. Design the gene expression algorithm that does this transformation.

8.5 Advantages in Using Gene Expression

As will be demonstrated in the sections to follow, using a gene expression algorithm has significant
advantages:
• Separation of Design Concerns: The binary representation of a feasible solution for an optimization
problem may be unsuitable for genetic algorithm operators, such as crossover. Using the gene
expression strategy, the parameters of the optimization heuristics can be encoded in a binary
representation that is completely different from the binary representation of the phenotype to be
used for a fitness evaluation. This facilitates the design of representations that are appropriate for
their utilization: the gene for crossover, the phenotype for fitness evaluation.
• Generation of Feasible Phenotypes: Frequently, the designer of a genetic algorithm must take special
pains to ensure that a crossover operation generates a child gene that is a valid member of the search
space. Since a gene expression algorithm is responsible for the generation of a phenotype, we can
demand that the constraint of feasibility be a part of the gene expression algorithm, the onus of
this responsibility being taken away from the reproductive subsystem that must build a valid gene.
• Using Optimization Heuristics to Extend Genetic Algorithms: Researchers studying optimization and
search strategies for hard NP-complete problems have developed sophisticated “problem specific”
heuristics that, in typical applications, generate very reasonable approximations. Gene expression
algorithms can often take advantage of these techniques thus incorporating solution strategies that
are not typically a part of the traditional approach provided by a simple genetic algorithm.
• Using Genetic Algorithms to Extend Optimization Heuristics: We may gain a mutual benefit by
combining problem specific heuristics with a genetic algorithm approach. Often, the specification
of the heuristic is highly deterministic but will perform better when a stochastic element is provided,
in this case, through the use of an evolutionary computation environment.

8.6 Prior Research in Gene Expression Algorithms

In addition to the work of Julstrom that is mentioned later, we discuss two other research efforts dealing
with gene expression algorithms. Ferreira (2001) introduces a methodology called gene expression pro-
gramming. The main idea is that a binary gene is transformed into an expression tree. The expression
tree (ET) is an infix expression for a program that may be evaluated when actual values are assigned to the
variables located in the leaves of the binary ET. During a training phase, a fitness function compares
the evaluation done by the tree with prescribed target values. The expression algorithm determines
how the ET is constructed from any valid gene made up of variable names and arithmetic operators.
The strategy is applied to problems such as symbolic regression, planning, Boolean concept learning, and
cellular automata rules.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 125 — #5

8-126 Handbook of Bioinspired Algorithms and Applications

Another interesting set of gene expression papers has been written by Hillol Kargupta. He uses the idea of
genetic code-like transformations. Kargupta (1999) emphasizes the importance of learning functions from
data. This would have application in areas such as inductive learning, statistics, data mining, and optimiza-
tion. In this approach, a function is learnt or induced by generating the coefficients in its Fourier expansion.
In general, a function (e.g., a fitness function) defined on an n-bit binary string, requires O(2n ) Fourier
coefficients. Hence, this inductive procedure is computationally inefficient. However, it is sometimes
possible to find very reasonable approximations of a function if the set of Fourier coefficients has a
smaller subset of “large” coefficients that are polynomial in number. In such a case, it must be demon-
strated that the power spectrum, defined by the set of coefficients, has a high concentration of energy in a
few coefficients (O(n k ) in number) with an exponential drop-off in the magnitude of all other nonzero
coefficients. This is equivalent to the demand that, the “small” coefficients (O(2n ) in number) be expo-
nentially small so that, cumulatively, they do not amount to any significant sum. The contribution of this
chapter to gene expression resides in the observation that for some functions with O(2n ) large coefficients,
it is possible to use a “genetic code-like” transformation of the data that will transform the function into
a new function that has only O(n k ) large coefficients. The strategy is considered to involve a gene expres-
sion algorithm because we assume that the given function is defined on a phenotype space and the gene
expression algorithm is used to derive a possibly degenerate transformation that establishes a mapping
between a phenotype string and its gene image, which typically has a length that is a small integer multiple
of the phenotype string length. This is similar to a natural system in which the coding part of the gene
has a string length that is three times the length of the amino acid sequence that it encodes. After such a
transformation of data, the fitness function can be seen to have the gene space as its domain and if the
genetic code-like transformation is defined in the appropriate manner, then the fitness function will have
O(n k ) large Fourier coefficients. These are two other papers (Kargupta and Park, 2000; Kargupta and
Ghosh, 2002) that deal with there related issues.

8.7 Gene Expression for the TSP

In the rest of this chapter we review the work presented by Burkowski (2003) which demonstrates the
application of the gene expression algorithm to the traveling salesperson problem (TSP). We will assume
that the reader is familiar with the basic notion of the TSP. We will also limit our attention to the Euclidean
case of finding a minimum length tour through n cities assuming both, the triangle inequality and the
symmetry of the intercity distances.
It should be stressed that we are not advocating the use of this algorithm as a serious performance
contender in the solution of a TSP. There are currently many highly successful optimization strategies
to derive near-optimal solutions of the TSP for problem instances involving several hundred or even
thousands of cities. When compared with any algorithm, that is a variant of a genetic algorithm, they are
typically much faster and produce better results. Our choice of the TSP as a problem to be solved by a gene
expression algorithm is mainly being done to illustrate the techniques employed by this methodology with
the goal of showing that it can meet the demands of a difficult problem.
Numerous papers have been written in this field of research and a fine bibliography can be found in
Reinelt (1994). Reinelt provides the reader with a useful categorization of TSP heuristics: Construction
heuristics are used to rapidly derive a somewhat coarse approximation to an optimal tour while improve-
ment heuristics are used to modify a given tour in an effort to shorten its length, hopefully deriving an
optimal tour. The improvement heuristics include the well-known 2-Opt, 3-Opt, and Lin–Kernighan
heuristics commonly used in practice.

8.7.1 Construction Heuristics

Among the construction heuristics, we can find various insertion strategies that typically start with a short
subtour going through a small subset of the cities and subsequently employ some technique to enlarge the
tour by iteratively adding more cities until a full tour is generated. As described by Ausiello et al. (1999)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 126 — #6

Optimization via Gene Expression Algorithms 8-127

we can derive a tour T from a set of cities C by doing tour extensions of a tour that initially starts as some
single city in C:

begin

Select any city c in C;

T := <c>;

C := C - {c};

while C != φ do

begin

Let c = f(C, T) in C be the city

meeting the insertion criterion;

Insert c into T;

C := C- {c};

end

end

There is a variety of choice for the city selection function f. For example, we may select a city using one
of the following predefined criteria:

• Nearest Neighbor: Select the city that is nearest to the last city inserted in T.
• Nearest Insertion: Select the city that has the minimal distance from any city already included in the
subtour T.
• Farthest Insertion: Select the city whose distance from the nearest city in T is maximal.
• Cheapest Insertion: Select the city whose insertion in T involves the minimum increase in path
length.

8.7.2 Genetic Algorithms and the TSP

The application of genetic algorithms to the TSP has been met with mixed success by many research
studies. A key issue is the quest for an effective gene representation. A good introduction to these issues
can be found in the work done by Michalewicz (1992). Further research in Tao and Michalewicz (1998)
report on the inver-over operator for fast, high quality solutions to the TSP. In Nagata and Kobayashi
(1997) a technique called edge assembly crossover is used.
An interesting strategy in applying GAs to the TSP has been done by Julstrom (1999). He uses an
insertion heuristic that is guided by selection priorities, which are specified by means of a binary gene.
This is an interesting strategy because the gene is not attempting to directly describe a feasible solution.
Instead, it essentially defines a plan of action that leads to a feasible solution. Of course, in a natural setting
the DNA of an animal is similarly utilized. DNA is used to specify a collection of proteins that function
in the environment of a cell, leading to the characteristics of a phenotype that then exhibits an observable
fitness. Our optimization strategies mimic this flow of information: the facilitation of gene expression
followed by phenotype generation leading to fitness evaluation.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 127 — #7

8-128 Handbook of Bioinspired Algorithms and Applications

8.7.3 The TSP Gene Expression Algorithm

Our construction heuristic is similar in spirit to the insertion strategies mentioned earlier; however, admits
more flexibility in which it involves more choice at each step of the construction. This extra choice at each
step allows us to formulate a strategy that involves a stochastic component. This is done by using a priority
vector that will specify the sequence of choices made in the complete construction of a feasible tour. The
stochastic part of this strategy arises because such a priority vector is derived from a gene in a population
that is subjected to an evolutionary process. We now describe this in more detail using the terminology
introduced earlier:

• The gene representation will be an inversion vector (to be defined later).

• The intermediate representation is a priority vector.
• The phenotype representation is a sequence of cities defining a tour.

We start by describing the conversion of a priority vector to a tour and then describe the inversion vector.

8.7.3.1 A Priority Vector is an Intermediate Representation

We assume that, our TSP instance involves n cities labeled 1, 2, 3, . . . , n. We define an intermediate rep-
resentation p to be a vector p = [a1 , a2 , . . . , an ]. Such a vector is intended to represent some permutation
of the set {1, 2, 3, . . . , n}. In other words, p will be the numbers 1, 2, 3, . . . , n in some “scrambled” order.
As described later, the entries in this vector will represent the priority order in which a set of subtours is
coalesced to form a final full tour.

8.7.3.2 Merging of Subtours to Derive the Phenotype

The phenotype corresponding to this intermediate representation will be a Hamiltonian tour represented
by a sequence of cities. It will have the same appearance as a priority vector but the significance of its
components will be very different. To build the phenotype corresponding to some particular priority
vector, we go through the following n + 1 steps:

Step(0): Build a set of subtours initialized to be n elementary subtours each containing a single city.
We then process the entries of the priority vector in a consecutive order.
Step(i): Find the subtour containing the city labeled ai and merge this subtour with its closest
neighboring subtour.

On completion of the final step, we will have a full n-city tour. To derive the phenotype, we simply read
out the cities in the final tour that is just constructed. To provide more details about step(i) we describe
the merge of subtours as follows: Two subtours are merged by making two cuts (removing an edge from
each subtour), thus opening up both subtours, and then reconnecting them to make one large subtour
(Figure 8.5).
Given an arbitrary subtour SA , the closest neighboring subtour SC is defined to be the subtour providing
the lowest merge cost. The merge cost of any two subtours is the minimal merge cost calculated by

FIGURE 8.5 Merge of two subtours.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 128 — #8

Optimization via Gene Expression Algorithms 8-129

evaluating all possible cut pairs, one cut from each subtour. We define the cost of a merge as follows:

cost of merge = total length of edges added − total length of edges deleted.

Merging two single city subtours will produce a simple two city loop and merging a single city subtour
with a multi-city subtour is essentially the insertion strategy described earlier.
It should be noted that, this strategy does more than simply generating feasible “single loop” tours.
Each tour constructed is typically a reasonable approximation. Experimentation with the TSPLIB test
suite (Reinelt, 1991) shows that an arbitrary priority vector (without any improvements facilitated by the
genetic algorithm described later) will generate a tour that is roughly within 15% to 20% of optimal.

8.7.3.3 Evaluation of Fitness

Once the phenotype (i.e., a full tour) is constructed, we calculate the sum of all intercity distances covered
by the tour and obtain the fitness calculation associated with the given priority vector.

8.8 A New Gene Representation

We now describe the binary gene representation. To the best of our knowledge, this is a novel data
representation for a TSP problem and we expect that future refinements of this strategy should produce
some interesting avenues of exploration.

8.8.1 The Inversion Vector

Let us consider a1 , a2 , . . . , an to be a permutation of the set {1, 2, . . . , n}. The set is deliberately chosen to
be a set of integers because we wish to have the ability to perform order comparisons on the given elements.
As in Knuth (1998)1 we define an inversion vector b1 , b2 , . . . , bn of the permutation a1 , a2 , . . . , an by letting
bj be the number of elements to the left of j that are greater than j. For example, the permutation:

a1 , a2 , . . . , an = 1 7 6 9 5 8 3 4 2

generates the inversion vector:

b1 , b2 , . . . , bn = 0 7 5 5 3 1 0 1 0.

To illustrate why this is true, note that b5 = 3 because there are 3 elements in the permutation, namely 7,
6, and 9 that are greater than 5 and situated to the left of 5.
Because of the manner in which the bj are defined, we will always have the following equations:

0 ≤ b1 ≤ n − 1,
0 ≤ b2 ≤ n − 2,
.. (8.1)
.
0 ≤ bn−1 ≤ 1,
bn = 0.

1 We have adopted the terminology inversion to pay proper respect to Knuth who has a prior use of this term in
the context of a permutation. For Knuth inversion is used in the sense of a two element swap. The reader should not
confuse our use of the term with the notion of inversion as used in genetics.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 129 — #9

8-130 Handbook of Bioinspired Algorithms and Applications

8.8.2 Recovering the Permutation

As noted by Marshall Hall (1956), there is a one-to-one correspondence between the set of all permutations
of {1, 2, . . . , n} and the set of inversion vectors that are constrained to follow the dictates of (1). In fact, if we
are given the inversion vector we can easily recover the original permutation by successively determining
the relative placement of the elements, considering them in the order of largest to smallest. In the given
example, we would write down 9 (the largest number in the city set), then place 8 to the right of 9, since the
inversion vector indicates that there is a number to the left of 8 that is larger than 8. Similarly, we place 7 to
the left of 9 since there is a 0 in the 7 position of the inversion vector, indicating that there are no elements
to the left of 7 that are larger than 7. The 1 in position 6 of the inversion vector will cause the 6 to be written
between the 7 and the 9. The permutation elaborated up to this point will be 7 6 9 8. By continuing in
this fashion, until all the numbers are written out, we can recover the original permutation. The following
table gives the order of element insertion and the results:

Next element to be put

into the permutation Result
9 9
8 98
7 798
6 7698
5 76958
4 769584
3 7695834
2 76958342
1 176958342

8.8.3 Using the Inversion Vector

Why go through the trouble of expressing a permutation as an inversion vector? We do this because
the representation is wonderfully compatible with a crossover operation. Note that any two properly
formed inversion vectors will follow the constraints dictated by Equations (8.1). Furthermore, a crossover
operation between these two inversion vectors will preserve the relative positions of entries in the inversion
vectors produced by the crossover. Hence, a child is immediately a legitimate representation of some
inversion vector since it too will adhere to the constraints imposed by Equation (8.1). By following the
recovery instructions described in Section 8.10, the inversion vector of the child can be converted to a
priority vector and we can commence the generation of a phenotype.
Most importantly, there is no need for any further “postcrossover” modifications of the representation
such as we see with the adjacency representation or the PMX representation (Michalewicz, 1992: 168–172).
It should be noted that the ordinal representation discussed by Michalewicz is also “crossover compatible”;
however, in practice, the representation shows poor experimental results since there is little evidence
of good parents producing good children due to the disruptive effects of crossover with this type of
representation. If the reader indulges in some “back-of-the-envelope” experimentation, you will discover
that the inversion vector representation does exhibit some degree of priority preservation from parent to
child when a crossover is applied. This is fairly easy to appreciate since the inversion vector, by counting
elements larger than a given element, will tend to place a city in a particular comparison relation with the
other cities. This is a quality that is not supported by the ordinal representation.
Since a suitable TSP tour representation has been something of a Holy Grail in the GA community,
we initially went through the trouble of designing an experiment that used the inversion vector as the
representation of the permutation of cities in a tour. Although results were somewhat encouraging, there
was still too much disruption, and experiments demonstrated a disappointing lack of inheritance of high

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 130 — #10

Optimization via Gene Expression Algorithms 8-131

TABLE 8.1 Experimental Results (Population Size = 256)

Number CPU
Problem Opt. Tour Quality of tour time
instance length length of tour evals. sec

att48 10,628 10,653 1.0024 640 9

berlin52 7,542 7,718 1.0233 1,024 12
ch130 6,110 6,336 1.0370 640 49
ch150 6,528 6,910 1.0585 768 89
eil101 629 647 1.0286 896 40
eil51 426 432 1.0140 896 9
kroA100 21,282 22,043 1.0358 640 26
lin105 14,379 14,747 1.0256 1,024 43
tsp225 3,919 4,002 1.0212 1,536 420

TABLE 8.2 Experimental Results

(Population Size = 1024)

Number CPU
Problem Opt. Tour Quality of tour time
instance length length of tour evals. sec

att48 10,628 10,628 1.0 2,560 25

berlin52 7,542 7,542 1.0 3,072 39
ch130 6,110 6,227 1.0191 6,144 483
ch150 6,528 6,711 1.0280 4,096 503
eil101 629 641 1.0191 3,584 169
eil51 426 427 1.0023 3,072 34
kroA100 21,282 21,831 1.0258 4,608 199
lin105 14,379 14,379 1.0 6,144 266
tsp225 3,919 3,952 1.0084 8,704 2,745

quality paths. As a consequence, we have chosen to work with the permutation as representing not the
order of the cities but rather the order of merging subtours as described earlier.

8.9 Experimental Results

All experiments were done using problem instances taken from TSPLIB (Reinelt, 1991). The quality of
the final approximations is presented as the ratio of final tour length divided by the published optimal
tour length. A population size of 256 members was used. Each new generation of 128 genes replaced
the least fit half of the population. A roulette strategy was used for selection of parent genes. If two
successive generations produced no improvement in the best fitness evaluation then the production of a
new generation was stopped, mutation was not used. Results are presented in Table 8.1. See Table 8.2 for
the same experiment run with a population size of 1024. Processing times (600 MHz Pentium III) are in
seconds.
The objective of the experiment was both to test the feasibility of the gene expression model and to
verify the utility of the inversion vector strategy. We contend that our results are very encouraging and
provide empirical evidence that the gene expression approach is a viable strategy.
Figures 8.6 through 8.10 show the progress of the algorithm in the solution of the Berlin52 problem
from TSPLIB (Reinelt, 1991). In all figures except for the last one, the final tour is provided as a lighter
line. In Figure 8.6, we see that the first ten merge operations produce short subtours that are in line with
the final tour. Figures 8.7 to 8.9 show the solution as a growing set of subtours appearing as small loops
that undergo various merging operations. Figure 8.10 is the final tour represented in dark black lines.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 131 — #11

8-132 Handbook of Bioinspired Algorithms and Applications

FIGURE 8.6 After 10 merge operations.

FIGURE 8.7 After 20 merge operations.

FIGURE 8.8 After 30 merge operations.

8.10 Final Discussion

In this chapter, we have described a mathematical model for an optimization strategy that is inspired
by gene expression in a living cell. We have argued that, there are various advantages in using separate
representations for genes and phenotypes with the specific intention that the gene will provide a set of

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 132 — #12

Optimization via Gene Expression Algorithms 8-133

FIGURE 8.9 After 40 merge operations.

FIGURE 8.10 The optimal tour.

parameters that define the construction of a phenotype that will then be subjected to a fitness evaluation.
In particular:

• We have a gene representation that is compatible with crossover and a phenotype representation
with the usual appearance of a Hamiltonian path.
• All phenotypes generated by the gene expression algorithm are valid representations suitable for
fitness evaluation.
• Our experiments show that, utilization of a sophisticated algorithm such as subtour merging gives
the evolutionary computation a “head start” in the construction of good approximations to the
optimal tour.
• Our experiments further demonstrate that a TSP construction heuristic will do better when
combined with a subsystem for evolutionary computation.

We have referred to the heuristic strategy used to construct a phenotype as a gene expression algorithm
since gene expression in the living cell is responsible for utilizing the gene when protein products contrib-
uting to the phenotype need to be synthesized. While there is no guarantee that such a gene expression
algorithm can ever produce an optimal solution for an NP-complete problem our empirical evidence
shows that optimal solutions for a TSP instance can at times be constructed using a subtour merging
heuristic adapted to work with a genetic algorithm. This at least holds the promise of providing a viable
evolutionary computation environment for future avenues of research.

References
Ausiello G., Crescenzi P., G. Gambosi, V. Kann, A. Marchetti-Spaccamela, and M. Protasi (1999) Complexity
and Approximation. Springer-Verlag, Berlin.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 133 — #13

8-134 Handbook of Bioinspired Algorithms and Applications

J.D. Bagley (1967) The behaviour of adaptive systems which employ genetic and correlation algorithms.
Dissertation Abstracts International, 28, 5106B (University Microfilms No. 68-7556).
F.J. Burkowski (2003) Proximity and priority: Applying a gene expression algorithm to the traveling
salesperson problem, paper presented at NIDISC’03 (The Sixth International Workshop on Nature
Inspired Distributed Computing) Nice, France, April 22–26, 2003.
C. Ferreira (2001) Gene expression programming: A new adaptive algorithm for solving problems.
Complex Systems, 13, 87–129.
D.B. Fogel (1994) An introduction to simulated evolutionary optimization. IEEE Transactions on Neural
Networks, 5, 3–14.
D. Goldberg (1989) Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley
Publishing, Reading, MA.
D.E. Goldberg, B. Korb, and K. Deb (1989) Messy genetic algorithms: Motivation, analysis, and first
results. Complex Systems, 3, 493–530.
J.H. Holland (1992) Adaptation in Natural and Artificial Systems. MIT Press, Cambridge, MA.
B. Julstrom (1999) Coding TSP tours as permutations via an insertion heuristic. In SAC ’99, Proceedings
of the 1999 ACM Symposium on Applied Computing. ACM Press, New York, pp. 297–301.
H. Kargupta (1999) A striking property of genetic code-like transformations. School of EECS Technical
report EECS-99-004, Washington State University, Pullman, WA.
H. Kargupta and S. Ghosh (2002) Toward machine learning through genetic code-like transformations.
Genetic Programming and Evolvable Machines, 3, 231–258.
H. Kargupta and B.H. Park (2000) Gene expression and fast construction of distributed evolutionary
representation. Evolutionary Computation, 9, 45–68.
D. Knuth (1998) The Art of Computer Programming, Sorting and Searching, Vol. 3, 2nd ed. Addison-Wesley,
Reading, MA.
B. Lewin (1997) Genes VI. Oxford University Press, New York.
Z. Michalewicz (1992) Genetic Algorithms + Data Structures = Evolution Programs. Springer-Verlag,
Berlin.
M. Mitchell (1996) An Introduction to Genetic Algorithms, MIT Press, Cambridge, MA.
Y. Nagata and S. Kobayashi (1997) Edge assembly crossover: A high-power genetic algorithm for the travel-
ing salesman problem. In Proceedings of the Seventh International Conference on Genetic Algorithms,
T. Bäck (Ed.), pp. 450–457.
G. Reinelt (1991) TSPLIB — A Traveling Salesman Problem Library, ORSA Journal on Computing, 3,
376–384. See also: http://softlib.rice.edu/softlib/tsplib/
G. Reinelt (1994) The Traveling Salesman. Springer-Verlag, Berlin.
H. Sawai and A. Adachi (1999) Genetic algorithms inspired by gene duplication. Congress on Evolutionary
Computation, July. 1999, IEEE Press, Washington, D.C., pp. 480–487.
R.E. Smith (1988) An investigation of diploid genetic algorithms for adaptive search of non-stationary
functions, TCGA report No. 88001, University of Alabama, The Clearinghouse for Genetic
Algorithms, Tuscaloosa.
G. Tao and Z. Michalewicz (1998) Inver-over operator for the TSP. In Proceedings of the 5th Parallel Problem
Solving from Nature, T. Baeck, A. Eiben, M. Schoenauer, and H. Schwefel (Eds.), Lecture Notes in
Computer Science. Springer-Verlag, Amsterdam, pp. 803–812.
A. Wu and R. Lindsay (1995) Empirical studies of the genetic algorithm with non-coding segments.
Journal a Evolutionary Computation, 3, 121–147.
A. Wu and R. Lindsay (1996) A survey of intron research in genetics. Parallel Problem Solving from Nature
(PPSN IV), H. Voigt, W. Ebeling, I. Rechenberg, and H. Schwefel (Eds.), Springer-Verlag, Berlin,
pp. 101–110.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C008” — 2005/7/20 — 19:53 — page 134 — #14

 9
Dynamic Updating
DNA Computing
Algorithms

9.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-135

Motivation • Our New Model
9.2 The New Fundamental Algorithm . . . . . . . . . . . . . . . . . . . . . . 9-138
Coloring Problem • The New Algorithm
9.3 Dynamically Updating the Answers . . . . . . . . . . . . . . . . . . . . 9-140
Zhiquan Frank Qiu 9.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-144
Mi Lu References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-144

9.1 Introduction
It has been clearly shown that DNA computing can be used to solve those problems that are currently
intractable on even the fastest electronic computers. For methods to design the algorithms for DNA
computing, however, is not straightforward. To develop efficient DNA computing algorithms requires a
strong background in both DNA molecule and computer engineering. All of these algorithms need to start
over from the very beginning when their initial conditions change. It is very frustrating, especially when
the initial condition change is very small. The existing models based on which a few DNA computing
algorithms were developed are not able to accomplish the dynamic updating.
People have been talking about the huge memory made possible through DNA computing due to
the fact that each strand can be treated as both storage media and processor for a long time. Currently,
there is no existing application that has used this huge memory because although it is easy to read
from this memory, it is extremely hard to store data in it. Memories can only be ready after data has been
stored.
A new DNA computing model is introduced based on which new algorithms are developed to solve
the 3-coloring problem. These algorithms are presented as vehicles to demonstrate the advantages of the
new model. They can be expanded to solve other NP-complete problems. They have the advantage of
dynamic updating, so answers can be changed when the initial conditions are modified. The new model
takes advantage of this huge memory by generating “lookup tables” during the process of implementing
the algorithms. When the initial conditions change, the answers are changed accordingly. The new model
can be used to solve computationally intense problems both efficiently and attractively.

9-135

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 135 — #1

9-136 Handbook of Bioinspired Algorithms and Applications

9.1.1 Motivation
A strand of DNA is composed of four different base nucleotides: A (adenine), C (cytosine), G (guanine),
and T (thymine). When attached to deoxyribose, these base nucleotides can be strung together to form a
strand. Because DNA strand can be used to encode information and DNA bio-operations are completely
based on the interactions between strands, each DNA strand can be counted as a processor as well as
storage media. Numerous strands are involved in DNA bio-operations and the interactions between one
another occur simultaneously. This, then, can be viewed as a realization of massive parallel processing.
Since Adleman [1] solved a 7-vertex instance of the Hamiltonian Path Problem, a well-known repre-
sentative of NP-complete problems, the major goal of subsequent research in the area of DNA computing
has been to develop new techniques for solving NP-complete problems. NP-complete problems are those
problems for which no polynomial-time algorithm has yet been discovered, in contrast to polynomial-time
algorithms whose worst-case run time is O(n k ) for some constant k, where n is the size of the problem.
Consider that 1 litre of water can hold 1022 DNA strands. The potential computing power is significant,
and this recognition raises the hope of solving problems currently intractable on electronic computers.
Rather than using electronic computers upon which the time needed to solve NP-complete problems
grows exponentially with the size of the problem, DNA computing technology can be used to solve
these problems within a time proportional to the problem size. An NP-complete problem that may take
thousands of years for current electronic computers to solve would take a few months, if the existing DNA
computing techniques were adopted.
As indicated in several articles [2–7], most DNA computing algorithms are based on certain developed
DNA computing models. The most popular models are the sticker based model [8,9], the surface-based
model [10,11], and the self-assembly based model [12,13]. The problem with the sticker-based model is
that the stickers annealed to the long strand may fall off during bio-operations, thus causing a very high rate
of error. The limitation of the surface based model is that the scale of computation is severely restricted by
the two-dimensional nature of surface based computations. The shortcoming of the self-assembly based
model is that it makes use of biological operations that are not yet matured.
While the theory of molecular computation has developed rapidly, most of these algorithms usually
take months to solve problems that may take thousands of years to solve with electronic computers. The
problem is that when the initial condition changes, the algorithms have to start over again. Here, a new
DNA computing model, which can eliminate this problem, is introduced. Based on this model, algorithms
can be designed to dynamically update the answer. When the initial condition changes, the new algorithms
can continue with the current process, and the solution for the new problem can be generated by a few extra
processes. In addition, this new model can also be used to solve several similar problems simultaneously.

9.1.2 Our New Model

Our new model uses only the DNA biological operations that are matured [1,6]. The following are the
basic principle operations: synthesis, ligation, separation, combination, and detection that are picked for
the new model.
Synthesis I(P, π): this operation is used to generate a pool of coded strands, P, following the predefined
criteria π . Different applications code strands differently use the four base nucleotides: A, G, T, and C.
A group of strands are defined as one set and a pool is defined as the container holding a set of strands.
If the criteria are the colors of a node in a graph, then a pool of strands coding all the possible colors
for the nodes is expected after synthesis. In the graph coloring problem, the colors of a number of
nodes are encoded by the strands. A few consecutive nucleotides on the strand coded for the color of
one node form a region. For example, in Figure 9.1, one strand consists of three regions such that
s = {RBR} where (CCAAG), (AATTC), and (CCAAG) each represents the color for one node as R(Red),
B(Blue), and R(Red), respectively. Ligation L(P3 , P1 , P2 ): this operation is used to bind strands in P1 with
strands in P2 . Each code s1i in P1 , is ligated to every other code s2j in P2 . Assuming before the ligation,
strands in P1 represent the codes {s1i |i = 1, 2, . . . , c, where s1i ∈ P1 } and those in P2 represent the code

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 136 — #2

Dynamic Updating DNA Computing Algorithms 9-137

Node #2 colored as Blue

G A A T C A A T T C

A T G G G C C A A G G G C C T C C A A G

Node #1 colored as Red Node #3 colored as Red

FIGURE 9.1

Temperature control fluid

One pool

Gel running buffer

The other pool

Capture layers

FIGURE 9.2

{s2j |j = 1, 2, . . . , c, where s2j ∈ P2 }, after the ligation, the ligated strands are stored in P3 and they represent
the code {s1i |i = 1, 2, . . . , c, where s1i ∈ P1 } where sk = s1i s2j , for k = i + ( j − 1) × c.
In separation operation, many identical short strands defined as probes are attached to magnetic beads.
These probes are then sent into the pool containing the strands to be separated. Every probe will be
paired up with a complementary strand and together they form a double helix. Such pair-ups occur only
under the Watson–Crick complement rule: C only pairs with G and T only pairs with A. For example,
in Figure 9.1, if the strands containing the region for node 1 colored as ‘R’ need to be separated, the
DNA short strands TACCCGGTTC should be used as a probe because TACCCGGTTC complements
ATGGGCCAAG. Also, the double helix can be separated by heating in order to have the paired strands
part from each other without breaking the chemical bonds that hold the nucleotides together inside the
single strand. The strands in the pool containing a region that complements the probes will be hybridized
to, and captured by, the probe while all those without the region will remain in the pool [8].
A gel-based separation technique for DNA computing [14] has been developed that uses gel-layer
probes instead of the bead to capture the strands. The capture layer only retains the strand with a region
that complements the probe when it is cooled down, and will let all strands pass when the layer is heated
up. The advantage of using gel-based probes over bead-based probes is that the gel-based method is more
accurate when capturing DNA molecules. In Figure 9.2, which illustrates the gel-based separation, a set
of strands runs from the left side buffer to the right. At each capture layer, the temperature is cold in
order to capture the desired strands, and all unwanted strands are passed through into one pool. Then the
temperature is raised to let all desired strands in the layer pass into another pool. The strands from the
left buffer are separated and stored in two different pools:
Combination B(P, P1 , P2 ): to pour two pools, P1 and P2 , together to form a new pool, P.
Detection D(P): to check if there is any strand left in the pool, P. If the answer is “yes,” the strands in
the pool may need to be decoded.
The rest of this chapter is organized as follows: Section 9.2 gives an introduction to our new algorithm
and how, based on the new model we purposed, it solves the 3-coloring problem. The complexity analysis
of the new algorithm is provided in Section 9.3, which shows how dynamic updating can be accomplished.
The last section makes the conclusion.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 137 — #3

9-138 Handbook of Bioinspired Algorithms and Applications

9.2 The New Fundamental Algorithm

This new algorithm for the 3-coloring problem has been developed based on the newly developed DNA
computing model. The basic algorithm that will generate the answer to the 3-coloring problem of a given
graph will be introduced in this section. The algorithm will be advanced to show how the answer can be
dynamically updated and this will be shown in Section 9.2.1.

9.2.1 Coloring Problem

The 3-coloring problem is a special case of the k-coloring problem where k = 3, is a well-known
representative of the class of NP-complete problems. A new algorithm for solving the 3-coloring problem
will be introduced, and will simplify the explanation of our new DNA computing model. These when
developed can be expanded to solve the k-coloring problem and hereby can be generalized to solve many
other NP-complete problems.
k-coloring problem: k-coloring problems require the coloring of an undirected graph G = (V , E) in
such a way that no two adjacent vertices share the same color [15]. The two nodes connected by an edge
are referred to as adjacent vertices. The solution is a function c: V → 1, 2, . . . , k such that c(u)  = c(v)
for every edge (u, v) ∈ E. In other words, the numbers 1, 2, . . . , k represent the k colors, and the adjacent
vertices must have different colors. The k-coloring problem determines whether k colors are enough to
color a given graph [16].
A simple example graph with 10 nodes and 10 edges, G(10, 10), is given in Figure 9.3. From the graph,
it is clearly shown that the graph can be colored if k ≥ 3.
In order to solve this 3-coloring problem, it is necessary to generate a pool of encoded DNA strands to
represent all the possible color patterns of the n-node graph where each color pattern is an assignment of
colors of nodes [17,18]. For example, for nodes n1 n2 n3 n4 , “BBRG” is one pattern, which assigns n1 with
color Blue, n2 with color Blue, n3 with color Red and n4 with color Green. “RGBB” is another pattern
with given color Red, Green, Blue, and Blue to n1 n2 n3 n4 , respectively. After the strands are generated
and stored in one pool, the color patterns with no color conflict need to be separated. Two nodes
along an edge are defined as having color conflicts when they have the same color. For the color patters
with any color conflicts along some edges of the graph, the corresponding strands must be filtered from
the pool.
The new algorithm will be introduced in Section 9.2.2. The dynamically updating algorithm and the
advantages of the new algorithms will be described in Section 9.2.3.

R
G

G B
G

R
R

FIGURE 9.3

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 138 — #4

Dynamic Updating DNA Computing Algorithms 9-139

G Recursive dividing step #

G1 G2

G3 G4 G5 G6 1

G2b–1

G2b+1 b
G2b+1 G2b+2

FIGURE 9.4

9.2.2 The New Algorithm

Given a graph G = (V , E), V = {vi | i = 1, 2, . . . , n} is a set of nodes and E = {ej | j = 1, 2, . . . , m} is
the corresponding set of edges. The new approach to solving the 3-coloring problem for such graph is
divide and merge. Partition graph G into two subgraphs: G1 = (V1 , E1 ) and G2 = (V2 , E2 ) such that
V1 ∪ V2 = V , V1 ∩ V2 = φ and |V1 | ≈ |V2 | by eliminating all edges (u, v) such that u ∈ V1 and v ∈ V2 .
This set of edges, called C, is referred to as the cut set of G [15,19]. The partition process can be executed
recursively. That means, subgraphs Gi can be partitioned into G2i+1 and G2i+2 , until each subgraph
contains only one vertex, and n subgraphs exist in total as shown in Figure 9.4.
After partitioning the graph G into n subgraphs, the algorithms start to merge every two subgraphs
recursively and in parallel. Before each merge, every subgraph is colored with three colors. During the
merge process, the color patterns of the two subgraphs are combined together. The merge operation
continues until the original graph G is reestablished. To merge two subgraphs, the edges in the cut sets
earlier eliminated for partitioning the subgraphs will be added back. Each addition of such edges will
introduce some color conflicts if the nodes linked by the edge are of the same color. The color patterns
that worked for the subgraphs may not work for the merged graphs after the subgraphs are combined.
Some combined color patterns will be eliminated. This elimination process will continue until the color
patterns that are legitimate for the graph are found.
The algorithm to solve the 3-coloring on a sparse graph is shown in Figure 9.5. The first for loop is used
to generate n pools of strands representing all possible color patterns for n subgraphs while initially each
subgraph contains only one node.
The function of the while loop is first to merge the pairs of the two subgraphs. The bio-operation used
to merge two subgraphs is ligation. This step ligates the strands in two pools in order to form longer
strands. If the color pattern for the first subgraph is si and that for the second graph is sj , all the si should
be ligated to sj . The strands for one color pattern of a subgraph is replicated and each duplicated copy is
ligated to those strands representing the color patterns of the other subgraph. All the color patterns of the
merged graph will be represented by the ligated longer strands after the merge operation.
Inside the while loop, multiple copies of all strands in every pool need to be prepared for the following
round of ligation. The duplication can be accomplished by using the polymerase chain reaction (PCR)
process [17,20].
After the merge, some ligated strands encoding the color patterns with color conflicts introduced by
those edges in all cut sets are eliminated in the partition step. The next task is to investigate all edges in the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 139 — #5

9-140 Handbook of Bioinspired Algorithms and Applications

Algorithm 1.
For i = 1 to n do
In parallel (I(Pi, color of node i))
End
F=n
While f ≠ 1 do
In parallel (Make multiple copies of strands in all pools)
For all odd I do
In parallel ( L (Pi, Pi, Pi+1 ))
In parallel (relabel all pools 1 to f/2)
End
f = f/2
End
S (P, Pt, Pf1, θ), θi is color conflicts along ei
K = 1;
For i = 2 to m do
S (Pt, P1t, P1f, θ), θi is color conflicts along ei
For j = 1 to k do
In parallel { S (Pfj, Pjt, Pjf, θi) }, θi is color conflicts along ei
End
For j = 1 to k do
In parallel do ( (B(Pfj, Pj-1f, Pjt))
End
B (Pt, P1t, θ)
B (Pfk+1, Pkf, θ)
K=k+1
End
Check if P is empty to return ‘‘yes’’ or ‘‘no’’ accordingly.

FIGURE 9.5

cut sets and detect all the color conflicts caused. The task can be accomplished by the separation operation
to filter out all strands that contain any color conflict from the pool. Two nodes, i and j, are connected by
an edge. The pool is first separated into three pools while each pool contains the strands coloring node i
as R, G, and B. The strands having node j colored as R, G, and B are filtered out by using the separation
operation.
The answer to the 3-coloring problem is “yes” if there is any strand left in the final pool. The answer is
“no” and the graph cannot be colored by only three colors if there is no strand left.

9.3 Dynamically Updating the Answers

Once a solution to the 3-coloring problem of the graph is obtained, given minor changes in the initial
condition, it is significant to have a method that can quickly update the solution without restarting the
algorithm and completely recalculating. What is presented next is an effort to make such dynamically
updating solution both realistic and efficient.
For 3-coloring problem, four possible changes may occur with the initial condition: nodes and edges are
inserted or removed. Based on the original solution “yes” or “no” to the original graph, different strategies
need to be considered to update the answer.
Beginning with the easiest updating strategies, if the original answer is “yes” and an edge or node is
removed from the original graph, the answer will remain “yes.”
If the original answer is “no,” it will remain “no” if a node or edge is added.
If the original answer is“yes,” it can be changed to“no”after a node is inserted into the graph. An example
is shown in Figure 9.6. The graph shown in Figure 9.6(a) has the answer “yes” for the 3-coloring problem

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 140 — #6

Dynamic Updating DNA Computing Algorithms 9-141

(a) (b)

FIGURE 9.6

(a) (b)

FIGURE 9.7

of the given graph. The answer can be changed to “yes” after one node is inserted to the graph as shown
in Figure 9.6(b).
If the original answer is“no,” it can be changed to“yes”after a node is removed from the edge. An example
is shown in Figure 9.7. Figure 9.7(a) contains the graph with the answer “no” for the 3-coloring problem.
The answer is changed to “yes” after one node is removed from the graph as shown in Figure 9.7(b).
If an edge is removed or inserted to the graph, it can be dealt with similarly because at least one edge
should be removed if a node is eliminated and at least one edge should be added when a node is inserted.
What is illustrated next shows how to dynamically update a solution when a node or edge is inserted
into the graph, following an original answer “yes.” The strands in the final set, Pt , are checked for possible
new answers. The final set is the only set that can be used because it is the only set that contains the
strands representing all possible coloring solution that do not have any color conflicts among all the nodes
except the newly added one. Only the color conflicts that occur between the newly added node and nodes
connected with it need to be checked based on these sets. Only the newly added edges are checked for
color conflicts.
The most difficult case occurs when a node or edge is removed from a graph with an original answer of
“no.” The answer to the new graph may be either “yes” or “no.” To remove a node includes removing both
the node itself and all the edges connecting the new node to the graph. What follows is the dynamically
updating algorithm for this case. The DNA computing result that reflects an original answer of “no” is

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 141 — #7

9-142 Handbook of Bioinspired Algorithms and Applications

Algorithm 2.
For i = 1 to α do
In parallel (S (Pfi, Pnewi, Pfi, θi), θi is color conflicts along exact i # of edges)
End
B (Pnew, φ, φ)
For i = 1 to α do
In parallel ( B (Pnew, Pnew, Pnewi))
End
B (Pnew, Pt, Pnew)
For j = 1 to β do
S (Pnew, Pnew, Pnewf, wi), wj is the colored conflicts based on edge ej
End
Check if Pnew is empty to return ‘‘yes’’ or ‘‘no’’ accordingly.

FIGURE 9.8

represented by an empty Pt set with no strand. All other sets represent the coloring patterns of the original
graph with the color conflicts. After removing the nodes or edges, some coloring patterns may no longer
have conflicts. The task now is to identify those patterns represented by DNA strands. Here, the strand
sets that need to be examined is limited. Only those strands representing color patterns with color conflict
involving the pair of nodes connected by the edges being removed are checked. Finding the above strand
sets takes O(α) steps, where α is the number of edges being removed. This process is much less expensive
than recomputing the updated graph from the very beginning when α is not large.
The detailed algorithm needed to find the answer for the new graph with the removed edges, based on
the original “no” answer, is illustrated in Figure 9.8.
When only one edge is removed from the original graph, pool Pf 1 is checked. This is because Pf 1
contains all the strands representing all the color combinations for the graph that have no color conflicts
along all edges, except one.
Assuming that the two nodes along the edge being removed are n1 and n2 , the strands that need to be
separated from the pool are those that have the two nodes colored as {RR}, {BB}, and {GG}. That means
that only those strands, which have two identically colored nodes are extracted to a new pool, Pnew . If Pnew
is not empty, the answer to the 3-coloring problem for the new graph is “yes,” which is different from the
original graph. Otherwise, the “no” answer remains.
When two edges are removed from the graph, both Pf 1 and Pf 2 need to be checked. This is because Pf 2
may contain strands that represent color combinations that have color conflicts along both the edges being
removed. Pf 1 may contain strands that represent color combinations of the graph with a color conflict
along only one of the two edges being removed. Suppose the two edges being removed are e1 and e2 .
Then, strands that need to be extracted from pool Pf 2 using the separation operation must represent the
color combinations of the graph having color conflicts along both edges. Strands that should be extracted
from Pf 1 are those representing color combinations with color conflict along either e1 or e2 . The extracted
strands are stored in a new pool, Pnew . If Pnew is not empty, the answer to the 3-coloring problem for the
new graph is “yes,” which is different from the original graph. Otherwise, the answer for the 3-coloring
problem to the new graph remains “no.”
When α different edges are removed from the original graph, α different pools should be checked.
These pools are Pf 1 , Pf 2 , . . . , Pf α . For different pools, different operations need to be undertaken. For
pool Pf 1 , all strands are left due to the color conflict along one edge. If the edge that caused the conflict
is removed, the answer will change to “yes.” Because of this, all strands in this pool representing those
color combinations with color conflicts along one of the α edges that have been removed should represent
the answer to the 3-coloring problem of the new graph. For pool Pf 2 , all strands representing the color
combinations have color conflicts along two, and only two, of the edges being removed, representing the
answers to the 3-coloring problem of the new graph. For pool Pft where t ≤ α, all strands representing
the color combinations having color conflicts along exactly t different edges being removed will generate
the answer to the 3-coloring problem for the new graph. All strands extracted from these sets are stored

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 142 — #8

Dynamic Updating DNA Computing Algorithms 9-143

in a new pool, Pnew . If Pnew is not empty, the answer to the 3-coloring problem for the new graph is “yes,”
and thus different from the original graph. The answer is “no” if Pnew is empty.
When the graph is changed by both removing and adding edges, multiple processing steps need to be
considered. Assuming that the number of edges being removed is α and the number of edges being added
is β, the strands with color conflicts along the removed edges should be found first. This will put the
strands to be considered for the following operations in one pool, Pnew , instead of involving several pools.
Those α edges should first be considered by using the method introduced above to go through α different
pools. Then, Pt is combined with Pnew and relabeled Pnew . This is due to the fact that those strands that
may generate the “yes” answer are distributed in α + 1 different pools. Collecting the strands in one pool
will save time and further operations as compared to working on these pools one at a time. If no strands
are left in pool Pnew , then the answer to the new graph is “no.” If there are some strands in pool Pnew after
α edges are removed, color conflicts along β edges are checked. This operation can be accomplished in a
manner similar to what has been described above for adding edges.
Compared to the existing algorithms, our new method can dynamically update the solution when the
initial condition changes for the 3-coloring problem of a graph. It can also solve the 3-coloring problem
for many graphs that are similar to each other. The complexity of the existing algorithm is O(m + n),
where n is the number of vertices and m is the number of edges [18]. If the updating process is not used,
any change to the initial condition will result in a restarting of the process. With our new algorithm,
the number of extra processes that need to be undertaken depends upon the significance of the changes.
The complexity of the updating process is O(α + β), where α is the number of edges being removed. β is
the number of edges being added.
When this method is used to solve the 3-coloring problem for multiple graphs that are similar to each
other, the time complexity is O(θ) after the solution for one graph is generated, where θ is the difference
between the number of edges of the two graphs.
It is necessary to check the extra space and effort that may be necessary for making dynamic updating
available. First, m additional containers are needed to keep all m extra sets of strands. Second, the extra
DNA material for generating these sets needs to be contained. Because strands are generated to represent all
color combinations for the graph before the separation process takes place, no extra material is necessary
as compared with the existing algorithms until the answer is generated for the original graph. The extra
material is only necessary if new solution needs to be formed for the modified graph when the edges
and/or nodes are added.
When the procedure for approaching a 3-coloring problem of a given graph is finished and a new graph
is provided, how can one determine whether to start again from the beginning or to use the dynamic
updating method to generate the new answer?
Assuming that the implementation of the algorithms introduced above for the 3-coloring problem of
the graph with n nodes and m edges has been finished, the 3-coloring problem of a new graph needs
to be solved. This new graph has N nodes and M edges. This graph can be converted from the existing
graph by first removing δ nodes and α edges, and then adding γ nodes and β edges. The new graph can
be generated by changing the original graph, or it can be treated as a totally new graph. In order to solve
the problem for the new graph, N ligation and M separation operations are necessary if the algorithm is
being restarted from the beginning. The total time necessary is:

T1 = N × l + M × s,

where l is the time for each ligation operation and s is the time necessary for each separation operation.
Here, combinations are ignored due to their simplicity because the time needed for the combination
operations is very short, as compared to the other operations used in DNA computing. When the answer
is generated based on the pools already generated using this new, dynamically updating strategy, the time
necessary for reaching the answer is

T2 = (α + β) × s + γ × l.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 143 — #9

9-144 Handbook of Bioinspired Algorithms and Applications

In order to take advantage of the new method, the time that is needed must be shorter than restarting
the algorithm from the beginning.

T 2 ≤ T1 ,
(α + β) × s + γ × l ≤ N × l + M × s,
(α + β) × s + γ × l ≤ (n + γ − δ) × l + (m + β − α) × s,

because N = n + γ − δ and M = m + β − α. It is easy to get

(m − α) × s + (n − δ) × l ≥ α × s,

as n − δ is always >0, the above condition can be tightly restrained as follows:

(m − α) × s > α × s.

So, α < m/2. The algorithm needs to be restarted from the beginning only when the change is significant,
which means, when more than half of the edges need to be removed to generate the new graph from the
original.
Given the above condition, it is clear that there is no need to retain all m sets. At least half of the pools
can be destroyed in order to save storage space. This will save the expenses once required for storing m sets
of strands and the material needed to work on them.

9.4 Conclusion
A new model for DNA computing is introduced. Based on the new model, our new algorithms for the
3-coloring problem have been presented. The new algorithms have the advantage of dynamic updating,
as compared to the existing algorithms. These new algorithms represent a huge improvement over the
existing algorithms.
Instead of restarting the DNA computing algorithm from the very beginning every time the initial
condition would change, this new method can generate the new solution through a few extra DNA
operations based on the existing answer. It can also quickly solve problems similar to those already solved.
No extra material is needed to prepare for the dynamically updating process. The only expense is some
extra storage containers for storing the additional pools of DNA strands. As compared to the existing
DNA computing algorithms, this new method can achieve a solution much more quickly after the answer
for the first problem is generated and it is very financially efficient. This will make DNA computing more
attractive to potential users who want to solve the problem that is currently unsolvable.

References
[1] L. Adleman. Molecular computation of solutions to combinatorial problems. Science, 1021–1024,
1994.
[2] Y. Gao, M. Garzon, R.C. Murphy, J.A. Rose, R. Deaton, D.R. Franceschetti, and S.E. Stevens Jr.
DNA implementation of nondeterminism. In Proceedings of the Third DMIACS Workshop on DNA
Based Computers, University of Pennsylvania, June 1997, pp. 204–211.
[3] G. Gloor, L. Kari, M. Gaasenbeek, and S. Yu. Towards a DNA solution to the shortest common
superstring problem. In Proceedings of the Fourth International Meeting on DNA Based Computers,
University of Pennsylvania, June 1998, pp. 111–116.
[4] V. Gupta, S. Parthasarathy, and M.J. Zaki. Arithmetic and logic operation with DNA. In Proceedings
of the Third DIMACS Workshop on DNA Based Computers, University of Pennsylvania, June 1997,
pp. 212–222.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 144 — #10

Dynamic Updating DNA Computing Algorithms 9-145

[5] P. Kaplan, D. Thaler, and A. Libchaber. Parallel overlap assembly of paths through a directed
graph. In Proceedings of the Third DIMACS Workshop on DNA Based Computers, University of
Pennsylvania, June 1997, pp. 127–141.
[6] R. Lipton. Using DNA to Solve SAT, 1995.
[7] Z.F. Qiu and M. Lu. Arithmetic and logic operations for DNA computers. In Parallel and Distributed
Computing and Networks (PDCN’98), IASTED, December 1998, pp. 481–486.
[8] S. Roweis, E. Winfree, R. Burgoyne, N. Chelyapov, M. Goodman, P.R. Othemund, and L. Adleman.
A sticker based architecture for DNA computation. In Proceedings of the Second Annual Meeting
on DNA Based Computers, Princeton University, June 1996, pp. 1–27.
[9] S. Roweis, E. Winfree, R. Burgoyne, N. Chelyapov, M. Goodman., P.R. Othemund, and L. Adleman.
A sticker based model for DNA computation. Journal of Computational Biology, 5: 615–629, 1998.
[10] Q. Liu, Z. Guo, A.E. Condon, R.M. Corn, M.G. Legally, and L.M. Smith. A surface-based approach
to DNA computation. In Proceedings of the Second Annual Meeting on DNA Based Computers,
Princeton University, June 1996, pp. 206–216.
[11] L. Wang, Q. Liu, A. Frutos, S. Gillmor, A. Thiel, T. Strother, A. Condon, R. Corn, M. Lagally, and
L. Smith. Surface-based DNA computing operations: Destroy and readout. In Proceedings of the
Fourth International Meeting on DNA Based Computers, University of Pennsylvania, June 1998,
pp. 247–248.
[12] E. Winfree. Proposed techniques. In Proceedings of the Fourth International Meeting on DNA Based
Computers, University of Pennsylvania, June 1998, pp. 175–188.
[13] E. Winfree, X. Yang, and N.C. Seeman. Universal computation via self-assembly of DNA: Some
theory and experiments. In Proceedings of the Second Annual Meeting on DNA Based Computers,
Princeton University, June 1996, pp. 172–190.
[14] R.S. Braich, C. Johnson, P.W.K. Rothemund, D. Hwang, N. Chelyapov, and L.M. Adleman. Solution
of a satisfiablility problem on a gel-based DNA computer. In Proceedings of the Sixth International
Meeting on DNA Based Computers, June 2000, pp 31–42.
[15] J. Clark and D.A. Holton. A First Look at Graph Theory. World Scientific, Singapore, 1991.
[16] T.H. Cormen, C.E. Leisenson, and R.L. Rivest. Introduction to Algorithms. MIT Press, Cambridge,
MA, 1990.
[17] L. Adleman. On constructing a molecular computer, 1995.
[18] E. Bach and A. Condon. DNA models and algorithms for NP-complete problem. Journal of
Computer and System Sciences, 57: 172–186, 1996.
[19] N. Christofides. Graph Theory: An Algorithmic Approach. Academic Press, New York, 1975.
[20] P.D. Kaplan, G. Cecchi, and A. Libchaber. DNA-based molecular computation: template-template
interactions in pcr. In Proceedings of the Second Annual Meeting on DNA Based Computers,
Princeton University, June 1996, pp. 159–171.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C009” — 2005/8/6 — 13:22 — page 145 — #11

 10
A Unified View on
Metaheuristics and
Their Hybridization

10.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-147

10.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-148
10.3 A Unified Framework for Iterative Search
Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-148
10.4 Some Thoughts about Memory. . . . . . . . . . . . . . . . . . . . . . . . .10-150
10.5 Combining Evolutionary Algorithms and Ant
Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-151
Basic Evolutionary Algorithm • Basic Ant Colony
Optimization • Hybrid Algorithms
Jürgen Branke 10.6 Empirical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-153
Michael Stein 10.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-155
Hartmut Schmeck References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-155

10.1 Introduction
Over the past decades, a multitude of new search heuristics, often called “metaheuristics” have been
proposed, many of them inspired by principles observed in nature. Common representatives include
evolutionary algorithms (GAs) [1], ant colony optimization (ACO) [2], simulated annealing [3], tabu
search [4], or estimation of distribution algorithms [5]. Besides the book at hand, overviews of several
such metaheuristics can be found, for example, in References 6 and 7 or 8.
Each of these metaheuristics has been proven successful on a variety of applications. Although there
have been attempts to compare their performance, the results are contradicting and inconclusive. There
does not seem to be a superior candidate that should generally be preferred over the others. Thus, it is
not surprising that recently, there has been a growing interest in hybridization of these metaheuristics
(cf. Section 10.2).
In this chapter, we propose a simple unified framework that describes the fundamental principle
common to all metaheuristics. The framework focuses on the commonalities rather than the differences
between search algorithms. Due to its simplicity and generality, it suggests a natural way for hybridization,
basically turning the variety of metaheuristics into one large toolbox from which an algorithm designer
can choose those parts that seem most appropriate for the application at hand. The power of the model

10-147

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 147 — #1

10-148 Handbook of Bioinspired Algorithms and Applications

to unify different metaheuristics will be demonstrated at the example of combining EAs and ACO, and we
will report on some preliminary empirical results on the performance of the hybrids so generated.
The chapter is structured as follows: in Section 10.2, we will survey some related work. Then, in
Section 10.3, we will describe the proposed unified framework. A specific aspect of that framework,
the organization of memory, is discussed in Section 10.4. Section 10.5 demonstrates the application
of the model to the hybridization of EAs and ACO. The resulting hybrids are compared empirically in
Section 10.6. The chapter concludes with a summary and some suggestions for future work.

10.2 Related Work

There have been numerous attempts to combine aspects of different metaheuristics, usually in the hope of
keeping the benefits and avoiding the pitfalls of the pure heuristics. Examples include, but are not limited
to, combinations of EAs and simulated annealing [9,10], EAs and tabu search [11], EAs and ACO [12],
ACO and tabu search [13], or EAs and particle swarm optimization [14]. Also, there is growing interest in
the field of memetic algorithms [15], which focuses on the combination of EAs and local search.
Calegari et al. [16] developed a taxonomy to describe iterative optimization heuristics, but as the goal of
a taxonomy is to differentiate rather than to unify, it generally hides the commonalities and opportunities
for hybridization. An overview on hybrid approaches involving biologically inspired heuristics, together
with a valuable classification and a grammar for hybridization schemes can be found in Reference 17.
An early attempt to create a general framework for designing metaheuristics can be found in
Reference 18. The suggested framework is more fine-grained and not quite as general as the frame-
work that we suggest. It attempts to incorporate not only metaheuristics, but also more classical search
methods like branch and bound. Implicit assumption of the framework seems to be that the algorithm
operates on (partial) solutions, that is, that the memory stores (partial) solutions.
After an extensive description and classification of different metaheuristics, Blum and Roli [19] develop
a unified view on the intensification and diversification aspects of algorithmic components. Analyzing
the signature of available algorithm components in the developed framework, and identifying suitable
combinations, seems to be a promising way toward the development of a systematic design approach for
hybrid metaheuristic algorithms.
A framework that is focused on algorithms that use an updated parameterized model to generate can-
didate solutions was presented in Reference 20. It offers a unified view on ACO, estimation of distribution
algorithms (EDAs) and related methods.

10.3 A Unified Framework for Iterative Search Heuristics

Most modern search heuristics like EAs, simulated annealing, ACO, or tabu search, are iterative and
repeatedly probe the search space at new locations.
What distinguishes them from random search is primarily that they maintain some sort of memory
of the information gathered during the search so far, and that they use this information to select the
location where the search space should be tested next. The proposed general model, first presented in
Reference 21, follows from this observation and is depicted in Figure 10.1: new solutions are constructed
based on information stored in the memory, possibly involving several construction operators that may
be applied in parallel (e.g., when different solutions are generated by different operators in each iteration)
or sequentially (e.g., when a local optimizer is applied to each solution generated). The construction
operators can be rather simple (as e.g., a single bit flip) or rather complicated (e.g., a local optimizer). The
new solutions are then evaluated and can be used to update the memory, after which the cycle repeats.
In the following sections, we will show in more detail how this general framework can be used to
describe some of the aforementioned metaheuristics:

• Evolutionary algorithms store information about the previous search in the form of a set of solutions
(population). New solutions are constructed by selecting two solutions (parents), combining them

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 148 — #2

A Unified View on Metaheuristics and Their Hybridization 10-149

Construction operators

Select one or
several operators

Memory

Tabu list Pheromone matrix

1 4 7 2 5 12 10 3 6 8 9 11 1 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7
4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6
1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11
3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9
5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 2 Construct New solutions
1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 1 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7
4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 new solutions
3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9
Population 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 2
1 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7
1 4 7 2 5 12 10 3 6 8 9 11 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 2 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 1 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 4 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 3 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 5 7 8 4 11 9 3 5 0 1 3 6 8 10 8 7 6 9 2 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 1 4 7 2 5 12 10 3 6 8 9 11
1 4 7 2 5 12 10 3 6 8 9 11 1 4 7 2 5 12 10 3 6 8 9 11

Probabilistic model
Other memory
operators
Memory update

Select one or
several operators

Update operators

FIGURE 10.1 Unified framework for iterative search algorithms.

in some way (crossover), and performing some local modifications (mutation). Then, the memory
is updated by inserting the new solutions into the population. Although there exist a variety of
EA variants, they all fit into this general framework. For example, evolution strategies with self-
adaptive mutation can be specified by extending the memory to also maintain information about
the strategy parameters. Steady state genetic algorithms update the population after every newly
generated solution, while genetic algorithms with generational reproduction generate a whole new
population of individuals before updating the memory.
• Simulated annealing only maintains a single solution in the memory. In addition to that, it keeps
track of time by a temperature variable. New solutions are created by local modifications more or
less equivalent to the mutation operator in EAs. The new solution replaces the current memory
solution depending on the quality difference and the temperature.
• Tabu search, just as simulated annealing, creates new solutions based on a single current solution in
the memory. Additionally, it maintains so-called tabu-lists to avoid revisiting previous solutions.
These tabu lists are generally recently visited solutions or recently performed move operations.
New solutions are created by local modifications, while taking into account the tabu list. More
advanced tabu search algorithms can comprise of a number of additional memory structures,
like for example, a long-term frequency memory that records the number of times a particular
component has appeared in a solution.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 149 — #3

10-150 Handbook of Bioinspired Algorithms and Applications

• Particle swarm optimization uses a swarm (set) of particles (current solutions). The search process
can be imagined as a parallel search of particles “moving” through the landscape defined by the
fitness function. In addition to their locations (solution characteristics), the memory contains for
each particle the personal best solution encountered so far and a velocity, which can be seen as a
kind of general accumulated search direction. In every iteration, new solutions are generated by
moving the particles according to their velocity, and a linear, spring-like attraction to their personal
best solution encountered and the overall best solution encountered by any of the swarm’s particles.
Memory update includes an update of the particle locations, the personal best solutions, and the
particles’ velocities.
• Ant colony optimization, when compared to the approaches outlined earlier, has a completely dif-
ferent way to store information about the search conducted so far. Instead of storing complete
solutions, it accumulates information about which partial decisions have been successful when
constructing solutions from scratch. For example, for the traveling salesperson problem, it main-
tains a (so-called pheromone) matrix indicating for each city how desirable it is to visit another city.
Using this matrix, new solutions are constructed systematically, starting at a random city, and iter-
atively selecting the next city probabilistically according to the relative preferences encoded in the
matrix. Usually, several new solutions are generated that way, and then the best solution found is
used to update the matrix, increasing the probability that future ants will make the same decisions.
An elitist ant (best solution found so far) can be modeled by an additional (complete) solution
stored in the memory.
• Estimation of distribution algorithms, similar to ACO, construct solutions based on a probabilistic
model, only that the probabilistic model is not necessarily stored in the form of a matrix. The
new solutions are then evaluated, and the information gained is used to update the probabilistic
model. The class of EDAs contains a multitude of different approaches that vary primarily in the
complexity of the probabilistic model (in particular whether they take variable dependencies into
account or not), and in the way the probabilistic model is updated (incrementally or reconstructed
every iteration based on the generated samples). Note that many EDA approaches actually do not
use a probabilistic model as main memory component, but instead rely on a population of solutions
as underlying memory structure, and construct a new probabilistic model in every iteration based
on the current population.

Given a description of the different metaheuristics in this general form has many benefits. First, it
creates a common language, which allows researchers from different fields to understand each other’s
approaches easily. Second, it moves the focus from the complete algorithms to the components. And
third, it provides the interfaces for the different components to work together.
Based on the presented unified framework, it is almost straightforward to combine different com-
ponents from different algorithmic paradigms: an algorithm designer can easily select a combination of
memorization features, choose a suitable set of construction operators or create new ones that make use
of the combined set of selected memorization features, and then decide how the memory is updated with
the newly generated information. The framework allows for a lot of freedom: new solutions may be con-
structed in different ways, using different information from the memory, the solutions thus constructed
using one part of the memory may be used to update another part of the memory, and so on.

10.4 Some Thoughts about Memory

Since it is basically the memory that differentiates metaheuristics from random search, the organization
of the memory seems to be a crucial aspect. Generally, it would be desirable to store in the memory all
the information gathered during the search so far. However, that is usually impossible, not only because
of memory space restrictions, but also because extracting relevant information from such a huge memory
would be prohibitively slow. Therefore, the information has to be condensed in some way.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 150 — #4

A Unified View on Metaheuristics and Their Hybridization 10-151

Depending on what information is stored in the memory, metaheuristics may be classified into solution-
based or model-based (cf. [20]).
The approaches from the former category primarily keep some of the solutions generated so far.
Simulated annealing just stores a single solution; EAs and tabu search store a set of solutions. Although
the way these solutions are selected is different for the different algorithms, the implicit assumption
always is that the stored solutions sufficiently represent the promising regions of the search space and
appropriately reflect the history of the search.
Ant Colony Optimization belongs to the latter category: it assumes that the problem is to make a
sequence of decisions, and then accumulates information about the desirability of making a certain
decision in a given situation (state). It builds a model of construction methods. The space and complexity
limitations are observed by restricting the number of states, and by ignoring interdependencies between
decisions. For example, the state usually considered when solving a traveling salesperson problem is the
current city, independent of the sequence of cities visited so far. Because usually, only complete solutions
(corresponding to a combination of decisions) can be evaluated, but the memory stores desirability of
decisions, ACO has to assign the credit for a good solution to the individual decisions. Currently, this is
done in a straightforward way by simply distributing the credit evenly.
The class of EDAs is rather broad, and by using a population of solutions as well as a probabilistic
model, different instantiations can be closer to either the solution-based or the model-based memory
category. They are more or less decision-based (i.e., construct solutions step by step) but may keep track
of variable dependencies through graphical models (e.g., Bayes networks or Gaussian networks).
Obviously, each of the above memorization schemes has its benefits and its drawbacks. Storing complete
solutions preserves all the interdependencies between decision variables and thus implicitly takes epistasis1
into account. However, it discards a lot of the solutions generated. On the other hand, the decision-
based memorization scheme used for example, in ACO integrates information about many generated
solutions (information is only slowly evaporated), at the expense of losing a lot of information about
interdependencies.
From the above considerations, it seems natural that a combination of these two fundamental mem-
orization schemes may be beneficial, which is one of the reasons why we decided to further explore the
combination of EAs and ACO using the framework proposed in Section 10.3.
Furthermore, it should be mentioned that besides information about the search space, it seems prom-
ising to also store information about algorithmic parameters or meta-information about the search like
some characteristics of the fitness values observed over time. Examples for such meta-knowledge are
the already mentioned self-adaptive mutation in evolution strategies, or the dependence of the tempera-
ture parameter on the fraction of accepted moves in simulated annealing. Tabu search may have several
additional memorized features like the frequency of certain moves in the past, etc. Another example, in
particular for hybrid approaches, would be to track the performance of different operators in order to
decide which operators should be used more often. Again, all those aspects easily fit into the proposed
framework.

10.5 Combining Evolutionary Algorithms and

Ant Colony Optimization
In this section, we propose a number of EA/ACO hybrids, which attempt to combine the two memorization
schemes. Before that, however, let us briefly present the pure EA and ACO we built on. The application
considered is the traveling salesperson problem (TSP).

1 Epistasis,
in the field of evolutionary computation, generally refers to the fact that different components of the
solution interact.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 151 — #5

10-152 Handbook of Bioinspired Algorithms and Applications

10.5.1 Basic Evolutionary Algorithm

The standard EA uses a population of n individuals as memory. Individuals are represented as permuta-
tions, and the population is initialized randomly. A new individual is constructed by first selecting two
parents according to linear ranking selection, applying crossover, and mutating the resulting offspring.
As crossover operator, one of the following operators was used:
• Order crossover (OX) selects a random connected part of one parent and fills the remaining places
with the missing cities in the order they occur in the other parent [22].
• Edge recombination crossover (ERX) attempts to preserve as many edges as possible. It iteratively
constructs a tour, starting from a random city. At each step, it first considers the (up to four) cities
that are neighbors (i.e., connected) to the current location in either of the two parents. If at least
one of those has not yet been visited, it selects the city that has the fewest yet unvisited other cities
as neighbors in the parents. Otherwise, a random successor is selected [23].
Mutation exchanges the position of two randomly selected cities. The mutation operator is called
repeatedly with probability pm , that is, an individual is mutated i times with probability (pm )i (1 − pm ).
In each cycle, k individuals are constructed (usually k = 1) and, to update the memory, the worst
individuals in the population are replaced.

10.5.2 Basic Ant Colony Optimization

The memory structure of ACO is an m × m pheromone matrix with m being the number of cities in the
problem. Initially, all values τij of the matrix are set to 0.5. New individuals are constructed by starting at
a random city i, and choosing the following city j probabilistically according to the relative pheromone
values in row i of the pheromone matrix, that is, with probability

τijα
Pij =
α,
l∈U τil

where U being the set of yet unvisited cities and α being a user-defined weighting parameter.
After the k new solutions of one cycle have been constructed, pheromone is evaporated by multiplying
every element τij of the pheromone matrix with (1 − ρ) where ρ is the pheromone evaporation rate. Then,
the best individual found in the iteration is used to update the pheromone matrix by adding 1/L to all
matrix elements τij where the edge from city i to city j is part of the corresponding tour with length L.
Furthermore, the algorithm maintains the overall best solution as an elite that is also used to update
the matrix in the above way in every cycle.
Note that for many optimization problems, the construction procedure as used by ACO allows to
incorporate heuristic knowledge in a very elegant way. In the case of TSP, for example, this can be done
by simply preferring close cities in the selection process.
The selection probabilities then become

β
τijα ηij
Pij =
,
α β
l∈U τil ηil

where ηij = 1/dij being the reciprocal of the distance dij between cities i and j and β being the weight for
the heuristic information.
Naturally, a direct comparison of algorithms that use heuristic knowledge with algorithms that do not
is not fair. Since the focus of this chapter is not to construct the best algorithm to solve TSP, but rather
to study the effect of different memorization schemes and algorithmic hybrids, most of the approaches
suggested later do not incorporate problem-specific knowledge. Nevertheless, we will also briefly report
on the performance of ACO and a hybrid operator when they incorporate heuristic knowledge.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 152 — #6

A Unified View on Metaheuristics and Their Hybridization 10-153

10.5.3 Hybrid Algorithms

50/50: Maybe the most straightforward combination of EAs and ACO is to simply use both basic algorithms
to generate a portion of the new solutions each. More specifically, in every cycle, we generate 50% of the
k new solutions on the basis of the pheromone matrix, while the remaining 50% are generated using the
edge recombination operator. The complete set of k new solutions is then used in the standard way to
update the pheromone matrix as well as the population.
Pheromone Completion (PC) crossover: This operator at the same time uses pheromone matrix and
population to create a new individual. First, an individual is selected from the population by rank-based
selection. A random connected part of that individual is then chosen, and this partial permutation
is completed using the standard ACO construction operator, that is, probabilistically according to the
pheromone matrix. After k individuals have been created that way, the new individuals are used to
update the pheromone matrix as well as the population. Note that this operator is somewhat similar to
the approach suggested by Miagkikh and Punch [24]. However, while we are proposing to use separate
memory structures for the population and the pheromone matrix, they propose to use a population of
“agents,” each agent consisting of a solution and an individual pheromone matrix. It is difficult to reason
whether using a global pheromone matrix or individual pheromone matrices is more promising. A global
pheromone matrix collects more information and will, therefore, perhaps be a better guide in particular in
short runs. Individual pheromone matrices, on the other hand, allow for different solutions to be encoded
simultaneously, which may be more beneficial for long optimization runs, when diversity is a major
issue.
Pheromone-supported Edge Recombination (PsER) crossover: It may happen that all the four edges used
by the edge recombination operator to select the next city lead to cities that have already been visited.
In these cases, edge recombination selects a city randomly. The pheromone-supported edge recombination
operator suggested here instead uses the probabilistic selection based on the pheromone matrix. Again,
the resulting individual is subject to mutation, and after k individuals have been created, all of them are
used to update both the pheromone matrix as well as the population.
Mutating Ants (MutA): ACO maintains diversity by choosing cities probabilistically in every step, and
thus does not seem to require additional mutations. However, a simple change, like swapping two cities
(i.e., what mutation does) is very unlikely to be produced by the probabilistic construction procedure.
Therefore, we here suggest to mutate the solutions created by the ACO, adding a different kind of change.
Ant-based crossover (ABX): The idea of this hybrid is to combine ACO’s sequential construction and
elegant integration of heuristic knowledge with the population-based memory of EAs. ABX selects parents
from the population just as an ordinary EA. These parents are used to construct a temporary pheromone
matrix, by initializing all pheromones to the same small value, and allowing the parents to place addi-
tional pheromone on the solutions (paths) they represent. New solutions are then constructed in an ACO
way, based on the temporary pheromone matrix and possible heuristic knowledge. Only the best of the
generated solutions is returned as child and used to update the population (memory). It is straightfor-
ward to extend this idea by running an ACO on the temporary pheromone matrix for a few iterations
(allowing the generated solutions to update the temporary pheromone matrix before generating some
more solutions). ABX has first been proposed in Reference 25, and for the tests given later, we simply
used the same parameter settings that had shown to be successful in that paper: two parents are selected,
and the ACO is run on the temporary matrix for 5 iterations with 12 ants each. Only the best solution
found is returned as a child.

10.6 Empirical Evaluation

Since ACO is primarily designed for permutation problems, we chose a simple symmetric Euclidean TSP
with 100 cities to compare the different algorithms. Three problem instances of varying difficulty have
been created: in problem P0, all cities are located equally spaced on a unit circle. To generate problem
instances P3 and P6, the location of each of P0’s city has been moved in a random direction by a distance

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 153 — #7

10-154 Handbook of Bioinspired Algorithms and Applications

of 0.3 or 0.6, respectively. The three resulting problems are visualized in Figure 10.2. Independent of the
problem instance, each algorithm was allowed to create and evaluate 200,000 solutions.
Comparing different algorithms is a tricky business, because the results are highly dependent on the
parameters used. Therefore, we tested all possible combinations of the parameter settings depicted in
Table 10.1.
A comparison of the performance of the different algorithms on all three problems is presented in
Table 10.2. The table reports the performance of the best parameter setting for each algorithm. All results
are averaged over 30 runs with different random seeds. ACO with incorporated heuristic knowledge
(β = 5), and ABX is also included for comparison.
Looking at the basic algorithms first, it is obvious that the edge recombination crossover yields signifi-
cantly better results than the order crossover, which again performs better than ACO. The EA is able to
find the optimal solution for the simple problem (which is approximately the circumference of a unit

(a) (b) (c)

1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
–0.5 –0.5 –0.5
–1 –1 –1
–1.5 –1.5 –1.5
–1.5 –1 –0.5 0 0.5 1 1.5 –1.5 –1 –0.5 0 0.5 1 1.5 –1.5 –1 –0.5 0 0.5 1 1.5

FIGURE 10.2 Location of cities in problem instance (a) P0, (b) P3, and (c) P6.

TABLE 10.1 Parameter Settings Tested

α k ρ n pm s
ACO 1, 5, 10 10, 20, 50, 100 0.005, 0.01, 0.05 — — —
EA with ERX — — — 100, 200, 300 0.3, 0.5 1.25, 1.5, 2.0
EA with OX — — — 100, 200, 300 0.3, 0.5 1.25, 1.5, 2.0
50/50 1, 5 10, 20, 50 0.005, 0.01, 0.05 200, 300 0.3, 0.5 1.5
PsER crossover 1, 5 10, 20, 50 0.005, 0.01, 0.05 200, 300 0.3, 0.5 1.25, 1.5, 2.0
PC crossover 1, 5 10, 20, 50 0.005, 0.01, 0.05 200, 300 0.3, 0.5 1.25, 1.5, 2.0
MutA 1, 5, 10 10, 20, 50, 100 0.005, 0.01, 0.05 — 0.3, 0.5 —

TABLE 10.2 Tour Length of the Best Solution Found by the Different
Algorithms

Problem instance

Algorithm 0.0 0.3 0.6

Pure ACO 14.22 24.54 29.76

EA with ERX 6.28 18.29 23.22
EA with OX 12.25 22.30 28.72
Hybrids 50/50 6.58 18.17 23.92
PC crossover 12.61 22.21 28.56
PsER crossover 6.28 17.59 22.88
MutA 8.51 18.05 24.21
With heuristic knowledge ACO + heuristic 6.28 16.11 21.80
ABX 6.28 16.11 21.80

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 154 — #8

A Unified View on Metaheuristics and Their Hybridization 10-155

circle, i.e., 2π ), but deviates from the optimal solution for the other two problems (the optimum is not
known, but ACO with heuristic knowledge achieves better results). ACO, on the other hand, seems to be
fairly unaffected by the structure of the problem.
As to the hybrids, only the pheromone-supported ER crossover was a clear winner. Simple mutation
of the solutions constructed by ACO also performed very well, and consistently outperformed the basic
ACO approach, indicating that the algorithm indeed benefited from the mix of operators. The pheromone
completion crossover was disappointing (although still better than the basic ACO). However, these results
are preliminary, and may be due to the specific problem instances chosen. Certainly, there are problems
where ACO prevails, while for other problems, EAs perform better. Yet other problems may require a
hybrid approach to be solved.
As expected, heuristic domain knowledge is able to drastically improve performance. The ACO with
heuristic knowledge as well as ABX generate equally good (presumably optimal) results on all problem
instances, outperforming all methods without heuristic knowledge. Note that besides the idea of ABX,
incorporation of domain knowledge into the EA is not as straightforward, and we have not been able to
produce similar results, for example, by seeding the population with a heuristic (results not reported).
As the results show, the problem instances examined are too simple if heuristic knowledge is incorporated.
On the larger problem instances used in Reference 25, ACO and ABX clearly outperformed a standard EA
with ERX, and ABX outperformed ACO by ∼1.3%.

10.7 Conclusion
In this chapter we have presented a unified framework for iterative search heuristics. According to the
framework, each search heuristic maintains some sort of memory of the search history, which is used to
construct new solutions, which are then evaluated and used to update the memory. Furthermore, we have
argued that different memory schemes have different advantages, and that a search heuristic should benefit
from combining different memorization paradigms.
The presented unified framework suggests a natural way for hybridization, and we have demonstrated
its usefulness by deriving several interesting combinations of EAs and ACO and conducting a preliminary
empirical evaluation of the resulting hybrids. A closer look at the compatibility of different memory
schemes, and how they are best combined, is subject to future research. For the algorithm designer,
of course, it would be invaluable to know which operators and memory schemes are most promising
depending on the application at hand. However, that assumes a useful categorization of problems, and is
thus several steps in the future. Overall, we hope that this chapter helps to gain a general understanding
of different metaheuristics and of the way they interact.

References
[1] K.A. DeJong. Evolutionary Computation. MIT Press, Cambridege, MA, 2002.
[2] E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence: From Natural to Artificial Systems.
Oxford University Press, Oxford, 1999.
[3] E. Aarts and J. Korst. Simulated Annealing and Boltzmann Machines. John Wiley & Sons, New York
1989.
[4] F. Glover. Tabu search — p. I. ORSA Journal of Computing, 1: 190–206, 1989.
[5] P. Larrañaga and J.A. Lozano, Eds. Estimation of Distribution Algorithms. Kluwer Academic,
New York, 2002.
[6] C. Reeves, Ed. Modern Heuristic Techniques for Combinatorial Optimization. McGraw-Hill,
New York, 1995.
[7] E.L. Aarts and J.K. Lenstra, Eds. Local Search in Combinatorial Optimization. Wiley, Chichester,
1997.
[8] Z. Michalewicz and D.B. Fogel. How to Solve It: Modern Heuristics. Springer, New York, 1999.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 155 — #9

10-156 Handbook of Bioinspired Algorithms and Applications

[9] D.E. Brown, C.L. Huntley, and A.R. Spillane. A parallel genetic heuristic for the quadratic
assignment problem. In International Conference on Genetic Algorithms, Morgan Kaufmann, San
Francisco, CA, 1989, pp. 406–415.
[10] S.W. Mahfoud and D.E. Goldberg. Parallel recombinative simulated annealing: A genetic
algorithm. Parallel Computing, 21: 1–28, 1995.
[11] C. Fleurent and J. Ferland. Genetic and hybrid algorithms for graph coloring. Technical report,
Departement d‘Informatique, Montreal, Canada, 1994.
[12] V.V. Miagkikh and W.F. Punch. An approach to solving combinatorial optimization problems
using a population of reinforcement learning agents. In Genetic and Evolutionary Computations
Conference, Morgan Kaufmann, San Francisco, CA, 1999, pp. 1358–1365.
[13] E.-G. Talbi, O. Roux, C. Fonlupt, and D. Robilliard. Parallel ant colonies for the quadratic
assignment problem. Future Generation Computer Systems, 17: 441–449, 2001.
[14] T. Krink and M. Lovbjerg. The lifecycle model: Combining particle swarm optimisation, genetic
algorithms and hillclimbers. In J.J. Merelo, P. Adamidis, H.-G. Beyer, J.-L. Fernandez-Villacanas,
and H.-P. Schwefel, Eds., Parallel Problem Solving from Nature, Vol. 2439 of Lecture Notes in
Computer Science, Springer, New York, 2002, pp. 621–630.
[15] P. Moscato. Memetic algorithms: A short introduction. In D. Corne, M. Dorigo, and F. Glover,
Eds., New Ideas in Optimization, McGraw Hill, New York, 1999, chap. 14, pp. 219–234.
[16] P. Calegari, G. Coray, A. Hertz, D. Kobler, and P. Kuonen. A taxonomy of evolutionary algorithms
in combinatorial optimization. Journal of Heuristics, 5: 145–158, 1999.
[17] E.-G. Talbi. A taxonomy of hybrid metaheuristics. Journal of Heuristics, 8: 541–564, 2002.
[18] R. Poli and B. Logan. The evolutionary computation cookbook: Recipes for designing new
algorithms. In Online Workshop on Evolutionary Computation, 1996, pp. 33–36.
[19] C. Blum and A. Roli. Metaheurisitics in combinatorial optimization: Overview and conceptual
comparison. ACM Computer Survey, 35: 268–308, 2003.
[20] M. Zlochin, M. Birattari, N. Meuleau, and M. Dorigo. Model-based search for combinatorial
optimization. Technical report TR/IRIDIA/2000-15, INRIDIA, Universite Libre de Bruxelles, 2001.
[21] J. Branke, M. Stein, and H. Schmeck. A Unified Framework for Metaheuristics. Technical
report 417, University of Karlsruhe, Institute AIFB, Karlsruhe, Germany, 2002.
[22] L. Davis. Applying adaptive algorithms to epistatic domains. In International Joint Conference on
Artificial Intelligence, Morgan Kaufmann, San Francisco, CA, 1985, pp. 162–164.
[23] D. Whitley, T. Starkweather, and D’A. Fuquay. Scheduling problems and traveling salesman:
The genetic edge recombination operator. In J. Schaffer, Ed., International Conference on Genetic
Algorithms, Morgan Kaufmann, San Francisco, CA, 1989, pp. 133–140.
[24] V.V. Miagkikh and W.F. Punch. Global search in combinatorial optimization using reinforce-
ment learning algorithms. In Congress on Evolutionary Computation, IEEE, Piscataway, 1999,
pp. 189–196.
[25] J. Branke, C. Barz, and I. Behrens. Ant-based crossover for permutation problems. In E. Cantu-Paz,
Ed., Genetic and Evolutionary Computation Conference, Vol. 2273 of Lecture Notes in Computer
Science, Springer, New York, 2003, pp. 754–765.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C010” — 2005/8/8 — 21:40 — page 156 — #10

 11
The Foundations of
Autonomic
Computing

11.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-157

11.2 Autonomic Nervous System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-158
Ashby’s Ultrastable System • The Nervous System as a
Subsystem of Ashby’s Ultrastable System
11.3 Autonomic Computing Paradigm . . . . . . . . . . . . . . . . . . . . . 11-161
Architecture of an Autonomic Component
11.4 Autonomic Computing Systems . . . . . . . . . . . . . . . . . . . . . . . . 11-163
11.5 Illustrative Example: Distributed Cellular DEVS
Dynamic Forest Fire Simulation . . . . . . . . . . . . . . . . . . . . . . . . 11-164
Forest Fire Cell as an Autonomic Component • Forest Fire
Cell Space as Autonomic Coupled Component • Dynamic
Composition and Self-Management of Forest Fire Application
Salim Hariri 11.6 The Autonomic Computing Landscape . . . . . . . . . . . . . . . . 11-171
Bithika Khargaria AutoMate — Enabling Autonomic Applications • Autonomia
Manish Parashar 11.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-174
Zhen Li References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-174

11.1 Introduction
The advances in computing and communication technologies and software tools have resulted in an
explosive growth in networked applications and information services that cover all aspects of our life. These
services and applications are inherently complex, dynamic, and heterogeneous. In a similar way, the under-
lying information infrastructure, for example, the Internet, is large, complex, heterogeneous, and dynamic,
globally aggregating large numbers of independent computing and communication resources, data stores,
and sensor networks. The combination of the two results in application development, configuration, and
management complexities that break current computing paradigms, which are based on static behaviors,
interactions, and compositions of components and services. As a result, applications, programming
environments, and information infrastructures are rapidly becoming brittle, unmanageable, and insecure.
This has led researchers to consider alternative programming paradigms and management techniques that
are based on strategies used by biological systems to deal with complexity, dynamism, heterogeneity, and
uncertainty.

11-157

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 157 — #1

11-158 Handbook of Bioinspired Algorithms and Applications

Autonomic computing is inspired by the human autonomic nervous system, which has developed
strategies and algorithms to handle complexity and uncertainties, and aims at realizing computing systems
and applications capable of managing themselves with minimum human intervention. In this chapter,
we first give an overview of the architecture of the nervous system and use it to motivate the autonomic
computing paradigm. We then illustrate how this paradigm can be used to build and manage complex
applications. Finally, we present an overview of existing autonomic computing systems and applications
and highlight two such systems.

11.2 Autonomic Nervous System

The human nervous system is, to the best of our knowledge, the most sophisticated example of autonomic
behavior existing in nature today [1]. It is the body’s master controller that monitors changes inside and
outside the body, integrates sensory input, and effects appropriate response. In conjunction with the
endocrine system, which is the body’s second important regulating system, the nervous system is able to
constantly regulate and maintain homeostasis. Homeostasis is one of the most remarkable properties of
highly complex systems. A homeostatic system (e.g., a large organization, an industrial firm, a cell) is an
open system that maintains its structure and functions by means of a multiplicity of dynamic equilibriums
that are rigorously controlled by interdependent regulation mechanisms. Such a system reacts to every
change in the environment, or to every random disturbance, through a series of modifications that are
equal in size and opposite in direction to those that created the disturbance. The goal of these modifications
is to maintain internal balances.
The manifestation of the phenomenon of homeostasis is widespread in the human system. As an
example, consider the mechanisms that maintain the concentration of glucose in the blood within
limits — if the concentration should fall below about 0.06%, the tissues will be starved of their chief
source of energy; if the concentration should rise above about 0.18%, other undesirable effects will occur.
If the blood-glucose concentration falls below about 0.07%, the adrenal glands secrete adrenaline, which
causes the liver to turn its stores of glycogen into glucose. This passes into the blood and the blood-glucose
concentration drop is opposed. Further, a falling blood-glucose also stimulates appetite causing food
intake, which after digestion provides glucose. On the other hand, if the blood-glucose concentration rises
excessively, the secretion of insulin by the pancreas is increased, causing the liver to remove the excess
glucose from the blood. Excess glucose is also removed by muscles and skin, and if the blood-glucose con-
centration exceeds 0.18%, the kidneys excrete excess glucose into the urine. Thus, there are five activities
that counter harmful fluctuations in blood-glucose concentration [2].
The above example focuses on the maintenance of the blood-glucose concentration within safe or
operational limits that have been “predetermined” for the species. Similar control systems exist for other
parameters such as systolic blood pressure, structural integrity of the medulla oblongata, severe pressure
of heat on the skin, and so on. All these parameters have a bearing on the survivability of the organism,
which in this case is the human body. However, all parameters are not uniform in their urgency or their
relations to lethality [2]. Parameters that are closely linked to survival and are closely linked to each other
so that marked changes in one leads sooner or later to marked changes in the others, have been termed as
essential variables by Ashby in his study of the design for a brain [2], which is discussed later.

11.2.1 Ashby’s Ultrastable System

Every real machine embodies no less than an infinite number of variables [2], and for our discussion
we can safely think of the human system as represented by a similar sets of variables, of which we will
consider a few. In order for an organism to survive, its essential variables must be kept within viable
limits (see Figure 11.1). Otherwise the organism faces the possibility of disintegration and loss of identity
(i.e., dissolution or death) [2].
The body’s internal mechanisms continuously work together to maintain its essential variables within
their limits. Ashby’s definition of adaptive behavior as demonstrated by the human body follows from

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 158 — #2

The Foundations of Autonomic Computing 11-159

Viability zone
EV2

EV1

FIGURE 11.1 Essential variables.

Environment Essential variables

Motor Sensor
channels channels

Reacting part R Step mechanisms/input parameter S

FIGURE 11.2 The ultrastable system architecture [2].

this observation. He states that a form of behavior is adaptive if it maintains the essential variables within
physiological limits [2] that define the viability zone. Two important observations can be made:

• The goal of the adaptive behavior is directly linked with the survivability of the system.
• If the external or internal environment pushes the system outside its physiological equilibrium
state, the system will always work toward coming back to the original equilibrium state.

Ashby observed that many organisms undergo two forms of disturbance: (1) frequent small impulses
to main variables and (2) occasional step changes to its parameters. Based on this observation, he devised
the architecture of the Ultrastable system that consists of two closed loops (see Figure 11.2): one that
controls small disturbances and a second that is responsible for longer disturbances.
As shown in Figure 11.2, the ultrastable system consists of two subsystems, the environment and the
reacting part (R). R represents a subsystem of the organism that is responsible for overt behavior or
perception. It uses the sensor channels as part of its perception capability and motor channels to respond
to the changes impacted by the environment. These set of sensors and motor channels constitute the
primary feedback between R and the environment. We can think of R as a set of behaviors of the organism
that gets triggered based on the changes affected by the environment. S represents the set of parameters
that triggers changes in relevant features of this behavior set. Note that in Figure 11.2, S triggers changes
only when the environment affects the essential variables in a way that causes them to go outside their
physiological limits. As mentioned earlier, these variables need to be maintained within physiological
limits for any adaptive system/organism to survive. Thus we can view this secondary feedback between
the environment and R as responsible for triggering the adaptive behavior of the organism. When the
changes impacted by the environment on the organism are large enough to throw the essential variables
out of their physiological limits, the secondary feedback becomes active and changes the existing behavior
sets of the organism to adapt to these new changes. Notice that any changes in the environment tend to
push an otherwise stable system to an unstable state. The objective of the whole system is to maintain
the subsystems (the environment and R) in a state of stable equilibrium. The primary feedback handles

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 159 — #3

11-160 Handbook of Bioinspired Algorithms and Applications

finer changes in the environment with the existing behavior sets to bring the whole system to stable
equilibrium. The secondary feedback handles coarser and long-term changes in the environment by
changing its existing behavior sets and eventually brings back the whole system to stable equilibrium state.
Hence, in a nutshell, the environment and the organism always exist in a state of stable equilibrium and
any activity of the organism is triggered to maintain this equilibrium.

11.2.2 The Nervous System as a Subsystem of Ashby’s Ultrastable System

The human nervous system is adaptive in nature. In this section we apply the concepts underlying the
Ashby’s ultrastable system to the human nervous system. The goal is to enhance the understanding of
an adaptive system and help extract essential concepts that can be applied to the autonomic computing
paradigm presented in the following sections.
As shown in Figure 11.3, the nervous system is divided into the Peripheral Nervous System (PNS) and
the Central Nervous System (CNS). The PNS consists of sensory neurons running from stimulus receptors
that inform the CNS of the stimuli and motor neurons running from the CNS to the muscles and glands,
called effectors, which take action. CNS is further divided into two parts: sensory–somatic nervous system
and the autonomic nervous system. Figure 11.4 shows the architecture of the autonomic nervous system
as an Ashby utrastable system.
As shown in Figure 11.4, the sensory and motor neurons constitute the sensor and motor channels
of the ultrastable system. The triggering of essential variables, selection of the input parameter S and
translation of these parameters to the reacting part R constitute the workings of the nervous system.
Revisiting the management of blood-glucose concentration within physiological limits discussed earlier,
the five mechanisms that get triggered when the essential variable (i.e., concentration of glucose in blood)

Sensory neurons Sensory neurons

Autonomic Central Sensory–

Internal nervous somatic External
nervous
environment system (CNS) nervous environment
system
system

Motor neurons Motor neurons

FIGURE 11.3 Organization of the nervous system.

Sensor channels
Sensory neurons

Internal External
environment Reacting part R environment

Motor neurons
Essential variables (EV)
Motor channels

Environment
S = f (change in EV)

Step mechanisms/input parameter S

FIGURE 11.4 Nervous system as part of an ultrastable system.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 160 — #4

The Foundations of Autonomic Computing 11-161

goes out of the physiological limits change the normal behavior of the system such that the reacting part R
works to bring the essential variable back within limits. It uses its motor channels to effect changes so that
the internal environment and the system (organism) come into the state of stable equilibrium. It should be
noted that the environment here is divided into the internal environment and external environment. The
internal environment represents changes impacted internally within the human system and the external
environment represents changes impacted by the external world. However, the goal of the organism is to
maintain the equilibrium of the entire system where all the subsystems (the organism or system itself, and
the internal and external environments) are in stable equilibrium.

11.3 Autonomic Computing Paradigm

An autonomic computing paradigm, modeled after the autonomic nervous system, must have a mech-
anism whereby changes in its essential variables can trigger changes in the behavior of the computing
system such that the system is brought back into equilibrium with respect to the environment. This state
of stable equilibrium is a necessary condition for the survivability of the organism. In the case of an
autonomic computing system, we can think of survivability as the system’s ability to protect itself, recover
from faults, reconfigure as required by changes in the environment, and always maintain its operations at
a near optimal performance. Its equilibrium is impacted by both the internal environment (e.g., excessive
CPU utilization) and the external environment (e.g., protection from an external attack). The autonomic
computing system requires: (1) sensor channels to sense the changes in the internal and external envi-
ronment, and (2) motor channels to react to and counter the effects of the changes in the environment
by changing the system and maintaining equilibrium. The changes sensed by the sensor channels have
to be analyzed to determine if any of the essential variables has gone out of their viability limits. If so,
it has to trigger some kind of planning to determine what changes to inject into the current behavior of
the system such that it returns to the equilibrium state within the new environment. This planning would
require knowledge to select the right behavior from a large set of possible behaviors to counter the change.
Finally, the motor neurons execute the selected change. “Sensing,” “Analyzing,” “Planning,” “Knowledge,”
and “Execute” are in fact the keywords used to identify an autonomic system [3]. We use these concepts
to present the architecture of an autonomic component that represents the smallest unit of an autonomic
application or a system with self-managing capabilities.

11.3.1 Architecture of an Autonomic Component

An autonomic component (see Figure 11.5) is the smallest unit of an autonomic application or system.
Multiple autonomic components (units) can be composed to form an autonomic application or system,
which is a self-contained software module or system with specified input/output interfaces and explicit
context dependencies. It also has embedded mechanisms for self-management responsible for providing
functionalities, exporting constraints, managing its own behavior in accordance with context and policies,
and interacting with other components. For example, an autonomic component consists of the following
parts:

Managed Element: This is the smallest unit of the application and it contains executable code
(e.g., numerical model of a physical process) and a data structure that defines the executable
code’s attributes (e.g., its purpose, operation, input and output requirements, criteria for when and
how to control it). At runtime, the managed element can be affected in different ways, for example,
it can encounter a failure during execution, it can be externally attacked, or it may slow down and
affect the performance of the entire application.
Environment: The environment represents all the factors that can impact the managed element. The
environment and the managed element can be two subsystems forming a stable system. Any
change in the environment causes the whole system to go from a stable state to an unstable state.
This change is then offset by reactive changes in the managed element causing the system to move

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 161 — #5

11-162 Handbook of Bioinspired Algorithms and Applications

Input ports Output ports

Control
PE PE

KE KE
programmed autonomic
behavior behavior
M & A Cardinals

G cardinal
L

M&A E E M&A
S A S
Managed element

Environment
Internal External

FIGURE 11.5 An automatic component.

back from the unstable state to a different stable state. Notice that the environment consists of
two parts — internal and external. The internal environment consists of changes internal to the
managed element, which can be looked at as reflecting the state of the application. The external
environment can be thought of as reflecting the state of the execution environment.
Control: Each autonomous component has its own manager that: (1) accepts user-specified require-
ments (fault tolerance, performance, security, etc.) (2) interrogates the data structure that
characterizes the executable code (3) senses the state of the overall computation (4) determines the
nature and instantaneous state of the overall computational environment and (5) uses this infor-
mation to control the operation of its associated executable code within the overall system in order
to effectively achieve the user-specified requirements. This process is accomplished on-the-fly and
continuously throughout the execution of the overall computation. As is evident from Figure 11.5,
the control part consists of two control loops — the local loop and the global loop.
The local loop can only handle known environment states. Its knowledge engine contains the
mapping of environment states to behaviors. For example, when the load on the local system goes
above the threshold value, the local control loop will work toward balancing the load by either con-
trolling the local resources available to the managed element or by reducing the size of the problem
handled by this element. This will work only if the local resources can handle the computational
requirements. However, the local loop is blind to the overall behavior of the entire application or
system and thus cannot achieve the desired global objectives. In a scenario where the entire system
is affected, the local loop will continue repeating local optimization that may lead to degradation
in performance and result in unadapted or chaotic behavior. At some point, one of the essential
variables of the system (in this case, a performance cardinal) overshoots its limits. This is when the
global loop comes into action.
The global loop can handle unknown environment states and may involve machine learning.
It uses four cardinals for the monitoring and analysis of the managed elements. These are
performance, configuration, protection, and security. These cardinals are like the essential variables
described in Ashby’s ultrastable system. This control loop acts toward changing the existing behavior
of the managed element such that it can adapt itself to changes in the environment. For example, in
load-balancing, the desired behavior of the managed element (as directed by the local loop) requires
its local load to be within prescribed limits. However, the local loop might not be able to maintain
the local load within these acceptable limits, which in turn might degrade the performance of the
overall system. Consequently, this change in the overall performance cardinal triggers the global

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 162 — #6

The Foundations of Autonomic Computing 11-163

loop, which then selects an alternate behavior pattern from the pool of behavior patterns for the
managed element. This analysis and planning uses its knowledge engine. Finally, the new plan is
executed to adapt the behavior of the managed element to the new environment conditions.
Input and Output Ports: Many interacting autonomous components may be composed to form a
complex application. These autonomic components use the input and output ports for such a
composition.

11.4 Autonomic Computing Systems

An autonomic computing system can be a collection of autonomic components, which can manage their
internal behaviors and relationships with others in accordance to high-level policies. The principles that
govern all such systems have been summarized as eight defining characteristics [4]:

Self-Awareness: an autonomic system knows itself and is aware of its state and its behaviors.
Self-Protecting: an autonomic system is equally prone to attacks and hence it should be capable of
detecting and protecting its resources from both internal and external attack and of maintaining
overall system security and integrity.
Self-Optimizing: an autonomic system should be able to detect suboptimal behaviors and intelligently
perform self-optimization functions.
Self-Healing: an autonomic system must be aware of potential problems and should have the ability to
reconfigure itself to continue to function smoothly.
Self-Configuring: an autonomic system must have the ability to dynamically adjust its resources based
on its state and the state of its execution environment.
Contextually Aware: an autonomic system must be aware of its execution environment and be able to
react to changes in the environment.
Open: an autonomic system must be portable across multiple hardware and software architectures, and
consequently it must be built on standard and open protocols and interfaces.
Anticipatory: an autonomic system must be able to anticipate, to the extent possible, its needs and
behaviors and those of its context, and be able to manage itself proactively.

Sample self-managing system/application behaviors include installing software when it detects that
the software is missing (self-configuration), restarting a failed element (self-healing), adjusting current
workload when it observes an increase in capacity (self-optimization), and taking resources offline if it
detects an intrusion attempt (self-protecting).
Each of the attributes listed above are active research areas toward realizing autonomic systems and
applications. Generally, self-management is addressed in four primary system/application aspects, that is,
configuration, optimization, protection, and healing. Further, self-management solutions typically consist
of the steps outlined earlier: (1) the application and underlying information infrastructure provide infor-
mation to enable context and self-awareness (2) system/application events trigger analysis, deduction, and
planning using system knowledge and (3) plans are executed using the adaptive capabilities of the system.
An autonomic system implements self-managing attributes using the control loops described earlier to
collect information, makes decisions, and adapt, as necessary.
Autonomic components need to collaborate to achieve coherent autonomic behaviors at the application
level. This requires a common set of underlying capabilities including representations and mechanisms for
solution knowledge, system administration, problem determination, monitoring and analysis, and policy
definition, enforcement, and transaction measurements [5]. For example, a common solution knowledge
capability captures installation, configuration, and maintenance information in a consistent manner,
and eliminates the complexity introduced by heterogeneous tools and formats. Common administrative
console functions ranging from setup and configuration to solution runtime monitoring and control
provide a single platform to host administrative functions across systems and applications, allowing users
to manage solutions rather than managing individual systems/applications. Problem determination is one

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 163 — #7

11-164 Handbook of Bioinspired Algorithms and Applications

of the most basic capabilities of an autonomic element and enables it to decide on appropriate actions when
healing, optimizing, configuring, or protecting itself. Autonomic monitoring is a capability that provides
an extensible runtime environment to support the gathering and filtering of data obtained through sensors.
Complex analysis methodologies and tools provide the power and flexibility required to perform a range
of analyses of sensor data, including deriving information about resource configuration, status, offered
workload, and throughput. A uniform approach to defining the policies is necessary to support adaptations
and govern decision-making required by the autonomic system. Transaction measurements are needed to
understand how the resources of heterogeneous systems combine into a distributed transaction execution
environment. Using these measurements, analysis and plans can be derived to change resource allocations
to optimize performance across these multiple systems as well as determine potential bottlenecks in the
system.

11.5 Illustrative Example: Distributed Cellular DEVS Dynamic

Forest Fire Simulation
One important feature of autonomic applications is their ability to change the components and the
structure that interconnects these components at runtime as required by the self-control and management
algorithms. For example, in fire forest simulation, you might need to change the type of components used
to simulate the fire depending on the current atmosphere conditions (rain, dry, speed of wind), and the
vegetation type at each computational grid. The runtime system framework will then modify the data
structures of the autonomic components in order to add or remove components.
The forest fire simulation model has been developed using the concepts of cellular automata. A cellular
automaton is an array of identically programmed automata, or “cells,” which interact with one another.
Essentially, it is a one-dimensional (1D) string of cells, a 2D Grid or a 3D solid and has three important
features — state, neighborhood, and its program. Just as every living cell contains all of the instructions
for its replication and operation, each individual cell in a cellular automaton can be programmed with a
set of rules that defines how its state changes in response to its current state and that of the neighbors.
In the forest fire model, the forest is represented as a 2D cell space composed of cells of dimensions l × b
(l: length, b: breadth). For each cell there are eight major wind directions N, NE, NW, S, SE, SW, E, W as
shown in Figure 11.6 and Figure 11.7
A group of such individual cells will together constitute, a Virtual Computational Unit (VCU). The
weather and vegetation conditions are assumed to be uniform within a cell, but may vary in the entire
cell space. A cell interacts with its neighbors along all the eight directions as listed earlier, using input and

NW N NE

W E

SW SE
S

FIGURE 11.6 Cell with major wind directions.

NW N NE

W E

SW S SE

FIGURE 11.7 Fire directions after ignition.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 164 — #8

The Foundations of Autonomic Computing 11-165

Input ports Output ports

Autonomic manager
Analyze Plan
Knowledge
Monitor Execute

S A

Forest fire cell

FIGURE 11.8 A forest fire cell expressed as an autonomic component.

output ports (the DEVS-Java model). A cell is programmed to undergo state changes from “unburned”
to “burning” if it is hit by an igniter or gets a notification message to compute its fire-line intensity value.
The cell changes state from “unburned” to “burning” only if the computed fire-line intensity is above a
threshold value for “burning.” During the “burning” phase, the cell propagates to eight different fire
components along the eight directions (refer to Figure 11.5). The direction and value of maximum fire
spread is computed using Rothermel’s fire spread model [6]. The remaining seven components are then
derived using a different decomposition algorithm. Rothermel’s model takes into account the wind–speed
and direction, the vegetation type, the calorific value of the fuel, and terrain type in calculating the fire
spread.

11.5.1 Forest Fire Cell as an Autonomic Component

A single forest fire cell is modeled by an Autonomic Component (Figure 11.8), which consists of:
1. Specification: get notification messages from neighbor cell or igniter and send notification messages
to the neighbor cells.
2. States: windew , windns , vegh , vegm , vegl , unburned, burning, burned, blocked.
3. Actions: partition cell space horizontally/vertically, allocate/de-allocate resources, migrate compon-
ent, add/delete coupling, start, terminate, pause, resume, and so on.
4. Policies: for self-optimizing, self-healing, self-configuration, and self-protecting.
The set of all possible actions on the component, are distributed into different policies that can be
activated at runtime based on the component current state and the state of the physical resources. The
policies are then used to map the states to the actions, as shown in Figure 11.9. In the following, we will
explain few of these mappings with example scenarios.
Case 1: State change from windew to windns triggers actions guided by the self-optimizing policy.
The entire forest cell space is to be distributed into smaller units consisting of individual forest cells
such that each group of forest cells executes on a single compute node. Since this application involves a
lot of interactions between neighboring cells, an optimal cell space partitioning plan must ensure that
neighbor cells do not spend too much time communicating with each other. One strategy is to partition
cells based on wind direction. As is evident from Figure 11.10, the wind direction determines the direction
of the communications among neighboring cells in the cell space. Thus, a partitioning strategy based
on wind direction will ensure that most of the communication occurs within the cells assigned to a single
compute node.
As seen from Figure 11.11, as the wind direction changes from east–west to north–south, the state of
the fire cell changes from windew to windns after a certain time. At the end of the time advance, an output

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 165 — #9

11-166 Handbook of Bioinspired Algorithms and Applications

Self-optimizing
actions
Self-
optimizing Pause
policies
Resume
Self-protecting Partition cell space
actions horizontally
Self- States
protecting Partition cell space vertically
policies
windw Allocate resources

windn De-allocate resources

vegh Migrate component
vegm
Start
Self-healing vegl
actions Unburned
Self-
healing Self-configuring
Burning
Terminate policies Self- actions
Burned configuring
Resume policies
Blocked
Pause
Ready
Migrate Delete coupling
component Failed Add coupling
Stopped

FIGURE 11.9 Mapping from states to actions governed by policies.

(a)
1

2
... North or south
1 2 N
wind directions
...

East or west
wind directions

(b)
1 1

2 2 North–east or south–west
... ... wind direction

N N

North–west or south–east
wind direction

FIGURE 11.10 Partitioning of cell space based on wind direction.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 166 — #10

The Foundations of Autonomic Computing 11-167

Repartition
cell space
vertically

1
2
... 1 2 ... N

<Cact: partition cell space vertically>

<Cpol: self-optimization policy >

Generate output

Time advance

Cstate: windew Cstate: windns

State
transition

FIGURE 11.11 State transition guided by the self-optimization policy.

Cact: pause Cact: migrate component Cact: resume

Cstate: burning Cstate: failed Cstate: stopped Cstate: ready Cstate: burning

Internal transition
due to
node failure

FIGURE 11.12 Sets of state transitions and actions guided by self-healing policies.

is generated, which involves a self-optimization action. In this case, the requested action is to repartition
the cell space vertically. Referring to Figure 11.11, we see that this action is mapped with the state windns
The autonomic manager uses its monitoring and analysis engines to detect this state change. The planning
engine then generates the appropriate action as shown in Figure 11.11. The appropriate action (mapping
function) is a part of the knowledge stored in the knowledge engine. Finally, the execution engine executes
this action.
Case 2: State change from burning to failed triggers actions guided by the self-healing policy.
A scenario where a component stops running due to a failure in the node is shown in Figure 11.12.
A series of state changes generates the corresponding actions as shown in Figure 11.12. Finally the
component resumes execution from the last check-pointed state and its state changes into burning.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 167 — #11

11-168 Handbook of Bioinspired Algorithms and Applications

Input ports Output ports

Autonomic manager
Analyze Plan Group of forest
Knowledge cells
Monitor Execute

S A
VCU

Input ports
Output ports Input ports Output ports Input ports Output ports

Autonomic manager Autonomic manager Autonomic manager

Analyze Plan Analyze Plan Analyze Plan
Knowledge Knowledge Knowledge

Monitor Execute Monitor Execute Monitor Execute

S A S A S A

Forest fire cell 1 Forest fire cell 2 Forest fire cell 3

FIGURE 11.13 A group of forest fire cells expressed as an autonomic coupled component.

11.5.2 Forest Fire Cell Space as Autonomic Coupled Component

The forest fire cell space is nothing but a collection of forest cells that are connected in a particular manner
to create the entire forest application. In terms of our autonomic component formalisms, we can view the
forest cell space as a collection of autonomic forest cells (Figure 11.13).
A forest fire autonomic coupled component consists of: (1) specification: a union of a group individual
forest cell autonomic components with connected input, output ports. (2) actions: repartition cell space,
migrate, pause, resume. (3) global policies: for self-optimization, self-configuration, and so on to manage
the autonomic capabilities of the coupled component which, in our case, is the collection of forest fire
cells. The interaction between the states, actions, and policies are very much similar as explained earlier.
In what follows, we will explain this interaction in the context of the forest fire application.
Load balanced execution is crucial to the performance of the forest fire application because the
entire application proceeds at the speed of the slowest problem piece (VCU) in the entire cell space.
In Figure 11.14, we depict a scenario where the complexity of VCU1 (number of forest cells) is
reduced by removing some cells from that VCU and adding them to another VCU2 that is lightly
loaded.
VCU1 starts at loadHigh. Note that this state is a union of the states of the cells in that VCU. Hence
loadHigh is the union of burning1, burning2, unburned3, . . . , burningn . This could be a possible scenario
because cells perform the major amount of computation during the burning state. Hence, if the complexity
of VCU1 is high and in addition, most of the cells in it are burning, then it can go to the loadHigh state.
The self-optimization policy for this VCU then selects the defined individual cell action to migrate cells.
This causes the VCU to transition to state paused. The next action is to select cells for migration. Again, the
self-optimizing policy uses its criteria to guide the selection of cells for migration. For example, it might
pick cells with the objective of minimal increase in communication overhead after the migration. At this
point the VCU goes to state reconfigure. The corresponding action is to remove couplings for the selected
cells. After this action is applied on the VCU, it goes back to state ready when it is ready to run again
with the load reduced. Its execution is then resumed when it changes state to the loadMid state and starts
execution.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 168 — #12

 The Foundations of Autonomic Computing
CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 169 — #13

CCact: migrate cells CCact: remove couplings

CCstate: loadHigh CCstate: paused CCstate: reconfigure CCstate: ready CCstate: loadMid

VCU2
CCact: select cells for migration CCact: resume

VCU1 before repartitioning

VCU1 after
repartitioning

FIGURE 11.14 Sets of state transitions and actions guided by self-optimizing policies bring a component from heavily loaded to lightly loaded state.

11-169
© 2006 by Taylor & Francis Group, LLC
 11-170
Input interfaces Output interfaces Input interfaces Output interfaces Input interfaces Output interfaces

Autonomic manager Autonomic manager Autonomic manager

Analyze Plan Analyze Plan Analyze Plan
CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 170 — #14

Knowledge Knowledge Knowledge

Self-healing Self-healing Self-healing
Self-configuring Self-configuring Self-configuring

Monitor Self-optimizing Execute Monitor Self-optimizing Execute Monitor Self-optimizing Execute

S E S E S E
Forest fire cell burning Forest fire cell burning Forest fire cell burning
1. Data 1. Data 1. Data
2. Operational rules 2. Operational rules 2. Operational rules
3. Neighbor information 3. Neighbor information 3. Neighbor information
4. Local computation 4. Local computation 4. Local computation

Handbook of Bioinspired Algorithms and Applications

Input interfaces
Input interfaces Output interfaces
X Input interfaces Output interfaces
X Output interfaces

Autonomic manager Autonomic manager Autonomic manager

Analyze Plan Analyze Plan Analyze Plan

Knowledge Knowledge Knowledge

Self-healing Self-healing Self-healing

Self-configuring Self-configuring Self-configuring

Monitor Self-optimizing Execute Monitor Self-optimizing Execute Monitor Self-optimizing Execute

S E S E S E
Forest fire cell burning Forest fire cell burning Forest fire cell burning
1. Data 1. Data 1. Data
2. Operational rules 2. Operational rules 2. Operational rules
3. Neighbor information 3. Neighbor information 3. Neighbor information
4. Local computation 4. Local computation 4. Local computation

FIGURE 11.15 Forest cell space dynamically reconfigures itself.

© 2006 by Taylor & Francis Group, LLC

The Foundations of Autonomic Computing 11-171

11.5.3 Dynamic Composition and Self-Management of Forest Fire

Application
As discussed previously, we might encounter scenarios where the cells need to be dynamically deleted
from or added to the cell space. In this section, we will discuss how our framework achieves this objective.
As shown in Figure 11.15, two cells have burned and the other cells are burning. Hence, the couplings of
the burned cells are to be dynamically removed from this group of cells at runtime (denoted by the “cross”
in Figure 11.15). The monitoring engine senses this state change from burning to burned. It notifies the
analysis engine and planning engine. The planning engine then looks up the appropriate action for this
state by using the information stored in the self-configuration policy. In this case, the action is to remove
the couplings of the burned cells. Now, the execution engine deletes the specific ports and then sets the
new state of the VCU as ready and the VCU resumes execution.

11.6 The Autonomic Computing Landscape

There have been a number of research efforts in both academia and industry addressing autonomic
computing concepts and investigating the issues outlined earlier. Existing projects and products can
be broadly classified as: (1) Systems that incorporate mechanisms to address autonomic properties for
problem determination, autonomic monitoring, complex analysis, policies for autonomic managers, and
transaction measurements. (2) Systems that investigate models, programming paradigms and develop-
ment environments to support the development of autonomic applications and systems. Systems in the
former category are summarized in Table 11.1, while systems in the latter category are summarized in
Table 11.2.

TABLE 11.1 Systems Incorporating Autonomic Properties

System Application area Addressed autonomic issues

addressed/Key issues

OceanStore [7,8] Global, consistent, highly-available Self-healing, self-optimization,

persistent data storage self-configuration, self-protection
Policy-based caching, routing substrate
adaptation, autonomic replication,
continuous monitoring, testing,
and repairing
Storage Tank [9] Multi-platform, universally Self-optimization, self-healing
accessible storage management Policy-based storage and data management,
server redirection and
log-based recovery
Océano [10] Cost effective scalable management of computing Self-optimization, self-awareness
resources for software farms Autonomic demands distribution,
constant component monitoring
SMART DB2 [11] Reduction of human intervention & cost for DB2 Self-optimization, self-configuration
Autonomic index determination,
disaster recovery, continuous monitoring
of DB2’s health and alerting the DBA
AutoAdmin [12] Reduction of total cost of ownership (TCO) Self-tuning, self-administration
Usage tracking, index tuning and
recommending based on workload
Sabio [13] Autonomic classification of documents Self-organization, self-awareness
Group documents according to the
word and phrase usage
Q-Fabric [14] System support for continuous online management Self-organization
Continuous online quality management
through “customizability” of each
application’s QoS

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 171 — #15

11-172 Handbook of Bioinspired Algorithms and Applications

TABLE 11.2 Systems Support Development of Autonomic Applications and Systems

System Focus Autonomic issues addressed

KX (Kinesthetics eXtreme) [15] Retrofitting automicity Enabling autonomic properties

in legacy systems
Anthill [16] P2P systems based on Ant colonies Complex adaptive behavior
of P2P systems
Astrolabe [17] Distributed information management Self-configuration, monitoring
and to control adaptation
Gryphon [18] Publish/subscribe middleware Large communication
Smart Grid [19] Autonomic principles applied Autonomic Grid computing
to solve Grid problems
Autonomia [20] Model and infrastructure Autonomic applications
for enabling autonomic
applications
AutoMate [21] Execution environment for autonomic applications Autonomic applications

Two projects, AutoMate [21] and Autonomia [20], belonging to the second category, directly investig-
ate the key issues of autonomic component/service definition and construction, autonomic application
construction, execution and management, and autonomic middleware services. These systems are briefly
described below.

11.6.1 AutoMate — Enabling Autonomic Applications

Project AutoMate (TASSL, Rutgers University) [21] investigates autonomic strategies to deal with the
challenges of complexity, dynamism, heterogeneity, and uncertainty in Grid computing systems, and
enables systems and applications that are capable of managing (i.e., configuring, adapting, optimizing,
protecting, healing) themselves. The overall goal of Project AutoMate is to investigate conceptual models
and implementation architectures that can enable the development and execution of such self-managing
Grid applications. Specifically, it investigates programming models, frameworks, and middleware services
that support the definition of autonomic elements, the development of autonomic applications as the
dynamic and opportunistic composition of these autonomic elements, and the policy, content, and
context driven definition, execution, and management of these applications.
A schematic overview of AutoMate is presented in Figure 11.16. AutoMate builds on emerging
Grid/P2P middleware services to define and manage virtual organizations. Its key components include
the Accord [22] programming system, the Rudder [23] decentralized coordination framework and
agent-based deductive engine, and the Meteor [24] content-based middleware providing support for
content-based routing, discovery and associative messaging. Project AutoMate additionally includes the
Sesame context-based access control infrastructure, the DAIS cooperative-protection services and the
Discover collaboratory [25] services for collaborative monitoring, interaction, and control.
The core components of AutoMate have been prototyped and are being currently used to enable
self-managing applications in science and engineering (e.g., autonomic oil reservoir optimizations,
autonomic runtime management of adaptive simulations, etc.) and to enable sensor-based perva-
sive applications. Further information about AutoMate and its components can be obtained from
http://automate.rutgers.edu/.

11.6.2 Autonomia
Autonomia (University of Arizona) provides application developers with the tools required to specify the
appropriate control and management schemes, the services to deploy and configure required software
and hardware resources, and to run applications. Autonomia can efficiently support the development
of pervasive systems and services, and provides an environment to make the control and management

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 172 — #16

The Foundations of Autonomic Computing 11-173

Autonomic grid applications

programming
framework
Accord
Programming system

Sesame/DAIS protection service

autonomic components, dynamic composition,
opportunistic interactions, collaborative monitoring/
control

Ontology, taxonomy

coordination
middleware
Decentralized coordination engine

Rudder
agent framework,
decentralized reactive tuple space

Semantic middleware services

content-based
Meteor/squid

middleware

content-based discovery, associative messaging

Content overlay
content-based routing engine,
self-organizing overlay

FIGURE 11.16 AutoMate architecture.

User’s Application
application management
editor

AMS

Autonomous run-time engines

ASIK repository
Self- Self- Self- Self- Policy Security
Event deploying protecting healing optimizing repository repository
server
Policy Engine Resource Component
repository repository

Application autonomic manger (AAM)

Analyze Plan
Repository interface
Knowledge
Monitor Execute

Managed elements
Host 1 Host 2 Host N
Interface Interface Interface
Component 1 Component 2 Component N
Mobile agent 1 Mobile agent 2 Mobile agent N

FIGURE 11.17 Autonomia architecture.

of large-scale parallel and distributed applications autonomic. Autonomia provides online monitoring
and management to maintain the desired autonomic attributes of applications as well as system services,
achieving self-deployment, self-configuration, self-optimization, self-healing, and self-protection by the
policy engines.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 173 — #17

11-174 Handbook of Bioinspired Algorithms and Applications

The main modules of Autonomia include Application Management Editor (AME), Autonomic
Middleware Services (AMS) and Application Autonomic Manager (AAM). The AME is a graphical user
interface for developing an application using pre-developed and standard components and specifying the
management requirements for each application component. The AMS provides common middleware
services and tools needed by applications and systems to automate the required control and management
functions. The AAM mainly focuses on setting up the application execution environment. It acts as the
application administrator that is responsible for allocating the appropriate resources to run the application
and maintaining the application requirements at runtime.

11.7 Summary
In this chapter, we presented the autonomic computing paradigm, which is inspired by biological systems
such as the autonomic human nervous system and which enables the development of self-managing
computing systems. These systems use autonomic strategies and algorithms to handle complexity and
uncertainties with minimum human interference, thus shifting the burden of managing systems from
people to technologies. An autonomic computing system is a collection of autonomic components, which
implement an intelligent control loop to monitor, analyze, plan, and execute using knowledge of the
environment.
Several research efforts focused on enabling the autonomic properties address four main areas: (1) A self-
healing system could be expected to heal program parts that malfunction. (2) Self-protection of systems
to prevent large-scale correlated attacks or cascading failures from permanently damaging valuable infor-
mation and critical system functions. (3) Self-configuration that involves automatic incorporation of
new components and automatic component adjustment to new conditions. (4) Self-optimization on
a system level addressing automatic parameter tuning. Projects in both industry and academia have
addressed autonomic behaviors at all levels of system management, from the lowest levels of the hardware
to the highest levels of software systems and applications. At the hardware level, systems are dynamically
upgradable [26], while at the operating system level, active operating system code is replaced dynamically
[27]. Efforts have also focused on autonomic middleware, programming systems, and runtime [21, 28].
At the application level, self-optimizing databases and web servers dynamically reconfigure to adapt service
performance. The challenges to achieve true autonomic computing still exist, which will be accomplished
through a combination of process changes, skills evolution, new technologies and architecture, and open
industry standards.

References
[1] Autonomic Nervous System. http://users.rcn.com/jkimball.ma.ultranet/BiologyPages/P/PNS.
html#autonomic
[2] W. Ross Ashby. Design for a Brain, 2nd ed., Revised, Chapman & Hall Ltd, London, 1960.
[3] J.O. Kephart and D.M. Chess. The vision of autonomic computing. IEEE Computer, 36, 41–50,
2003.
[4] IBM Corporation. Autonomic computing concepts. http://www-3.ibm.com/autonomic/library.
shtml, 2001.
[5] IBM. An Architectural Blueprint for Autonomic Computing, April 2003.
[6] R. Rothermel. A mathematical model for predicting fire spread in wildland fuels. Research paper
INT-115. Ogden, UT: U.S. Department of Agriculture, Forest Service, Intermountain Forest and
Range Experiment Station, 1972.
[7] Oceanstore. http://oceanstore.cs.berkeley.edu, July 8, 2002.
[8] J. Kubiatowicz. OceanStore: Global-scale persistent storage. Stanford Seminar Series, Stanford
University, 2001.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 174 — #18

The Foundations of Autonomic Computing 11-175

[9] IBM Almaden Research. IBM storage tank — A distributed storage system WhitePaper. January 24,
2002.
[10] The Océano Project, http://www.research.ibm.com/oceanoproject, IBM Corporation.
[11] Guy M. Lohman and Sam Lightstone. SMART: Making DB2 (More) Autonomic. In Very Large
Data Bases (VLDB) Conference 2002.
[12] S. Chaudhuri. AutoAdmin: Self-tuning and self-administering databases. http://research.
microsoft.com/research/dmx/autoadmin, Microsoft Research Center.
[13] R. Pool. Natural selection, a new computer program classifies documents automatically, 2002.
[14] Q-fabirc. http://www.cc.gatech.edu/systems/projects/ELinux/qfabric.html.
[15] G. Kaiser, P. Gross, G. Kc, J. Parekh, and G. Valetto. An approach to autonomizing legacy systems.
In Workshop on Self-healing, Adaptive and Self-managed Systems, New York City, NY, June 23, 2002.
[16] Anthill. http://www.cs.unibo.it/projects/anthill/index.html.
[17] Robbert van Renesse, Kenneth Birman, and Werner Vogels. Astrolabe: A robust and scalable
technology for distributed system monitoring, management, and data mining. ACM Transactions
on Computer Systems, 21, 164–206, 2003.
[18] Gryphon. http://www.research.ibm.com/gryphon/gryphon.html
[19] Smart Grid, http://www.ldeo.columbia.edu/res/pi/4d4/testbeds/
[20] S. Hariri, Lizhi Xue, Huoping Chen, Ming Zhang, S. Pavuluri, and S. Rao. Autonomia:
an autonomic computing environment. In Proceedings of the Performance, Computing, and
Communications Conference, IEEE International, April 9–11, 2003.
[21] M. Agarwal, V. Bhat, H. Liu, et al. AutoMate: Enabling autonomic applications on the grid. In
Autonomic Computing Workshop Fifth Annual International Workshop on Active Middleware Services
(AMS’03), June 25–25, 2003.
[22] H. Liu, M. Parashar, and S. Hariri. A component-based programming framework for autonomic
applications. In Proceedings of 1st IEEE International Conference on Autonomic Computing
(ICAC-04), IEEE Computer Society Press, Washington, 2004, pp. 278–279.
[23] Z. Li and M. Parashar. Rudder: A rule-based multi-agent infrastructure for supporting autonomic
grid applications. In Proceedings of 1st IEEE International Conference on Autonomic Computing
(ICAC-04),” May 2004, pp. 10–17.
[24] N. Jiang, C. Schmidt, V. Matossian, and M. Parashar. Content-based Middleware for Decoupled Inter-
actions in Pervasive Environments, Rutgers University, Wireless Information Network Laboratory
(WINLAB), Piscataway, NJ, USA, 2004.
[25] V. Bhat and M. Parashar. Discover middleware substrate for integrating services on the grid.
In Proceedings of 10th International Conference on High Performance Computing (HiPC 2003),
Springer-Verlag, Heidelberg, December 2003, pp. 373–382.
[26] J. Jann, L.M. Browning, and R.S. Burgula. Dynamic reconfiguration: Basic building blocks for
autonomic computing on ibm pseries servers. IBM Systems Journal, 2003.
[27] J. Appavoo, K. Hui, C.A.N. Soules, R.W. Wisniewski, D.M. Da Silva, O. Krieger, M.A. Auslander,
D.J. Edelsohn, B. Gamsa, G.R. Ganger, P. McKenney, M. Ostrowski, B. Rosenburg, M. Stumm,
and J. Xenidis. Enabling autonomic behavior in systems software with hot swapping. IBM Systems
Journal, 2003.
[28] James Kaufman and Toby Lehman. Optimal grid: Grid middleware for high performance
computational biology. Research report, IBM Almaden Research Center.
[29] P. Horn. Autonomic Computing: IBM’s Perspective on the State of Information Technology,
http://www.research.ibm.com/autonomic/, October 2001.
[30] Adaptive Systems. http://www.cogs.susx.ac.uk/users/ezequiel/AS/lectures.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C011” — 2005/8/6 — 13:23 — page 175 — #19

 II
Application Domains

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 177 — #1

 Setting Parameter
12
Values for Parallel
Genetic Algorithms:
Scheduling Tasks on
a Cluster

12.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-179

12.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-181
12.3 The Task-Scheduling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 12-181
Definitions • The Computational Model • Complexity of
the Problem • The Optimal Algorithm • P-GOA Time and
Space Complexity
12.4 The Genetic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-183
GAs and Optimization • The Genetic Metaphor
12.5 Genetic Operators and Parameters . . . . . . . . . . . . . . . . . . . . . 12-184
Encoding and Initialization • The Sizing Equations and
Population Distribution • Fitness and the Objective
Function • Survival, Mating, and Mutation • Migration
Frequency, Policy, and Rate
12.6 Experiment Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-186
The Data Sets • The Test Programs • Sizing Equation
Variable Value Computation • Scaling Factor
12.7 Results and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-189
Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-191
Michelle Moore References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-191

12.1 Introduction
Optimization problems are widespread and appear frequently in a variety of common, everyday applica-
tions. For example, a shipping company handles containers of varying sizes and shapes, and wants to pack
the maximum possible number into a fixed space. This packing plan must be generated for each truck. An
airport wants to determine the fastest pattern for their fleet of snow plows and dump trucks for clearing

12-179

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 179 — #3

12-180 Handbook of Bioinspired Algorithms and Applications

snow from the runways. Since heavier snowfall will require more trips by the dump trucks, they also need
the plowing pattern generator to include the rate of snowfall in the computations. Finally, consider a
school district that needs to reduce the amount of fuel used by its 40 buses, and wants to determine the
shortest route that will allow children to be picked up at their homes. The district has a call-in system
for children who need not be picked up on a particular day, so the route plan has to be updated every
morning.
In each of these examples, the result must be determined quickly. Unfortunately, in order to determine
the very best answer in each case, all possible solutions must be considered. For example, in order to
determine the best bus route to pick up 50 children for one bus in a fixed area requires that 50! routes
be examined. That is 3.04e64 routes for a single bus. For 40 buses of 50 children each, there are 2000!
possible schedules! Because of the amount of time taken to find an optimal answer to these types of
computationally intractable problems, approximation algorithms are developed to find acceptably good
answers in a reasonable amount of time.
The decision to use an approximation algorithm raises additional questions. How close will a particular
approximation algorithm’s answer be to the optimal answer? Will the approximation algorithm be able
to find this answer fast enough? In all the examples given, there is a limit to the amount of time that can
be spent to find a solution. The program implementing the algorithm must guarantee a solution before
the time limit is reached.
Biologically inspired optimization methods such Genetic Optimization Algorithms (GOAs) offer the
advantage of an immediate approximate solution. The available time is then used to refine that solution
to bring it as close as possible to the optimal. Of course, in real applications, the actual optimal solution
is not known. However, where optimal solutions are known, GOA approximations frequently find them.
When the optimal is not found, the approximation is usually very close. GOAs can be written to execute in
parallel to speed up the process of finding an optimal or approximate solution. Communication of local
results among Parallel GOAs (P-GOAs) can often improve both the quality of the approximations and the
speed at which they are found.
The likelihood that a very good solution will be found when using a P-GOA depends largely on the values
of the parameters used by the program. These parameters include population size, deme size (the number
of individuals on each processor), number of generations, rate of communication, and communication
frequency. This chapter discusses a mathematical method to determine appropriate P-GOA values for
these parameters for a multiprocessor scheduler. The computations required to determine proper values
for the P-GOA parameters are shown in detail. In addition, known optimal schedules are compared with
schedules generated using the P-GOA and the results are shown.
The analysis and experimentation discussed here represents ongoing investigation. The results presented
here represent the culmination of an initial attempt to find an “all purpose” sizing formula for task
scheduling in a computational science programming environment. Using the work by Cantú-Paz [1] and
Goldberg et al. [2] as a theoretical and practical foundation, numerous computations were explored,
and trials were run using a variety of equations. An “all purpose” formula was not found; however, it is
now clear what can be accomplished using this sort of methodology, and what problem qualities suggest
a very different approach. The material in some of the sections of this chapter has appeared in previous
discussions of the progress in this area [3–6]. It has been repeated here so that this chapter may be read
without the need to refer to the earlier work.
The rest of this chapter is organized as follows. Section 12.2 outlines the background of the sample
problem and of previous work related to parameter value selection for parallel Genetic Algorithms (GAs).
Section 12.3 presents the problem-specific scheduling model and the complexity of the problem. This
material has appeared in previous publications. Section 12.4 introduces GAs and the application of the
genetic metaphor to the scheduling problem, which has appeared in previous publications. Section 12.5
describes the genetic operators and the parameter variables used in the scheduler. This section is similar
to discussions in Reference 6, but provides the critical modifications that produced a successful deme
sizing equation for the task scheduling P-GOA. Readers familiar with the previous reports might begin
their reading in Section 12.5. Section 12.6 describes the design of the experiments and describes how

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 180 — #4

Setting Parameter Values for Parallel Genetic Algorithms 12-181

particular values were chosen for the sizing equations. Section 12.7 details the results of the experiments
and summarizes the conclusions.

12.2 Background
A multiprocessor scheduling problem is used to illustrate appropriate calculations for determining
several important parameter values for a P-GOA solution approximation. The schedule to be developed
by the P-GOA specifies the processor on which each of the tasks required for a particular application
is to run. This application schedules tasks on a cluster of homogeneous processors with an Ethernet
connection. The task execution times and communication times are known in advance. This models
the scheduling requirements that may be encountered when performing distributed database queries or
when doing “production runs” in computational science applications with known computation times and
varying data. In applications such as these, execution and communication times can be known (or at least
estimated very closely) in advance.
Parameter sizing methods were reported by Goldberg [7], Goldberg et al. [2], and others. Precise analysis
of these methods for selected problems and processor topologies were presented by Cantú-Paz [1]. In his
forward [1], Goldberg states, “I believe it is fair to say that prior to this work, the design of parallel GAs
was something of an empirical black art, guided largely by ad hoc trial and error experimentation.” The
statistical analysis presented in these works had their basis in schema theory. Schema theory attempts
to provide an explanation for the behavior of genetic algorithms. Analysis of this explanation of GA
behavior resulted in precise formulae for determining, a priori, the values of GA input parameters that
would produce a particular solution quality.
A number of objections to schema theory have been published. A recent book by Reeves and Rowe [8]
summarizes many of these objections, and provides sufficient bibliographic information to begin a more
in-depth investigation into the reasoning of schema theory detractors. They suggest that schema theory-
based analysis can be fruitful for specific categories of GAs. However, they insist that the behavior of other
types of GAs differs vastly from that described by schema theory. Nevertheless, the benefits of calculating
accurate GA parameters in advance, in terms of time saved and solution quality, can be sizable for some
applications. In this chapter, effective parameters are found for a task-scheduling problem by applying
appropriately adapted analysis.
In Reference 3 these concepts were first applied to this scheduling problem. In References 4 to 6, an
in-depth examination of three deme sizing equations developed over time by Cantú-Paz, Goldberg, and
others was reported in terms of their applicability to the scheduling problem. Information in Reference 6
outlined how to apply the equations to the cluster-scheduling problem, specifically detailing the appro-
priate determination of values for the equation variables. This chapter explains refinements developed
through analysis of the most promising equations, and provides the rationale for making those refine-
ments. Justification of approximations used in parameter computations and modifications to the original
equations are explained.

12.3 The Task-Scheduling Problem

The task-scheduling problem presented here has been the sample problem during the various stages
of analysis of the parameter sizing computation development. This definition and description of the
scheduling problem has appeared, in slightly varying form in References 3 to 6.

12.3.1 Definitions
A task ti is represented as a taskpair (ei , ci ), consisting of an execution time ei and a communication
time ci . The total number of tasks to schedule is n. The symbol pi represents one of the m processors
available to the scheduler. The makespan of a schedule is the time at which the execution of all taskpairs

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 181 — #5

12-182 Handbook of Bioinspired Algorithms and Applications

is completed. The optimal makespan is the shortest possible period in which a given set of taskpairs can
execute on the available multiprocessor system. The goal of the scheduler is to produce schedules with
makespans as close to optimal as possible within a predictable and practical amount of time. In summary,
the set of tasks {t0 , . . . , tn−1 } is scheduled to execute on a system of m processors {p0 , . . . , pm−1 }.
Tasks are created and scheduled by an initial processor designated p0 . The time p0 uses to create
a schedule for the given set of tasks is not considered as a part of the makespan. The time required to
send the task assignments to the other processors is assumed constant and is therefore not considered.
Any final computations that must occur after all results are communicated back to p0 are unaffected
by the particular schedule, and are not included in the schedule quality evaluations. Finally, messages
and processes for control or monitoring are assumed to have no effect on the relative efficiencies of the
schedules, so these values are also not considered in the evaluation of schedule quality.
The time required for each task to execute and communicate a result (or estimations of these times)
is available in advance. Dependent computations are grouped into a single task so that each task is
independent. Consequently, there are no precedence constraints on the tasks. However, the computation
portion of each task must complete execution before its corresponding message is sent.

12.3.2 The Computational Model

The target hardware environment is a cluster system using message passing. The cluster consists of
m processors {p0 , . . . , pm−1 } connected by a shared Ethernet bus. All processors are identical and the
communication bandwidth between {p1 , . . . , pm−1 } and p0 is constant. The processor that creates the set
of tasks and the schedule is represented as p0 . This process requires no communication time in order to
complete its tasks. Processors {p1 , . . . , pm−1 } represent the additional processors to which p0 sends tasks or
control messages, and from which p0 receives the computation results. Tasks scheduled on {p1 , . . . , pm−1 }
are not considered complete until the result message has left the communication channel. After all n tasks
have been completed, p0 may use the results obtained to execute a final computation. However, since
this final execution time is the same regardless of the manner in which the tasks are scheduled, its value
is ignored during schedule quality evaluation.
Only one message may be transmitted to or from a processor and only one task may execute on a
processor at a time. All schedules and tasks are non-preemptive. If there is a task available to execute, and
a processor is available at that time, the task will be scheduled. If the communication channel is avail-
able, the message will be transmitted immediately after the associated task is completed. Otherwise, the
message will be transmitted as soon as messages associated with any previously executed tasks leave the
communication bus.

12.3.3 Complexity of the Problem

In References 8 to 10 and 17, GAs were shown to provide effective approximations for combinatorial
optimization problems. The combinatorial optimization problem of finding a schedule on m > 2 identical
processors that minimizes a finish time for independent tasks consisting of execution times has been shown
to be in class NP [11,12,19]. In addition to execution times, ei , the amount of time, ci , required to return
the result of each computation over the communication channel is also considered. Therefore, in order
to schedule n taskpairs {t0 , . . . , tn−1 }, the values of {(e0 , c0 ), . . . , (en−1 , cn−1 )} must be considered. This
problem can be transformed into the above scheduling problem by setting its communication times to
zero, and is clearly in NP. When communication times are nonzero, the communication channel is an
additional resource that must be scheduled, which increases problem complexity. The quality or “fitness”
of the schedule is the total time required for all tasks to complete execution. An exhaustive search of
all possible schedules will require time bounded by (m n ), where n is the number of tasks to schedule,
and m is the number of processors to schedule. The addition of a communication bus that must also be
“scheduled” increases the complexity of the problem. Since the problem is NP-complete, approximation
algorithms for scheduling attempt to create schedules as close to the optimal as possible.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 182 — #6

Setting Parameter Values for Parallel Genetic Algorithms 12-183

12.3.4 The Optimal Algorithm

In order to accurately evaluate the schedules produced by the P-GOA, optimal schedules were generated
for a series of small scheduling problems. It was necessary to keep the problems fairly small due to the
complexity of an optimal algorithm. Consider if m is the number of processors and n is the number of tasks,
an exhaustive search will require m n time to evaluate every schedule. If a branch and bound algorithm
is used to avoid unfruitful partial schedules, the worst case will occur if data requires all branches to be
searched, giving a worst-case complexity of O(m n ). A parallel version of the optimal algorithm executed
on r processors has complexity O(m n /r), again an exponential time complexity measure. Because of the
time required to obtain comparison values using an optimal algorithm, the size of the test cases was kept
small.

12.3.5 P-GOA Time and Space Complexity

The space requirements of the sequential version of the genetic schedule optimization algorithm are linear,

population_size × number_of _tasks × 2

and the execution time complexity of the sequential version of the genetic schedule optimization
algorithm is

num_generations × population_size × number_of _tasks × number_of _processors.

The scheduler is cost optimal in that the space complexity of the P-GOA is

(population_size × number_of _tasks × 2)/r

and the execution time complexity of the P-GOA is

(num_generations × population_size × number_of _tasks × number_of _processors)/r,

where r is the number of processors on which the scheduler is executed.

12.4 The Genetic Approach

12.4.1 GAs and Optimization
In 1975, Holland [13] introduced the idea of combining directed randomness with adaptation as a search
and optimization technique. Since that time, GAs have been used in a wide array of applications. In
addition, the characteristics of GAs have been studied extensively, and diverse theoretical explanations
for the ability of these complex systems to generate solutions have appeared [18]. GAs have been used to
approximate solutions for a wide variety of NP-complete problems [2,9].
A GA begins with a population of initial coded “guesses” (chromosomes) at a solution. These guesses
may be randomly generated or produced using a problem specific heuristic. An application specific
fitness or objective function is applied to each chromosome, and the “better” individuals are selected to
“survive.” Survivor individuals are then merged together to form a new generation (mating). Occasionally,
portions of the chromosomes of an individual are randomly altered (mutation). This process of fitness
determination, mating, and mutation is repeated for a given number of generations, or until the individuals
improve sufficiently to achieve a predefined goal. The simplest algorithms often view an individual as
a single chromosome. However, applications with multiple constraints may employ individuals with
multiple chromosomes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 183 — #7

12-184 Handbook of Bioinspired Algorithms and Applications

12.4.2 The Genetic Metaphor

The P-GOA employs methods and metaphors from nature and genetics to “evolve” an initial population
of solutions (schedules in this case) into high quality solutions. The initial population of schedules is
randomly generated. The fitness, or quality, of each schedule is evaluated. The “fitter” schedules are
saved and used as a “mating pool.” Finally, these schedules are paired randomly to create new schedules
until there are enough “offspring” to return the population to its original size. The P-GOA divides the
population of schedules into local demes and distributes them over the available processors. Periodically,
individual schedules are allowed to migrate to other processors in hopes of speeding the rate of schedule
improvement.
The P-GOA represents each schedule as a single chromosome. Each individual schedule consists of
lval genes, where lval is the number of tasks to be scheduled. Each position of the chromosome corresponds
to a task, and the values at each position correspond to a processor number. A schedule for six tasks on
five processors might appear as (1 3 4 2 3 0), indicating that t0 is executed on p1 , t1 is executed on p3 , and
so on. Chromosomes are also referred to as strings. The cardinality of the string alphabet is the number
of digit values that may appear on the chromosome. Many GAs use binary alphabets, but in general, we
may say that a chromosome string has a χ-ary alphabet where the cardinality is χ. In the given example,
the cardinality of the alphabet (χ ) is equal to five.

12.5 Genetic Operators and Parameters

12.5.1 Encoding and Initialization
For scheduling in the P-GOA, each chromosome encodes a schedule solution and each gene represents
a scheduled task. The number of genes in a chromosome equals n, which is the number of tasks to be
scheduled. In the P-GOA, allele values represent the processor to which a task is assigned. Allele values
range from 0 to m − 1, where m equals to the number of processors available. For example, given the
following taskpair set:

{(7, 16), (11, 22), (12, 40), (15, 22), (17, 23),
(17, 23), (19, 23), (20, 28), (20, 27), (26, 27),
(28, 31), (36, 37), (31, 29), (28, 22), (23, 19),
(22, 18), (22, 17), (29, 16), (27, 16), (35, 15)},

the optimal schedule on three processors would appear as

1 0 0 0 0 0 0 0 2 0 0 0 1 1 1 1 1 2 1 2.

The optimal makespan for this sequence of tasks is 202.

Various population “seeding” techniques were examined in Reference 14. However, for general
GA schedulers, random initialization of the population of schedules has proven to be the most effective
initialization technique and is used here.

12.5.2 The Sizing Equations and Population Distribution

As mentioned previously, the determination of population and deme sizes for practical applications of
parallel GAs required a great deal of what was frequently termed “empiricism,” that is, trial and error.
Cantú-Paz [1] offered a way to analyze a particular problem so that a very precise population and deme
size, an efficient migration strategy, and an effective processor topology could be determined in advance.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 184 — #8

Setting Parameter Values for Parallel Genetic Algorithms 12-185

The analysis here considers only deme sizing, sets migration to the maximum possible value, and uses
a fully connected system.
Consideration of collateral noise (the noise of other partitions when deciding between best and second
best building block) is built into the sizing equation. External noise may exist if fitness cannot be directly
calculated. This was not the case for the scheduling application, so the sizing computation does not require
adjustments for external noise. Each of the processors used in finding the schedule is assigned nd schedules
to evolve locally.
Analysis produced a sizing equation that is identical to the one given in Reference 1. However, many of
the variables could not be calculated in the same manner as in Reference 1 and required approximations.
In addition, experiments revealed that a scaling factor was needed to compensate for a crucial difference
in this application. Unlike the applications in Reference 1, the scheduler does not need to converge. It is
only necessary to find one schedule of the required quality. The scaling factor will be discussed in detail
in Section 12.6.4. The equation, minus scaling, is given below:

nd = sqrt(−χ k ln(1 − Pbb ))

, (12.1)
(2p − 1)/p

where

χ = cardinality of the chromosome string alphabet, the number of processors

k = building block size, the number of fixed positions in the schema
Pbb = probability that one partition (one building block) in one deme is correct, probability of success
per deme ≈ Q/m − sqrt(ln(r))/sqrt(2m)
Q = required problem solution quality = (1 − α)m
α = the probability of making the wrong choice between two competing schema
m = number of partitions in the string, number of building blocks lval/k
lval = length of the chromosome (schema), the number of tasks
r = the number of processors used to create the schedule
p = (1 − q), ≈1/2 + ψ/sqrt(2π )
π = the constant value 3.14159
ψ = d/(σbb sqrt(2m ))
d = signal difference for peak to peak (mean to mean) function difference between the best and the
second best competing individuals, this affects the likelihood of choosing the better individual
σbb = sqrt(σf2 ) the average building block standard deviation


σf2 = overall variance of all schemas for each fixed position = m i=1 σfi
2

( fmax − fmin )2
≈ , (12.2)
12

fmax = maximum fitness function value possible

fmin = minimum fitness function value possible

12.5.3 Fitness and the Objective Function

Tasks assigned to processors other than p0 , will have an associated message that must be scheduled. The
total message time will therefore be longer than the task execution time on p1,...,n−1 . Tasks assigned to p0
do not require message time, so the greater of the communication completion time and the execution
time on p0 , will be the makespan

makespan = MAX( p0 , message time).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 185 — #9

12-186 Handbook of Bioinspired Algorithms and Applications

Because the goal of the scheduler is to minimize the makespan, the fitness of an individual chromosome
string is evaluated as

fitness = −1 × makespan.

12.5.4 Survival, Mating, and Mutation

Each generation, the average of all the individual makespans is calculated. Schedules with makespans
less than or equal to the average are placed in the survivor pool. Each individual schedule with a fitness
greater than or equal to the average fitness is allowed to reproduce and survive to the next generation.
This rank-based selection method combines truncation selection and (γ + λ) selection. The top 1/ν of
the population is selected. Then the γ = 1/ν parents mate to create λ offspring. The union of γ and λ
constitute the next generation. This allows very fit individuals to survive and mate repeatedly.
Survivors are randomly chosen, with replacement, to produce offspring until the size of the population
returns to its original level. Each mating produces a single offspring. As with sexual reproduction, an
offspring is created from the genes of the two parents. Each allele is randomly chosen from one parent
or the other. Mating is accomplished using uniform crossover with a 0.5 crossover probability [15,16].
In other words, when creating an offspring from a pair of schedules, the offspring’s processor assignment
for each task is equally likely to match either parent.
In addition, a 2.5% mutation probability is applied for each allele. This means that after crossover, the
GA allows for a .025 probability for each task being reassigned to a randomly chosen processor. This is
a relatively high mutation rate, but was chosen to prevent premature convergence from occurring due to
small deme sizes. The process of selection, mating, crossover, and mutation occurred for 1000 generations.
This is a relatively small number of generations. The goal of the scheduler is to find near optimal schedules
in a very short amount of time. A small number of generations, combined with a deme size no larger than
necessary, allows this to be accomplished.

12.5.5 Migration Frequency, Policy, and Rate

The initial population of schedules is distributed among the available cluster nodes. After each generation,
the more fit individuals of each deme are copied and distributed among the other demes, replacing the
least fit individuals. The P-GOA uses the migration rate necessary to fill the positions left open by the unfit
individuals. Migration occurs after selection and before mating so that the new migrants can contribute to
the next generation. After each migration, the local iterations of selection, mating, crossover, and mutation
resume. When the execution of 1000 generations completes, the P-GOA outputs the best schedule found
so far, along with its makespan.
Using the results from the migration policy analysis and experiments in Reference 1, migration
follows what is called a “best–worst policy,” since migrants take the place of the least fit individu-
als in each deme. The proportion of migrants to the population as a whole is the migration rate.
Since migrants are sent to replace individuals below average, the rate of migration decreases as the
fitness values begin to converge. Since the individual schedules remain on their original processors and
copies are sent to other processors, this process is sometimes referred to as “pollination” rather than
migration.

12.6 Experiment Design

The experiments performed by Goldberg et al. [2,7] and Cantú-Paz [1] to verify the given calculations and
guidelines, involved problems qualitatively different from the scheduling problem described here. Their
experiments involved binary encodings only. The schedule encoding is m-ary, where m is the number of
processors. For the scheduling problem, this is the same as χ-ary where χ is the alphabet size. The fitness
of their experimental individuals involved functional calculations that directly evaluated the encodings.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 186 — #10

Setting Parameter Values for Parallel Genetic Algorithms 12-187

In the scheduling application, fitness is an indirect computation requiring an evaluation of the meaning
and implications (i.e., effect on communication time) of the encoding. When they compared expected
confidence levels with experimental results, the degree of correctness was defined as the percentage of
alleles possessing the correct value when the algorithm converged. In the scheduling application, the degree
of correctness is defined as the “nearness” to the optimal schedule that can be obtained in a limited amount
of time by the best individual. These differences were overcome by problem specific interpretations of
the meanings of the equation variables and by employing a scaling factor to adjust the output of the
equation.

12.6.1 The Data Sets

The initial data set consisted of taskpair sets, each containing 20 normally distributed, randomly generated
(e, c) task pairs. The mean execution time was 25 with a standard deviation of 2%. Three communication
means were used, 10, 25, and 40, (std = 0.02) to vary the relative effects of communication overhead.
This data set was used throughout the experiments with various sizing equations. The final deme sizing
equation reported here was also tested using the original data sets as well as data sets of exponentially
distributed task execution and mean times, taskpair sets of varying sizes, and data sets with larger or
smaller means and standard deviations.

12.6.2 The Test Programs

The parallel scheduling program was written in C using LAM MPI constructs to execute on multiple
cluster processors. The migration rate was set to the maximum that could occur for the deme with the
most survivors after selection. Migration occurred at every generation.
In order to evaluate the quality of the schedules that were produced, actual optimal schedules were
found for the data sets. Because of the exponentially increasing amount of time required to find the
comparison optimal schedules, the number of tasks in the tests sets was kept small. A sizing program was
written to calculate population sizes that corresponded to varying predicted levels of schedule accuracy.
The sizing program generated population and deme sizes using probabilities of success of 0.80, 0.85, 0.90,
0.95, and 0.99. Deme sizes were generated at each confidence level with each of the three communication
means. To reduce the effect of stochastic variance, five runs were averaged for each communication mean
at each confidence level. The program to find the optimal schedule for each of these cases was also executed
five times, and the results were averaged for comparison. An expected quality of 80% meant that when
deme size was calculated at the 80% confidence level. They were expected to produce schedules that
would allow all tasks to complete execution in an amount of time exceeding the optimal by no more
than 20%.

12.6.3 Sizing Equation Variable Value Computation

The graphs shown represent results when the following variable values were set to the given values:
r = 12, the number of processors used to create the schedule, the number of cluster nodes running the
P-GOA scheduler.
χ = 5, the number of processors to be scheduled, the cardinality of the alphabet, the number of
possible allele values.
k = 1, the size of a building block.
lval = 15, the number of tasks to schedule, the number of genes in the chromosome.
m = the number of partitions, the number of building block sets in the chromosome string, since the
building block size is 1, lval/k = 15.
m = m − 1.
Signal difference is defined in Reference 1 as the difference in the means of the fitness distri-
bution produced by the best building block and that produced by the second best building block.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 187 — #11

12-188 Handbook of Bioinspired Algorithms and Applications

The fitness distribution associated with a particular building block in a task schedule cannot be
directly determined from the value assigned to the building block. Fitness calculations must consider
the meaning of the allele value in terms of previous assignments and the effect on communication time.
Consequently, the P-GOA scheduler considers the smallest possible difference in average makespans
and the average execution time. The average of these two values is used as an estimate for signal
difference.

d ≈ (1 + µe )/2.

This value will vary in relationship with the means of the data set values. Data sets with larger execution
times will have larger signal difference values.
The worst possible schedule for a given data set could be generated if tasks were scheduled to
maximize the makespan. This worst fitness for the specific scheduling problems is calculated as
follows:

fmax ≈ if µc < µe lval µe + µc which in this example is 15 × 25 + 10 = 385,

if µc = µe lval µe + µc which in this example is 15 × 25 + 25 = 400,

if µc > µe lval µc + µe which in this example is 15 × 40 + 25 = 625.

The minimum makespans were estimated using observations of task placements found by the optimal
algorithm

1
fmin ≈ if µc
µe (1/num_procs) lval µe + 2µc × 15 × 25 + 10 = 135,
3
1
if µc > µe (1/num_procs) lval µc + µe × 15 × 40 + 25 = 225.
3

Next, the probability of retaining the best building block between generations on at least one deme is

1
p≈ + ψ/sqrt(2π ) where ψ = d/(σbb sqrt(2m )),
2
1
≈ + (13/(σbb sqrt(28))/sqrt(2π )),
2
≈ 0.5 + (13/σbb × 5.292)/0.251,

≈ 0.5 + 2.457/(σbb × 0.251),

≈ 0.5 + 9.789/σbb ,

where

σbb = sqrt(σf2 )

and

σf2 ≈ ( fmax − fmin )2 /12,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 188 — #12

Setting Parameter Values for Parallel Genetic Algorithms 12-189

then

if µc < µe sqrt(σf2 ) ≈ sqrt((385 − 135)2 /12) = 72.17

p ≈ 0.5 + 9.789/72.17 = 0.636,
if µc = µe sqrt(σf2 ) ≈ sqrt((400 − 135)2 /12) = 76.50
p ≈ 0.5 + 9.789/76.50 = 0.628,
if µc > µe sqrt(σf2 ) ≈ sqrt((625 − 225)2 /12) = 115.47
p ≈ 0.5 + 9.789/115.47 = 0.585.

12.6.4 Scaling Factor

Because the P-GOA does not need to converge on a schedule of the required quality, merely find at least
one, the population size can be scaled down. Previous calculations produced population sizes that met
or exceeded predicted quality at the 95 to 99% confidence levels. This was the primary goal; however,
it was discouraging that the deme sizes produced for the lower confidence levels produced schedules
whose quality far exceeded the predicted quality. This “excess” quality was not a problem in practical
terms; however, it indicated that the sizing equations were not capturing an important feature in the
behavior of the scheduler. It is well known that fitness in GAs tends to improve at a logarithmic rate. Early
improvements come quickly, but later ones appear more slowly. Since the deme sizes produced accurate
results at the higher range, a scaling factor was needed to make the deme sizes reflect the rate of increase
at the lower quality levels. Deme sizes were scaled so that they used the calculated size for the highest
confidence level. These sizes were decreased at an exponential rate until they reached the lowest confidence
level.

12.7 Results and Conclusions

The sizing equation produced very accurate deme sizes for the scheduling problem where the task sizes
were normally distributed. Table 12.1 shows the accuracy levels for each population size for averaged
results with all communication means. Figure 12.1 illustrates how closely the actual makespans generated
track the makespans predicted by the equation.
Tests run with larger data sets, varying data means, and varying standard deviations gave similar results.
The deme sizing equation with the scaling factor applied provides sizes that allow the P-GOA scheduler
to produce schedules with makespans very close to a predetermined accuracy. In addition, the very small
populations at the lower confidence levels allows for trade-offs between accuracy and speed. Clearly, this
deme sizing equation captures the behavior of the P-GOA for this scheduling application.
In fact, Tables 12.2 to 12.4 show the best approximate makespan found for each of the five tests that
were run for each communication mean. These approximate makespans are shown with the optimal
makespans for each run and with the generation number in which the best approximate makespan was

TABLE 12.1 Accuracy Levels for Each

Population Size — Normal Data

Deme size Actual Predicted

7 0.805 0.80
10 0.878 0.85
14 0.950 0.90
24 0.975 0.95
50 0.997 0.99

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 189 — #13

12-190 Handbook of Bioinspired Algorithms and Applications

All means — normal

1.1

Nearness to optimal
0.9

0.8
actual
0.7 predicted

0.6
7 10 14 24 50
Deme size

FIGURE 12.1 Average results, all communication means — normal data.

TABLE 12.2

Optimal Normal 10 Generations

115 115 78
117 117 404
118 119 462
118 118 40
114 114 620

TABLE 12.3

Optimal Normal 25 Generations

194 195 112

193 195 124
195 197 630
194 194 609
191 191 40

TABLE 12.4

Optimal Normal 40 Generations

230 231 800

236 237 366
242 242 18
244 245 126
237 237 13

found. The tables illustrate that the P-GOA found the optimal makespan 53.33% of the time. When the
optimal was not found, the approximation was very close.
However, experiments with exponentially distributed data yielded very disappointing results.
Makespans in all experiments were far from optimal. The distribution of the data is undeniably a strong
factor in the applicability of the sizing equations. This is a serious limitation, but it should be kept in
perspective. The results of the sizing equation were disappointing, but the scheduler was able to pro-
duce schedules that were <3% away from the predicted quality at the high range (99%). In practical
terms, this is quite useful and the scheduler is ready for experimental incorporation into a cluster system.
Table 12.5 provides the average quality measures compared with the predicted quality at each population
size calculated. Figure 12.2 illustrates the disappointing performance.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 190 — #14

Setting Parameter Values for Parallel Genetic Algorithms 12-191

TABLE 12.5 Accuracy Levels for Each

Population Size — Exponential Data

Deme size Actual Predicted

7 0.436 0.80
10 0.541 0.85
14 0.668 0.90
24 0.842 0.95
50 0.962 0.99

Exponential — all means

1.1
1
Nearness to optimal

0.9
0.8
0.7
0.6
actual
0.5 predicted
0.4
7 10 14 24 50
Deme size

FIGURE 12.2 Average results, all communication means — exponential data.

A great deal of analytical work is needed before an “all purpose” parameter sizing methodology can be
found for P-GOAs, but it is encouraging that a “special purpose” parameter sizing methodology has been
proven for at least one complex optimization problem. The only data specific variables needed by the
P-GOA scheduler are the means of the task execution times and communication times. For the type of
ongoing computational science application that this scheduler was designed to work with, this information
is either available or easily estimated.

Acknowledgments
Research Assistants John Picarazzi, Shelia Poorman, Brian McCord, Tzintzuni Garcia, William Jackson,
Simon San Miguel, Jason Picarazzi, and Lucas Wilson have contributed or are currently contributing to
this project. This work is supported by NASA Grant NAG9-1401 and NSF Grant NSF 01-171.

References
[1] Cantú-Paz, E. Efficient and Accurate Parallel Genetic Algorithms, Kluwer Academic Publishers,
Dordrecht, 2000.
[2] Goldberg, D., Deb, K., and Clark, J. Genetic algorithms, noise, and the sizing of populations.
Complex Systems, 6, 1992, 332–362.
[3] Moore, M. Parallel genetic algorithms to find near optimal schedules for tasks on multiprocessor
architectures. In Proceedings of Communicating Process Architectures, Bristol, UK, September 2001,
pp. 27–36.
[4] Moore, M. An accurate and efficient parallel genetic algorithm to schedule tasks on a cluster. In
Proceedings of the International Parallel and Distributing Processing Symposium, Nature Inspired
Distributed Computing Workshop, Nice, France, April 2003.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 191 — #15

12-192 Handbook of Bioinspired Algorithms and Applications

[5] Moore, M. Accurate calculation of deme sizes for a parallel genetic scheduling algorithm.
In Proceedings of Communicating Process Architectures, Enschede, NL, September 2003,
pp. 305–314.
[6] Moore, M. An accurate parallel genetic algorithm to schedule tasks on a cluster. Parallel Computing,
30(5–6), 2004, 567–583.
[7] Goldberg, D. Sizing populations for serial and parallel genetic algorithms. In Proceedings of the
Third International Conference on Genetic Algorithms, Fairfax, VA, USA, June 1989, pp. 70–79.
[8] Reeves, C. and Rowe, K., Genetic Algorihtms: Principles and Perspectives, Kluwer Academic
Publishers Group, Boston, MA, 2003.
[9] De Jong, K. and Spears, W. Using genetic algorithms to solve NP-complete problems. In Pro-
ceedings of the Third International Conference on Genetic Algorithms, Fairfax, VA, USA, June 1989,
pp. 124–132.
[10] Hou, E., Hong, R., and Ansari, N. Efficient multiprocessor scheduling based on genetic algorithms.
In Proceedings of the 16th Annual Conference of the IEEE Industrial Electronics Society, Asilomar,
CA, USA, November 1990, pp. 1239–1243.
[11] Coffman, E. Introduction to Deterministic Scheduling Theory, Computer and Job-Shop Scheduling
Theory. John Wiley & Sons, New York, 1976.
[12] Horowitz, E. and Sahni, S. Exact and approximate algorithms for scheduling non-identical
processors. Journal of the ACM, 23, 1976, 317–327.
[13] Holland, J. Adaptation in Natural and Artificial Systems. University of Michigan Press, Ann Arbor,
MI, 1975.
[14] Kidwell (Moore), M. Using genetic algorithms to schedule distributed tasks on a bus-based system.
In Proceedings of the Fifth International Conference on Genetic Algorithms, Urbana-Champaign, IL,
USA, July 1993, pp. 368–374.
[15] Eshelman, L., Caruana, R., and Schaffer, J. Biases in the crossover landscape. In Proceedings of the
Third International Conference on Genetic Algorithms, Fairfax, VA, USA, June 1989, pp. 10–19.
[16] Syswerda, G. Uniform crossover in genetic algorithms. In Proceedings of the Third International
Conference on Genetic Algorithms, 1989, pp. 2–9.
[17] Goldberg, D. Genetic Algorithms Search, Optimization, and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
[18] Mitchell, M. An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA, 1996.
[19] Salleh, S. and Zomaya, A. Scheduling in Parallel Computing Systems. Kluwer Academic Publishers
Group, Dordrecht, The Netherlands, 2000.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C012” — 2005/7/20 — 19:53 — page 192 — #16

 13
Genetic Algorithms
for Scheduling in
Grid Computing
Environments:
A Case Study

13.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-194

13.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-194
List Scheduling Heuristics • GridSim • Task Graphs for Free
• Genetic Algorithms
13.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-197
A Comparison Study of Static Mapping Heuristics •
A Dynamic Matching and Scheduling Algorithm for
HC Systems
13.4 Development Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-197
DAG Support • The TaskBroker • Mapping Heuristic
Extensions • GA Mapping Heuristic
13.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-201
Methodology • Comparisons of the Objective Functions
of GA • GA Onset Trials • Large-Scale DAGs •
Fast Greedy Variants
13.6 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-204
Network Latency • Multiple User Environments •
Randomized Heterogeneous Processor Environments •
Inconsistent Execution
13.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-205
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-205
Kris Crnomarkovic Appendix A. Test Processor Configuration . . . . . . . . . . . . . . . . . . . . 13-206
Albert Y. Zomaya Appendix B. Sample Task Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-207

13-193

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 193 — #1

13-194 Handbook of Bioinspired Algorithms and Applications

13.1 Introduction
The proliferation of the Internet and the availability of powerful computers and high-speed networks as
low-cost commodity components are changing the way computing is done today. The interest in cou-
pling geographically distributed computational resources is also growing for solving large-scale problems,
leading to what is popularly called Grid Computing. Grids enable the sharing, selection, and aggregation
of suitable computational and data resources for solving large-scale data intensive problems in science,
engineering, and commerce [1–3].
An important issue for Grid and other Heterogeneous Computing environments is how to assign
tasks to resources and order execution of the tasks to maximize some performance criterion of the Grid
environment. These procedures are termed matching and scheduling, and taken together, are known as
mapping. There are two different types of mapping: static and dynamic. Static mapping is performed
when the applications are mapped in an offline planning phase, for example, planning the schedule for a
set of production jobs. Dynamic mapping is performed when the applications are mapped in an online
fashion, for example, when tasks arrive at unknown intervals and are mapped as they arrive (the workload
is not known a priori) [4,5]. In both cases, this generalized mapping problem has been shown to be
NP-hard (e.g., in References 6 to 8).
The goal of this study has been to investigate classes of scheduling algorithms for service-based Grid
Environments, that is, where a known request is served to a user or users. Such an environment may
involve request scheduling, which is a task scheduling architecture that uses a service request as the minimal
scheduling unit. A service request is considered to be finer-grained than a job request. A single server would
be able to handle multiple requests [9].
This chapter examines some variations of conventional schedulers for dynamic mapping of depen-
dent tasks. Section 13.2 describes further background material. Section 13.3 discusses related work
in the literature. Section 13.4 describes enhancements made to a grid simulation toolkit and intro-
duces a genetic mapping heuristic. Section 13.5 gives the results from the study. Section 13.6
describes some further work that might be considered, and Section 13.7 draws some conclusions from
the work.

13.2 Background
13.2.1 List Scheduling Heuristics
As described in Section 13.1, this chapter focuses on the dynamic scheduling of tasks on a service-oriented
Grid, a network of heterogeneous machines. The service to be scheduled may be represented by a task
graph or directed acyclic graph (DAG). The DAG specifies the tasks that make up the service, as well as the
dependencies between tasks [10].
The scheduling of DAGs may be done using a number of broad strategies, including Clustering
Algorithms [11]. Clustering Algorithms are not considered in this chapter, but involve mapping sub-
sets of tasks that have large inter-task communications to a set of processors with high bandwidth and
low latency (e.g., see Reference 12).
An example of a generational algorithm is a List Scheduler. A List Scheduler performs mappings based
upon a subset of the tasks in the DAG. It only attempts to schedule those tasks that have had all dependency
relationships fulfilled. This subset of tasks is called a meta-task and consists of tasks that are independent
with respect to each other. This greatly simplifies the mapping process and allows it to be done dynamically.
An auxiliary algorithm then maps this meta-task to available resources. When a rescheduling event (such
as a task finishing) occurs, a new meta-task is constructed and scheduled. The expected execution time of
each task on every resource is considered to be known a priori. This assumption is typically made when
conducting mapping research [11,13].
The mapping of tasks to resources, whether static or dynamic, is an NP-hard problem. Finding an
optimal solution is intractable. Therefore, a heuristic is normally used to choose a mapping.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 194 — #2

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-195

Some examples [1,4,14,15] of simple mapping heuristics are:

• OLB (Opportunistic Load Balancing): OLB assigns each task, in arbitrary order, to the next available
machine, regardless of the task’s expected execution time on that machine.
• UDA (User Defined Assignment): UDA assigns each task, in arbitrary order, to the machine with the
best expected execution time for that task, regardless of that machine’s availability.
• Fast Greedy: Each task is assigned to the resource with the minimum completion time. The tasks
are assigned in an arbitrary order.
• Min–Min: The Min–Min heuristic begins with the set U of all unmapped tasks. For each task in
U , the minimum completion time on all machines is calculated. The task with the earliest overall
minimum completion time is selected and assigned to the machine that yielded that minimum
completion time. This task is removed from U , and the process reiterated until all tasks have been
mapped. Min–Min attempts to map as many tasks as possible to their first choice of machine,
under the assumption that this will result in a shorter makespan. Min–Min has been reported as
superior to many other simple mapping heuristics [14].
• Max–Min: Max–Min is similar to Min–Min except that the task with the latest minimum comple-
tion time is selected and mapped. Max–Min attempts to minimize the penalties incurred by the
scheduling of long-running tasks.

13.2.2 GridSim
GridSim [3,16,17] (http://www.gridbus.org/gridsim/) is a toolkit that supports the modeling and simula-
tion of parallel and distributed computing environments. Entities such as users, applications, resources,
and schedulers may be incorporated, primarily to aid in the design and evaluation of scheduling
algorithms. The features of GridSim are described in Reference 17.
GridSim embodies a layered and modular architecture [17] to make use of existing infrastructure such
as the open-source discrete-event simulator SimJava [18], which itself runs in a java virtual machine. The
layered structure is as follows:

• The gridbroker package provides high-level support for Schedulers or Grid Resource Brokers.
Some of the main classes are:
• Broker, which encapsulates a Scheduler — The Broker incorporates a variant of the Fast
Greedy heuristic within its default time-optimized job scheduler.
• BrokerResource, which embodies a Resource, as known to the Broker.
• UserEntity, which represents a user as known to the Broker.
• Experiment, which manages the work done by the Broker for the UserEntity. It includes
a GridletList object that represents the tasks to be scheduled.
• The GridSim package provides the basic grid infrastructure used by gridbroker. It includes
the following primary classes:
• GridSim. This class represents a grid entity and gives it communication and simulation support.
• Gridlet. A class that represents a single independent task. It includes attributes such as
execution size and size of input and output files. As it models an independent task, there is no
support for multiple input and output file sizes.
• GridletList. A class that encapsulates a set of independent tasks. It simply extends the
java utility class LinkedList, while adding support for sorting component Gridlet objects by
execution size.
• GridResource. A class that embodies a grid computing resource. Its attributes are specified
in the member class ResourceCharacteristics.
• ResourceCharacteristics. A class that embodies the attributes of a grid computing
resource. It is used to specify characteristics such as CPU speed, baud rate, availability, and cost.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 195 — #3

13-196 Handbook of Bioinspired Algorithms and Applications

• The SimJava package is a general-purpose discrete-event simulation package implemented in Java.

Simulations in SimJava contain a number of entities each of which runs in parallel in its own
thread. An entity’s behavior is encoded in Java using its body( ) method [17].
• The Java Virtual Machine (JVM) provides the runtime environment for SimJava, and hence
GridSim. Since the JVM essentially interprets code that has been compiled for it, it does not
provide the raw performance of a fully compiled language; however, implementations of JVMs
exist for multiprocessor systems and clusters [19], and so it does provide scalable support for grid
simulations.

13.2.3 Task Graphs for Free

Task Graphs for Free (TGFF) [20] is an open-source package that generates random or semi-random task
graphs, or DAGs. Such task graphs are ideal for scheduler simulation research. The task graphs may be
generated to have a series-parallel-like structure or may be more purely random.
A generated graph consists of tasks and arcs. Arcs are the dependencies that link tasks. A defini-
tions file (i.e., a .tgffopt file) is created that specifies the parameters for construction of the graph
or graphs, as well as the corresponding tables that define task attributes and arc attributes. Upon
processing by the TGFF utility, a text file (a .tgff file) is produced that fully specifies the task
graph [20].

13.2.4 Genetic Algorithms

Genetic Algorithms (GAs) are often used to solve difficult optimization problems heuristically. GAs seek
to make use of the fundamental principle of biological evolution, Natural Selection. A GA mirrors the
same mechanisms by evolving a population of fixed size through a large number of generations. The
population consists of randomly generated candidate solutions encoded as a string of bits or numbers,
called a chromosome. An objective function is used to evaluate the chromosomes yielding a fitness value that
reflects the suitability of the candidate solution. The fitness value is used to select a subset of individuals.
This subset is then subjected to the processes of mutation and crossover, resulting in the next generation
of individuals. Mutation and crossover help the GA to escape from local optima in the solution space and
find near-optimal solutions to a problem.
Selection of parents for the next generation uses a competitive evaluation of the objective function
for members of the population. The degree to which the fit individuals are likely to be selected is called
the selective pressure. Surprisingly, too high a selective pressure is undesirable, since it may lead to the
population converging overly quickly to a local optimum and not finding a global optimum. Various
mechanisms have been used to moderate the selective pressure, an example of which is tournament
selection. In tournament selection, a set of individuals (generally two) is randomly selected from the
population. The fittest individual is then selected as a parent [21].
Note that, in tournament selection, the fittest individual in the population may well not proceed to the
next generation. However, if the GA is engineered such that such an individual is guaranteed to proceed
to the next generation, then the GA is said to have the quality of elitism.
Mutation involves a random change to a chromosome and so has the chance of introducing great
novelty into the population. However, when the mutation rate is too high, good solutions tend to be
overwritten, and the population tends not to converge to a solution.
Crossover emulates the genetic mixing effects of biological sex. It allows successful individuals to share
their genetic material, and may therefore result in offspring that combine the virtues of both parents.
Crossover is a considerably more important optimization mechanism than mutation [21]. Crossover may
involve a division in one or many places in the chromosome. A crossover with a single point of division is
termed Single-Point Crossover.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 196 — #4

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-197

13.3 Related Work

13.3.1 A Comparison Study of Static Mapping Heuristics
Genetic Algorithms may be used in the static mapping of independent tasks to resources. In [1] the
GA was used to find near-optimal mappings using populations of 200 chromosomes. The chromosomes
were encoded mappings, represented by a vector of t integers, where the ith integer represents the ith
task, and its value represents the machine to which it has been mapped. The GA used the makespan as
the objective function. Elitism was used to retain the best solution between generations. The GA used
one of three termination criteria, either stopping after 1000 iterations, or after no change in the elite
chromosome for 150 generations, or after all chromosomes have converged. The termination criterion
that was generally triggered was the second criterion, that is, no change in the elite chromosome for
150 generations.
For each optimization task, the GA was executed eight times: four times with a random initial pop-
ulation, and four times with populations seeded with the Min–Min solution. Note that the best GA
solution always came from a population seeded with a Min–Min solution. The GA always improved on
the Min–Min solution, providing an improvement of up to 10%.
It was decided to attempt a variant of this optimization technique, for a dynamic generational scheduler
operating on a set of dependent tasks. The GA was used to find a near-optimal mapping for each meta-
task during each scheduling event. The intention was to improve upon the performance of a Min–Min
generational List Scheduler.

13.3.2 A Dynamic Matching and Scheduling Algorithm for HC Systems

In the paper [13], the critical path length for a task is defined as the longest path from the task node to an exit
node. The paper describes a hybrid remapper, which uses a combination of static and run-time generated
information. It uses the critical path length as a motivating factor in its Minimum Completion Time Static
Priority Algorithm. The concept of a Critical Path was used in the FAST_GREEDY_CRITICAL mapping
heuristic and the SUM_CRITICAL_FINISH objective function.

13.4 Development Work

While GridSim is a versatile and powerful simulation toolkit, it is not perfectly suited to the needs of this
project. It was therefore extended in a number of ways.
First, GridSim is designed for the scheduling of independent tasks, that is, there is no support for task
dependencies because GridSim is principally designed for job scheduling. Second, GridSim is intended
as an application level scheduler, whereas the project requirement was for a scheduler at a service-level.
Third, there is no facility in GridSim for flexible configuration of brokers or execution environment.
It was therefore extended to provide xml-based support for broker configuration and task and resource
specification. (Similarly, TGFF was later integrated, so that the broker might directly parse .tgff definition
files.)
Various other extensions were also made to the broker, such as the implementation of several alternate
mapping heuristics, in particular, one based on a genetic algorithm.
Where possible, all extensions were made by the creation of the new packages:

• gridscheduler
• gridscheduler.ga

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 197 — #5

13-198 Handbook of Bioinspired Algorithms and Applications

Only one class in the gridsim and gridbroker packages was directly amended. All the required
functionality was generally accommodated by writing new subclasses of the existing gridsim or
gridbroker classes.

13.4.1 DAG Support

Since GridSim is not designed for the scheduling of dependent tasks, a number of classes required extension
to permit scheduling simulation of DAGs. The class GridSim.Gridlet required support to specify
its ancestor and dependent tasks. It was decided not to model the data size of these dependencies. As the
project was a simulation of a service-based broker, it was accepted that communications would, in fact, be
centralized via the broker itself. Therefore, the data sizes of the various dependencies could be represented
by the input and output data sizes of the task itself. Gridlet was, therefore, extended by the new
gridscheduler class Tasklet. The new class has additional attributes to identify its predecessors
and successors, as well as its critical path length. The critical path length was an attribute added late in the
development phase of the project, and is intended to support a new heuristic. It is a measure of the size of
the execution path from the task to the latest subsequent point of termination [13].
Gridlet itself was the only class in the GridSim packages that was directly altered. Some small changes
were made to support a new status: PREPARED. This new status indicates that the task has no further
dependencies, and may now be executed by the processor. It follows CREATED in the life cycle of a
Gridlet, and precedes READY, QUEUED, INEXEC, SUCCESS, and FAILED. READY means that the
task has been assigned to a resource. QUEUED means that the task has been committed to a resource and is
in its input queue. The meanings of the other statuses are as implied by the labels. The PREPARED status
is essential in the operation of a generational scheduler.
Similarly, the class GridSim.GridletList represents a meta-task, a set of independent tasks.
While it did not require extension for any fundamental new functionality, it was extended by the class
TaskGraph to provide support for a number of methods relating to sets of tasks. The getWaiting( )
method was provided to return those tasks in the TaskGraph in a state of PREPARED. The
setPredecessors( ) method was added to set the predecessors for all tasks in a TaskGraph, given
that all successors had already been set — the rationale being that fewer errors would then be made in the
xml definition files of TaskGraphs. This class was made redundant with the introduction of support for
TGFF definition files. Two inner classes were created to provide support for sorting the TaskGraph by
fanout size or CriticalPathLength. The setCriticalPathLengths( ) method was added in
order to set all critical path lengths for tasks in the TaskGraph. This method delegates to the Tasklet
method setAndGetCriticalPathLength( ). Setting of the critical path length (cpl) uses the
following simple recursive algorithm, which is executed for the initiating task or tasks:

cpl = execution size of this task

if this task has dependent tasks
cpl + = max(cpl of all dependent tasks)

The GridSim class ResourceCharacteristics was extended by the class TaskResource-

Characteristics, in order to allow the baud-rate of the resource to be visible by the broker. This
visibility was needed so that the broker would be able to include an estimate for the communication delay
in its mapping heuristic calculations.
For this reason, the gridbroker class BrokerResource was extended by the class
TaskBrokerResource, and the method getExpectedCompletionTime( ) was overridden
to add a term for the communication delay to the estimate. The expected completion time is given by:

completion time = exec + max(avail, comms)

avail = earliest time at which a resource is available
exec = anticipated execution time for a task
comms = expected communication delay

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 198 — #6

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-199

The use of the max( ) function reflects the assumption that communications transmission may proceed
in parallel with waiting for the processor to be available.
If the task in question has no predecessors, or is colocated with all such predecessors, then the commu-
nication delay is considered to be only a nominal amount — reflecting only an initiating communication
from the scheduler. A corresponding method isCoLocated( ) was created to test whether the task is
colocated. This method is also used by the TaskBroker class (see Section 4.2) when tasks are forwarded
to resources; if a task is colocated, then the input file size of the Tasklet is also set to a nominal value.
A simple and popular model of the communication delay for message passing is:

message time = latency + message size/bandwidth

This model includes a term for the latency of the communication link. However, a bandwidth-only
model has been found to lead to better predictions of communication delay, at least in some applications
[22]. It was therefore considered not essential to add a network latency component to the GridSim
communication delay methods.

13.4.2 The TaskBroker

The gridbroker class Broker was replaced by the new gridscheduler class TaskBroker. Broker
was designed for the scheduling of independent tasks using one of several economic- or deadline-based
optimization schemes, based on cost, time, or combined cost/time optimizations. The architecture is
described in Reference 17.
For the purposes of this investigation, the economic and deadline support of the Broker class was
ignored. The base time optimization scheduling was instead extended to incorporate a Generational
Scheduler for DAGs. Time-shared processors have not been supported.
The scheduleAdviser( ) method, which builds a list of those tasks that are to be scheduled,
was amended to only schedule PREPARED tasks (i.e., those tasks for which all parent tasks have com-
pleted successfully). It does this by moving such tasks from the existing glPreparedList to the
new glWaitingList collection (note that glPreparedList actually holds those tasks that are
CREATED, rather than PREPARED, and so is something of a misnomer). It then invokes one of two
new methods: scheduleWith_ListInsertion( ) or scheduleWith_ListInsertionFG( );
these methods use various heuristics to map the available tasks to resources.

13.4.3 Mapping Heuristic Extensions

The existing broker used a variant of the Fast Greedy mapping heuristic (Section 13.2.1). In Fast Greedy,
each task is assigned to the resource with the minimum completion time. The tasks are assigned in
an arbitrary order. The existing broker instead assigned tasks in increasing order of execution size, in
an attempt to maximize packing of tasks.
TaskBroker applies various different heuristics. All may be defined in the broker configuration. The
configuration parameter <heuristic> defines the desired heuristic.
The scheduleWith_ListInsertionFG( ) method applies various variants of the Fast Greedy
heuristic:
• FAST_GREEDY_BASE. The base Fast Greedy heuristic.
• FAST_GREEDY. Fast Greedy as originally used by GridSim, that is, with tasks mapped in order of
execution size.
• FAST_GREEDY_FANOUT. Fast Greedy, but with the tasks mapped in reverse order of the number
of dependent tasks. This is an attempt to execute those tasks that may have many child tasks first.
• FAST_GREEDY_CRITICAL. Fast Greedy, but with the tasks mapped in reverse order of the
critical path length. This is an attempt to map those tasks that are in the critical path of a task graph
first.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 199 — #7

13-200 Handbook of Bioinspired Algorithms and Applications

The scheduleWith_ListInsertion( ) method applies the Min_Min heuristic or the GA

heuristic. The configuration parameter <minmin_limit> governs when the GA is invoked for any
particular scheduling event. When the size of the meta-task (the set of tasks to be scheduled) is less than
or equal to the minmin_limit, then Min_Min is used to schedule the meta-task, otherwise the GA is
invoked to find a near-optimal mapping.

13.4.4 GA Mapping Heuristic

The GA described in Reference 1 was used as the model for the GA heuristic used in this project to map
the meta-task arising at each scheduling event. The GA heuristic was designed to be entirely configurable
(see Section 13.4.2).
The functions of the various GA configuration parameters are as follows:

• heuristic. See Section 13.4.3.

• minmin_limit. See Section 13.4.3.
• minmin_seed. Indicates whether the initial population will be seeded with the Min_Min
solution.
• population_size. The size of the population selected for breeding.
• max_generations. The GA stops after this number of generations.
• max_winning_run. The GA stops if the same Chromosome is the elite value for this number
of generations.
• crossover_rate. The proportion of parent matings that result in crossover, and therefore
recombine genetic information.
• mutation_rate. The proportion of Chromosomes that will have a gene randomly changed.
• fertility_rate. The proportion of the population that attempts to breed each generation.
• random_injection. A proportional injection of random Chromosomes injected every
generation. Not yet implemented.
• elitism. Whether or not the elite Chromosome is guaranteed to survive to the next generation.
• objective. The objective function used by the GA. In this implementation, the value returned
by the objective function is actually the inverse fitness — for example, in the case of MAKESPAN,
we wish it minimized. The following objective functions are available:
◦ MAX_FINISH (a.k.a. MAKESPAN). The makespan of a mapping, which is the latest time of
completion of the mapped tasks (see Section 13.4.1).
◦ SUM_TASK_FINISH. The sum of all completion times of the mapped tasks.
◦ SUM_PE_FINISH. The sum of all last completion times of all processors.
◦ SUM_TASK_CRITICAL. The sum, for all mapped tasks, of the products of their completion
times and the log of their critical path lengths.

The classes in the gridscheduler.ga package are:

• Evolution. Encapsulates the scheduler GA itself. Contains the method evolve( ), which
performs the complete evolutionary process and returns the solution as a Chromosome.
• Population. The set of Chromosomes that encapsulate the mappings currently under
consideration.
• Chromosome. Specifies the processor mappings for all tasks currently ready for execution.
A SingleGene object specifies each task mapping. The Chromosome contains a collection
of these SingleGene objects.
• SingleGene. A container for a gene within a Chromosome. Each gene specifies the mapping
of a task to a processor [resource] and the priority of the mapping to the resource. The priority is
encoded as an integer value. If another SingleGene in the Chromosome specifies that its task
is to be mapped to the same resource with a higher priority, then the other task will be scheduled
before this one.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 200 — #8

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-201

• Evaluation. A container for the result of an evolutionary generation, that is, whether the
evolution has completed; why it has completed; how many generations the result took; and the
elite Chromosome itself. Completion may have occurred for one of three criteria described in
Section 13.3.1.
• Mating. Encapsulates the genetic pairing of two Chromosomes, which may share genetic
information to produce two [genetically related] offspring.
• GaRandom. A simple class with static methods to return random numbers efficiently.

13.5 Results
13.5.1 Methodology
The GridSim distribution, including its example source files, was used as the starting point for this project.
A number of incremental changes were then made in order to provide needed support (see Section 13.4).
With the implementation of the GA mapping heuristic, a large number of tests were carried out using
a TGFF generated task graph. These tests indicated that the GA heuristic gave a nearly 8% improvement
over a Min–Min heuristic for the task graph in question. Interestingly, the GA heuristic displayed a large
degree of variability in its performance. In a typical set of 100 simulations for the task graph, the GA
heuristic gave an improvement of 7.8%, but with a standard deviation of 5%. However, the vast majority
of results were an improvement over the control Min–Min result.
An identical heterogeneous processor arrangement was used for this and all subsequent tests. The
configuration file for this is given in Appendix A.
In recognition that this result was for only a single task graph, a script was designed to generate
multiple variants of similar task graphs and compare the performance of a trial heuristic against the
Min–Min heuristic for all of these task graphs.
Fortunately, TGFF includes excellent support for just this scenario, as it accepts a random number
generator seed parameter. This seed affects all the randomized aspects of the generated task graph.
Varying the seed while holding all other parameters constant generates task graph families containing an
arbitrary number of task graphs. Given an identical seed, an identical task graph is produced [20].
The script invokes TGFF with an incremented seed, starting from a value of zero (the default value),
generating a family of n task graphs. For each generated task graph, first the Min-Min heuristic is run
and then the trial heuristic is run m times. Therefore, each test involved running the trial heuristic n × m
times.
These task graph family tests revealed that the previously recorded 8% improvement had merely been
due to a fortunate choice of task graph. The results of other task graphs in the same family led to an overall
ambivalent result. Further tests were therefore conducted in an attempt to find circumstances in which
the GA might perform with more consistency.
This further testing at first indicated that the GA heuristic does give a significant improvement for
a family of task graphs. However, this improvement vanished when further members of the task graph
family were evaluated.

13.5.2 Comparisons of the Objective Functions of GA

The various objective functions described in Section 13.4.4 were developed with differing rationales.

• MAX_FINISH (a.k.a. MAKESPAN). This was the objective function used in Reference 1. It is
ideal for a static mapping scenario; but in a generational scheduler it does not allow the GA to
discriminate between solutions on the basis of the completion times of all tasks — not just those
tasks on the processor that completes last. (The critical path of the mapped meta-task.)
• SUM_TASK_FINISH. This function seeks to achieve a better result for all tasks in the meta-task,
not just those that execute on the critical path of the mapped meta-task. This function differentiates

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 201 — #9

13-202 Handbook of Bioinspired Algorithms and Applications

between solutions with an identical makespan, but with poorer task completion times on other
resources. For example, if one solution mapped those tasks not contributing to the makespan on
a single processor, and another solution mapped such tasks to multiple processors — then this
function would discriminate between the two solutions. MAKESPAN gives the same score to both
solutions.
• SUM_PE_FINISH. This applies a variant of the logic of the previous heuristic, but is likely not to
be as discriminating.
• SUM_TASK_CRITICAL. This objective function hopes to introduce some non-locality into the
GA. In other objective functions, the position of the task in the task graph plays no part in how well
it might be mapped. This function seeks to preferentially map those tasks that are most critical to
a total reduction in the makespan of the task graph.

Each GA objective function was tested against 100 task graphs, with each task graph having 100 nodes,
and an execution size of 2 to 8 × 109 instructions and a communication size of 0.5 to 1.5 × 105 bytes.
Each task graph was a recursive series-parallel graph, with a series length of 1–3 and series width of 1–3.
Each task graph also had a local crossover of 20 (see Reference 20 for a full description of parameters).
Appendix B depicts the seed task graph (0) resulting from these parameters.
Other broker configuration parameters were:

<broker_properties>
<heuristic>min_min</heuristic>
<minmin_limit>1</minmin_limit>
<minmin_seed>true</minmin_seed>
<population_size>100</population_size>
<max_generations>500</max_generations>
<max_winning_run>50</max_winning_run>
<crossover_rate>1.0</crossover_rate>
<mutation_rate>0.01</mutation_rate>
<fertility_rate>2.0</fertility_rate>
<random_injection>0.0</random_injection>
<elitism>true</elitism>
..<max_task_per_PE>1</max_task_per_PE>
</broker_properties>

Each test consisted of a family of 10 task graphs, with 20 scheduling runs per task graph, for a total of
800 scheduling runs.
The results indicate that, for this task graph family subset, the GA mapping heuristic gives an average
6% improvement — when the SUM_TASK_CRITICAL objective function is used (Table 13.1). The
SUM_TASK_FINISH function performs nearly as well, while the MAKESPAN function gives less than
half the same improvement. This result is in line with the expectations of the objective functions. The
SUM_TASK_FINISH function, with its contribution of all task finish times, gives a greatly improved
result over the MAKESPAN. The SUM_TASK_CRITICAL function, with its non-localized contribution,
gives a further improved result.
Note that this improvement, relative to Min–Min, vanished when further members of the task graph
family were examined. However, the relative results of the different objective functions still hold.

TABLE 13.1 GA Objective Function Comparisons

Objective function sum_pe_finish makespan sum_task_critical sum_task_finish

Proportional increase to min-min 0.015 0.024 0.06 0.05

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 202 — #10

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-203

13.5.3 GA Onset Trials

The configuration parameter MINMIN_LIMIT (Section 13.4.3) specifies a meta-task size limit for which
the Min–Min heuristic should be used rather than the GA heuristic. This functionality was designed since
it is clear that for a single task, the optimal mapping returned by the GA would, at best, simply be that
resource that would first complete the task, which is precisely the mapping that a Min–Min heuristic
would result in.
Two series of tests were conducted to investigate the effect of increasing the MINMIN_LIMIT. The
tests were discontinued at the point where, for a large number of the tasks tested, the GA result was
identical to a simple Min–Min heuristic. This occurs when the size of meta-tasks rarely, if ever, exceeds
the MINMIN_LIMIT.
In the first series of tests, the task family used in the previous test was tested with the
SUM_TASK_FINISH objective function, with MINMIN_LIMIT varied from one to three. Other broker
configuration parameters were identical to the previous test.
The second series of tests was similar to the first, with the task graph family having similar charac-
teristics to the previous test, but with a series width of two to eight. The value of the broker parameter
MAX_GENERATIONS was 100. Other broker configuration parameters were identical to the previous test.
The MINMIN_LIMIT varied between one and five.
Each test consisted of a family of 10 task graphs, with 20 scheduling runs each, for a total of 400
scheduling runs.
The results in Table 13.2 indicate that the favorable results obtained in Section 13.5.2 for the
SUM_TASK_FINISH objective function steadily degrade as the MINMIN_LIMIT is increased and more
of the mapping is made using Min–Min. This indicates that for these particular task graphs, the GA heuristic
was more effective than a Min–Min mapping.
The results in Table 13.3 do not demonstrate a consistent pattern. The GA appears not to perform
inconsistently given the characteristics of the task graph family.

13.5.4 Large-Scale DAGs

An attempt was made to see whether the GA might be more advantageous for large-scale task graphs.
Each GA objective function was tested against 2 to 5 task graphs, with each task graph having 1000
nodes, with an execution size of 2 to 8 × 109 instructions and communication size of 0.5 to 1.5 × 105
bytes. The task graphs were random, not series parallel. The test has not yet been performed against a
larger number of task graphs because of the time taken to perform the test itself.
Each test consisted of a family of 2 to 5 task graphs, with 10 scheduling runs per task graph, for a total
of 110 scheduling runs.
The results for SUM_TASK_FINISH and SUM_TASK_CRITICAL appear to approximate that of a
Min–Min mapping, while the other objective functions do considerably worse. This again demonstrates
the superiority of these two objective functions to SUM_PE_FINISH and MAKESPAN (Table 13.2)

TABLE 13.2 GA Onset

minmin_limit 1 2 3

Proportional increase to minmin 0.05 0.025 0.001

Low-width series parallel.

TABLE 13.3 GA Onset

minmin_limit 1 2 3 4 5
Proportional increase to minmin −0.017 0.012 0.006 0.019 0.005

Medium degree series parallel.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 203 — #11

13-204 Handbook of Bioinspired Algorithms and Applications

TABLE 13.4 Large-Scale DAGs

Objective sum_pe_finish makespan sum_task_critical sum_task_finish

Result versus minmin −0.094 −0.038 −0.01 −0.006

TABLE 13.5 Comparison of Fast Greedy Heuristic Variants with Min–Min

2–5 8–12 14–17

Task_degree
Heuristic (fg) Exec Fanout Critical Exec Fanout Critical Exec Fanout Critical

Improvement versus
unsorted fg 0.073 0.035 0.051 0.043 0.036 0.029 0.040 0.034 0.033

13.5.5 Fast Greedy Variants

A number of trials were also conducted of the various variants of the Fast Greedy heuristic. Each was
tested against a total of 300 task graphs, in 3 families differentiated by task degree, that is, 100 task graphs
of low degree, 100 of medium degree, and a further 100 of high degree.
All task graphs had 100 nodes, with an execution size of 2 to 8 × 109 instructions and communication
size of 0.5 to 1.5 × 105 bytes. The task graphs were random, not series parallel.
Each test consisted of a family of 100 task graphs, with 1 scheduling run each, for a total of 1500
scheduling runs (including Min–Min runs).
The results indicate that, for all task degrees, performance of the Fast Greedy heuristic is improved by
assigning the tasks in order of execution size, order of fanout, or order of critical path size (Table 13.5). The
effectiveness of the simple order of execution size case is explained by considering that this approximates
to a Min–Min mapping, at least for consistent processors and no communication costs.

13.6 Further Work

13.6.1 Network Latency
As indicated in Section 13.4.1, a generalized model of communication delay for message passing is:

message time = latency + message size/bandwidth

This chapter assumed a latency of zero, which was considered acceptable, given the effectiveness of a
bandwidth-only model [22] and GridSim’s lack of support for network latency. However, incorporation
of non-zero latencies might provide for a more accurate model of a grid environment. This would require
some refactoring of several GridSim classes, but particularly IO_Data, Input, and Output. It is expected
that this work would have little bearing upon the GA mapping heuristic.

13.6.2 Multiple User Environments

The GA heuristic has only been tested in a single-user environment. It would be of great interest to
investigate performance in a true multi-user service-based environment. In such an environment, the
load on the grid would be variable, due to tasks being processed for other users. It might then be that
in this circumstance, the local nature of the meta-task mapping of the GA heuristic would be of less
consequence.

13.6.3 Randomized Heterogeneous Processor Environments

This project has only investigated a single example of a heterogeneous computing environment,
in a more complete investigation; a randomized heterogeneous processor environment could be

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 204 — #12

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-205

evaluated. The TGFF table generating functionality can be used to generate such randomized grid
environments.

13.6.4 Inconsistent Execution

This chapter has only investigated execution of the GA heuristic in a consistent processor environment.
That is, while the processors have been heterogeneous in performance and available bandwidth, they
have behaved consistently for all tasks. However, it would be interesting to evaluate performance in
an inconsistent environment. This would necessitate adding a TaskType member to the Tasklet
class (which extends Gridlet), and supporting differing execution rates for the PE class for different
TaskType objects.

13.7 Conclusions
This chapter has described several variants of an existing dynamic generational mapping heuristic, the
Fast Greedy heuristic. It has been shown that simply sorting the meta-task in order of execution size may
make substantial improvements to the Fast Greedy heuristic. Other orderings, such as fanout size and
critical-path length, appear not to be as advantageous.
This chapter has also described a new dynamic generational mapping heuristic, the GA mapping
heuristic. It has been shown that the total makespan of the heuristic is roughly comparable to that of the
Min-Min heuristic, with an added variance. The variance is introduced by the non-deterministic nature of
the GA (The execution time of the scheduler itself will, of course, be substantially greater). It is suspected
that this result might be enhanced with an improvement in the objective function of the GA. Most
evaluated objective functions only make use of the current meta-task, and have no broader knowledge
of the task graph. The exception to this was the MAX-TASK-CRITICAL function, which attempted to
factor the critical path of a task into the fitness value. This GA appeared to be the best performing (along
with the SUM_TASK-FINISH function). One difficulty in improving the GA is that it may only use the
fitness value to choose between solutions. It may not use any other contextual information, and the fitness
is only a simple scalar value.
It is also suggested that performance of the GA heuristic would be relatively greater in a multiuser or
inconsistent processor environment. In such an environment, the local nature of the GA should be less of
a constraint on performance.

References
[1] Braun, T., Siegel, H., et al. A Comparison study of static mapping heuristics for a class of meta-tasks
on heterogeneous computing systems. In 8th IEEE Heterogeneous Computing Workshop (HCW ’99),
1999.
[2] Foster, I. and Kesselman, C., Ed. The Grid: Blueprint for a New Computing Infrastructure. Morgan
Kaufmann, San Francisco, CA, 1998.
[3] Murshed, M. and Buyya, R. Using the GridSim Toolkit for Enabling Grid Computing Education,
http://citeseer.nj.nec.com/574558.html.
[4] Maheswaran, M., Ali, S., et al. Dynamic matching and scheduling of a class of independent
tasks onto heterogeneous computing systems. In 8th IEEE Heterogeneous Computing Workshop
(HCW ’99), April 1999, pp. 30–44.
[5] Kim, J., Shivle, S., et al. Dynamic mapping in a heterogeneous environment with tasks having
priorities and multiple deadlines. In International Parallel and Distributed Processing Symposium
(IPDPS’03), April 2003.
[6] Coffman, E., Jr., Ed. Computer and Job-shop Scheduling Theory. John Wiley & Sons, New York,
1976.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 205 — #13

13-206 Handbook of Bioinspired Algorithms and Applications

[7] Fernandez-Baca, D. Allocating modules to processors in a Distributed System. IEEE Transaction

on Software Engineering, 15, 1989, 1427–1436.
[8] Ibarra, O. and Kim, C. Heuristic algorithms for scheduling independent tasks on non-identical
processors. Journal of the ACM, 24, 1977, 280–289.
[9] Novaes, M. and Challenger, J., Request Level Scheduling for the Grid. IBM T.J. Watson Research
Center, Yorktown Heights, N.Y.
[10] Cirou, B. and Jeannot, E., “Triplet: a clustering scheduling algorithm for heterogeneous systems.
IEEE ICPP International Workshop on Metacomputing Systems and Applications (MSA’2001),
September 2001, Valencia, Spain.
[11] Freund, R., Gherrity, M., et al. Scheduling resources in multi-user, heterogeneous, comput-
ing environments with SmartNet. In 7th IEEE Heterogeneous Computing Workshop (HCW ’98),
March 1998, pp. 184–199.
[12] Taura, K. and Chien, A. A heuristic algorithm for mapping communicating tasks on heterogeneous
resources. In Heterogeneous Computing Workshop, May 1, 2000.
[13] Maheswaran, M. and Siegel, H. A dynamic matching and scheduling algorithm for heterogeneous
computing systems. In 7th IEEE Heterogeneous Computing Workshop (HCW ’98).
[14] Holenarspiur, P., Yarmolenko, V., et al. Characterisation and Enhancement of Static Mapping
Heuristics for Heterogeneous Systems, Technical report OSU-CISRC-2/00-TR07, Department of
Computer and Information Science, Ohio State University, 2000.
[15] Siegel, H. and Ali, S. Techniques for mapping tasks to machines in heterogeneous computing
systems. Euromicro Journal of Systems Architecture, Special Issue on Heterogeneous Distributed
and Parallel Architectures: Hardware, Software and Design Tools, 46, 2000, 627–639.
[16] Sherwani, J., Ali, N., et al. Libra: An economy driven job scheduling system for clusters.
In Proceedings of the 6th International Conference on High Performance Computing in Asia-Pacific
Region (HPC Asia 2002), Bangalore, India, 2002.
[17] Buyya, R. and Murshed, M. GridSim: a toolkit for the modeling and simulation of distrib-
uted resource management and scheduling for grid computing. The Journal of Concurrency and
Computation: Practice and Experience (CCPE), 14(13–15), Nov–Dec 2002, 1175–1220.
[18] Howell, F. and McNab, R. Simjava: a discrete event simulation package for java with applications
in computer systems modelling. In First International Conference on Web-based Modelling and
Simulation. San Diego, CA, Society for Computer Simulation, January 1998.
[19] Aridor, Y., Factor, M., and Teperman, A. cJVM: a single system image of a JVM on a cluster.
In Proceedings of the 29th International Conference on Parallel Process (ICPP’99), 1999.
[20] Dick, R., Rhodes, D., and Wolf, W. TGFF: Task Graphs for Free. In Proceedings of the 5th
International Workshop Hardware/Software Co-Design/Codes/CASHE’97, March 1998, pp. 97–100.
[21] Wilkinson, B. and Allen, M., Parallel Programming. Prentice Hall, 1999, pp. 372–377.
[22] Dail, H., Berman, F., and Casanova, H. A decoupled scheduling approach for grid application
development environments. Journal of Parallel and Distributed Computing, 63(5), May 2003,
505–524.

Appendix A. Test Processor Configuration

<?xml version="1.0" encoding="utf-8" standalone="yes" ?>
<grid>
<resource name="resource0">
<sharing>space</sharing>
<cost>1</cost>
<baud>100</baud>
<machine>
<pe mips="300"/>
</machine>

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 206 — #14

Genetic Algorithms for Scheduling in Grid Computing Environments: A Case Study 13-207

</resource>
<resource name="resource1">
<baud>80</baud>
<machine>
<pe mips="250"/>
</machine>
</resource>
<resource name="resource2">
<baud>150</baud>
<machine>
<pe mips="68"/>
</machine>
</resource>
<resource name="resource3">
<baud>80</baud>
<machine>
<pe mips="500"/>
</machine>
</resource>
<resource name="resource4">
<baud>100</baud>
<machine>
<pe mips="120"/>
</machine>
</resource>
</grid>

This configuration specifies that the available grid environment consists of five machines, each having a
bandwidth ranging from 80 to 150 KBpsec. All are space-sharing architectures, rather than time-sharing,
and have a nominal cost. Each of the machines have a single processor element, with speed ratings of
68 to 500 mips.

Appendix B. Sample Task Graph

The .tgffopt definition that follows, given a seed value of zero, results in the task graph depicted
in Figure 13.B1. Resultant Task sizes are 2000 to 8000 M instructions, and communication costs are
500 to 1500 KB.

tg_cnt 1
gen_series_parallel true
series_must_rejoin true
series_len 2 1
series_wid 2 1
series_local_xover 20
task_cnt 100 1
task_degree 2 4
tg_write
eps_write
table_label PROCESS_SIZE
table_cnt 1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 207 — #15

13-208 Handbook of Bioinspired Algorithms and Applications

TEST:_GRAPH 0
Period = 3000
In/Out Degree Limits = 2/4
0

2 4

18 20 22

23 26 28 70 72

24 46 66 68 74 76 98 100

42 44 47 48 67 69 77 99 101

29 43 50 58 62 64 71 75

27 51 52 60 63 73 86 88

19 58 59 87 89 90 92

54 61 65 91 93

21 45 49 55 56

25 57

3 5

30 32

31 33

82 84

83 85

8 78 80

10 12 79 81

11 14 16

17 38 40 94 96

30 41 95 97

15

30 36

35 37

13

1
d=3000

FIGURE 13.B1 Sample task graph.

type_attrib exec_time 5000 3000

pe_write
table_label COMMS_SIZE
table_cnt 1
type_attrib comms_size 1000 500
trans_write

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C013” — 2005/8/6 — 13:24 — page 208 — #16

 14
Minimization of
SADMs in
Unidirectional
SONET/WDM Rings
Using Genetic
Algorithms

14.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-209

14.2 Example of SADM Minimization . . . . . . . . . . . . . . . . . . . . . . 14-210
14.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-212
14.4 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-213
14.5 Genetic Algorithms for Traffic Grooming Problem . . . 14-213
Solution Representation • Initial Population • Selection
Operator • Crossover • Mutation • Acceptance Policy •
The Final Genetic Algorithm
Anirban Mukhopadhyay 14.6 An Improvement to the Genetic Algorithm for
Utpal Biswas All-To-All Unitary Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-215
Mrinal Kanti Naskar 14.7 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-215
Ujjwal Maulik 14.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-217
Sanghamitra Bandyopadhyay References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-217

14.1 Introduction
Most of today’s optical networks are hierarchies of SONET rings. This optical fiber communication is
employed with WDM technology, where the whole bandwidth of an optical fiber is divided among a
number of nonoverlapping wavelengths, each of which is capable of carrying high-speed optical data. In
recent years, the bandwidth of a wavelength has increased from 2.5 Gbps (OC-48) to 10 Gbps (OC-192)
and is likely to go upto 40 Gbps (OC-768) in the near future [1,2]. Thus, the bandwidth capacity on
a wavelength is too large for certain traffic requirements. An approach to provide fractional wavelength
capacity is to split a wavelength into multiple time slots and multiplex traffic on the wavelength. Therefore,

14-209

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 209 — #1

14-210 Handbook of Bioinspired Algorithms and Applications

each wavelength running at the line rate of OC-N can carry several low-speed OC-M (M < N ) traffic
channels in TDM fashion. For example, an OC-48 line can carry 16 OC-3 channels. The resulting networks
are called WDM–TDM networks or WDM traffic grooming networks. The ratio of N to the smallest value
of M is called the grooming ratio [1–5].
Using WDM technology, multiple rings can be supported on a single fiber ring. In this architecture,
each wavelength independently carries a SONET ring. Each SONET ring can further support multiple
low-speed streams. At every node, a Wavelength Add/Drop Multiplexer (WADM) adds and drops or
bypasses traffic on any wavelength. At each node, there are SONET add/drop multiplexers (SADM) on
each wavelength to add/drop low-speed streams. So the number of SADMs per node will increase linearly
with the number of wavelengths that a single fiber ring can carry. The cost of SADMs dominates the total
cost of the optical network. But in fact, it is not necessary for each node to be equipped with SADMs on
each wavelength. An SADM on a wavelength at a node is needed only if there is traffic terminating at this
node on this wavelength. So, the problem is to combine different low-speed traffic streams into high-speed
traffic streams in such a way that the number of SADMs is minimized [3,4]. This problem is proven to
be NP-complete [3]. As far as our knowledge goes, we are the first to propose a genetic algorithm (GA)
solution to this problem for unidirectional ring topologies. Here, we have restricted ourselves to consider
static traffic pattern only. We have proposed an ILP formulation for the problem and solved it using a
GA. We have shown that our algorithm produces better results compared to those found in some recent
literature.
The rest of the chapter is organized as follows: Section 14.2 presents an example for SADM min-
imization problem. Section 14.3 proposes the ILP formulation of the problem. A brief introduction
to GA is presented in Section 14.4. Section 14.5 gives an overview of the proposed GA to solve
the problem. The results obtained are discussed in Section 14.6. Finally, Section 14.7 concludes the
chapter.

14.2 Example of SADM Minimization

In this section, we will present an example to show that careful grooming can reduce the number of
SADMs. Consider a ring network with four nodes. Each wavelength can carry two traffic streams. The
traffic pattern for this example is unidirectional uniform traffic. Figure 14.1 shows the traffic matrix.
Without traffic grooming, every node requires two SADMs, one for each wavelength. Thus, without
grooming, a total of 4 × 2 = 8 SADMs are required.
The traffic assignment as shown in Table 14.1 will result in Figure 14.2, where we can see that though
we have used traffic grooming, still the number of SADMs required is eight. This is due to the lack of
careful grooming.
But if we assign traffic as shown in Table 14.2, node 2 will bypass traffic on the wavelength λ2
(Figure 14.3); hence saving an SADM, as node 2 will have no SADM for wavelength λ2. Thus, in this
case, the total number of SADMs required is seven.

0 1 1 1

0 0 1 1
T=
0 0 0 1

0 0 0 0

FIGURE 14.1 The traffic matrix.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 210 — #2

Minimization of SADMs in Unidirectional SONET/WDM Rings Using GAs 14-211

TABLE 14.1 Careless Traffic Assignment

Wavelength Time slots Traffic

1

 1 → 2, 2→4
λ1
2 −→ 1 → 3, 3→4
1

 1→4
λ2
2 −→ 2→3

l1 l2

l1 l1
4 2
l2 l2

l1 l2

FIGURE 14.2 Traffic configuration requiring eight SADMs.

TABLE 14.2 Careful Traffic Assignment

Wavelength Time slots Traffic

1

 1 → 2, 2→4
λ1
2 −→ 2 → 3, 3 → 4
1

 1→4
λ2
2 −→ 1→3

l1 l2

l1 l1
4 2
l2

l1 l2

FIGURE 14.3 Traffic configuration requiring seven SADMs.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 211 — #3

14-212 Handbook of Bioinspired Algorithms and Applications

14.3 Problem Formulation

Here we have considered a unidirectional ring network. Let the number of nodes in the ring be N and L,
the total number of wavelengths available in a link; g is the grooming ratio, that is, the maximum number
of low-speed traffic streams that can be multiplexed onto a single wavelength. An SADM is needed at a
node for some wavelength only if some of the traffic contained in that wavelength is originated from, or
destined to, that node.
Suppose N nodes in the network are labeled from 0 to N − 1 in a clockwise direction. Wavelengths
available are 0 to L − 1. We are given a set of source–destination node pairs, R = {(si , di )|0 ≤ si , di ≤
N − 1, 0 ≤ i ≤ |R| − 1}, where each pair (si , di ) denotes the requirement of a low-speed, unidirectional
traffic stream from node si to di .
The solution is to find a suitable wavelength assignment W = {li |0 ≤ li ≤ L − 1, 0 ≤ i ≤ |R| − 1},
where wavelength li is assigned to traffic stream (si , di ), such that total number of SADMs needed is
minimized.
For a wavelength assignment {li }, we can find the total number of SADMs needed. Let SADMnw be a
binary variable that is true if an SADM is required at node n for wavelength w, where 0 ≤ n ≤ N − 1, 0 ≤
w ≤ L − 1.


1 if node n needs SADM for wavelength w,
SADMnw =
0 if node n does not need SADM for wavelength w.

We know that an SADM for wavelength w is needed at node n only if node n is the source or destination
of any traffic stream using wavelength w. Hence, we can compute SADMnw for each (n, w) pair in the
following way.


1 if [si = n or di = n] and li = w,
SADMnw =
0 otherwise,

where (si , di ) ∈ R, 0 ≤ li ≤ L − 1, 0 ≤ i ≤ |R| − 1. Therefore, the total number of SADMs in the ring is
N −1 L−1
given by n=0 w=0 SADMnw . The objective is to minimize the number of SADMs, that is, minimize
N −1 L−1
n=0 w=0 SADMnw .
There are certain constraints that we have to consider. First, 0 ≤ SADMnw ≤ 1, and SADMnw is an
integer variable. The other constraint is that no wavelength at any link should be overloaded, that is, the
maximum number of traffic streams that pass through it should be less than or equal to the grooming
ratio g .
The link from node n to ((n + 1) mod N ) is traversed by all low-speed traffic streams (si , di ) ∈ R,
where (si ≤ n < di ), or (n < di < si ), or (di < si ≤ n). Therefore, the load on the wavelength w on link
n → ((n + 1) mod N ) is given by the cardinality of the set
 
(si , di ) | [(0 ≤ si ≤ n < di ≤ N − 1) or (0 ≤ n < di < si ≤ N − 1) 

LDnw = or (0 ≤ di < si ≤ n ≤ N − 1)] and [li = w], .
 
(si , di ) ∈ R, 0 ≤ li ≤ L − 1, 0 ≤ i ≤ |R| − 1

Hence, the load constraint will be ∀n, w(0 ≤ n ≤ N − 1, 0 ≤ w ≤ L − 1) ⇒ |LDnw | ≤ g . Thus,

the problem of minimization of SADMs in unidirectional SONET/WDM ring via traffic grooming can
be given as:
Given integers N > 0, L > 0, g > 0 and request set R = {(si , di )}(0 ≤ si , di ≤ N − 1, 0 ≤ i ≤ |R| − 1),
find a suitable wavelength assignment W = {li |0 ≤ li ≤ L − 1, 0 ≤ i ≤ |R| − 1}, so that,

N −1 L−1
SADMnw is minimized.
n=0 w=0

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 212 — #4

Minimization of SADMs in Unidirectional SONET/WDM Rings Using GAs 14-213

Subject to:

1. 0 ≤ SADMnw ≤ 1, and SADMnw is an integer variable

2. ∀n, w (0 ≤ n ≤ N − 1, 0 ≤ w ≤ L − 1) ⇒ |LDnw | ≤ g

14.4 Genetic Algorithm

Genetic algorithms, introduced by John Holland, are efficient methods for the solution of many search
and optimization problems. Genetic algorithms are robust and stochastic search procedures based on the
principle of natural genetics and evolutionary theory of genes [6,7]. They follow the evolutionary process
as stated by Darwin. The algorithm starts by initializing a population of potential solutions encoded as
strings called chromosomes. Each solution has some fitness value, which indicates the goodness of the
encoded solution. Based on the fitness values, the parents that would be used for reproduction are selected
(survival of the fittest ). The new generation is created by applying genetic operators such as crossover
(exchange of information among parents) and mutation (sudden small change in a parent) on selected
parents. Thus the quality of population is improved as the number of generations increases. The process
continues until some specific criterion is met or the solution converges to some optimized value.
Though GAs sound complicated, one does not have to be a biologist to understand its mechanism. The
main advantage of GAs is their robustness, and hence they can be applied to a wide range of problems, for
example, VLSI desing [8,9], pattern recognition and image processing [10–12], bioinformatics [13,14].
Genetic algorithms are not guaranteed to find the global optimum solution to a problem in a small number
of generations, but they are generally good at finding “reasonably good” solutions to problems “acceptably
quickly”. In this regard, GAs have been found to outperform many other search processes.

14.5 Genetic Algorithms for Traffic Grooming Problem

In this section, we will present a GA to solve the problem stated in the previous section. We will discuss
each step of the algorithm and at last give the final GA for the problem of traffic grooming in unidirectional
SONET/WDM ring networks to minimize the number of SADMs.

14.5.1 Solution Representation

Here, we have encoded the solution to integer strings. In the previous section we showed that the input set
of source–destination pairs is the set R = {(si , di )|0 ≤ si , di ≤ N − 1, 0 ≤ i ≤ |R| − 1}, where each pair
(si , di ) denotes the requirement of a low-speed, unidirectional traffic stream from node si to di . We have
to find a suitable wavelength assignment such that the total number of SADMs required is minimized.
We can represent the solution as an integer string W = {li |0 ≤ li ≤ L − 1, 0 ≤ i ≤ |R| − 1}, where
wavelength li is assigned to traffic stream (si , di ). The fitness value (objective function in the ILP), that is,
the total number of SADMs required, for each solution can be calculated as shown in Section 14.2. We
do not explicitly check the validity of a string, that is, whether it satisfies the constraint that a wavelength
on a link can support maximum g (grooming ratio) numbers of low-speed traffic streams; rather, the
fitness function is converted into an unconstrained form, by adding a penalty term greater than L × N
(maximum number of SADMs in the ring) times the number of violations, to the objective function. Thus
while minimizing the fitness values, the solutions violating the constraint are automatically ruled out.

Example 14.1
Input set: R = {(2, 3), (6, 2), (1, 8), (5, 6), (2, 4), (3, 6)}
Solution representation: W = { 3 8 3 2 2 1 }

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 213 — #5

14-214 Handbook of Bioinspired Algorithms and Applications

14.5.2 Initial Population

The initial population is generated by creating a number of individual solutions (strings). Each string is
generated by random assignment of the wavelengths to each traffic stream of the input set R.

14.5.3 Selection Operator

We use the tournament method of selection, in which a number of strings are picked randomly from the
current generation; from among them we choose the best solution to act as a parent. The other parent
is always the best solution of the previous generation. Genetic operators are applied on the parents to
produce the offspring solutions.

14.5.4 Crossover
The crossover operation is used to exchange genetic materials between two parents. Here we use single-
point crossover, which produces two children from two parents. A random crossover point is generated
between 0 and the maximum chromosome length of the solution. This point divides each solution into
two parts. The corresponding parts of the parent chromosomes are swapped to produce two new offspring.
Example 14.2
Parent1: 1 3 5 2 3 6 7 4 1
Parent2: 3 4 5 2 2 5 7 1 5
Crossover point: 5
Offspring1: 1 3 5 2 3 | 5 7 1 5
Offspring2: 3 4 5 2 2 | 6 7 4 1

14.5.5 Mutation
A random mutation point between 0 and maximum chromosome length is selected, and another
wavelength is assigned instead of the current wavelength at that point with some probability called the
mutation probability.
Example 14.3
Chromosome: 1 3 5 2 3 6 7 4 1
Mutation point: 4
New assignment to 4th wavelength (2): 8
New chromosome: 1 3 5 8 3 6 7 4 1

14.5.6 Acceptance Policy

A new offspring solution obtained by crossover or mutation operation is retained in the population,
if its fitness value is better than the worst fitted solution in the population, otherwise the solution is
rejected [15].

14.5.7 The Final Genetic Algorithm

Example 14.4

Procedure GA
Begin
Generate initial population as per Section 5.2
Initialize number_of_generations = 0
While (number_of_generations < max _generation) do
Select the parents for crossover as per Section 5.3

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 214 — #6

Minimization of SADMs in Unidirectional SONET/WDM Rings Using GAs 14-215

Apply crossover operator as per Section 5.4

Apply mutation operator as per Section 5.5
Accept offspring solution as per section 5.6
Number_of_generations = number_of_generations + 1
End while
Output the solution and its fitness value
End GA

14.6 An Improvement to the Genetic Algorithm for

All-To-All Unitary Traffic
In this section, we will present an improvement over the aforementioned GA by introducing an extra
constraint. This improvement works best in the case of all-to-all unitary traffic, that is, when exactly one
connection is to be established between each pair of nodes. In such a scenario, if the input set R contains
a source–destination pair (s, d), then it will also contain the pair (d, s).
The additional constraint is that both the pairs, (s, d) and (d, s), are assigned the same wavelength.
The rationale being that this assignment will form a circle and SADM saving is maximum when such a
circle is created, because a circle needs two SADMs, one at node s and the other at node d (Figure 14.4).
If two different wavelengths are used, then a total of four SADMs are needed, two SADMs at each node
(Figure 14.5). Hence, while generating a string, we apply this additional constraint with the expectation
of getting a better performance. Let us denote this improved algorithm as GA∗ .

14.7 Results and Discussions

In this section, we will present the results of our algorithm and make comparisons with two other results
found in References 3 and 4. In Reference 3, Modiano presented a heuristic algorithm for the same

s

2

2
2

2
d

FIGURE 14.4 A total of two SADMs are needed.

1,
2

1
2

1,
2

FIGURE 14.5 A total of four SADMs are needed.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 215 — #7

14-216 Handbook of Bioinspired Algorithms and Applications

140

120 Modiano

GA

Number of SADMs
100
GA*
80
RLS
60 Lower
bound
40

20
8 9 10 11 12 13 14 15 16
Number of nodes

FIGURE 14.6 Performance of algorithms for g = 4.

problem, whereas in Reference 4, a Reactive Local Search (RLS) algorithm has been presented. We have
tested the performances of the algorithms for all-to-all unitary traffic, that is, traffic requirement is exactly
one for each pair of nodes. Hence the input set of traffic requirements R can be calculated as:

Example 14.5
For s = 0, 1, 2, 3, . . . , N − 1
For d = 0, 1, 2, 3, . . . , N − 1
If s = d then R = R ∪ {(s, d)}
End for
End for

For example, for a SONET ring with three nodes, the input set of connection requirements will be:

R = {(1, 2), (2, 1), (1, 0), (0, 1), (2, 0), (0, 2)}.

We have run our algorithm upto 10,000 generations to get the results, which takes close to 1 min on a
1.7 GHz Pentium IV computer with 128 MB RAM running Windows Me operating system. The execution
time does not depend on the number of nodes of the ring.
Figure 14.6 is a comparison of Modiano’s algorithm, the RLS algorithm, our GA, and GA∗ for all-
to-all unitary traffic in unidirectional rings. The graphs represent the number of SADMs required for a
certain number of nodes. The number of nodes varies from 8 to 16, as most of the SONET ring contains
a maximum of 16 nodes. We have shown the graphs for grooming ratio g = 4. We have also given the
graphs for the optimal number of SADMs required. From the graph it is clear that our algorithm performs
better than Modiano’s algorithm in all the cases. GA∗ and RLS give the optimal result for g = 4.
Figure 14.7 shows the graphs comparing the four algorithms for grooming ratio g = 16. Here we can
see that the improved GA, that is, GA∗ performs better than the other algorithms and its performance is
near optimal.
The effectiveness of the proposed improved GA, that is, GA∗ can be viewed better if we compare the
values for the number of SADMs obtained using the algorithm GA∗ with its lower bound. In Table 14.3,
we have shown the optimum solution and the solution obtained by algorithm GA∗ for different numbers
of nodes and for grooming ratio 4 and 16. It is clear that for grooming ratio 4, GA∗ gives the optimum
result. For larger grooming ratios also, GA∗ provides near-optimal results for a small number of nodes.
Undoubtedly, GA∗ gives the best result among all the algorithms for the case of all-to-all unitary traffic.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 216 — #8

Minimization of SADMs in Unidirectional SONET/WDM Rings Using GAs 14-217

70

80

Modiano
60

Number of SADMs
GA

40 GA*

RLS
30
Lower
bound
20

10
8 9 10 11 12 13 14 15 16
Number of nodes

FIGURE 14.7 Performance of algorithms for g = 16.

TABLE 14.3 Comparison of results of GA∗ with optimal

value

No. of Grooming ratio g = 4 g = 16

nodes
GA∗ Lower bound GA∗ Lower bound
8 28 28 14 14
9 36 36 18 18
10 45 45 20 20
11 55 55 26 26
12 66 66 32 32
13 78 78 37 36
14 91 91 42 41
15 105 105 47 46
16 120 120 56 54

14.8 Conclusions
In this chapter, we have presented a GA for the traffic grooming problem in unidirectional SONET/WDM
ring networks to minimize the number of SADMs required. We have shown that our algorithm performs
better than other heuristic and reactive search techniques found in recent literature. The GA also takes
lesser time to reach the optimal result; this time is independent of the number of nodes as the number of
generations has been fixed at 10,000. In most of the cases, the optimum value is found with a much lesser
number of generations.

References
[1] A. Mukhopadhyay, J.K. Singh, U. Biswas, and M.K. Naskar, Distributed approaches for dynamic
traffic grooming in WDM optical networks. In Proceedings of International Conference on CODEC-
04, Kolkata, India, January, 2004, p. P-55.
[2] A. Mukhopadhyay, J.K. Singh, U. Biswas, and M.K. Naskar, Improved distributed approaches for
dynamic traffic grooming in WDM optical networks. In Proceedings of Conference on DPN-04, IIT
Kharagpur, India, January, 2004, pp. 92–96.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 217 — #9

14-218 Handbook of Bioinspired Algorithms and Applications

[3] A.L. Chiu and E.H. Modiano, Traffic grooming algorithms for reducing electronic multiplexing
costs in WDM ring networks. IEEE Journal of Lightwave Technology, 18, 2–12, 2000.
[4] R. Battiti and M. Brunato, Reactive search for traffic grooming in WDM networks. In S. Palazzo,
Ed., Proceedings of IWDC2001, Lecture Notes in Computer Science, Springer-Verlag, Heidelberg,
2001.
[5] E.H. Modiano and P.J. Lin, Traffic grooming in WDM networks. IEEE Communications Magazine,
39, 124–129, 2001.
[6] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley,
New York, 1989.
[7] M. Mitchell, An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA, 1996.
[8] P. Mazumder and E.M. Rudnick, Genetic Algorithms for VLSI design, Layout & Test Automation.
Prentice Hall, New York, 1999.
[9] S. Roy, S. Bandyopadhyay, U. Maulik, and B.K. Sikdar, A genetic algorithm based state assignment
scheme for synthesis of testable FSMs targeting area and delay. International Journal of Engineering
Intelligent Systems, 10, 45–52, 2002.
[10] U. Maulik and S. Bandyopadhyay, Fuzzy partitioning using real coded variable length genetic
algorithm for pixel classification. IEEE Transactions on Geosciences and Remote Sensing, 41,
1075–1081, 2003.
[11] U. Maulik and S. Bandyopadhyay, Genetic algorithm based clustering technique. Pattern
Recognition, 33, 1455–1465, 2000.
[12] S. Bandyopadhyay and U. Maulik, Non-parametric genetic clustering: Comparison of validity
indices. IEEE Transactions on Systems, Man, and Cybernetics Part-C, 31, 120–125, 2001.
[13] G.B. Fogel and D.W. Corne, Evolutionary Computation in Bioinformatics. Morgan Kaufmann, San
Francisco, CA, 2003.
[14] S. Bandyopadhyay, An efficient technique for superfamily classification of amino acid sequences:
Feature extraction, fuzzy clustering and prototype selection. Fuzzy Sets and Systems, 152, 5–16,
2005.
[15] K. Dev and C. Siva Ram Murthy, A genetic algorithm for the knowledge base partitioning problem.
Pattern Recognition Letters, 16, 873–879, 1995.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C014” — 2005/8/6 — 14:08 — page 218 — #10

 15
Solving Optimization
Problems in Wireless
Networks Using
Genetic Algorithms

15.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-219

15.2 Basics of Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-220
15.3 Location Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-221
15.4 Channel Assignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-224
15.5 Call Admission Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-227
15.6 QoS-based Multicast Routing Protocol . . . . . . . . . . . . . . . . 15-228
Sajal K. Das Underlying Routing Algorithm • Improvement
Nilanjan Banerjee 15.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-231
Abhishek Roy References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-232

15.1 Introduction
The surging demand for mobile communications and the emerging field of ubiquitous computing require
enormous development in the realm of wireless networking. Although many problems in mobile wireless
networking can be successfully solved using techniques borrowed from wireline networks, there exists
some problems that are very specific to the wireless domain and often computationally very difficult or
sometimes even intractable. This is primarily due to the inherent limitations of wireless communications,
such as scarce bandwidth, high bit error rate, or location uncertainty. Most of these problems can be
mapped to classical optimization problems, with the goal of optimizing some objective functions, say
resource utilization, while satisfying all the constraints imposed by the wireless communication systems.
The constraints make most of the problems NP-complete or NP-hard [1]. There are two broad approaches
to solve such problems. One is to directly compute the exact solution based on the constraints, using a brute
force technique. However, this approach is often infeasible in a large-scale problem domain. The other
approach is to use a heuristic-based solution that can be computed in feasible time. Although a heuristic
may not yield the optimal solution, a carefully designed heuristic may produce a near-optimal solution
with low computational complexity. Various attempts have been made to develop efficient heuristics that
range from calculus-based search methods to random search techniques. The paradigm of “evolutionary

15-219

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 219 — #1

15-220 Handbook of Bioinspired Algorithms and Applications

computing or programming” essentially provides such heuristics for solving computationally difficult
problems, by mimicking the evolution process that the nature uses to eliminate weak forms of life.
An important example of evolutionary computing is genetic algorithm (GA), which is a guided, random
search technique. In GA, the less fit (or incompetent) solutions are replaced by better solutions produced
by applying some genetic operators on the existing ones in the solution space. This evolutionary learning
process progressively refines the search space and makes the problem computationally manageable, thus
enabling the search procedure to converge quickly and resulting in a near-optimal solution in feasible time.
GA has been successfully applied to solve several optimization problems in the wireless domain [2–4]. The
common objective of such problems is the optimal usage of scarce and hence costly wireless resources,
such as bandwidth. For example, in References 3 and 5, GA has been used to find an optimal location
management strategy for cellular networks. The goal here is to determine an optimal policy for location
updates of a mobile host, which ensures minimal paging and signaling overhead, thereby conserving the
costly wireless bandwidth. Whereas this work considers only a single optimization criteria (the location
management cost), a multi-objective GA-based location management framework has been proposed in
Referece 6 that considers other optimization factors, such as the load balancing between various mobility
routers in an Universal Mobile Telecommunication Systems (UMTS) [7] network. Channel assignment is
another challenging problem in cellular networks. The entire spectrum of wireless bandwidth is divided
into several channels and assigned to the cells, with the constraint that no two adjacent cells are assigned the
same channel, which ensures that there is no co-channel interference. A single cell may have more than one
channels. The channel assignment problem can be shown to be equivalent to the graph coloring problem,
and hence NP-hard [1]. GA-based channel assignment algorithms [2] provide a scalable mechanism to
determine the least number of channels required for a particular cellular network, while satisfying the
constraint related to the co-channel interference problem in adjacent cells. Due to the limited resource
available in wireless networks, the admission of new calls must be monitored to ensure a minimum quality
of service (QoS) for the already admitted calls. This is done by call admission control algorithms, which
essentially select, from a large number of possible admission policies, the optimal one with overall service
quality improvement and minimum call blocking rate. However, the number of possible policies can be
very large such that finding the optimal policy can be computationally intractable. Again, GA provides
a heuristic-based approach [4] to yield a near-optimal call admission policy that ensures QoS for the
admitted calls while minimizing the call blocking rate. GA has also been used in solving QoS routing
problems in networks. It can be shown that unicast routing with two or more QoS constraints can be
mapped to a NP-hard problem. Naturally, the problem is also NP-hard for multicast routing. The imprecise
network state information in dynamic wireless networks further complicates the problem. In Reference 8,
GA has been used to produce near-optimal solutions in computationally feasible time, for multicast
routing problem in networks with imprecise state information, while satisfying more than one QoS
constraints.
The goal of this chapter is to present a review of GA-based solutions of the above problems. The rest of
this chapter is organized as follows. Section 15.2 reviews the basic concept of GAs and their interpretation.
The location management problem is discussed in Section 15.3. Section 15.4 presents a GA-based solution
for optimal channel allocation in cellular networks. Efficient call admission control in cellular networks
is described in Section 15.5, while the design and performance analysis of an efficient multicast routing
protocol is described in Section 15.6. Finally, Section 15.7 concludes the chapter.

15.2 Basics of Genetic Algorithms

GAs are stochastic algorithms, whose search methods model a natural phenomenon, genetic evolution.
In evolution, the problem each species faces is one of searching for beneficial adaptations to complex
changing environments. The “knowledge” that each species has gained is embodied in the makeup of the
chromosomes of its members. In this light, GAs can also be classified as a learning technique. They have

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 220 — #2

Solving Optimization Problems in Wireless Networks Using GAs 15-221

FIGURE 15.1 A generic framework for genetic algorithms.

been successfully applied to tackle optimization problems such as scheduling [9], adaptive control [10],
game playing [11], cognitive modeling [12], transportation problems [13], traveling salesman problems
[14], database query optimization [15], and so on. Although GAs belong to the class of probabilistic
algorithms, they differ from random search algorithms because they combine elements of directed and
stochastic search. Therefore, GAs are also more robust than purely directed search techniques [16].
As an evolutionary programming model, a GA for a particular problem must have the following five
properties:
• A genetic representation of the potential solution space to a symbolic domain, such as a strings of
symbols, which form the so-called chromosomes.
• A methodology to create an initial population of potential solutions.
• An evolution function that plays the role of the environment, rating solutions in terms of their
“fitness.”
• Genetic operators that alter the structure of chromosomes.
• Values for various parameters such as population size, probabilities of applying genetic operat-
ors, etc.
Figure 15.1 shows a pseudo-code depicting the generic framework of any GA. An initial random
population is created and each member of the population is evaluated using a fitness function. Three
major genetic operators, selection, crossover, and mutation, are then successively applied to the members
of the population in each generation until the stopping criteria is met or there is little change in the quality
of the best solution in the entire population. Selection is the operation that increases the number of good
solutions or chromosomes in a population; usually the best chromosome in each phase of evolution is
selected. The crossover operation generates new representative chromosomes containing the properties
of more than two chromosomes. Finally, mutation changes or mutates a portion of a chromosome to
introduce a new chromosome in the population. When there is little or no change in the quality (or
fitness) of the chromosomes in the population with successive application of GA operators, the best one
is selected to represent the final solution.

15.3 Location Management

The location management problem is defined as the tracking of a mobile user in cellular wireless networks.
As shown in Figure 15.2, a cellular network is divided into discrete cells, each served by a base station
(BS). Mobile users communicate with a BS using wireless links. A number of such BSs are linked to
a mobile switching center (MSC), which is connected to the existing wireline networks. The location
management problem has two flavors: one on the part of the mobile device and the other on the part
of the network system. In a stand-by mode, every mobile device is made to report its location in regular
intervals. This process is known as registration or update. On the arrival of a new call to the mobile
device, the network system initiates a search for the device by polling the cells, where it could be possibly

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 221 — #3

15-222 Handbook of Bioinspired Algorithms and Applications

Wired backbone

Cells Base station (BS)

Wired connection Wireless links

Mobile switch center (MSC)

FIGURE 15.2 Overview of a cellular system.

located. This procedure is known as paging. By giving a low upper bound on the maximum number of
cells that can be polled, the paging process reduces the paging cost, but requires more frequent updates,
thus increasing the update cost. On the other hand, a reduction of number of updates essentially decreases
the update cost, but increases the location uncertainty and the subsequent paging cost. The essence of an
optimal location tracking thus lies in minimizing the combined update and paging costs associated with
a particular mobile device. The concept of using a suitable GA-based approach to minimize the average
per-user location management cost was first proposed in Reference 3.
A cellular network can be represented by a graph G(V , E), where a node represents the location areas
(LA — a logical representation of one or more cells) and an edge represents the access paths between a pair
of location areas. At this point, the most interesting question is whether a mobile device will issue an update
or not upon entering a new LA. Let δi represent an update decision variable for a user in LAi (1 ≤ i ≤ M ),
such that δi = 1 or 0 depending on whether or not an update has been issued. Now, if Cost p (i) and Costu
respectively denote the cost associated with paging LAi and issuing a single update, then the total location
management cost (LMC) is determined by taking a weighted average of the LMCs in the individual LAs.

(δi )
This is mathematically represented by: LMC = M i=1 i × LMi , where i is the normalized weight
associated with the LMC, LMi , of LAi . The average LMCs for both cases (δi = 1 and δi = 0) depend on
the individual update cost, paging cost, call arrival rate, and the user’s residence time in the LA. Assuming
a Poisson arrival with rate λ, geometrically distributed (with parameter pi ) residence time, it has been
(1) (0)
shown in Reference 3 that LMi = Costu + ( pλi )Costp (i) and LMi ≈ Cost0p (i) + ((λ/pi ) − 1)Costp (i).
It is now clear that an update strategy Us = [δi ] for the user constitutes a vector of decision variables,
having values 0 or 1 for all the LAs. The objective is to obtain optimal strategy Us∗ such that the LMC
is minimized. While enumerating all possible update strategies, the state space of the solution increases
exponentially with the number of LAs. Therefore, a GA is proposed in Reference 3 to obtain a (near-)
optimal solution for the location management problem.
As mentioned in Section 15.2, the first step in a GA-based approach is to map the state space into a
symbolic domain. The most obvious way is to represent each bit-string associated with a strategy Us by a
single chromosome (or genome). The length of every chromosome is equal to the number of LAs. A group
of strategies is chosen to form the initial population. For faster convergence, the relative proportion of

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 222 — #4

Solving Optimization Problems in Wireless Networks Using GAs 15-223

the number of 0s and 1s in the bit-string is chosen based on the call arrival rate and update cost. For
a relatively low call arrival rate and high update costs, it is wise to issue less frequent updates, thereby
resulting in more 0s than 1s in the chromosomes. Since, at every iteration, the GA inherently attempts
to increase the associated fitness values, the fitness function is chosen to be reciprocal of the total LMC,
that is, 1/LMC. The roulette wheel spinning selection [16] is used with elitism so that better chromosomes
will survive for the next iteration. After this selection, the crossover and mutation functions are executed
with probabilities 0.8 and 0.01 respectively. The fitness of the children are now evaluated and the entire
process is repeated.

Illustrative Example: At each iteration of the GA, the best chromosome, from the initialization phase till
that iteration cycle, is tracked. This gives the optimal (or near-optimal) solution at the termination of
the algorithm. The population size for each generation is kept constant at 50 and the number of bits
(i.e., the number of LAs) in the chromosome is chosen as 8. The cost function LMC is computed using the
steady-state transition probabilities between any two LAs. It was found that the GA converged very fast
to the near-optimal solution. The population size is kept constant at 20 and in all the cases the algorithm
converges to the optimal solution within 1000 generations. The best and average values of the fitness of
the chromosomes as well as the standard deviations are computed for each generation. A sample run of
the entire process with different generations having best and average fitness values is shown in Table 15.1,
which results in 11011 as the best chromosome having fitness 0.127 corresponding to LMC = 7.87 units.
Table 15.2, on the other hand, represents the optimal update strategies for different call arrival rates and
update/paging cost ratios.
Recently, the location management problem is also investigated in Reference 5 using a combination of


N −1
cellular automata (CA) and GAs. The total LMC is estimated as r i∈C wmi + j=0 wcj .v(j), where r
is the update-to-paging cost ratio, C is total number of reporting cells, that is, the number of cells from
which at least one update is issued, wmi represents frequency of movement into a cell i, wcj represents
frequency of call arrival within a cell j, and v(j) represents the vicinity (neighborhood) of cell j. Note that

TABLE 15.1 Sample Run of Genetic

Algorithms

Generation no. Best fitness Average

1 0.102 0.102
3 0.106 0.103
7 0.107 0.105
15 0.120 0.106
20 0.127 0.116
35 0.127 0.116
50 0.127 0.125
100 0.127 0.122
200 0.127 0.126
300 0.127 0.127

TABLE 15.2 Optimal Update

Strategies

Cu /Cp Optimal strategy

λ = 0.3 λ = 0.5 λ = 0.8
0.10 00000 11111 11111
0.33 00000 11111 11011
0.50 00000 01101 10011
0.67 00000 01110 10001
0.75 00000 01110 10001
0.90 00000 11010 10001
1.00 00000 11010 10001

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 223 — #5

15-224 Handbook of Bioinspired Algorithms and Applications

States 0, 1 or 2

States 0 or 1

FIGURE 15.3 Location management using cellular automata.

cellular automata is a decentralized, discrete space–time system and the state of every cell is governed by
its surrounding cells. Each cellular unit of cellular automata is associated with each cell in the network.
Each cell is represented either by “1” or “0” depending on whether or not it is a reporting cell. This leads
to two possible states for each cellular unit in the CA. Considering the hexagonal cells of the network,
the maximum neighborhood of every cell is taken as 6. For cells with less than 6 neighbors, dummy cells,
represented by “2” are added. Thus, as shown in Figure 15.3, each cell itself has 2 possible states (0 or 1) and
each of its 6 neighbors has 3 possible states (0, 1, or 2), implying a total number of 37 × 2 = 1458 possible
neighborhood states. Hence, the corresponding rule is of length 1458 bits and there can be a total of 21458
transition functions. Genetic algorithms are used to search for the best one from this exponential number
of transition rules. In Reference 5, an initial population of 1000 rules are created with a random value
for each rule. At every iteration, a new set of CA test data is generated. The set consists of four randomly
generated reporting cells configuration. The fitness function is chosen as the sum of the LMCs of all these
four configurations. The selection strategy is used to get the minimum fitness values (minimum location
management costs). A two-point crossover [16] with probability 0.8 and a mutation with probability 0.01
is used to achieve better solutions at every iteration. A set of 80 rules (chromosomes) have been used for
elitism, while crossover and mutation have been applied on the rest of the 920 rules. Simulation results
on different 4 × 4 and 5 × 5 cellular networks result in a cost per call arrival between 12.25 and 16 within
200 generations.
A careful look at both of the above mentioned location management strategies reveals that the only
objective of both the schemes is to reduce the overall LMC. However, more recently, the problem of
location management is combined with load balancing constraints, thereby introducing the notion of
multi-objective optimization techniques. Genetic algorithms can be easily extended to solve such multi-
objective optimization problems. A multi-objective, hierarchical mobility management optimization for
UMTS networks has been proposed in Reference 6, which balances the load among various unique logical
areas in the network, such as the LA, and the mobility routers, such as the mobile switch centers (MSCs),
Radio Network Controller (RNC), and Serving GPRS Node (SGSN), while minimizing the signaling
cost for location update. A schema-based niched Pareto GA [17] has been used that deals with multiple
objectives by incorporating the concept of Pareto domination in its selection operator, and applying
a niching pressure to spread out its population along the Pareto optimal tradeoff surface. The fitness
function or the cost functions used are the RA load balancing and the intra-SGSN signaling cost, which
covers the intra-SGSN routing area update and paging cost.

15.4 Channel Assignment

A major challenge associated with cellular wireless networks is the limitation on available bandwidth. In a
time division multiple access (TDMA) scheme, the wireless bandwidth is divided into different channels

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 224 — #6

Solving Optimization Problems in Wireless Networks Using GAs 15-225

F2
F7 F3
F1

F6 F4
F2 F5
F7 F2
F3
F1 F7 F3
F6 F4 F1

F5 F6 F4
F5

FIGURE 15.4 Channel reuse in cellular systems.

and each cell is allocated to a certain number of channels depending on its traffic density. In the cellular
system, the same frequency cannot be used in the adjacent cells, as there will be co-channel interference.
The hexagonal cell structure creates a cluster of 7 cells, where the frequencies will differ from each other.
Thus, channels used by a cell can be reused by a different cell belonging to a different cluster (sufficiently
far apart), so that the interference is bounded by some specific value. This is demonstrated in Figure 15.4,
where Fi represents the different frequencies used in the corresponding cell.
The channel assignment technique [2] can be static or dynamic. In a static assignment, a fixed num-
ber of channels are allocated to each cell and are estimated by the traffic demand in that cell. On the
other hand, in a dynamic allocation strategy, the channels are assigned on demand, and no cell has
exclusive control on any channel. Channel assignment strategies need to tackle two types of channel
interferences: (i) co-channel interference and (ii) adjacent channel interference. Co-channel interference
occurs when a different signal is received in the same channel as the original signal and therefore can-
not be eliminated by the receiver itself. Adjacent channel interference, on the other hand, is caused
by inadequate/incomplete filtering of unwanted modulation products in frequency modulation systems,
improper tuning, or poor frequency control in either the reference channel or the interfering chan-
nel, or both. The channel assignment strategy needs to maintain both the co-channel and the adjacent
channel interference below a tolerable limit. Most of the research in this area is focussed on finding
the minimum bandwidth spectrum to satisfy a given traffic load, while maintaining the interference
constraint.
Let the available frequency spectrum consist of M consecutive channels f1 , f2 , . . . , fM . Also, let N , di ,
and cij respectively represent the number of cells, channel demand of cell i, and frequency separation
needed between any two channels allotted to a pair of cells i and j. The matrices formed by the elements
cij and di , (1 ≤ i, j ≤ N ), are denoted as C and D, respectively. Intuitively, cij = 1 and cij = 2, respect-
ively mean that the same and adjacent frequencies cannot be assigned to certain channels. The channel
assignment strategy now needs to find out a suitable frequency assignment matrix F = [fmi ]M×N , where
fmi is binary-valued, with 1 indicating that the frequency fm is assigned to the ith cell and 0 indicat-
ing that it is not. The problem now boils down to minimize the value of M such that the channels
can be allocated to all the cells without interference, that is, when fmi = flj = 1, |fmi − flj | ≥ cij and

M
m=1 fmi = di . It has been shown in Reference 18 that the channel assignment problem is equivalent
to the graph coloring problem (NP-hard) [1], even when only the co-channel constraints are considered,
that is, when the matrix C is binary-valued. The actual channel assignment problem considered here is
even more complex and the solution space increases exponentially with the increasing number of cells.
The role of evolutionary algorithm now comes into play to obtain a (near-) optimal solution [2] in a
polynomial time.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 225 — #7

15-226 Handbook of Bioinspired Algorithms and Applications

The first step to solve this problem using GA is again to generate the initial pool of encoded strings or
chromosomes. Let, S1 , S2 , . . . , SP represent the P strings of an initial population. Each Si is an M × N
matrix, where M and N are number of frequencies and number of cells, respectively. The elements of the
matrix Si can have values 0, 1, −1, or 9. The interpretation of these four different values are enumerated
here:
1. 0: Cell (given by the column number) is not using the frequency (row number), and even if it uses
that frequency, there will be no conflict with other existing allocations.
2. 1: Cell is using the particular frequency.
3. −1: Cell is not using the frequency (row number), and cannot use that frequency for possible
interference.
4. 9: It is used at the head of all unused channels (rows).
Illustrative Example: Table 15.3 demonstrates an example of a valid solution for the four-node channel
allocation problem shown in Figure 15.5. The initial population has been created by using different
permutations of the nodes in the chromosomes, e.g., 1, 2, 3, 4, or 1, 3, 2, 4, or 3, 4, 1, 2, or 3, 1, 2, 4. The
fitness function is decided based on the total number of channels allocated, that is, the value of M. The
objective is to minimize this value of M. For chromosomes with equal M values, the one with more 0s is
selected as the better one. The reason is that the chromosomes having more 0s allow more channels to be
added, while satisfying the interference constraint. Now, out of the P × N columns, P × N × ρ columns
are selected at random, where ρ is the probability of mutation. From a selected column, randomly a
0 and 1 is chosen and flipped. An unavailability of a column with 0 or 1 leads to a mutation failure. If
the mutation results in a row with all 0 values, a leading 9 is placed, reducing the bandwidth requirement
by 1. Simulation results show that the lower bound for the minimum number of frequencies range from
∼300 to 500 for different types of demands. The GA approach takes only 1 to 3 generations (fraction of a
second) to converge in a near-optimal solution.

TABLE 15.3 Examples of Valid Solutions

solution
+1 −1 +1 −1
−1 −1 −1 −1
−1 −1 −1 +1
−1 −1 −1 −1
−1 +1 −1 −1
−1 −1 0 −1
−1 −1 −1 −1
−1 −1 −1 +1
0 −1 −1 −1
0 0 0 −1
0 0 0 −1
0 0 −1 −1
0 −1 −1 +1
9 0 −1 −1
9 0 0 −1

5 4 0 0 1
4 5 0 0 1
C= D=
0 0 5 0 1
0 1 2 5 3

FIGURE 15.5 A four-node channel allocation problem.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 226 — #8

Solving Optimization Problems in Wireless Networks Using GAs 15-227

15.5 Call Admission Control

An important issue in cellular systems is to maintain a good service quality. Admission control policies
are designed to meet this end. A wise admission decision for new calls can improve the overall service
quality by guaranteeing QoS to the already admitted calls, while having a minimal average blocking rate
for new calls. An admission policy is a binary decision controlling the admission criteria. Thus, any such
policy can be represented by a finite number of logic functions forming the policy space. However, the
enormously large policy state space often needs some compact coding [4] to restrict the blow up of search
space. There is a finite possibility for the coded policies to get stuck at local optimal solutions. Multimodal
optimization strategies are required to solve such problems. Genetic algorithms provide such a robust
optimization strategy that yields quite satisfactory results even when little is known about the search
space.
Following Reference 4, a network system S with n-dimensional state vector s = (s1 , . . . , sn ) is con-
sidered. The entire set of customers is divided into different classes depending on their relative importance.
A state represents the number of customers of each class in the system. It has been shown in Reference 4
that for a system described by a finite dimensional Markovian state vector, the call admission control
can be formulated as a Markov Decision Process (MDP). Such a model assumes a set of state-dependent
controls. Fixing a particular policy actually specifies the state transition matrix, a set of controls and
rewards depending on both the system state and state transitions. At every iteration, the current policy is
evaluated and using this evaluation a better policy is obtained. Each step however requires solving a set
of n linear equations. Each possible decision for each possible state is also evaluated. For this admission
control, the strategy needs O(n) evaluations per iteration. The computational effort becomes excessive
when the number of states is very high. Hence, GA has been used to reduce the computational complexity
and yield moderately good results. For this, a suitable coding is needed to represent such policy-based
decisions suitable for using in GAs.
If η represents an admission-decision and η
denotes the number of bits necessary to represent a
policy, then the total number of possible policies is 2η
. For smaller values, this policy can be directly
represented by the string called direct coding. However, for any practical system, the enormous size of
search space makes such direct coding impossible and intractable. A compact coding strategy of good
strings, which is also theoretically capable of coding any arbitrary string, is presented in Reference 4. It
introduces two different policy coding procedures. In block coding, a table of entries is built, with each
entry corresponding to a K -bit pattern. A number of such K -bit-patterns can be concatenated to form
full policy strings of length
η. The policy is now specified by the sequence of table indices, rather than
sequence of K bits. For any G < K a compression is achieved, and the total number of entries is 2G . In
program coding the states of the policies are assumed as inputs to a computer program. The policy coding
now specifies the bits of the program. Assuming infinite time and memory, the existence of a universal
computer or a Finite State Machine (FSM) can be modeled, which will provide the decision as the output.
These coded strings are represented by the chromosomes. A policy search scheme for two different
classes of policies are modeled using Markovian queuing systems [4]. If λ1 , λ2 , bk1 , bk2 , represents the
arrival rates and blocking probability of two different classes of calls, then for any policy k, the average
blocking is given by Pr[block] = (λ1 bk1 + λ2 bk2 )/(λ1 + λ2 ). Minimizing this Pr[block] is essentially
maximizing (1 − Pr[block]). This measure provides the performance of any policy i. For each generation
g , the performance performi of a policy i in the population is evaluated and the minimum performmin is
computed. The fitness function of policy i used is given by fitnessi = exp[−α(performi − performmin )],
where α > 0 is a constant scale factor constant that emphasizes or de-emphasizes deviation from the best
population performance, performmin . Experimental results show that it is useful to increase α with the
increase of g . Normal crossover (swapping the portions of parent policies) and mutation (flipping the
bits) operations are used to generate the offspring policies. The algorithm is capable of obtaining (near-)
optimal solutions within g = 1000 generations. Table 15.4 provides some sample results of direct, block,
and program codes for 10 and 50 cells respectively. The table demonstrates that the GA approach achieves
a near-optimal value of the blocking probability in almost all cases. Obviously, the time, iteration and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 227 — #9

15-228 Handbook of Bioinspired Algorithms and Applications

TABLE 15.4 Results of Admission Control

Cells Null policy Best policy Direct code (GA) Block code (GA) Program code (GA)
Pr[B] bits G Pr[B] bits G Pr[B] bits G
10 0.121 0.0753 0.076 400 132 0.077 200 100 0.076 24 < 40
50 0.216 0.072 0.097 2652 3600 0.078 510 800 0.072 48 < 40

length of strings varies for different coding strategies. Program codes perform the best in terms of the
number of generations and string lengths, followed by the block codes and direct codes.
An adaptive resource allocation and call admission control scheme, based on GA, was proposed in
Reference 19 for wireless ATM networks. Multimedia calls have their own distinct QoS requirements (e.g.,
cell loss rate, delay, jitter, etc.) for each of their substreams. The network usually allocates an appropriate
amount of resource that constitutes a certain QoS level, which remains fixed during a call. But such static
schemes are inefficient in terms of resource utilization. In the adaptive algorithm proposed in Reference 19,
each substream declares a range of acceptable QoS levels (e.g., high, medium, low) instead of just a single
one. With the variation of network resources, the algorithm selects the best possible QoS level that each
substream can obtain, while achieving maximum utilization of the resources and their fair distribution
among the calls. For example, in case of congestion, the algorithm tries to free up some resources by
degrading the QoS levels of some of the existing calls. The problem essentially boils down to finding the
best QoS levels for all existing calls amidst a large search space. Thus, if three types of streams, namely
audio, video, and data are considered with four possible QoS levels — “High,”“Medium,”“Low,” and “No”
Component — then a total number of 43 = 64 QoS levels are possible, and if we consider N existing calls,
the search space is given by 64N . For N = 10, the dimension of search space will be 6410 = 1.153 × 1018 .
A GA has been provided in Reference 19, which is a tool for searching an optimal solution in this huge
search space. Simulation results show that the algorithm is capable of searching solutions with a huge gain
(250%) in terms of the number of admitted calls, while achieving a resource utilization ranging between
85.87% and 99.985%.

15.6 QoS-based Multicast Routing Protocol

A majority of multimedia applications, such as video on demand and conferencing, depend on effi-
cient multicast protocols. Multicasting is essentially a selective broadcast of information from a source
to a given set of destinations. The information is delivered from the source to the destinations using
a multicast routing tree computed by multicast routing algorithm. Since the basic unicast routing
(route computation from a single source to a single destination) with multiple QoS constraints is
an NP-complete problem [20], the computation of a multicast route that satisfies certain QoS con-
straints is also computationally intractable. In wireless networks, fluctuations in available resources
complicates the problem further. Thus, multicast routing in wireless networks, with multiple QoS
constraints and imprecise information on the available network resources need some kind of heuristic-
based technique to compute an efficient multicast routing tree. In this section, we describe the design
and performance analysis of a multicast routing QoS-based multicast routing scheme [21], that relies
on the link-state advertisements and uses a GA framework for near-optimal multicast delivery tree
computation.
The problem thus is to compute a source-specific tree-based route, spanning the source and a given set
of destination nodes, such that the route satisfies multiple QoS parameters, such as minimum bandwidth
and end-to-end delay, and also complies with policy requirements such as minimum total bandwidth
utilization. Modeling the input network as a graph, with the routers as the nodes and the network links
as the edges of the graph, maps the QoS-based multicast routing problem to a constrained Steiner tree
computation problem, which is a well-known NP-hard problem [1]. In case of wireless networks, it
is rather difficult to obtain accurate values for these QoS parameters at any instant of time due to the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 228 — #10

Solving Optimization Problems in Wireless Networks Using GAs 15-229

time-varying nature of the wireless medium and the dynamism in the available resources. However, the
moments of their probabilistic distributions can be obtained by studying the history of their change over
a certain period of time, broadcasted with link state advertisements. Proper modeling with probability
distributions for each QoS parameter enables us to estimate the probabilistic bound corresponding to
a particular value of that parameter. This bound serves as the guarantee provided by the network, for
satisfying a particular value of a QoS parameter. Additionally, the tree should admit the maximum
number of sessions with such guarantees. This is achieved by efficient resource and traffic management
mechanisms.

15.6.1 Underlying Routing Algorithm

The network is represented by a graph G = (V , E), where V is a set of nodes and E is a set of links between
node pairs. A logical path between two nodes ν1 and νn is represented by ν1 , ν2 , . . . , νn , where νi ∈ V
for 1 ≤ i ≤ n. There can be multiple such paths between each pair of nodes. The primary objective is
to develop an efficient scheme that will find a near-optimal multicast routing tree with source as the root
and the destination as the leaves, such that
• The bandwidth provisioning for the specified QoS requirement is guaranteed.
• The end-to-end delay requirement is satisfied with maximal guarantee.
• The amount of bandwidth utilized is minimal.
The GA-based solution is encoded as follows:
• All the possible paths between the source and each destination are first computed and stored in a
pool. Note that a path is a sequence of network nodes. Among these paths, we select those paths
that conform to the bandwidth requirement.
• From each created pool, one solution is randomly chosen and concatenated with each other to
obtain a multicast solution.
In order to achieve optimality in terms of bandwidth availability, bandwidth utilization, and the end-to-
end delay guarantees, we devise an optimization function, which the GA can use in its course of action. The
transmission delay along each link is assumed to follow an exponential distribution. Hence, the available
bandwidth has been assumed to follow Poisson distribution [22] and the end-to-end delay on a path from
the source to any particular destination is assumed to follow Erlang-K distribution [22]. Based on these
assumptions, the probability distribution function for the end-to-end delay along a path from the source
to a single destination is given as Dpath (t ) = (d̄ κ t κ−1 e −d̄t /(κ − 1)!), where d̄ is the average delay on each
link and κ is the number of links on the path. Similarly, the probability of getting a bandwidth of B over a
link l is given by the probability density function, βl (B) = (b̄ B e −b̄ /B!), where b̄ is the average bandwidth
on each link. In addition to these measures, we have considered residual bandwidth as the third measure


of optimization. This is given by l∈T (ςl − σl ), where ςl is the capacity of a link l ∈ T , and σl is the
bandwidth allocated for all the paths in the multicast tree T , along the link l. The optimization function
is chosen as the linear combination of the probabilistic measure of the three QoS parameters. Hence the
fitness function is given by:


l∈T (ςl − σl )
F = path∈T Dpath (t ) + l∈T βl (B) +
.
l∈T ςl

Genetic algorithm-based operations, namely selection, crossover, and mutation operations are then
successively applied to the members of the population until there is little change in the quality of the
best solution in the entire population. The best solution replaces the worst solution of the previous
population. In crossover operations, the corresponding parts of two randomly selected multicast routing
trees are concatenated to obtain new routing trees. In mutation operations, a path in a multicast routing
tree is randomly selected and is replaced by another valid path between the same source and destination.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 229 — #11

15-230 Handbook of Bioinspired Algorithms and Applications

1 4 5

2 3 6

FIGURE 15.6 Example scenario.

Routes from 1 Routes from 1 Routes from 1 Routes from 1

to 4 to 5 to 6 to 7

12734 185 1236 127

1234 1845 1846 1237

12345 1276 18437

184

1854 12765 12736 12367

12364 18546 18467

184 185 1236 127

Solution string representing multicast route

FIGURE 15.7 Multicast solution selection strategy.

Illustrative Example: To illustrate the coding scheme of potential solutions, we study a network of eight
nodes, shown in Figure 15.6, where node 1 is the source of multicast delivery and 4, 5, 6, and 7 are
the destination nodes. Then the pools of valid paths for each source–destination pair are as shown
in Figure 15.7. The initial population, as required by the GA, is created as follows. Each member of
the population is formed by randomly selecting a path between each source–destination pair and then
concatenating them to represent a multicast tree spanning the source node and the set of destination
nodes. Figure 15.8 depicts the multicast delivery tree computed by the underlying routing algorithm for
the sample network, shown in Figure 15.8, with node 1 as the source and nodes 4, 5, 6, and 7 as the
destination nodes.

15.6.2 Improvement
The GA framework described here, combines the three QoS objectives into a single linear fitness function.
This scheme works well when only one solution is needed. But when multiple, mutually conflicting optim-
ization parameters are involved, it cannot yield solutions, which are better than the other with respect to a

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 230 — #12

Solving Optimization Problems in Wireless Networks Using GAs 15-231

Source node
1 4 5

2 3 6

Destination nodes (MSC)

Network links which belongs to the
multicast tree

FIGURE 15.8 Multicast delivery tree.

single optimization parameter, but not the superior when all the optimization parameters are considered.
In such a case, no solution is dominated by others when all the parameters are considered. These are
generally termed as nondominated or pareto-optimal solutions, generated by using a multi-objective GA.
The GA-based multicast routing algorithm has been extended to incorporate a multi-objective QoS-
optimization mechanism [8]. The procedure does not combine the three predefined QoS parameters into
a single objective function but attempts to optimize each parameter individually, thereby providing a near-
optimal and non dominated set of solutions (i.e., multicast trees). The solution set consists of not only
those trees that offer best delay, bandwidth requirement, and residual bandwidth guarantee individually,
but also a set of trees compromising fairly between the three optimization parameters.
Simulation results demonstrate that the algorithm is capable of obtaining a near-optimal multicast tree
in reasonable time. With a session arrival rate of 5–10 multicast sessions, the average session blocking
rate is only 2–5%. The multi-objective GA improves the flexibility of this scheme by offering a set of
nondominated solutions. The user now has the flexibility to choose his/her favorable solution from this
nondominated set. The dynamism and fluctuation of networks always creates resource uncertainty. This
might result in the unavailability of a particular QoS-constrained path. The nondominated solutions aids
in offering an alternate QoS-guaranteed path, thereby resulting in the graceful degradation scheme of QoS
provisioning. With multi-objective GA, it becomes possible to sustain more calls with their minimum level
of QoS.

15.7 Conclusion
In this chapter, we have presented a survey of the computationally difficult problems specific to wireless
networking that have been solved using GA, a bio-inspired optimization algorithm. The primary objective
of each of the problems is to derive a near-optimal solution in a computationally feasible time. GA provided
a tool for deriving such solutions. Relevant performance results of the solutions for the problems have
been been presented to illustrate the efficiency of GAs in solving the complex problems. Potentially, GA
can be used in any of the optimization problems faced in the design and operation of wireless networks.
While, in this chapter, we have predominantly discussed problems from the cellular network domain,
recent developments in the area of ad hoc and sensor networks are posing new challenges, a majority of
which are difficult optimization problems. As a result, there have been some research attempts [23–25]

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 231 — #13

15-232 Handbook of Bioinspired Algorithms and Applications

to solve these problems using GA. We hope that this chapter will help in providing a clear idea on the
methodology of solving complex, real-life problems in the wireless domain using GAs.

References
[1] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of
NP-Completeness, W.H. Freeman and Company, San Francisco, CA, 1983.
[2] G. Chakraborty and B. Chakraborty. A genetic algorithm approach to solve channel assignment
problem in cellular radio networks. IEEE Midnight-Sun Workshop on Soft Computing Methods in
Industrial Applications, June 1999.
[3] S.K. Sen, A. Bhattacharya, and S.K. Das. A selective update strategy for PCS users. Wireless Networks,
5, 313–326, 1999.
[4] A. Yener and C. Rose. Genetic algorithms applied to cellular call admission: Local policies. IEEE
Transactions on Vehicular Technology, 46, 72–79, 1997.
[5] R. Subrata and A.Y. Zomaya. Evolving cellular automata for location management in mobile
computing networks. IEEE Transactions on Parallel and Distributed Computing, 14, 13–26, 2003.
[6] T. Ozugur, A. Bellary, and F. Sarkar. Multiobjective hierarchical 2G/3G mobility manage-
ment optimization: niched Pareto genetic algorithm. In Proceedings of Globecom, Vol 6, 2001,
pp. 3681–3685.
[7] 3GPP - UMTS Standards. http://www.3gpp.org/ftp/tsg_cn/ TSG_CN/TSGN_03/Docs
[8] A. Roy and S.K. Das. QM 2 RP: A QoS-based Mobile Multicast Routing Protocol. ACM/Springer
Wireless Networks (WINET), 10(3), 271–286, 2004.
[9] S.J. Beaty. Genetic algorithms and instruction scheduling. In Proceedings of the 24th annual
international symposium on Microarchitecture, pp. 206–211
[10] K.Ng and Y.Li. Design of sophisticated fuzzy logic controllers using genetic algorithms. In
Proceedings of the IEEE World Congress on Computational Intelligence, 1994.
[11] C.T. Sun and M.D. Wu. Multi-stage genetic algorithm learning in game playing. In Proceedings
of the First International Joint Conference of the North American Fuzzy Information Processing
Society, the Industrial Fuzzy Control and Intelligent Systems, and the NASA Joint Technology, 1994,
pp. 223–227.
[12] C.L. Karr. Genetic algorithms for modelling, design, and process control. In Proceedings of the
second international conference on Information and knowledge management, 1993, pp. 233–238.
[13] Z. Michalewich. A genetic algorithm for the linear transportation problem. IEEE Transactions on
Systems, Man, and Cybernetics, 21, 445–452, 1991.
[14] C.A.R. Jahuria. Hybrid genetic algorithm with exact techniques applied to TSP. In proceedings of
the Second International Workshop on Intelligent Systems Design and Application, 2002, pp. 119–124.
[15] J.J. yang and R.R. Korfhage. Query optimization in information retrieval using genetic algorithms:
Report on the experiments of the TREC project. In Proceedings of TREC’1. NIST, 1993, pp. 31–58.
[16] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning ”, Addison-Wesley
Longman Publishing Co., Reading, MA, 1991.
[17] J. Horn, N. Nafpliotis, and D.E. Goldberg, “ A niched pareto genetic algorithm for mutiobjective
optimization. In IEEE Conference on Evolutionary Computation, New Jersey, vol. 1, 1994, pp. 82–87.
[18] W.K. Hale. Frequency assignment: Theory and applications. Proceedings of IEEE, 68, 1497–1514,
1980.
[19] M. Sherif, I. Habib, M. Naghshineh, and P. Kermani. An Adaptive Resource Allocation and
Call Admission Control Scheme for Wireless ATM Using Genetic Algorithms. In Proceedings
of Globecom, 1999, pp. 1500–1504.
[20] S. Chen and K. Nahrstedt. An overview of quality of service routing for next-generation high-speed
networks: Problems and solutions. IEEE Network, Special Issue on Transmission and Distribution
of Digital Video, 12, 64–79, 1998.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 232 — #14

Solving Optimization Problems in Wireless Networks Using GAs 15-233

[21] N. Banerjee and S.K. Das. Fast determination of QoS-based multicast routes in wireless networks
using Genetic Algorithm. IEEE International Conference on Communications (ICC), 8, 2588–2592,
2001.
[22] R. Nelson. Probability, Stochastic Process and Queuing Theory. Springer-Verlag, Heidelberg, 1995.
[23] L. Barolli, A. Koyama, and N. Shiratori. A QoS routing method for ad-hoc networks based
on genetic algorithm. In Proceedings of International Workshop on Database and Expert Systems
Applications, 2003, pp. 175–179.
[24] D. Turgut, S.K. Das, R. Elmasri, and B. Turgut. Optimizing clustering algorithm in mobile ad
hoc networks using genetic algorithmic approach. In Proceedings of IEEE GLOBECOM, 2002,
pp. 62–66.
[25] Q. Wu, S.S. Iyengar, N.S.V. Rao, J. Barhen, V.K. Vaishnavi, H. Qi, and K. Chakrabarty. On comput-
ing the mobile agent routes for data fusion in a distributed sensor network. IEEE Transactions on
Knowledge and Data Engineering, 1(6), 740–753, 2004.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C015” — 2005/8/8 — 17:14 — page 233 — #15

 16
Medical Imaging and
Diagnosis Using
Genetic Algorithms

16.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-235

16.2 Preliminaries on Genetic Algorithms. . . . . . . . . . . . . . . . . . . 16-236
Basic Principles and Features • Encoding Strategy and
Population • Evaluation Technique • Genetic Operators •
Parameters of GA
16.3 Genetic Algorithms for Equipment Design for
Medical Image Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-239
16.4 Genetic Algorithms for Medical Diagnosis . . . . . . . . . . . . 16-240
Image Driven Diagnosis • Data Driven Diagnosis
Ujjwal Maulik 16.5 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-247
Sanghamitra Bandyopadhyay Acknowledgment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-247
Sajal K. Das References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-247

16.1 Introduction
The last half of the twentieth century has seen a vigorous growth in the field of digital image processing
(DIP) and its potential applications. DIP deals with the manipulation and analysis of images that are
generated by discretizing the continuous signals. One important area of application that has evolved from
the 1970s is that of medical images. Rapid development in different areas of image processing, computer
vision, pattern recognition, and imaging technology, and the transfer of technology from these areas to
the medical domain has changed the entire way of looking at clinical routine, diagnosis, and therapy. Also,
the need for more effective and less (or non) invasive treatment has led to a large amount of research for
developing what may be called computer aided medicine.
Most modern medical data are expressed as images or other types of digital signals. The explosion in
computer technology in recent years introduced new imaging modalities such as x-rays, magnetic reson-
ance imaging (MRI), computer tomography (CT), positron emission tomography (PET), single photon
emission computed tomography (SPECT), electrical impedance tomography (EIT), ultrasound, and so on.
These images are noninvasive and offer high spatial resolution. Thus the acquisition of a large number of
such sophisticated image data has given rise to the development of quantitative and automatic processing
and analysis of medical images (as opposed to the manual qualitative assessment done earlier). Moreover,
the use of new, enhanced, and efficient computational models and techniques has also become necessary.

16-235

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 235 — #1

16-236 Handbook of Bioinspired Algorithms and Applications

A large amount of research is being devoted to the various domains of medical image processing, and
some surveys are already published [1,2]. However, in view of the vastness of the field, it has become
necessary to specialize any further survey work that is undertaken in this area, so that it can become
manageable and can be of more benefit to researchers/users. Some such attempts have already been made,
for example, specialization in terms of period of publication [3], image segmentation [4], registration [5],
virtual reality, and surgical simulation [6,7].
The area of designing equipments for better imaging and hence improvement in subsequent processing
tasks has also received the attention of researchers. The design problem has been viewed as one of
optimization, and therefore the use of efficient search strategies has been studied. The application of
genetic algorithms, a well-known class of search and optimization strategies, is also one of the important
areas that has been investigated in this regard.
Genetic Algorithms (GAs) [8,9] are randomized search and optimization techniques guided by the
principles of evolution and natural genetics, and have a large amount of implicit parallelism. They provide
near optimal solutions of an objective or fitness function in complex, large, and multimodal landscapes.
In GAs, the parameters of the search space are encoded in the form of strings called chromosomes. A fitness
function is associated with each string that represents the degree of goodness of the solution encoded in
it. Biologically-inspired operators such as selection, crossover, and mutation are used over a number of
evolutions (generations) for generating potentially better strings.
The important fallout of (semi-) automated medical image processing tasks is enhanced diagnosis.
Several tasks in the area of medical diagnosis have also been modeled as an optimization problem, and
researchers have used GAs for solving them. In this chapter, we attempt to provide a state-of-the-art
survey in the application of the principles of GAs, an important component of evolutionary computation,
for improving medical imaging and diagnosis tasks. Section 16.2 describes the basic principles of GAs.
Thereafter, the use of GAs in improving equipment design has been studied. Finally, the application
of GAs for computer aided diagnosis, including schemes driven by both image and data (consisting of
information not derived from images), is provided.

16.2 Preliminaries on Genetic Algorithms

Genetic algorithms [8–10] are efficient, adaptive, and robust search and optimization processes that
are usually applied to very large, complex, and multimodal search spaces. They are modeled on the
principles of natural genetic systems, in which the genetic information of each individual or potential
solution is encoded in structures called chromosomes. They use some domain or problem-dependent
knowledge for directing the search in more promising areas of the solution space; this is known as the
fitness function. Each individual or chromosome has an associated fitness function, which indicates its
degree of goodness with respect to the solution it represents. Various biologically-inspired operators such
as selection, crossover, and mutation are applied on the chromosomes to yield potentially better solutions.

16.2.1 Basic Principles and Features

Genetic algorithms emulate biological principles to solve complex optimization problems. It essentially
comprises a set of individual solutions or chromosomes (called the population), and some biologically-
inspired operators that create a new (and potentially better) population from an old one. According to the
theory of evolution, only those individuals in a population who are better suited to the environment are
likely to survive and generate offspring, thereby transmitting their superior genetic information to new
generations.
The essential components of GAs are the following:

• A representation strategy that determines the way in which potential solutions will be encoded to
form string like structures called chromosomes.
• A population of chromosomes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 236 — #2

Medical Imaging and Diagnosis Using Genetic Algorithms 16-237

Initialize the population

Perform the job with decoded versions of the strings

Compute fitness values

Yes
Termination criterion attained? Stop

No

Reproduce/select strings to create new mating pool

Generate new population by crossover and mutation

FIGURE 16.1 Basic steps of a genetic algorithm.

• Mechanism for evaluating each string (fitness function).

• Selection/reproduction procedure.
• Genetic operators (crossover and mutation).
• Probabilities to perform genetic operations.

A schematic diagram of the basic structure of a GA is shown in Figure 16.1. The components of GAs
are described in the following sections.

16.2.2 Encoding Strategy and Population

To solve an optimization problem, GAs start with the chromosomal representation of a parameter set,
which is to be encoded as a finite size string over an alphabet of finite length. Usually, the chromosomes
are strings of 0s and 1s. For example, the string

1 0 0 1 1 0 1 0

is a binary chromosome of length 8. It is evident that the number of different chromosomes (or strings)
is 2l , where l is the string length. Each chromosome actually refers to a coded possible solution. A set of
such chromosomes in a generation is called a population, the size of which may be constant or may vary
from one generation to another. A common practice is to choose the initial population randomly.

16.2.3 Evaluation Technique

The fitness/objective function is chosen depending on the problem to be solved, in such a way that the
strings (possible solutions) representing good points in the search space have high fitness values. This is
the only information (also known as the payoff information) that GAs use while searching for possible
solutions.

16.2.4 Genetic Operators

The frequently used genetic operators are selection, crossover, and mutation operators. These are applied
to a population of chromosomes to yield potentially new offspring. The operators are described in Sections
16.2.4.1 to 16.2.4.3.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 237 — #3

16-238 Handbook of Bioinspired Algorithms and Applications

6
7

5 8

4 3

FIGURE 16.2 Roulette wheel selection.

16.2.4.1 Selection
The selection/reproduction process copies individual strings (called parent chromosomes) into a tentative
new population (known as mating pool) for genetic operations. The number of copies that an individual
receives for the next generation is usually taken to be directly proportional to its fitness value; thereby
mimicking the natural selection procedure to some extent. This scheme is commonly called the propor-
tional selection scheme. Roulette wheel parent selection, stochastic universal selection, and binary tournament
selection [8,10] are some of the most frequently used selection procedures. Figure 16.2 demonstrates the
roulette wheel selection. The wheel has as many slots as the population size P, where the size of a slot is
proportional to the relative fitness of the corresponding chromosome in the population. An individual is
selected by spinning the roulette, and noting the position of the marker when the roulette stops. Therefore,
the number of times that an individual will be selected is proportional to its fitness (or, the size of the slot)
in the population. In the commonly used elitist model of GAs, thereby providing what is called an elitist
GA (EGA), the best chromosome seen up to the present generation is retained either in the population,
or in a location outside it.

16.2.4.2 Crossover
The main purpose of crossover is to exchange information between randomly selected parent chromo-
somes by recombining parts of their genetic information. It combines parts of two parent chromosomes
to produce offspring for the next generation. Single-point crossover is one of the most commonly used
schemes. Here, first of all, the members of the selected strings in the mating pool are paired at random.
Then each pair of chromosomes is subjected to crossover with a probability µc where an integer position
k (known as the crossover point) is selected uniformly at random between 1 and l − 1 (l > 1 is the string
length). Two new strings are created by swapping all characters from position (k + 1) to l. For example,
let the two parents and the crossover points be as shown below.

1 0 0 1 1| 0 1 0
0 0 1 0 1| 1 0 0

Then, after crossover the offspring will be the following:

1 0 0 1 1 1 0 0
0 0 1 0 1 0 1 0

Other common crossover techniques are two-point crossover, multiple point crossover, shuffle-
exchange crossover, and uniform crossover [9].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 238 — #4

Medical Imaging and Diagnosis Using Genetic Algorithms 16-239

The successful operation of GAs depends, to a great extent, on the coding technique used to represent
the problem variables [11,12]. The building block hypothesis indicates that GAs work by identifying good
building blocks, and by eventually combining them to get larger building blocks [8,13,14]. Unless good
building blocks are coded tightly, the crossover operation cannot combine them [15,16]. Thus coding–
crossover interaction is important for the successful operation of GAs. The problem of tight or loose
coding of problem variables is largely known as the linkage problem [17]. Recent work on linkage learning
GAs that exploits the concept of gene expression can be found in References 18 to 20.

16.2.4.3 Mutation
Mutation is the process by which a random alteration in the genetic structure of a chromosome takes
place. Its main objective is to introduce genetic diversity into the population. It may so happen that
the optimal solution resides in a portion of the search space that is not represented in the population’s
genetic structure. The process will therefore be unable to attain the global optima. In such situations,
only mutation can possibly direct the population to the optimal section of the search space by randomly
altering the information in a chromosome. Mutating a binary gene involves simple negation of the bit,
whereas that for real coded genes are defined in a variety of ways [10,21]. Here, we discuss the binary
bit-by-bit mutation, where every bit in a chromosome is subject to mutation with a probability µm . The
result of applying the bit-by-bit mutation on positions 3 and 7 of a chromosome is shown here.

1 0 0 1 1 0 1 0
1 0 1 1 1 0 0 0

16.2.5 Parameters of GA
There are several parameters in GAs that have to be manually tuned and fixed by the programmer.
Among these are the population size, probabilities of performing crossover and mutation, and the ter-
mination criteria. Several other things must also be determined by the programmer. For example, one
must decide whether to use the generational replacement strategy, in which the entire population is
replaced by a new population, or the steady state replacement policy where only the less fit individuals are
replaced. Most such parameters in GAs are problem dependent, and no guidelines for their choice exist
in the literature. Therefore, several researchers have also kept some of the GA parameters variable and/or
adaptive [22–24].
As shown in Figure 16.1, the cycle of selection, crossover, and mutation is repeated a number of times
till one of the following occurs:

1. The average fitness value of a population becomes more or less constant over a specified number
of generations.
2. A desired objective function value is attained by at least one string in the population.
3. The number of generations (or iterations) is greater than some threshold.

16.3 Genetic Algorithms for Equipment Design for

Medical Image Acquisition
Magnetic resonance imaging (MRI) is one of the most commonly used medical imaging techniques used
for the diagnosis of several ailments including multiple sclerosis, strokes, tumors, and other infections of
the brain, bones, spine, or joints, and for visualizing torn ligaments, soft tissues of the body, etc. Magnetic
resonance imaging is based on the phenomenon of nuclear magnetic resonance (NMR) discovered by
Felix Bloch and Edward Powell, for which they received the Nobel Prize in 1952. In an MRI scan, the
patient is placed inside the bore of a magnet, which is usually cubic in shape. The body part to be scanned

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 239 — #5

16-240 Handbook of Bioinspired Algorithms and Applications

is placed at the center (or, isocenter) of the magnetic field inside the bore. In this noninvasive procedure,
strong magnetic fields along with radio waves are used to visualize the structure of a particular body part.
The design of appropriate equipments for the purpose of good imaging may be considered as the first
step in medical image processing. For example, large superconducting solenoids with apertures of typically
1 m, highly uniform (20 ppm) central fields of 1–2T, and low fringe fields (5 Gauss at 5 m) are required
in clinical MRI. However, these magnets, which are now available in the clinical market, have deep bores,
typically between 1.8 and 2 m length, and have a number of disadvantages such as patient claustrophobia
and limited access for intervention. In order to overcome the limitations and evolve good designs for the
magnets, researchers have mapped the design problem to one of optimization and have investigated the
use of computational methods [25,26], including GAs [27–29], for this purpose.
Analytical techniques have been the preferred approach to design such magnets and gradient sets for
MRI. Such technique are computationally efficient but are approximate, particularly away from the axis
of symmetry. In Reference 30, an attempt has been made, which uses GA running on massively parallel
computers to design an actively shielder whole-body MRI solenoid magnet with a bore of 1 m. The task
is to optimize a cost function based on the magnetic field generated by a set of upto 20 circular coils, each
having upto 2000 turns. The coils are constrained to be concentric with the magnet, and are arranged in
symmetric pairs. A single coil is described by five parameters:

• Two parameters describing its position in the X –Z plane

• Depth of the coil
• Width of the coil
• The direction of the current

A chromosome encodes these five parameters per coil, for upto 20 coils. Since the coils are arranged in
pairs that are symmetric about the central X –Y plane of the magnet, only 10 of the coils are independent.
Thus a chromosome encodes upto 50 parameters that have to be optimally tuned. The magnetic field is
computed using the Biot–Savart law. The fitness function incorporates terms for uniformity of field in
the region of interest (ROI), and the smallness of the fringe field. The field is calculated by summing over
the contributions from the turns in each coil. Recombination is performed as a two-stage process. In the
first stage, a parent subset, which is of half the size of the population, is created by performing binary
tournament selection in the top eighth of the population. In the second stage, pairs of chromosomes from
the parent subset are picked at random and mated to produce a single offspring. The parents are replaced
in the subset, and this process is continued till the next generation is completed. Three types of mutations
are considered, which takes care of small perturbations, perturbations to add or remove a coil from the
design, and a drastic perturbation for adding extra diversity to the population. The initial result provided
in Reference 30 demonstrates the effectiveness of GAs for producing shorter whole-body magnet designs,
which are original and innovative.
Two-dimensional ultrasonic arrays provide the possibility of three-dimensional electronic focusing and
beam-steering and thus three-dimensional imaging. In simulation studies, it has been demonstrated [31]
that reducing the number of elements of a two-dimensional matrix array down to order eight, keeps
resolution and leads still to sufficient contrast. These random arrays are usually obtained by generating a
random pattern with the desired number of elements. In Reference 32, simulation is presented to show
that the imaging quality of a sparse tree can be improved by optimizing the random choice of elements.
The optimization is done using GA.

16.4 Genetic Algorithms for Medical Diagnosis

For diagnosis of diseases, several tests are performed on the patient, some of which may involve taking
various images such as x-rays, MRI scan, CT scan, etc., and some others that involve pathological and
other tests. Data from these tests may become quite large and conflicting, and manual interpretation by
collecting all such information may become difficult. This has therefore given rise to the development of

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 240 — #6

Medical Imaging and Diagnosis Using Genetic Algorithms 16-241

Low level image processing

Feature extraction
Medical image (e.g., noise removal, enhancement,
process
acquisition edge detection, segmentation etc.)

Knowledge encoding and/or

Domain expertise rule base design

Evaluation and decision

Interface with the

end user

FIGURE 16.3 Block diagram of a CAD system.

computer assisted diagnostic tools, which are intended to help the medical practitioners make sense out
of a large amount of data, and make diagnosis and therapy decisions. Figure 16.3 shows a block diagram
of a computer aided diagnosis system.

16.4.1 Image Driven Diagnosis

Computer aided detection and classification schemes have the potential of increasing diagnostic accuracy
in medical imaging. This may be done, for instance, by alerting radiologists to lesions that they initially
overlooked, or assisting in the classification of detected lesions [33]. As with any complicated pattern
recognition system, these schemes, generally referred to as computer aided diagnosis (CAD) schemes,
typically employ multiple parameters such as threshold values, filter weights, and ROI sizes to arrive at
a detection or classification decision. For the system to have a high performance, the values of these
parameters need to be set optimally. In general, the optimal set of parameters may change when a
component of the imaging chain is modified or changed. When the number of parameters is large, it
becomes very difficult to manually determine the optimal choice of parameter values, as some of the values
may be correlated in some unknown manner. However, conventional optimizing techniques, including
GAs, are designed to optimize a scalar objective function, and the task of optimizing the performance of
a diagnostic decision-making system is clearly multiobjective. Therefore, two objectives, increasing the
sensitivity and reducing the false-positive rate of the system, are described by a single objective function.
This objective function is defined as the weighted sum of the sensitivity and false-positive rate or using
area under the receiver operating characteristic (ROC) curve. The use of a niched pareto genetic algorithm
(NP-GA) [34] in training two popular diagnostic classifiers for optimizing their performance, has been
studied in Reference 35. Unlike conventional classifier techniques that formulate the problem as the
solution to a scalar optimization, the NP-GA explicitly addresses the multiobjective nature of the training
task. Traditional techniques of classifier training attempt to combine several objective functions into one,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 241 — #7

16-242 Handbook of Bioinspired Algorithms and Applications

so that conventional scalar optimization technique can be utilized. This involves incorporating a priori
information into the aggregation method so that the resulting performance of the classifier is satisfactory
for the task at hand. It has been shown in Reference 35 that the multiobjective genetic approach removes
the ambiguity associated with defining a scalar measure of classifier performance and that it returns a set
of optimal solutions that are equivalent in the absence of any information regarding the preference of
the objectives, that is, sensitivity and specificity. The a priori knowledge that is used for aggregating the
objective functions in conventional classifier training may instead be applied for post-optimization to select
from one of the series of solutions returned from multiobjective genetic optimization. This technique is
applied in Reference 35 to train a linear classifier and an artificial neural network, using simulated datasets.
The performance of the solutions returned from the multiobjective genetic optimization represents a series
of optimal pairs, sensitivity and specificity, which can be thought of as operating points on an ROC curve.
It is observed that all possible ROC curves for a given dataset and classifier are less than or equal to the ROC
curve generated by the NP-GA optimization. In Reference 36, a multiobjective approach for optimizing the
performance of two rule-based CAD schemes has been proposed. One of these CAD schemes is designed
to detect clustered microcalcifications in digitized mammograms, while the other scheme is developed to
detect the breast masses.
Diagnosis and follow up of pigmented skin lesions is an important step toward early diagnosis of skin
cancer [37]. For this purpose, digitized epiluminescence microscope (ELM) [38] images of pigmented
skin lesions is used. Epiluminescence microscopy is a noninvasive technique that uses an oil immersion to
render the skin translucent and make pigmented structure visible. During clinical diagnosis of pigmented
skin lesions, one of the main features is the lesion symmetry, which should be evaluated according to
its shape, color, and texture. This may be evaluated by drawing two orthogonal axes that maximize the
perceived symmetry [39]. The evaluation is binary, that is, the lesion is either symmetrical or asymmetrical.
In addition to the small number of possible outcomes, the evaluation is highly subjective and depends on
the physicians’ experience. As a result, the development of automatic techniques for the quantification of
symmetry and the detection of symmetry axes is necessary. Other methods based on principal component
technique for computing the axes may be found in References 40 and 41.
Reference 42 has proposed a GA-based technique and an optimization scheme derived from the self-
organizing maps theory for the detection of symmetry axes. The notion of symmetry map has been
introduced, which allows an object to be mapped to a symmetry space where its symmetry properties
can be analyzed. The objective function (ψ) used in Reference 42 is based on a given symmetry measure,
which is a function of the mean-square error (MSE) between the original and the reflected images. The
MSE is defined as follows:

MSE = E[||(x, y) − (x  , y  )||2 ] (16.1)

where (x, y) : [0, c − 1] × [0, r − 1] → IR n is the vector valued r × c input image, where the pixel
values are in an n dimensional space, (x, y) and (x  , y  ) represent the pixel coordinates and the symmetry
coordinates respectively. The input image can be decomposed into symmetric (s (x, y)) and asymmetric
(a (x, y)) components. Therefore, (x, y) = s (x, y) + a (x, y). The asymmetric component can be
considered as symmetric noise, and the MSE is proportional to its energy. The distortion due to noise can
be measured through the peak signal-to-noise ratio (PSNR), and is given by


(Nq − 1)2
PSNR = 10 log10 , (16.2)
MSE

where Nq is the number of quantization levels in the image. The fitness function ψ is defined as

1
ψ =1− . (16.3)
1 + PSNR

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 242 — #8

Medical Imaging and Diagnosis Using Genetic Algorithms 16-243

Real encoding of chromosomes has been used in Reference 42 along with a modified version of simu-
lated binary crossover [43]. The proposed technique is applied for detection and diagnosis of malignant
melanoma.
The development of computer supported systems for melanoma diagnosis is of great importance to
dermatologists due to its clinical accuracy in identifying malignant melanomas [44,45]. Several techniques
for computerized melanoma diagnosis are based on color images making use of image analysis methods
to quantify visual features as described by the “ABCD” rule (Asymmetry, irregular Border, varying Color,
Diameter) [37,46,47]. Laser profilometry opens up new possibilities to improve tumor diagnostics in
dermatology [48].
The recognition task is to classify surface profiles of melanomas and nevi also called moles. Due to the
fact that all profiles contain regions with a structure similar to normal skin, each profile is subdivided
into 16 non-overlapping quadratic subprofiles and image analysis algorithms are applied to each sub-
profile separately. Subsequently, feature selection algorithms are applied to optimize the classification
performance of the recognition system.
An efficient computer supported technique that uses GAs for diagnosis of skin tumors in dermatology is
presented in Reference 49. High resolution skin surface profiles are analyzed to recognize malignant melan-
omas and nevocytic nevi. In the initial phase, several types of features are extracted by two-dimensional
image analysis techniques characterizing the structure of skin surface profiles: texture features based on
co-occurrence matrices [50], Fourier features [51], and fractal features [37,52,53]. Subsequently, several
feature selection algorithms based on heuristic strategies, greedy technique, and GA are applied to determ-
ine suitable feature subsets for the recognition process by describing it as an optimization problem. As
a quality measure for feature subsets, the classification rate of the nearest neighbor classifier computed
with the leaving-one-out method is used. Among the different techniques used, GAs show the best result.
Finally, neural networks with error back-propagation as learning paradigm are trained using the selected
feature sets. Different network topologies, learning parameters, and pruning algorithms are investigated
to optimize the classification performance of the neural classifier. With the optimized recognition system,
a classification performance of 97.7% is achieved.
In the 1980s, microwave imaging was thought to have great potential in developing medical diagnostic
tools. However, because of the problem of inverse scattering, good reconstruction of images was found
to be difficult. Carosi et al. [54] have used the approach of focused imaging, where only a part of
the body is subjected to the investigation, combined with the capabilities of global search techniques
like GAs for accurate reconstruction. For experimental purpose, a human abdomen is considered and
different electromagnetic sources operating at the working frequency of 433 MHz and 2.54 GHz are used.
Numerical investigations are performed to define the optimal dimensions of the reduced investigation
domain. To quantitatively evaluate the effects of the reduction of the original investigation domain on
the inversion data, suitable relative errors are defined. Once the reduced domain is defined, preliminary
reconstructions are performed aiming to evaluate the imaging capability of GAs when a focussed approach
is used for tomographic application.
One of the most important facial paralysis diagnosis techniques is quantitative assessment of patient’s
facial expression motion. Johnson et al. [55] proposed a technique for achieving maximal regional facial
motion while the rest of the face is held at rest based on Maximal Static Response Assay of facial nerve
function. This requires the placement of removable adhesive dots and a small adhesive ruler on the
face at predefined locations. However, many facts, such as misplacement of facial dots, misplacement
of the grid, and reading and entry errors, will cause an error in the assay. Helling et al. [56] used
region-specific, subtracted, digitized image light reflectance as a two-dimensional marker for the complex
three-dimensional surface deformations of the face during expression. The velocity of region-specific
facial motion is estimated from the facial motion image sequences using the optimal flow (OF) technique.
The computation of the OF field requires estimates of both the spatial gradient and spatial time derivative
at each pixel, and this time-consuming process often limits its use, especially in medical application. To
overcome this problem, an OF technique based on GA is proposed in Reference 57 to detect facial motions
from dynamic image sequences. Experimental results demonstrate that the proposed technique is very

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 243 — #9

16-244 Handbook of Bioinspired Algorithms and Applications

useful to diagnose the site of facial paralysis and assess progression or recovery profiles of patients when
combined with other diagnosis techniques.

16.4.2 Data Driven Diagnosis

The previous section reported some research on development of diagnostic tools in which one of the
inputs considered is an image. In this section, we deal with some other diagnostic tools that use other
types of input data, mostly numeric.
Electromyography (EMG) is the recording and study of the electrical activity of voluntary contract-
ing muscles. Clinical EMG findings provide useful information in the electrodiagnostic examination of
peripheral nerves and skeletal muscle, and in deciding the level of the lesion in patients suffering from
neuromuscular disorders. It is also useful in deciding whether the symptom of muscle weakness in the
assessment of neuromuscular disorders is myopathic or neurogenetic in origin. The advantages of auto-
mated EMG diagnostic systems can be found in Reference 58. Different approaches have been used to
address the problem of automated EMG diagnosis. The utility of artificial neural networks in classifying
EMG data trained with back propagation or Kohonen’s self-organizing feature map algorithm has recently
been demonstrated in References 59 and 60. In Reference 61, a study has been made to investigate how
genetics-based machine learning (GBML) can be applied for diagnosing certain neuro muscular disorder
based on EMG data. The effect of GBML control parameters on diagnostic performance is also examined.
Subsequently, a hybrid diagnostic system is introduced that combines both the neural network and GBML.
In References 62 and 63, a methodology based on GAs for the automatic induction of Bayesian networks
from a file containing cases and variables related to the problem of predicting survival in malignant skin
melanoma is described. The structure is learned by applying three different techniques: the Cooper and
Herskovits metric for a general Bayesian network [64], the Markov blanket approach, and the relaxed
Markov blanket method. The methodologies are applied to the problem of predicting survival of people
after 1, 3, and 5 years of being diagnosed as having malignant skin melanoma. The induced Bayesian
network is used for classifying patients according to their prognosis of survival. These results are compared
to those obtained by the Naive–Bayes paradigm. An empirical comparison of Bayesian networks learned
using GAs, rule induction, and logistic regression is carried out in Reference 65 where the task is to
predict the survival of women suffering from breast cancer. In a more recent attempt, Blanco et al. [66]
have studied the problem of learning Bayesian networks using two stochastic, population-based search
algorithms: the univariate marginal distribution algorithm and population-based incremental learning.
Comparison with the GA-based scheme is also carried out.
The problem of Wisconsin breast cancer diagnosis (WBCD) has been considered in References 67 and
68, by combining fuzzy systems and evolutionary algorithms to design an automatic diagnosis system. The
proposed fuzzy-genetic approach produces systems exhibiting two prime characteristics. They attain high
classification performance with the possibility of attributing a new confidence measure to the output dia-
gnosis. Moreover, the resulting systems involve a few simple rules, and are, therefore, human interpretable.
Another approach for diagnosis of breast cancer by Bayesian networks can be found in Reference 69.
In medicine, prognostic models may be used to assess the likely prognosis of a patient, which, in turn,
may determine the treatment of that patient. In Reference 70, a prognostic model based on diffusion GAs
is sought to determine whether or not patients suffering from an uncommon form of cancer will survive.
The problem considered is a multiobjective one, in which the three objectives are:

• Maximize the correct number of survival predictions

• Maximize the correct number of death predictions
• Minimize the number of factors used

The motivation behind this study is to accurately predict the outcome so that patients who are more
likely to die can be identified at diagnosis, and subjected to high dose aggressive chemotherapy (which
has several negative side effects). On the other hand, those with a high chance of survival can be spared
this treatment and its side effects. Given a set of case histories, a technique is proposed in Reference 70

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 244 — #10

Medical Imaging and Diagnosis Using Genetic Algorithms 16-245

TABLE 16.1 Different Criteria and Their Ranges Used in Reference 70

Number Name Ranges

1 Brain metastases Yes, No

2 Liver metastases Yes, No
3 Placental site trophoblastic tumors (PSTT) Yes, No
4 Prior chemotherapy Yes, No
5 Pregnancy to term Yes, No
6 Age <26, 26–30, 31–41, >41
7 Serum hCG level <800, 801–29570, 29571–181000, >181000
8 Interval between pregnancy and diagnosis <4, 5–9, 10–33, >33

that attempts to find the relative weights of the different factors that are used to describe the cases. The
eight factors that are used in the prognostic model of Reference 70 and their possible ranges are provided
in Table 16.1.
A diffusion GA is used for building the prognostic model that will simultaneously optimize the three
objectives given here. A model is represented as a chromosome by encoding the weights associated with
each factor. Boolean criteria (1–5 in Table 16.1) have single associated weights, while criteria having a
range of values (6–8 in Table 16.1) have one weight per subrange. In addition a combination weight is used
in Reference 70, which can incorporate the possibility that a combination of factors might be important.
Thus a total of 18 weights are used, 5 for the Boolean criteria, 12 for the subranged criteria, and the
combination weight. In the diffusion GA, the individuals in a population are arranged along the vertices
of a square lattice. During crossover, each individual chooses one of its four neighbors randomly as the
mate. Mutation can either increase or decrease a weight by 10%, or set it to zero. Only if the result of
crossover and mutation at a position is better, in the Pareto optimal sense, than the original, the latter will
be replaced. Marvin et al. [70] experimented with a population size of 169 placed on a 13 × 13 grid. The
weights are randomly set in the range [−2000, 2000], and the algorithm is executed for 1000 generations.
It is found to predict 90% of the survivals and 87% of the deaths, while using 6 of the 8 factors, and 13
of the possible 18 weights when the entire data set is used. For training using 90% data and testing using
the remaining 10%, several prognostic models are obtained each coming up with a different compromise
among the three objective values. Significantly, the method in Reference 70 enables a simple model to be
evolved, one that produces well-balanced predictions and one that is relatively easy for clinicians to use.
Ngan et al. [71] employ evolutionary programming (EP) and genetic programming (GP) in the domain
of knowledge discovery in medical systems. Evolutionary Programming is used to learn Bayesian networks,
which is known to be an intractable problem. Minimum description length principle is used to measure the
goodness of solution in EP followed by the use of GP to learn rules. The entire knowledge discovery system
is applied on limb fracture and scoliosis data, where it is able to detect many interesting patterns/rules
that were uncovered earlier. Ngan et al. had earlier used GP for discovering comprehensible rules in the
medical domain; they used grammar to restrict the seach space, and to ensure the syntactical correctness of
the rules [72]. The discovered rules were evaluated within the framework of support confidence proposed
for association rule mining. Here, a major limitation was that the grammar was application dependent,
and had to be written for each application domain.
Genetic programming is also used to discover comprehensible rules for predicting 12 different diseases
using 189 predicting attributes, or measurements [73]. The 12 diseases considered here are stable angina,
unstable angina, acute myocardial infarction, aortic dissection, cardiac tamponade, pulmonary embolism,
pneumothorax, acute pericarditis, peptic ulcer, esophageal pain, musculoskeletal disorders, and psycho-
genic chest pain, the characteristic of all of which was chest pain. All the 189 attributes are binary. Genetic
programming is used to learn rules expressed in a kind of first-order logic of the form <Atti Op Attj >,
where Atti and Attj are the predicting attributes, and Op is some relational operator. Genetic programming
evolves a population of “programs” candidate to the solution of a specific problem. Here, a program is
represented in the form of a tree, in which the internal nodes are functions (operators) and the leaf nodes

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 245 — #11

16-246 Handbook of Bioinspired Algorithms and Applications

are terminal symbols. In the GP formulation of the problem in Reference 73, the terminal set consists
of the 189 attributes, and the function set consists of {AND, OR, NOT}. The GP is executed once for
each class, with the appropriate rule for the ith class being evolved in the ith GP run. Thus, each run
consists of learning a two-class classification problem, in which the goal is to predict whether a patient
has a particular disease (class i) or not (NOT class i). For computing the fitness of a candidate (program
or rule), a (labeled) training set is used on which the the rule is tested. The size of the following sets are
computed:

• True positives (tp): the rule predicts that the patient has a given disease and the patient does
have it.
• False positives (fp): the rule predicts that the patient has a given disease and the patient does not
have it.
• True negative (tn): the rule predicts that the patient does not have a given disease and the patient
actually does not have it.
• False negative (fn): the rule predicts that the patient does not have a given disease but the patient
actually has it.

Thereafter, two measures are computed, the sensitivity (Se) and specificity (Sp):

tp
Se = (16.4)
tp + fn

tn
Sp = . (16.5)
tn + fp

The fitness function is taken to be the product of the two, namely,

fitness = Se × Sp. (16.6)

For the experiments, the data set consisted of 138 samples (patients), which were partitioned into a training
set with 90 samples, and a test set with 48 samples. The GP achieved an accuracy of 77.08% on. the test set.
Other related methods that used GP for classification problem can be found in References 74 to 78. On
similar lines, in Reference 79, GAs were used to discover comprehensible IF–THEN rules for the diagnosis
of dermatological diseases and prediction of the recurrence of breast cancer. Here, a chromosome is
of length n, where n is the number of attributes. The i, the gene corresponding to the ith attribute,
is divided into three fields: weight (Wi ), operator (Oi ), and value (Vi ). Each gene corresponds to one
condition in the IF part of the rule. The GA is executed once for each class, and therefore the THEN part
(indicating the class for which the GA was run) is not required to be encoded in the chromosome. The
weight value indicated whether the ith attribute, Ai , is at all present in the rule (if Wi > Limit) or not (if
Wi ≤ Limit). Limit was set to 0.3. The operator Oi could take values from {=, =}, if the corresponding
attribute is categorical, and from {≥, <}, if the corresponding attribute is continuous. The value Vi could
take values from the domain of the attribute Ai . Normal selection and crossover operators are used. Three
mutation operators are defined, namely weight mutation, operator mutation, and value mutation. The
fitness function of a chromosome was defined as fitness = Se ∗ Sp as in Reference 73. The dermatology
data set consists of the differential diagnosis of erythematisquamous. There are six different diagnoses (six
classes): psoriasis, seboreic dermatitis, lichen planus, pityriases rosea, chronic dermatitis, and pityriasis
rubra pilaris. The data set consists of 366 records with 34 attributes. The breast cancer data consists of
286 records with 9 attributes and 2 classes (recurrence and nonrecurrence of cancer). The accuracy rates
achieved in Reference 79 were 95% for the dermatological data and 67% for the cancer data. The resultant
rules were also found to be comprehensible, with one rule obtained per class.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 246 — #12

Medical Imaging and Diagnosis Using Genetic Algorithms 16-247

16.5 Discussion and Conclusions

This chapter provides a comprehensive survey of the application of GAs to the domain of designing of
equipments for medical image acquisition and medical diagnosis. In recent times, with the advent of
a variety of sophisticated imaging techniques, use of medical images in clinical diagnosis and therapy
has increased manifold. Therefore, attempts at increasing the resolution and quality of such images is
an important area of research. This, in turn, leads to research for designing equipments in such a way
that the imaging modality becomes faster and more informative. Computer aided diagnosis has also been
necessitated due to the large amount of data that is routinely collected for the patients. These problems
often turn out to be those of search and optimizations, requiring the use of good optimization tools
such as GAs.
The research works reviewed in this chapter have been reported in diverse journals and proceedings
like IEEE Transactions on Medical Imaging, IEEE Transactions on Pattern Analysis and Machine Intelligence,
IEEE Transactions on Neural Networks, Artificial Intelligence in Medicine, Medical Image Analysis, Magnetic
Resonance Imaging, Evolutionary Computation, International Conference on Evolutionary Computation,
International Conference on Genetic Algorithms, ACM SIGMOD Conference on Management of Data, and
so on. Encouraging results have been reported by researchers in this regard. This chapter presented a
methodical way in which a large number of such research activities have been compiled and reported
within a common platform, namely GA-based techniques.
It may be noted that the main challenges and issues in integrating GAs for solving optimization
problems in medical imaging and diagnosis are manifold. First, the encoding strategy must be suitably
defined so that it conforms to the building block hypothesis. According to this hypothesis, short low
order above average schema, or building blocks, should combine to yield potentially better solutions. Any
adhoc encoding strategy may not follow this hypothesis, and hence encoding strategy may not follow this
hypothesis, implying GAs may often be found to yield poor results in such situations. Second, the fitness
function must be adequately designed. Since fitness computation is often computation intensive, in the
medical domain one may need to go for parallel GAs working on massively parallel systems to get real
time response. This in turn incorporates an added level of difficulty to the problem. Finally, a still open
unsolved issue is the appropriate selection of the operator probabilities and the termination criterion, so
as to ensure good performance of the GA-based systems for medical imaging and diagnosis problems.

Acknowledgment
A part of this work was carried out when Dr. U. Maulik visited the University of Texas at Arlington, USA,
with the BOYSCAST fellowship provided by Department of Science and Technology, Government of
India, during 2001. Dr. S. Bandyopadhyay would like to acknowledge Indian National Science Academy,
Government of India sponsored project Soft computing for medical image segmentation and classification
No. BS/YSP/36/887 for providing partial support to carry out this work.

References
[1] G. Gerig, T. Pun, and O. Ratib. Image analysis and computer vision in medicine. Computerized
Medical Imaging and Graphics, 18, 85–96, 1994.
[2] N. Ayache. Medical computer vision, virtual reality and robotics. Image and Computer Vision, 13,
295–313, 1995.
[3] J.S. Duncan and N. Ayache. Medical image analysis: Progress over two decades and the challenges
ahead. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22, 85–108, 2000.
[4] T. Mclnerney and D. Terzopolous. Deformable models in medical image analysis: A survey. Medical
Image Analysis, 1, 91–108, 1996.
[5] J.B.A. Maintz and M.A. Viergever. A survey of medical image restoration. Medical Image Analysis,
2, 1–37, 1998.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 247 — #13

16-248 Handbook of Bioinspired Algorithms and Applications

[6] R. Shahidi, R. Tombropoulos, and R.P.A. Grzeszczuk. Clinical applications of three-dimensional

rendering of medical data sets. Proceedings of IEEE, 86, 555–568, 1998.
[7] S Lavallee, Registration for Computer Integrated Surgery: Methodology, and State of Art. MIT Press,
Cambridge, MA, 1996, pp. 77–97.
[8] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley,
New York, 1989.
[9] L. Davis, Ed., Handbook of Genetic Algorithms. Van Nostrand Reinhold, New York, 1991.
[10] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs. Springer-Verlag, New
York, 1992.
[11] H. Kargupta, K. Deb, and D.E. Goldberg. Ordering genetic algorithms and deception. In Pro-
ceedings of the Parallel Problem Solving from Nature (R. Manner and B. Manderick, Eds.).
North-Holland, Amsterdam, 1992, pp. 47–56.
[12] N.J. Radcliffe. Genetic set recombination. In Foundations of Genetic Algorithms 2 (L.D. Whitley,
Ed.). Morgan Kaufmann, San Mateo, CA, 1993, pp. 203–219.
[13] J.J. Grefenstette, R. Gopal, B. Rosmaita, and D. Van Gucht. Genetic algorithms for the travel-
ing salesman problem. In Proceedings of the 1st International Conference on Genetic Algorithms
(J.J. Grefenstette, Ed.). Lawrence Erlbaum Associates, Hillsdale, 1985, pp. 160–168.
[14] J.H. Holland, Adaptation in Natural and Artificial Systems. The University of Michigan Press, Ann
Arbor, MI, 1975.
[15] D.E. Goldberg, K. Deb, and B. Korb. Messy genetic algorithms: Motivation, analysis, and first
results. Complex Systems, 3, 493–530, 1989.
[16] D.E. Goldberg, K. Deb, and B. Korb. Do not worry, be messy. In Proceedings of the 4th International
Conference on Genetic Algorithms (R.K. Belew and J.B. Booker, Eds.). San Mateo, CA, Morgan
Kaufmann, 1991, pp. 24–30.
[17] D.E. Goldberg, K. Deb, H. Kargupta, and G. Harik. Rapid, accurate optimization of difficult
problems using fast messy genetic algorithms. In Proceedings of the 5th International Conference
on Genetic Algorithms (S. Forrest, Ed.), San Mateo, CA, Morgan Kaufmann, 1993, pp. 56–64.
[18] H. Kargupta and S. Bandyopadhyay. Further experimentations on the scalability of the GEMGA.
In Proceedings of the V Parallel Problem Solving from Nature (PPSN V), Lecture Notes in Computer
Science (T. Baeck, A. Eiben, M. Schoenauer, and H. Schwefel, Eds.), Vol. 1498. Springer-Verlag,
Amsterdam, The Netherlands, 1998, pp. 315–324.
[19] S. Bandyopadhyay, H. Kargupta, and G. Wang. Revisiting the GEMGA: Scalable evolutionary
optimization through linkage learning. In Proceedings of International Conference on Evolutionary
Computation, 1998, pp. 603–608.
[20] H. Kargupta and S. Bandyopadhyay. A perspective on the foundation and evolution of the link-
age learning genetic algorithms. The Journal of Computer Methods in Applied Mechanics and
Engineering, Special issue on Genetic Algorithms, 186, 266–294, 2000.
[21] L.J. Eshelman and J.D. Schaffer. Real-coded genetic algorithms and interval schemata. In
Foundations of Genetic Algorithms 2 (L. Whitley, Ed.). Morgan Kaufmann, San Mateo, CA, 1993,
pp. 187–202.
[22] J.E. Baker. Adaptive selection methods for genetic algorithms. In Proceedings of the 1st International
Conference on Genetic Algorithms (J.J. Grefenstette, Ed.). Hillsdale: Lawrence Erlbaum Associates,
1985, pp. 101–111.
[23] M. Srinivas and L.M. Patnaik. Adaptive probabilities of crossover and mutation in genetic
algorithm. IEEE Transaction on System, Man, Cybernetics, 24, 656–667, 1994.
[24] S. Bandyopadhyay, Pattern Classification using Genetic Algorithms. Ph.D. thesis, Machine
Intelligence Unit, Indian Statistical Institute, Calcutta, India, 1998.
[25] S. Crosier and D.M. Doddrell. Gradient-coil design by simulated annealing. Journal of Magnetic
Resonance, 103, 354–357, 1993.
[26] S. Crosier and D.M. Doddrell. Compact MRI magnet design by stochastic optimization. Journal
of Magnetic Resonance, 127, 233–237, 1997.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 248 — #14

Medical Imaging and Diagnosis Using Genetic Algorithms 16-249

[27] B.J. Fisher, N. Dillon, T.A. Carpenter, and L.D. Hall. Design by genetic algorithm of a z gradient
set for magnetic resonance imaging of the human brain. Measurement Science and Technology, 6,
904–909, 1995.
[28] B.J. Fisher, N. Dillon, T.A. Carpenter, and L.D. Hall. Design of a biplanar gradient coil using a
genetic algorithm. Magnetic Resonance Imaging, 15, 369–376, 1997.
[29] G.B. Wiliams, B.J. Fisher, C.L.-H. Huang, T.A. Carpenter, and L.D. Hall. Design of biplanar
gradient coils for magnetic resonance imaging of the human torso and limbs. Magnetic Resonance
Imaging, 17, 739–754, 1999.
[30] R.E. Ansorge, T.A. Carpenter, L.D. Hall, N.R. Shaw, and G.B. Williams. Use of parallel supercom-
puter to design magnetic resonance systems. IEEE Transactions on Applied Superconductivity, 10,
1368–1371, 2000.
[31] D.H. Turnbull, A.T. Kerr, and F.S. Foster. Simulation of B-Scan images from two dimensional
transducer arrays. In Proceedings of the 1990 IEEE Ultrasonic Symposium, 1990, pp. 769–773.
[32] P.K. Weber, R.M. Schmitt, B.D. Tylkowski, and J. Steak. Optimization of random sparse 2-D
transducer arrays for 3-D electronic beam steering and focusing. In Proceedings of the 1994 IEEE,
IEEE Ultrasonics Symposium, 1994, pp. 1503–1506.
[33] M.A. Anastasio, H. Yoshida, R. Nagel, R.M. Nishikawa, and K. Doi. A genetic algorithm-based
method for optimizing the performance of a computer-aided diagnosis scheme for detection of
clustered microcalcifications in mammograms. Medical Physics, 25, 1613–1620, 1998.
[34] J. Horn and N. Nafpliotis. Multiobjective Optimization using the Niched Pareto Genetic
Algorithms. illiGAL report no. 93005. University of Illinois at Urbana-Champaign, 1993.
[35] M.A. Kupinski and M.A. Anastasio. Multiobjective genetic optimization of diagnostic classifiers
with implications for generating receiver operating characteristic curves. IEEE Transactions on
Medical Imaging, 18, 675–685, 1999.
[36] M.A. Anastasio, M.A. Kupinski, R.M. Nishikawa, and M.L. Giger. A multiobjective approach
to optimizing computerized detection schemes. IEEE Nuclear Science Symposium, 3, 1879–1883,
1998.
[37] A. Green, N. Martin, J. Pfitzner, M. O’Rourke, and N. Knight. Computer image analysis in the
diagnosis of melanoma. Journal of American Academies of Dermatology, 31, 958–964, 1994.
[38] Z.B. Argenyi. Dermatoscopy (epiluminescence microscopy) of pigmented skin lesions. Dermato-
logy Clinical, 15, 79–95, 1997.
[39] W. Stolz, O. Braun-Falco, M. Landthaler, P. Bilek, and A.B. Cognetta, Color Atlas of Dermatoscopy.
Blackwell Science, Oxford, 1994.
[40] W.V. Stoecker, W.W. Li, and R.H. Moss. Automatic detection of asymmetry in skin tumors.
Computerized Medical Imagings and Graphics, 16, 191–197, 1992.
[41] D. Gutkowicz-Krusin, M. Elbaum, P. Szwaykowski, and A.W. Kopf. Can early malignant melanoma
be differentiated from a typical melanocytic nerves by in vivo techniques?, Part II, automatic
machine vision classification. Skin Research and Technology, Vol. 3, pp. 15–22, 1997.
[42] P. Schmid-Saugeon. Symmetry axis computation for almost-symmetrical and asymmetrical
objects: Application to pigmented skin lesions. Medical Image Analysis, 4, 269–282, 2000.
[43] K. Deb and A. Kumar. Simulated binary crossover for continuous search space. Complex Systsems,
9, 115–148, 1995.
[44] C.M. Balch and G.W. Milton, Hautmelanome Springer, Berlin, 1988.
[45] C.M. Grin, A.W. Kopf, B. Welkovich, R.S. Bart, and M.J. Levenstein. Accuracy in the clinical
diagnosis of malignant melanoma. Archives of Dermatology, 126, 763–766, 1990.
[46] J.E. Golston, W.V. Stoecker, R.H. Moss, and I.P.S. Dhillon. Automatic detection of irregular borders
in melanoma and other skin tumors. Computer Medical Imaging Graphics, 16, 163–177, 1992.
[47] A.J. Sober and J.M. Burstein. Computerized digital image analysis: An aid for melanoma diagnosis
preliminary investigations and brief review. Journal of Dermatology, 21, 885–890, 1994.
[48] K.P. Wilhelm, P. Elsner, E. Beradesca, and H. Baibach, Bioengineering of the Skin: Skin Surface
Imaging and Analysis. CRC Press, Boca Raton, FL, 1997.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 249 — #15

16-250 Handbook of Bioinspired Algorithms and Applications

[49] H. Handels, T. Rob, J. Kreusch, H.H. Wolff, and S.J. Poppl. Feature selection for optimized skin
tumor recognition using genetic algorithms. Artificial Intelligence in Medicine, 16, 283–297, 1999.
[50] R.M. Haralick, K. Shanmugam, and I. Dinstein. Texture features for image classification. IEEE
Transactions on System, Man, & Cybernetics, 3, 610–621, 1973.
[51] D.H. Ballard and M.B. Brown. Computer Vision. Prentice-Hall, Englewood Cliffs, NJ, 1982.
[52] K.J. Falconer. Fractal Geometry, Mathematical Foundations and Applications. Wiley, Chichester,
1990.
[53] H.O. Peitgen and D. Saupe. The Science of Fractal Images. Springer, Berlin, 1988.
[54] S. Carosi, A. Massa, and M. Pastorino. Numerical assessment concerning a focussed microwave
diagnostic method for medical application. IEEE Transactions of Microwave Theory and Techniques,
48, 1815–1830, 2000.
[55] P.C. Johnson, H. Brown, W.M. Kuzon, J.R. Ballit, J.L. Garrison, and J. Campbell. Simultaneous
quantitation of facial movements: The maximal static response assay of facial nerve function.
Annals of Plastic Surgery, 32, 171–179, 1994.
[56] T.D. Helling and M.J.G. Neely. Validation of objective measures for facial paralysis. Laryngoscope,
107, 1345–1349, 1997.
[57] Y. Cui, M. Wan, and J. Li. A new quantitative assessment method of facial paralysis based on motion
estimation. In Proceedings of the 20th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, Vol. 3. IEEE Press, Hong Kong, 1998, pp. 1412–1413.
[58] L.J. Dorfman and K.C. McGill. AAEE minimonograph #29: Automatic quantitative elec-
tromyography. Muscle Nerve, 11, 804–818, 1988.
[59] C.N. Schizas, C.S. Pattichis, I.S. Schofield, P.R. Fawcett, and L.T. Middleton. Artificial neural nets in
computer-aided macro motor unit potential classification. IEEE Engineering Medicine and Biology
Magazine, 9, 31–38, 1990.
[60] C.S. Pattichis. Artificial Neural Networks in Clinical Electromyography. Ph.D. thesis, Queen Mary
and Westfield College, University of London, UK, 1992.
[61] C.S. Pattichis and C.N. Schizas. Genetic-based machine learning for the assessment of certain
neuromuscular disorders. IEEE Transactions on Neural Networks, 7, 427–439, 1996.
[62] P. Larranaga, B.S.M.Y. Gallego, M.J. Michelena, and J.M. Pikaza. Learning bayesian networks by
genetic algorithms. A case study in the prediction of survival in malignant skin melanoma. In
Lecture Notes in Artificial Intelligence, Vol. 1211. Springer-Verlag, Heidelberg, 1997, pp. 261–272.
[63] B. Sierra and P. Larranaga. Predicting survival in malignant skin melanoma using Bayesian net-
works automatically induced by genetic algorithms. An empirical comparison between different
approaches. Artificial Intelligence in Medicine, 14, 215–230, 1998.
[64] F.V. Jensen, Introduction to Bayesian Networks. University College of London, UK, 1996.
[65] P. Larranaga, M.Y. Gallego, B. Sierra, L. Urkola, and M.J. Michelena. Bayesian networks, rule
induction and logistic regression in the prediction of women survival suffering from breast cancer.
In Lecture Notes in Artificial Intelligence, Vol. 1323. Springer-Verlag, Heidelberg, 1997, pp. 303–308.
[66] R. Blanco, I. Inza, and P. Larranaga. Learning bayesian networks in the space of structures by
estimation of distribution algorithms. International Journal of Intelligent Systems, 18, 205–220,
2003.
[67] C.A. Pena-Reyes and M. Sipper. Evolving fuzzy rules for breast cancer diagnosis. In Proceedings
of 1998 International Symposium on Nonlinear Theory and Applications (NOLTA ’98), Vol. 2.
Lausanne, Presses Polytechniques et Universitaires Romandes, pp. 369–372, 1998.
[68] C.A. Pena-Reyes and M. Sipper. A fuzzy-genetic approach to breast cancer diagnosis. Artificial
Intelligence in Medicine, 17, 131–155, 1999.
[69] C.J. Kahn, L. Roberts, K. Shaffer, and P. Haddawy. Construction of a bayesian network for
mammographic diagnosis of breast cancer. Computers in Biology and Medicine, 27, 19–29, 1997.
[70] N. Marvin, M. Bower, and J.E. Rowe. An evolutionary approach to constructing prognostic models.
Artificial Intelligence in Medicine, 15, 155–165, 1999.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 250 — #16

Medical Imaging and Diagnosis Using Genetic Algorithms 16-251

[71] S. Ngan, M.L. Wong, W. Lam, K.S. Leung, and J.C.Y. Cheng. Medical data mining using
evolutionary computation. Artificial Intelligence in Medicine, 16, 73–96, 1999.
[72] S. Ngan, M.L. Wong, and K.S. Leung. Using grammar based genetic programming for data mining
of medical knowledge. In Genetic Programming 1998: Proceedings of 3rd Annual Conference, Morgan
Kaufmann, San Mateo, CA, 1998, pp. 254–259.
[73] C.C. Bojarczuk, H.S. Lopes, and A.A. Freitas. Discovering comprehensible classification rules by
using genetic programming: A case study in a medical domain. In Proceedings of the Genetic and
Evolutionary Computation Conference (W. Banzhaf, J. Daida, A.E. Eiben, M.H. Garzon, V. Honavar,
M. Jakiela, and R.E. Smith, Eds.), Vol. 2, (Orlando, FL, USA), pp. 953–958, Morgan Kaufmann,
San Mateo, CA, 1999, pp. 13–17.
[74] A. Teller and M. Veloso. Program evolution for data mining. International Journal of Expert Systems,
8(3), 213–236, 1995.
[75] B. Marchesi, A.L. Stelle, and H.S. Lopes. Detection of epileptic events using genetic program-
ming. In Proceedings of the 19th International Conference IEEE/EMBS, pp. 1198–1201. IEEE Press,
Washington, 1997.
[76] J.R. Sherrah, R.E. Bogner, and A. Bouzerdoum. The evolutionary pre-processor: Automatic feature
extraction. In Genetic Programming: Proceedings of 2nd Annual Conference, Morgan Kaufmann,
San Mateo, CA, 1997, pp. 304–312.
[77] N.I. Nikolaev and V. Slavov. Inductive genetic programming with decision trees. In Proceedings of
the 1997 European Conference on Machine Learning (ECML), 1997.
[78] L. Martin, F. Moal, and C. Vrain. A relational data mining tool based on genetic programming.
In Principles of Data Mining and Knowledge Discovery: Proceedings of 2nd European Symposium
(LNAI), Vol. 1510 Springer-Verlag, Heidelberg, 1998, pp. 130–138.
[79] M.V. Fidelis, H.S. Lopes, and A.A. Freitas. Discovering comprehensible classification rules a genetic
algorithm. In Proceedings of the 2000 Congress on Evolutionary Computation CEC00, La Jolla
Marriott Hotel La Jolla, CA, USA, pp. 805–810, IEEE Press, Washington 2000, pp. 6–9.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C016” — 2005/8/8 — 17:15 — page 251 — #17

 17
Scheduling and
Rescheduling
with Use of
Cellular Automata

17.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-253

17.2 Multiprocessor Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-254
17.3 Cellular Automata (CA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-255
17.4 CA-Based Scheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . . 17-256
17.5 Two-Processor Scheduling with CA: Experimental
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-258
Scheduling Deterministic and Random Program Graphs •
Performance and Complexity Issues • Reusing Discovered
Rules
17.6 AISs and Rescheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-266
A Concept of AIS for Rescheduling • Experimental Results
Franciszek Seredynski 17.7 Scheduling with More than Two Processors . . . . . . . . . . . 17-269
Anna Świȩcicka 17.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-271
Albert Y. Zomaya References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-271

17.1 Introduction
Multiprocessor scheduling is one of the most challenging problems in parallel and distributed
computing [1]. It is known to be NP-complete in its general form [2]. Researchers have studied
restricted forms of the problem by constraining either a parallel program model or a multiprocessor
model. However, these special cases do not fully represent real-world systems. To solve the schedul-
ing problem in the general case, a number of heuristics based on different mathematical platforms and
metaheuristics based on mechanisms observed in nature, have been introduced. The commonly known
heuristics are list scheduling, critical path, or clustering [3,4]. Recently metaheuristics such as simulated
annealing (SA), genetic algorithms (GAs), ant colonies, tabu search (TS), or neural networks have been
successfully applied [5,6].
The main problem related to heuristics and metaheuristics is the existence of the scheduling overhead
represented by the cost of running the scheduler. Here, in the case of metaheuristics on which we focus

17-253

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 253 — #1

17-254 Handbook of Bioinspired Algorithms and Applications

our attention, the main source of the scheduling overhead is the necessity of calculation of a cost function
in subsequent iterations of a scheduling algorithm. One of the main sources of the scheduling overhead is
neglecting potential knowledge about the scheduling problem, which could be gained during solving its
instances. The prevailing number of scheduling algorithms does not extract, conserve, and reuse any
knowledge about the problem while solving instances of the scheduling problem.
The motivation of our work is to develop a framework for designing scheduling algorithms where
knowledge about scheduling process can be extracted and reused while solving new instances of the
scheduling problem. For this purpose we propose to use a recently emerged and very promising hybrid
technique combining cellular automata (CAs) and GA, and a computational paradigm of artificial immune
system (AIS) [7].
CAs are discrete dynamical systems made up of a large number of cells, which behave according to
local rules. It is an interesting feature of these systems that although cells interact only locally, in a
fully distributed manner, a complex global behavior can emerge. For this reason CAs are often used to
model real-world phenomena [8]. CAs are also considered as models of highly parallel and distributed
computations in multiprocessor and distributed systems [9]. They are used to find solutions of problems
such as scheduling and resource management [10].
The main problem related to CAs is a huge space of local CA rules representing possible solutions of
a problem. Therefore, most applications of CAs were a result of clever, but time-consuming hand designing
rather than an oriented search. Only recent works [11,12] on applying evolutionary computation to search
CA rules opened new possibilities. Results described in the literature show that CAs, combined with
evolutionary techniques for discovering local rules, can be effectively used to find parallel and distributed
solutions of complex global problems such as density classification task and synchronization task [11,12],
and location management in mobile computing [13] or cryptography [14]. Recently it has been shown
[15] that such a hybrid technique can be applied to discover scheduling algorithms. In this chapter,
we extend this methodology and additionally propose to use AIS as a support for reusing discovered CAs
scheduling rules to solve new instances of the scheduling problem.
The rest of this chapter is organized as follows. Section 17.2 presents the scheduling problem.
Section 17.3 provides a background on CAs. Section 17.4 contains the description of the proposed
CA-based scheduling system. Section 17.5 contains experimental results concerning CAs applied to
scheduling in two-processor systems. Section 17.6 describes AIS for reusing knowledge conserved in
CA rules while solving new instances of the scheduling problem. Section 17.7 presents the extension of
the proposed approach on the case of multiprocessor systems consisting of more than two processors.
Finally, Section 17.8 contains the conclusions.

17.2 Multiprocessor Scheduling

In our approach, we use some models of a multiprocessor system and a parallel program. They are both
represented by corresponding graphs. A multiprocessor system is represented by an undirected unweighted
graph Gs = (Vs , Es ), called a system graph. Vs is the set of Ns nodes representing processors with their
local memories. Es is the set of edges representing bidirectional channels between processors and defines
a topology of the multiprocessor system. Figure 17.1(a) shows an example of a system graph representing
a multiprocessor system consisting of two processors P0 and P1 . It is assumed that all processors that have
the same computational power and communication via links do not consume any processor time.
A parallel program is represented by a weighted, directed acyclic graph Gp = Vp , Ep , called a precedence
task graph or a program graph. Vp is the set of Np nodes of the graph representing elementary tasks. The
weight bk of the node k describes the processing time needed to execute a task k on any processor of
the system. Ep is the set of edges of the precedence task graph describing the communication patterns
between the tasks. The weight akl of the edge (k, l) describes the communication time between the pair of
tasks k and l when they are located in neighboring processors. If the tasks k and l are located in the same
processor than the communication time between them is equal to zero. Figure 17.1(a) shows an example

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 254 — #2

Scheduling and Rescheduling with Use of Cellular Automata 17-255

(a) P0 (b) 0
1

1 1

1 2
2 4

3
P1 2

FIGURE 17.1 Examples of a (a) system graph and (b) precedence task graph.

TABLE 17.1 An Example of a Rule Table for a

One-Dimensional, Binary State CA

η 000 001 010 011 100 101 110 111

a 0 1 0 1 1 0 1 0

of the system graph corresponding to the two-processor system. Figure 17.1(b) shows an example of the
program graph consisting of four tasks with their order numbers from 0 to 3. All communication costs of
the program graph are equal to one. Computational costs of tasks are 1, 2, 4, and 2, respectively.
The purpose of the scheduling is to distribute the tasks among the processors in such a way that the
precedence constraints are preserved and the response time T (the total execution time) is minimized.
Found optimal schedule is usually represented by a Gantt chart showing allocation of tasks to processors
and time when a given task starts and finishes execution.
The response time T for a given precedence task graph depends on an allocation of tasks in
multiprocessor topology and some scheduling policy [15], which defines the order of processing tasks,
ready to run in a given processor.

17.3 Cellular Automata (CA)

These [12,16] are dynamical systems in which space and time are discrete. Here, we shall concentrate on
one-dimensional CAs.
A one-dimensional CA consists of a spatial lattice of N cells, each of which, at time t , can be in one of k
states. In this chapter, we assume that each cell is a Boolean variable: ait ∈ {0, 1}, i = 0, 1, . . . , N − 1. The
collection of all local states is called the configuration. A CA has a single fixed rule, which is used to update
each cell. The rule maps from the states in a neighborhood of a cell to a single state, which is the updated
value for that cell. In a one-dimensional CA the neighborhood of a cell includes the cell itself and some
radius r of neighbors on either side of the cell.
The equations of motion for a CA are often expressed in the form of a rule table: a look-up table listing
in lexicographical order of each of the neighborhood patterns and the state to which the central cell in that
neighborhood is mapped. For example, Table 17.1 presents one possible rule table for a one-dimensional,
binary state CA with radius r = 1. Each possible neighborhood η is given, along with the output bit a to
which the central cell is updated.
The CA starts out with some initial configuration (IC) of cell states. To run the CA, the rule table is
applied, usually synchronously, to each neighborhood in the current lattice configuration, respecting the
choice of boundary conditions, to produce the configuration at the next time step. The most popular
boundary conditions are periodic boundary condition and null boundary condition. Periodic boundary

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 255 — #3

17-256 Handbook of Bioinspired Algorithms and Applications

1 2

3
3

FIGURE 17.2 Program graph and corresponding CA.

condition means that the leftmost cell is considered to be the right neighbor of the rightmost cell and
vice versa. Null boundary condition means that “absent” cells are always in state 0.
The behavior of CA is often illustrated using “space-time diagrams” in which the configuration of states
in n-dimensional lattice is plotted as a function of time (in most cases space-time diagrams are practical
only for n ≤ 2).

17.4 CA-Based Scheduling Algorithm

To design a scheduling algorithm with the use of CA we assume that a nonlinear structure of a program
graph is approximated by a one-dimensional CA of size Np , in contrast to the approach presented in
References 15 and 17, where a two-dimensional irregular structure of CA was used. The way of creating
local neighborhood with the use of two-dimensional CA and associating them with program graphs was
complex, and resulted in a constant size of the neighborhood that was equal to seven, a long CA rule with
the length equal to 250 bits, and a huge space of possible solutions. Therefore, the use of one-dimensional
CA seems to be a more simple and promising alternative. The choice of one-dimensional CA means that
with each task of a program graph an elementary cell is associated (see Figure 17.2). We assume that the
cell 0 is assigned to task 0, the cell 1 is assigned to task 1, etc. Dashed lines in Figure 17.2 mean dummy
cells in the state 0, which means null boundary conditions.
Let us consider the problem of scheduling in the two-processor system. In this case, cells of CA
corresponding to any program graph will be either in the state 0 or in the state 1. The state 0 or 1 of a cell
means that a corresponding task is allocated in the processor P0 or P1 , respectively. Each configuration of
CA corresponds to some allocation of tasks in the system graph.
CA corresponding to the program graph will evolve according to its rule. Initial states of CA correspond
to an initial allocation of tasks in the two-processor system. Changing states of CA will result in changing
an allocation of tasks in the system, and changing the response time T . We want to know if there exists a
rule for CA providing for any initial allocation of tasks converging CA to an allocation which minimizes T .
An architecture of CA-based scheduling system is presented in Figure 17.3. There are three modes of
work of the scheduling system: a mode of learning CA scheduling rules, a mode of normal operating
and a mode of reusing CA scheduling rules by applying AIS. CAs can update their states (CA mode)
sequentially (seq), in parallel (par) or sequentially with a random order of updating cells’ states (seq–ran).
In sequential mode the CA works asynchronously, that is, at a given moment of time only one cell updates
its state. An order of updating states by cells is defined by their order numbers corresponding to tasks in
the precedence task graph. A single step of running the CA is completed in Np steps. In parallel mode all
cells update their states synchronously. The third mode works in the same way as sequential mode, but an
order of updating cells’ states is random.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 256 — #4

Scheduling and Rescheduling with Use of Cellular Automata 17-257

The purpose of the learning mode (see, Figure 17.3, left) is to discover CA rules for scheduling.
Randomly generated allocations of program graph tasks into a system graph serve as initial states of CA,
which then runs according to a rule from GA population of rules. GA searches a CA rule, which will be
able to evolve to a final configuration corresponding to an optimal or suboptimal value of the response
time T , for the input allocation of the program tasks. The algorithm for discovering CA rules with use of
GA [11,18] is presented.

Learning mode: discovering CA rules using GA

BEGIN
create an initial population of rules of size P;
FOR l = 1 TO G DO
BEGIN
create a set of size I of test problems
FOR i = 1 TO P DO
BEGIN
Ti∗ = 0;
FOR j = 1 TO I DO
Ti∗ = Ti∗ +CA(rulei ,testj ,CA mode,M steps);
Ti∗ = Ti∗ /I ;
END;
sort current population of rules according to Ti∗ ;
move E of the best individuals to the next population;
FOR i = 1 TO (P − E)/2 DO
BEGIN
parent
rule1 =select();
parent
rule2 =select();
parent parent
(rule1child , rule2child ) =crossover(rule1 , rule2 );
mutation(rule1child , rule2child );
END;
END;
END;

GA begins with a population of P randomly generated CA rules. The length of the rule is calculated as
L = k 2r+1 . We then create a set (different for each generation of GA) of size I of test problems (ICs of
CA) by generating a set of random allocations of tasks in multiprocessor system for a given instance of a
problem. Each rule i is tested on a whole set of the set problems. The function CA() returns for a rule i
j
the response time Ti obtained by CA running M steps under given mode of CA (i.e., seq, seq–ran, or par)
on a test problem j. In our experiments, we set M ≈ 4 × Np . M can be calculated by some heuristic based
j
on many observations of behavior of CAs. Ti is calculated once, after the last step of running CA, when

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 257 — #5

17-258 Handbook of Bioinspired Algorithms and Applications

Program graph System graph New

program
graph

Initial (random)
allocation

CA-based T
scheduling
Learning algorithm
AIS
Normal operating
GA CA rule Rule

101.............001 10..............00 11..............00

011.............011 . . 10..............10
T* . . . . . .
. . . . . .
. . . . . .
000.............001 01..............01 01..............01

Population of rules Population of Library of rules

discovered rules found for different
program graphs

FIGURE 17.3 An architecture of CA-based scheduling system.

the CA final configuration of states corresponds a final allocation of tasks. As a rule fitness of the rule i,
j
a value Ti∗ is accepted, which is the sum of values Ti averaged over the number I of solved tests, that is,

j
Ti∗ = Ij=1 Ti /I .
After calculation of the fitness function for all rules, GA operators are used. A number E of the best
rules (“elite”) is copied without modification to the next generation. The remaining P − E rules for the
next generation are formed by two-point crossover operator applied to randomly chosen pairs of elite
rules, which are mutated next with probability pm . This process is continued over a predefined number
of generations G and when completed the discovered rules are stored.
In the normal operating mode (see Figure 17.3, middle), when a program graph is initially, randomly
allocated, CA is initiated and equipped with a rule taken from the set of discovered rules. In this mode
we expect that for any initial allocation of tasks of a given program graph, CA will be able to evolve very
fast, without time-consuming calculation of T , to a configuration corresponding to the minimal or near
minimal value of T .
In the third mode AIS (see Figure 17.3, right), enables a potential reusing of discovered knowledge
stored in CA rules. It has an important meaning when trying to reschedule. The concept of AIS, along
with experimental results, is presented in Section 17.6.

17.5 Two-Processor Scheduling with CA: Experimental Results

17.5.1 Scheduling Deterministic and Random Program Graphs
A number of experiments with program graphs that available in the literature (see Reference 15) has been
conducted. In addition, a number of random graphs were used, with 25 and 50 (in average) tasks. The

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 258 — #6

Scheduling and Rescheduling with Use of Cellular Automata 17-259

(a) 1 (b) 0 1
0

2 3 4 5 6 7 8
3 1 3 2 3 3 3 4 3 5 3 6 3 7 3 8 9 10

11 12 13 14 15 16 17 18 19
10 10 10 10
9 10 11 12

20 21 22 23 24 25 26 27 28

5 13 5 14 5 15 5 16 29 30 31 32 33 34 35 36 37

1 17 38 39

(c) (d)
0 0
8
12
4 1 4 2 4 3 4 5
4 4
8
8 8 8
1 2 8 6 6
12 12
12 12
3 7 3 8 3 9 3 10
12
8 8 8
8 4 11
3 4 5 6 12 12 12
12
2 12 2 13 2 14
8 8
8 2 15
7 8 9 10 11 12 13 14 12 12
1 16 1 17

FIGURE 17.4 Program graphs: (a) g 18, (b) g 40, (c) tree15, and (d) gauss18.

experiments were performed on PC, Celeron 745 MHz. We used CAs with a radius of neighborhood
r ∈ {1, 2, 3}, what corresponds to rule lengths equal to 8, 32, and 128, respectively. We used the following
parameters in the experiment: a rules population size P = 50 ÷ 200, elite size E = 15 ÷ 100, a size
of test problems I = 10 ÷ 50, a number of CA steps M = 4 × Np (Np — a number of tasks), and
mutation probability pm = .03. In this section, we describe in detail the results obtained for the following
deterministic program graphs: g 18, g 40, tree15, and gauss18 (see Figure 17.4) and for the randomly
generated program graph called Rnd25_5.
The first program graph used in experiments is tree15. It is a binary tree consisting of 15 tasks. All
computational and communication costs are the same and equal to one. The optimal response time T
for tree15 in the two-processor system is equal to nine. Experiments have shown that for program graphs
from family tree it is sufficient to use r = 1 to discover scheduling rules. Rules are discovered in a few
generations of GA for all three modes of the CA, that is, seq, par, and seq–ran. Figure 17.5 shows typical
runs of the CA-based scheduler with the best rule from the final generation, starting from randomly
generated initial configuration.
The next experiment is conducted with the program graph g 40 (see, e.g., Reference 15), which is
composed of 40 tasks, with computational and communication costs equal to four and one, respectively.
The optimal response time T for g 40 in the two-processor system is equal to 80. The minimal value of r,
which allows in the learning mode for g 40 converging to optimal value of T is r = 2. Figure 17.6(a) shows
a typical run of the scheduling system in the learning mode. One can see that for two modes of updating

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 259 — #7

17-260 Handbook of Bioinspired Algorithms and Applications

(a) 0

20
30
4 5 6 7 8 9 10 11 12
T=9

(b) 0

20
30
4 5 6 7 8 9 10 11 12 13
T=9

(c) 0

20

40

59
4 5 6 7 8 9 10 11 12 13 14
T=9

FIGURE 17.5 Space-time diagrams of CA-based scheduler for tree15: (a) sequential mode, (b) parallel mode, and
(c) sequential mode with a random order of updating cells’ states.

g40 g40
(a) 83 (b) 130

82.5
120
82 seq
Response time T

par
Average of T

81.5 seq – ran 110

81
100 1000 tests
80.5

80 90
79.5
80
79
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Generation Rules

FIGURE 17.6 Running CA-based scheduling system for program graph g 40: (a) learning mode and (b) normal
operating mode.

CA rules, seq and par, the system discovers rules converging CA to a configuration corresponding to
optimal value of T = 80, and for seq–ran mode the system converges to rules providing suboptimal value
of T . It takes <20 generations of GA.
After the run of the system in the learning mode the final population of GA contains rules suitable for
CA-based scheduling. We can find out a quality of these rules in the normal operating mode. We generate
1000 random ICs (program tasks’ allocations) and use them to schedule each test problem by each of the
rules found. Figure 17.6(b) presents the average value of T for each rule over a whole set of test problems.
One can see that about 60% of discovered rules are able to find an optimal schedule for each test problem.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 260 — #8

Scheduling and Rescheduling with Use of Cellular Automata 17-261

(a) 0

20
40
49
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94
T = 80

(b) 0
20
40
60
80
89
75 77 79 81 83 85 87 89 91 93 95 97 99 101 103
T = 80

(c) 0
20
40
49
76 78 80 82 84 86 88 90 92 94 96 98 100 102
T = 81

FIGURE 17.7 Space-time diagrams of CA-based scheduler for g 40: (a) CA in seq mode, (b) par mode, and
(c) seq–ran mode.

The scheduling in this mode is performed automatically, very quickly, without calculation of the cost
function T , and can be observed by watching a space-time diagram of CA.
Figures 17.7(a)–(c) show space-time diagrams for typical runs of the CA-based scheduler working
with the best rule found for g 40, and performing scheduling in the normal operating mode for a single,
randomly generated test problem, with CA working in seq, par, and seq–ran modes, respectively. Left
part of each figure presents a space-time diagram of the CA consisting of 40 cells and right part shows
graphical values of T corresponding to the allocations found in a given step. Let us consider Figure 17.7(a).
One can see that in the step 0 (the first row), cells of the CA are in some initial states corresponding to
the initial allocation of tasks (white cell — a corresponding task is allocated to P0 , black cell — a task
is allocated to P1 ) and the value of T corresponding to this allocation is equal to 94. Then the CA
starts to run, changing its states sequentially according to its rule, what results in changing values of T .
Each subsequent row of the figure shows a situation after changing states of 40 cells. One can see that
the CA needs 38 steps to converge to tasks’ allocation corresponding to the minimal value of T = 80.
During which CA running no calculation of a cost function is performed. The 38 steps of the CA requires
38 × 40 = 1520 time steps, in which sequentially performed elementary entries to a simple look-up
table are executed to update states of CA cells, and it corresponds to a very low computational cost (see,
Table 17.2).
Figure 17.7(b) shows a space-time diagram of CA-based scheduler working in par mode. In this case
the CA converges to the optimal value of T after 84 steps, which is equivalent to 84 time steps, because 40
updates of the look-up table are performed in parallel in each step. Figure 17.7(c) presents a typical run
of the CA-based scheduler working in seq–par mode, with the suboptimal value of T equal to 81 found in
20 steps of CA.
The next program graph used in experiments, referred as gauss18 (see, e.g., Reference 15), represents
the parallel Gaussian elimination algorithm consisting of 18 tasks. The optimal response time T for this
program graph in the two-processor system is equal to 44. Despite the lower number of tasks when
compared with the g 40, the gauss8 is much more difficult for the learning mode because of its nonregular

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 261 — #9

17-262 Handbook of Bioinspired Algorithms and Applications

TABLE 17.2 Performance of CA-Based Scheduling System

Learning mode Metaheuristics

Normal operating mode
Program graph seq CA par CA GA SA TS seq CA

tree15 9, 10 9, 55 9, 0.03, 0.003 9, 3.57 9, 0.086 9, -, 0.002

g 18 46, 9 46, 38 46, 0.04, 0.006 46, 2.93 46, 0.197 46, -, 0.002
g 40 80, 50 80, 125 80, 0.08, 0.03 80, 3.29 80, 0.665 80, -, 0.015
gauss18 44, 380 46, 1500 44, 1, 0.17 49, 2.32 49, 1.277 44, -, 0.003
Rnd25_1 (31) 495,1880 507, 5720 495, 0.4, 0.16 495, 1.74 495, 10.36 495, -, 0.011
Rnd25_5 (17) 94, 720 94, 1600 94, 41, 8.2 100, 3.29 95, 14.35 94, -, 0.005
Rnd25_10 (21) 62, 25 62, 131 62, 24.37, 0.58 62, 0.07 62, 0.79, - 62, -, 0.006
Rnd50_1 (59) 895, 4110 901, 9680 890, 4, 5.32 891, 2.64 890, 55.88 895, -, 0.023
Rnd50_5 (62) 225, 4365 238, 12790 203, 220, 290.4 228, 4.29 214, 70.68 225, -, 0.024
Rnd50_10 (53) 146, 6480 146, 21350 135, 29.33, 11.44 138, 6.19 145, 97.70 146, -, 0.019

(a) gauss 18 (b) Rnd 25_5

70 110
108
65
106 seq.
Response time T

Response time T

seq. par.
60 104 seq–ran.
par.
ser–ran. 102
55
100
50
98
45 96
94
40
0 100 200 300 400 500 600 700 800 900 1000 0 50 100 150 200 250
Generation Generation

FIGURE 17.8 Learning mode of the scheduling system: (a) for deterministic program graph gauss18 and (b) random
graph Rnd25_5.

structure. The minimal value of r, which allows for gauss18 converging in the learning mode to optimal
value of T = 3, what results in 128-bit long CA rule and a large space of possible rules.
Figure 17.8(a) shows typical runs of the GA in the learning mode for three modes of operating of
the CA. One can see that only for seq mode of CA the GA is able to find, after about 600 generations,
optimal CA rules. However, in the normal operating mode, the best-discovered rule needs only 18 steps
to schedule, so the performance in this mode is similar as the performance shown for the g 40.
Figure 17.9(a) shows a typical run of the CA-based scheduler working in sequential mode with the best
rule from the final generation, starting from randomly generated IC. One can see that the CA needs <20
time steps to converge to tasks’ allocation corresponding to the minimal value of T = 44. Figure 17.9(b)
shows space-time diagram of CA-based scheduler working in parallel mode. In this case the CA converges
to suboptimal value of T = 46. Figure 17.9(c) presents typical run of the CA-based scheduler working in
sequential mode with a random order of updating cells’ states, with the value of T = 51.
Last experiment presented in this section is performed on randomly generated program graph called
Rnd25_5 (with 17 tasks). The optimal response time T for Rnd25_5 in the two-processor system is equal
to 94. In the learning mode r = 3 is required. Figure 17.8(b) shows typical runs of the GA for three modes
of operating of the CA. For seq mode, the GA discovers optimal CA rules after about 50 generations.
For par mode, it takes above 150 generations. For seq–par mode, the GA does not discover optimal CA
scheduling rules. Figure 17.10 shows typical runs of the CA-based scheduler with the best rule from the
final generation, starting from randomly generated IC.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 262 — #10

Scheduling and Rescheduling with Use of Cellular Automata 17-263

(a) 0

20

40
49
39 43 47 51 55 59 63 67 71 75 79 83
T = 44

(b) 0

20

40
49
41 45 49 53 57 61 65 69 73 77 81 85 89
T = 46

(c) 0

20

40
49
46 50 54 58 62 66 70 74 78 82 86 90 94
T = 51

FIGURE 17.9 Space-time diagrams of CA-based scheduler for gauss18: (a) sequential mode, (b) parallel mode, and
(c) sequential mode with a random order of updating cells’ states.

(a) 0

20

40
49
89 94 99 104 109 114 119 124 129 134 139 144
T = 94

(b) 0

20

40
49
89 94 99 104 109 114 119 124 129 134 139 144
T = 94

(c) 0

20

40
49
95 100 105 110 115 120 125 130 135 140 145 150
T = 100

FIGURE 17.10 Space-time diagrams of CA-based scheduler for Rnd25_5: (a) sequential mode, (b) parallel mode,
and (c) sequential mode with a random order of updating cells’ states.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 263 — #11

17-264 Handbook of Bioinspired Algorithms and Applications

17.5.2 Performance and Complexity Issues

Table 17.2 summarizes results (averaged on ten runs) of experiments conducted with use of deterministic
and random program graphs (Rnd from 17 to 62 tasks), for the learning mode (seq and par mode of CA)
and normal operating modes (seq). The results are compared with the ones obtained with use of other
metaheuristics, such as standard GA, SA, and TS. Each result for a given program graph is represented
by a pair: the best found response time T , and a required number (in thousands) of calculation of cost
function T ; for GA and normal operating mode the time (in seconds) of running an algorithm on PC is
provided.
We can see that for a prevailing number of program graphs the scheduling system was able to discover
CA rules providing optimal or suboptimal value of T . Only for random program graphs we obtained the
worst results. These values are validated with use of metaheuristics.
The procedure of discovering CA rules is computationally, relatively expensive, because the learning
process requires the presentation of a number of test problems and calculation of the cost function T ,
which is the main source of the complexity of the learning algorithm. The complexity of the algorithm
expressed by a number of the cost function calculations depends (see Table 17.2) not only on a number of
tasks, but also on its topology and relation between communication and computational costs in a given
program graph. However, this procedure is used once. When learning is completed, discovered rules store
the knowledge, which can be reused while solving new instances of the scheduling problem. In general,
better results have been obtained for the sequential mode of CA. A reason for this is that a calculation
of the cost function T , which in fact is performed by a simulation of execution of a program graph in
a system graph, is a sequential algorithm defined by information flow in an oriented program graph.
A parallelization of this process is easier and computationally less expensive for graphs with a regular
structure. A cost of evolving rules for par mode of CA is a few times greater than for seq mode.
Metaheuristics were used to find an optimal value of T for different program graphs used in the learning
mode. One can see that none of the metaheuristics offer a dominating quality of solutions or performance,
which is a consequence of no free lunch theorem [19]. Note that the cost of running them defined by a
number of calculation of the cost function T is much lower than the cost of rules discovered in the learning
mode. However, this cost must be paid regularly each time the metaheuristics are used.
Results of experiments conducted in the normal operating mode (seq CA mode) clearly show the
advances of the proposed approach. Scheduling is performed without an expensive procedure of calcula-
tion of the cost function T and for larger program graphs it may be a few orders of magnitude faster than
for example, scheduling performed by GA. The scheduling algorithm is fully distributed — scheduling
is the result of local updating states of CA cells. The main component of the algorithm complexity is
a simple operation on the look-up table to read a decision concerning updating of a cell state, and this
complexity grows linearly with the number of tasks.
Proposed CA-based scheduling algorithm has an interesting feature. Since the coding CA rules does not
depend on a program graph, there is a technical possibility of reusing the discovered rules, in the normal
operating phase, on other program graphs.
The question that now arises is whether discovered CA rules have abilities of scheduling other program
graphs. The goal of experiments described in Section 17.5.3 is to examine this issue.

17.5.3 Reusing Discovered Rules

In the first experiment we used rules discovered for the program graph g 18 (see Figure 17.4[a] to find
solutions of other program graphs constructed on the base of g 18. The program graph called g 36 was
constructed through linking g 18 to its final task. The weight of the linking edge was set to zero. The program
graph called g 54 was constructed through linking g 18 to the final task of g 36 and setting the weight of
the linking edge to zero. In a similar way other program graphs (g 72, g 90, and g 108) were constructed.
Then we used five of the best CA rules found for the program graph g 18 in the normal operating phase
on 1000 random initial allocations of our new program graphs. In this experiment r = 2 and two modes

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 264 — #12

Scheduling and Rescheduling with Use of Cellular Automata 17-265

20

40
49
87 91 95 99 103 107 111 115 119 123 127
T = 92

FIGURE 17.11 Space-time diagrams of CA-based scheduler for g 36: sequential mode.

TABLE 17.3 The Best Response Time T Received for Different

Program Graphs

Rule\graph tree7 intree7 g 18 g 40 gauss18 Rnd25_5

tree7 5.0 5.0 48.04 85.9 64.58 119.7
intree7 5.0 5.0 47.0 81.0 63.35 124.33
g 18 5.02 5.08 46.0 84.17 63.93 122.49
g 40 5.0 5.0 49.01 80.0 61.32 113.65
gauss18 5.26 5.11 52.93 92.23 44.0 117.82
Rnd25_5 5.74 5.14 51.04 88.07 56.0 94.0
opt.T 5 5 46 80 44 94

of operating a CA were used: sequential and parallel mode. The experiments conducted have shown that
the best rules obtained for the program graph g 18 can successfully find an optimal scheduling for each
representative of the test. These solutions are obtained without a process of discovering CA rules and
without using GA to find optimal schedules. It means that the time required to find optimal solutions for
tested program graphs is significantly reduced.
Figure 17.11 shows typical run of the CA-based scheduler, working in sequential mode, for the program
graph g 36, with the best rule discovered for the program graph g 18. The optimal response time T for the
program graph g 36 is equal to 92.
The next question that arises is whether the discovered rules are sensitive to some modifications of
our program graph. To find out, discovered rules were used in the normal operating phase (assuming
sequential mode of a CA) to find solutions of some other program graphs. These graphs were constructed
from g 18 by introducing some random modifications to it. These modifications included changing the
values of the weights of some randomly chosen tasks or (and) edges. This way we obtained 30 new
program graphs. In most of the cases the best discovered rules were able to find an optimal (or near
optimal) scheduling for each representative of the test. These solutions are obtained without a process of
discovering rules and without using GA to find optimal schedules.
The last experiment had a more general character than the two previously described. We wanted to
know if the discovered rules had possibilities of scheduling other program graphs. We took the whole
population of rules discovered for the following program graphs: tree7, intree7, g 18, g 40, gauss18, and
Rnd25_5. These populations of rules were discovered assuming r = 3 and sequential mode of a CA. The
course of the experiment was identical as in the experiments described earlier. The populations of rules
discovered for these program graphs were tested (in the normal operating mode, assuming I = 1000) on
other program graphs the following way: “each population of rules on each program graph.” Obtained
results of the best rules are presented in Table 17.3. These are typical results obtained in a given test. The
last row of Table 17.3 contains the optimal response times T for tested program graphs.
Results presented in Table 17.3 indicate that in some cases rules discovered for a given program graph
have possibilities of scheduling other program graphs. The best rules in the tested populations can be used

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 265 — #13

17-266 Handbook of Bioinspired Algorithms and Applications

to obtain optimal (or near optimal) schedules of the program graph tree7 and intree7. This fact perhaps is
not surprising because tree-structured program graphs are the easiest instances. More interesting is that
rules discovered for tree-structured program graphs are suitable for scheduling program graphs g 18 and
g 40. On the other hand, rules discovered for gauss18 and Rnd25_5 operate on the program graphs (with
the exception of trees) weaker than other tested rules.
Presented results indicate that discovered CA rules store knowledge about a way of solving instances of
the scheduling problem. Typically, rules specialize in solving the scheduling problem for a specific program
graph. Some of them are “more general” than others. They have limited possibilities to schedule other
program graphs. These observations lead to a concept of AIS, which would support reusing knowledge
stored in discovered CA rules.

17.6 AISs and Rescheduling

17.6.1 A Concept of AIS for Rescheduling
Natural immune systems are adaptive systems in which learning takes place by evolutionary mechanisms
similar to biological evolution. AIS is a new computational paradigm based on natural immune system.
Different approaches to modeling AISs are presented in the literature [20–22]. We will use a simple
GA-based model of AIS studied by Forrest et al. [21]. They use two populations: antigens (foreign
material) and antibodies (the cell that perform the recognition). Binary strings represent individuals in
these populations and recognition is highly simplified and modeled as string matching. GA is used to
evolve population of antibodies that match specific antigens well.
Artificial immune systems (AISs) have been applied in many domain [7], for example, to create anti-
virus programs; detect intrusions in computer networks, in pattern recognition, data compression, and
optimization. In scheduling domain, an AIS, in conjunction with a GA, was used to produce effective
schedules in a factory environment [23].
In our approach rules discovered by the GA are interpreted as antibodies and program graphs — as
antigens. A given antibody (a rule) “recognizes” a specific antigen (a program graph) if it can find an
optimal or near optimal schedule for it. When a new scheduling instance arrives to the system, the AIS
will form a population of antibodies — rules from the stored library, and next run the population with
use of GA (the same as in the learning mode) to evolve faster scheduling rules for the new problem.
Figure 17.12 presents an idea of our AIS used for rescheduling. In our approach, rules discovered for
different program graphs (the whole populations) are stored in the library. When we want to discover

New program graph

Library of rules

101 .................................... 101

011 .................................... 011
..........................

..........................

101 ............. 101

011 .................................... 01 GA
...........

...........

101 ............. 011

Best rules +
random rules

FIGURE 17.12 An idea of the AIS for rescheduling.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 266 — #14

Scheduling and Rescheduling with Use of Cellular Automata 17-267

rules for a new program graphs, first, we choose from the library some number of the best rules (those
which the best “recognize” a new program graph). Then we add to these rules some number of random
rules.
An initial population of rules is created as follows: First, we run the CA-based system in the normal
operating mode applying stored rules from the library to evaluate schedules T proposed by them. As shown
earlier (see Table 17.2) this mode of work is characterized by a very low computational cost. Next, we
select the initial population of AIS d · P the best rules, where d ∈ (0, 1). Remaining (1 − d) · P rules are
randomly generated. Finally, the whole population of AIS rules is evolved by GA in the learning mode.
It is important to underline that all rules in the library have the same length (they were discovered
assuming r = 3). It enables applying genetic operators on rules from different program graphs.

17.6.2 Experimental Results

An experiment is conducted with the program graph called Rnd25_1g 18 shown in Figure 17.13. It was
constructed through linking the program graph g 18 to final tasks of Rnd25_1 and setting the weights of
linking edges to zero. Both components of Rnd25_1g 18 are known to the system, but their composition is
a new instance for scheduling. The optimal response time T for Rnd25_1g 18 is equal to 541. The library
of AIS rules is composed of already discovered rules (see Table 17.2).
Preliminarily, we test on Rnd25_1g 18 in the normal operating mode rules found earlier for Rnd25_1
and g 18. We can see (see Figure 17.14) that none of these rules provide an optimal solution. We also run the

0 1 2 3 4 5

6 7 8 9 10 11 12 13

14 15 16 17 18 19

20 21 22 23 24 25 26 27

28 29 30 31
0 0
0 0
32

33 34 35 36 37 38 39 40

41 42 43 44

45 46 47 48

49

FIGURE 17.13 Program graph Rnd25_1g 18.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 267 — #15

17-268 Handbook of Bioinspired Algorithms and Applications

Rnd 25_1g18 Rnd 25_1g18

(a) 620 (b)
610 640 1000 tests

600 1000 tests 620

Average of T

590

Average of T
600
580
570 580
560
560
550
540 540
0 20 40 60 80 100 120 140 160 180 0 5 10 15 20 25 30 35 40 45
Rules Rules

FIGURE 17.14 Normal operating phase for Rnd25_1g 18: (a) population of rules discovered for Rnd25_1,
(b) population of rules discovered for g 18.

(a) Rnd 25_1g18 (b) Rnd 25_1g18

600 600

590 590
Response time T

Response time T

580 580

570 Best 570

av. el.
av. pop.
560 560 Best
av. el.
av. pop.
550 550

540 540
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Generation Generation

FIGURE 17.15 Learning mode for Rnd25_1g 18: (a) randomly generated initial population (b) AIS with population
of rules from the library.

scheduling system in the standard learning mode with randomly created CA rules. We set P = 200, E = 50,
and pm = .03. We can see (Figure 17.15[a]) that the system can discover, during 200 generations, rules
providing only suboptimal solutions.
Finally, we run AIS in the learning mode. The initial population of AIS rules was composed of 60% of
rules from the library, after evaluating them in the normal operating mode, and the remaining rules were
randomly generated. We can see (Figure 17.15[b]) that the behavior of AIS is different when compared
with the behavior of the system in the standard learning mode. AIS that uses the knowledge stored in rules
library can discover the optimal CA scheduling rule after about 60 generations.
Figure 17.16 shows typical run of the CA-based scheduler with the best-discovered rule, starting from
randomly generated IC. One can see that in the step 0, cells of the CA are in some states correspond-
ing to the allocation of tasks and the value of T corresponding to this allocation is >656. Then the
CA starts to change its states sequentially, which results in changing values of T . One can see that
the CA needs <70 time steps to converge to tasks’ allocation corresponding to the minimal value of
T = 541.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 268 — #16

Scheduling and Rescheduling with Use of Cellular Automata 17-269

20

40

60

80

99
536 548 560 572 584 596 608 620 632 644 656 668 680
T = 541

FIGURE 17.16 Space-time diagram of CA-based scheduler for Rnd25_1g 18.

TABLE 17.4 The Best Response Time T for 3,

4, and 8 Processors

Program graph k=3 k=4 k=8

tree15 7.0 (7) 7.0 (7) 7.0 (7)

g 18 38.0 (36) 27.0 (26) 26.0 (24)
g 40 57.0 (57) 46.0 (45) 39.0 (33)
gauss18 48.0 (44) 52.0 (44) 52.0 (44)

Other conducted experiments have shown the advantage of AIS operating on rules from the library,
significantly accelerating the process of discovery rules for new program graphs, which were partially
presented earlier to the system.

17.7 Scheduling with More than Two Processors

When considering scheduling problem in the case of more than two processors, we are faced with the
major problem — the problem of the length of the rule. The length of the rule L is calculated as L = k 2r+1 .
We can see that the length of the rule grows very quickly with k and r. It results in increasing the search
space of possible solutions and creates a complex combinatorial optimization problem.
Table 17.4 presents the best response time T obtained for some program graphs (seq mode of CA) in
the case of 3, 4, and 8 processors (assuming a fully connected system). They are compared with results
obtained by a standard GA (in brackets).
One can see that the system can discover CA rules providing the optimal scheduling of tree15 for
all considered multiprocessor systems. For remaining program graphs, under given setting parameters
(r = 1, P = 300, E = 100, I = 50, and pm = .03), the system discovered rules providing suboptimal
solutions.
Figure 17.17 shows a typical run of GA for the program graph g 40 and multiprocessor system consisting
of eight processors. In this case the optimal response time T = 33. We can see that GA needs about 200
generations to find CA rules providing suboptimal scheduling with T = 39.
Figures 17.18 to 17.20 show typical runs of the CA-based scheduler in the normal operation mode, with
the best rule from the final generation, for the program graphs g 40 and 3-, 4-, and 8-processor systems,
respectively starting from randomly generated IC. Note that there is another important feature of the
CA-based scheduler in the normal operating mode. The system is not sensitive to the increasing number
of processors. Scheduling is performed in a number of steps similar to the case of two-processor system.
The ways of increasing the efficiency of the proposed scheduling system is the subject of our current
research. It includes using new definitions of CA rules and applying nonuniform CA [16].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 269 — #17

17-270 Handbook of Bioinspired Algorithms and Applications

g 40
46
45
44 Best

Response time T
av. el.
43 av. pop.

42
41

40
39
38
0 50 100 150 200 250
Generation

FIGURE 17.17 Learning mode for g 40 and eight processors.

20

40

59
52 54 56 58 60 62 64 66 68 70 72 74 76
T = 57

FIGURE 17.18 Space-time diagrams of CA-based scheduler for g 40 and three processors: sequential mode.

20

40
49
41 44 47 50 53 56 59 62 65 68 71 74 77
T = 46

FIGURE 17.19 Space-time diagrams of CA-based scheduler for g 40 and four processors: sequential mode.

20

40
49
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
T = 39

FIGURE 17.20 Space-time diagrams of CA-based scheduler for g 40 and eight processors: sequential mode.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 270 — #18

Scheduling and Rescheduling with Use of Cellular Automata 17-271

17.8 Conclusions
In this chapter we have presented the results of research on developing CA-based scheduler, working
in sequential and parallel modes. They show that the GA can discover CA rules suitable to solve the
scheduling problem for a given instance of the problem. This phase of the algorithm is called the learning
phase. In this phase, knowledge about solving a given instance of the scheduling problem is extracted and
coded into CA rules. Discovered rules are used in the normal operating phase by CA-based scheduler for
automatic scheduling.
However, the main effort was directed to the question of a potential reusing of discovered CA rules.
To solve this problem we proposed the immune system approach implemented by GA, which matches
the currently available knowledge in discovered and stored CA rules with new instances of the scheduling
problem and quickly recognizes familiar building blocks in these instances.
Proposed hybrid technique opens very promising possibilities in developing parallel and distributed
scheduling algorithms and reducing their complexity.
Future work will include the expanding of the immune system concept. Perhaps it is possible to keep
in the library not the whole chromosomes, as in the current approach, but only selected genes. It might
help in further reducing the complexity of the algorithm.

References
[1] J. Błażewicz, K.H. Ecker, G. Schmidt, and J. Wȩglarz, Scheduling in Computer and Manufacturing
Systems, Springer-Verlag, Heidelberg, 1994.
[2] M.R. Gary and D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP-
Completeness. W.H. Freeman and Company, San Francisco, CA, 1979.
[3] H. El-Rewini, T.G. Lewis, and H.H. Ali, Task Scheduling in Parallel and Distributed Systems. PTR
Prentice Hall, Englewood Cliffs, NJ, 1994.
[4] Y.K. Kwok and I. Ahmad, Benchmarking the task graph scheduling algorithms. In Proceedings of
1998 IPPS/SPDP Symposium, Orlando, FL, 1998, pp. 531–537.
[5] S. Saleh and A.Y. Zomaya, Multiprocessor scheduling using mean-field annealing. In Parallel and
Distributed Processing, Vol. 1388 of Lecture Notes in Computer Science Springer-Verlag, Heidelberg,
1998, pp. 288–296.
[6] L. Wang, H.J. Siegel, V.P. Roychowolhury, and A.A. Maciejewski, Task matching and scheduling
in heterogeneous computing environments using a genetic-algorithm-based approach. Journal of
Parallel and Distributed Computing, 47, 8–22, 1997.
[7] L.N. de Castro and J. Timmis, Artificial Immune Systems: A New Computational Intelligence
Approach, Springer-Verlag, Heidelberg, 2002.
[8] A. Schoneveld, Parallel Complex Systems Simulation, Ph.D. thesis, University of Amsterdam,
Holland, 1999 (http://www.science.uva.nl/research /pscs/papers/phd.html).
[9] M. Mitchell, Computation in cellular automata. In T. Gramb, S. Bornholdt, M. Grob, M. Mitchell,
and T. Pellizzari (Eds.), Non-Standard Computation, Wiley-VCH, Weinhein, Federal Republic of
Germany, 1998, pp. 95–140.
[10] B.J. Overeinder, Distributed Event-driven Simulation — Scheduling Strategies and
Resource Management, Ph.D. thesis, University of Amsterdam, Holland, 2000
(http://www.science.uva.nl/research /pscs/papers/phd.html).
[11] R. Das, M. Mitchell, and J.P. Crutchfield, A genetic algorithm discovers particle-based computa-
tion in cellular automata. In Y. Davidor, H.-P. Schwefel, and R. Männer (Eds.), Parallel Problem
Solving from Nature — PPSN III, Vol. 866 of Lecture Notes in Computer Science. Springer-Verlag,
Heidelberg, 1994, pp. 344–353.
[12] M. Sipper, Evolution of Parallel Cellular Machines. The Cellular Programming Approach, LNCS
1194, Springer-Verlag, Heidelberg, 1997.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 271 — #19

17-272 Handbook of Bioinspired Algorithms and Applications

[13] R. Subrata and A.Y. Zomaya, Evolving cellular automata for location management in mobile
computing. IEEE Transactions on Parallel and Distributed Systems, 14, 13–26, 2003.
[14] M. Tomassini, M. Sipper, and M. Perrenoud, On the generation of high-quality random numbers
by two-dimensional cellular automata. IEEE Transactions on Computers, 49, 1140–1151, 2000.
[15] F. Seredyński and A.Y. Zomaya, Sequential and parallel cellular automata-based scheduling
algorithms. IEEE Transactions on Parallel and Distributed Systems, 13, 1009–1023, 2002.
[16] S. Wolfram, Universality and complexity in cellular automata. Physica D, 10, 1–35, 1984.
[17] F. Seredyński, Scheduling tasks of a parallel program in two-processor systems with use of cellular
automata, Future Generation Computer Systems, 14, 351–364, 1998.
[18] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Springer-Verlag,
Heidelberg, 1992.
[19] D.H. Wolpert and W.G. Macready, No free lunch theorems for optimization. IEEE Transactions on
Evolutionary Computation, 1, 67–82, 1997.
[20] J.D. Farmer, N.H. Packard, and A.S. Perelson, The immune system, adaptation, and machine
learning. Physica D, 22, 187–204, 1986.
[21] S. Forrest, B. Javornik, R. Smith, and A.S. Perelson, Using genetic algorithms to explore pattern
recognition in the immune system. Evolutionary Computation, 1, 191–211, 1993.
[22] A.S. Perelson, Immune network theory. Immunological Review, 110, 5–36, 1989.
[23] E. Hart and P. Ross, An immune system approach to scheduling in changing environments. In
W. Banzhaf et al. (Eds.), GECCO-99: Proceedings of the Genetic and Evolutionary Computation
Conference. Morgan Kaufmann, San Mateo, CA, 1999, pp. 1559–1566.
[24] A.A. Khan, C.L. McCreary, and M.S. Jones, A Comparison of multiprocessor scheduling heuristics.
In Proceedings of International Conference on Parallel Processing, Vol. II, 1994, pp. 243–250.
[25] M. Mitchell and S. Forrest, Genetic algorithms and artificial life. In Ch.G. Langton (Ed.), Artificial
Life. An Overview. The MIT Press, Cambridge, MA, 1995.
[26] F. Seredyński and A. Świȩcicka, Immune-like system approach to multiprocessor scheduling, In
R. Wyrzykowski, J. Dongarra, M. Paprzycki, and J. Wasniewski (Eds.), Parallel Processing and
Applied Mathematics, Vol. 2328, of Lecture Notes in Computer Science. Springer-Verlag, Heidelberg,
2002, pp. 626–633.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C017” — 2005/8/6 — 14:09 — page 272 — #20

 18
Cellular Automata,
PDEs, and
Pattern Formation

18.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-273

18.2 Cellular Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-274
Fundamentals of Cellular Automaton • Finite-State Cellular
Automata • Stochastic Cellular Automata • Reversible
Cellular Automata
18.3 Cellular Automata for PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-276
Rules-Based System and Equation-Based System • Finite
Difference Scheme and Cellular Automata • Cellular Automata
for Reaction-Diffusion Systems • Cellular Automata for the
Wave Equation • Cellular Automata for Burgers Equation
with Noise
18.4 Differential Equations for Cellular Automata. . . . . . . . . . 18-279
Formulation of Differential Equations from Cellular
Automata • Stochastic Reaction-Diffusion
18.5 Pattern Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-280
Complexity and Pattern in Cellular Automata • Comparison
of Cellular Automata and PDEs • Pattern Formation in
Xin-She Yang Biology and Engineering
Y. Young References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-283

18.1 Introduction
A cellular automaton (CA) is a rule-based computing machine, which was first proposed by von Newmann
in early 1950s and systematic studies were pioneered by Wolfram in 1980s. Since a cellular automaton
consists of space and time, it is essentially equivalent to a dynamical system that is discrete in both space
and time. The evolution of such a discrete system is governed by certain updating rules rather than
differential equations. Although the updating rules can take many different forms, most common cellular
automata use relatively simple rules (Von Newmann, 1966; Wolfram, 1983). On the other hand, an
equation-based system such as the system of differential equations and partial differential equations also
describe the temporal evolution in the domain. Usually, differential equations can also take different forms
that describe various systems. Now one natural question is what is the relationship between a rule-based

18-273

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 273 — #1

18-274 Handbook of Bioinspired Algorithms and Applications

f1(t ) fi –2(t ) fi –1(t ) fi (t ) fi +1 (t) fi –2 (t ) fN (t )

••• •••
fi (t +1)

FIGURE 18.1 Diagram of a one-dimensional cellular automaton.

system and an equation-based system? Given differential equations, how can one construct a rule-based
cellular automaton, or vice versa? There has been substantial amount of research in these areas in the
past two decades. This chapter intends to summarize the results of the relationship among the cellular
automata, partial differential equations (PDEs), and pattern formations.

18.2 Cellular Automata

18.2.1 Fundamentals of Cellular Automaton
On a one-dimensional (1D) grid that consists of N consecutive cells, each cell i (i = 1, 2, . . . , N ) may be
at any of the finite number of states, k. At each time step, t , the next state of a cell is determined by its
present state and the states of its local neighbors. Generally speaking, the state φi at t + 1 is a function of
its 2r + 1 neighbors with r cells on the left of the concerned cell and r cells on its right (see Figure 18.1).
The parameter r is often referred to as the radius of the neighborhood.
The number of possible permutations for k finite states with a radius of r is p = k 2r+1 , thus the
number of all possible rules to generate the state of cells at next time step is k p , which is usually very large.
For example, if r = 2, k = 5, then p = 125 and the number of possible rules is 5125 ≈ 2.35 × 1087 , which
is much larger than the number of stars in the whole universe.
The rule determining the new state is often referred to as the transition rule or updating rule.
In principle, the state of a cell at next time step can be any function of the states of some neighbor
cells, and the function can be linear and nonlinear. There is a subclass of the possible rules according
to which the new state depends only on the sum of the states in a neighborhood, and this will simplify
the rules significantly. For this type of updating rules, the number of possible permutations for k states
and 2r + 1 neighbors is simply k(2r + 1), and thus the number of all possible rules is k k(2r+1) . Then
for the same parameter k = 5, r = 2, the number of possible rules is 515 = 3.05 × 1010 , which is much
smaller when compared with 5125 . This subclass of sum-rule or totalistic rule is especially important in the
popular cellular automata such as Conway’s Game of Life, and in the cellular automaton implementation
of partial differential equations.
The dynamics and complexity of cellular automata are extremely rich. Stephan Wolfram’s pioneering
research and mathematical analysis of cellular automata led to his famous classification of 1D cellular
automata:

Class 1: The first class of cellular automata always evolves after a finite number of steps from almost
all initial states to a homogeneous state where every cell is in the same state. This is something like
fixed point equilibrium in the dynamical system.
Class 2: Periodic structures with a fixed number of states occur in the second class of cellular automata.
Class 3: Aperiodic or “chaotic” structures appear from almost all possible initial states in this type of
cellular automata.
Class 4: Complex patterns with localized spatial structure propagate in the space as time evolves.
Eventually, these patterns will evolve to either homogeneous or periodic. It is suggested that this
class of cellular automata may be capable of universal computation.

Cellular automata can be formulated in higher dimensions such as 2D and 3D. One of the most
popular and yet very interesting 2D cellular automata using relatively simple updating rules is the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 274 — #2

Cellular Automata, PDEs, and Pattern Formation 18-275

Conway’s Game of Life. Each cell has only two states k = 2, and the states can be 0 and 1. With a
radius of r = 1 in the 2D case, each cell has eight neighbors, thus the new state of each cell depends on
total nine cells surrounding it. The boundary cells are treated as periodic. The updating rules are: if two
or three neighbors of a cell are alive (or 1) and it is currently alive, then it is alive at next time step; if three
neighbors of a cell are alive (or 0) and it is currently not alive its next state is alive; the next state is not
alive for all the other cases. It is suggested that this simple automaton may have the capability of universal
computation. There are many existing computer programs such as Life in Matlab and screen savers on all
computer platforms such as Windows and Unix.

18.2.2 Finite-State Cellular Automata

In general, we can define a finite-state cellular automaton with a transition rule G = [gij,...,l ], (i, j, . . . , l =
t +1
1, 2, . . . , N ) from one state t = [φij,...,l
t ] at time level n to a new state t +1 = [φij,...,l ] at a new time step
n + 1. The value of subscript (i, j, . . . , l) denotes the dimension, d, of the cellular automaton. Therefore,
a CA in the d-dimensional space has N d cells. For the 2D case, this can be written as

G : t  → t +1 , gij : φijt → φijt +1 , (i, j = 1, 2, . . . , N ).

In the case of sum-rule with a 4r + 1 neighbors, this becomes


r 
r 
φijt +1 = G t
aαβ φi+α, j+β , (i, j = 1, 2, . . . , N ),
α=−r β=−r

where aαβ (α, β = ±1, ±2, . . . , ±r) are the coefficients. The cellular automata with fixed rules defined
this way are deterministic cellular automata. In contrast, there exists another type, namely, the stochastic
cellular automata that arise naturally from the stochastic models for natural systems (Guinot, 2002;
Yang, 2003).

18.2.3 Stochastic Cellular Automata

When using cellular automata to simulate the phenomena with stochastic components or noise such as
percolation and stochastic process, the more effective way is to introduce some probability associated with
certain rules. Usually, there is a set of rules and each rule is applied with a probability (Guinot, 2002).
Another way is that the state of a cell is updated according to a rule only if certain conditions are met or
certain values are reached for some random variables. For example, the rule for 2D a cellular automaton
g (φijt ) = φijt +1 is applied at a cell only if a random variable v ≤ (φijt ) where the function  ∈ [0, 1].
At each time step, a random number v is generated for each cell (i, j), and the new state will be updated only
if the generated random number is greater than , otherwise, it remains unchanged. Cellular automata
constructed this way are called stochastic or probabilistic cellular automata. An example is given later in
the next section.

18.2.4 Reversible Cellular Automata

A cellular automaton with an updating rule φijt +1 = g (φijt ) is generally irreversible in the sense that it
is impossible to know the states of a region such as all zeros were the same at a previous time step or
not. However, certain class of rules will enable the automata to be reversible. For example, a simple finite
difference (FD) scheme for a dynamical system

u(t + 1) = g [u(t )] − u(t − 1) or u(t − 1) = g [u(t )] − u(t + 1),

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 275 — #3

18-276 Handbook of Bioinspired Algorithms and Applications

is reversible since for any function g (u), one can compute u(t + 1) from u(t ) and u(t − 1), and invert
u(t − 1) from u(t ) and u(t + 1). The automaton rule for 2D reversible automata can be similarly
constructed as

t +1 t t −1
ui,j = g (ui,j ) − ui,j ,

together with appropriate boundary conditions such as fixed-state boundary conditions (Margolus, 1984).

18.3 Cellular Automata for PDEs

Cellular automata have been used to study many phenomena (Vichniac, 1984; Wolfram, 1984a, 1994;
Weimar, 1997; Yang, 2002). In fact, many natural phenomena behave like finite-state cellular automata,
and these include self-organized systems such as pattern formation in biological system, insect colonies,
and ecosystem; multiple particle system such as the lattice-gas, granular material, and fluids; autocatalytic
systems such as enzyme functionality and mineral reactions; and even systems involving society and
culture interactions. However, most of these systems have been studied using continuum-based differential
equations. We now focus on the formulation of automaton rules from corresponding PDEs.

18.3.1 Rules-Based System and Equation-Based System

Equation-based relationships, often in the form of ordinary differential equations and partial differential
equations, form the continuum models of most physical, chemical, and biological processes. Differential
equations are suitable and work well for systems with only a small degree of freedom and evolution of
system variables in the continuous and smooth manner. There have been vast literatures on analysis and
solution technique of the differential models. On the other hand, cellular automata are often considered
as an alternative approach and may compliment the existing mathematical basis. The state variables are
always discrete, but the numbers of degree of freedom are large (Wolfram, 1984b).
Although continuum models have advantages such as high accuracy and conservation laws, mathe-
matical analysis are usually very difficult and the analytical solutions do not always exist. Only few
differential equations have a closed-form solution. In last several decades, the numerical methods have
become the essential parts of the solution and of the analysis of almost all problems in engineering,
physics, and biology. In fact, computational modeling and numerical computation have become the third
component, bridging the gap between theoretical models and experiments. As the computing speed and
memory of computers increase, computer simulations have become a daily routine in science and engi-
neering, especially in multidisciplinary research. There are also vast literatures concerning the numerical
algorithms, numerical solutions of partial differential equations, and others. These include the following
well-established methods: finite difference method, finite element method, finite volume method, cellular
automata, lattice-gas, Monte-Carlo method, and genetic algorithms.
Finite difference methods work very well for many problems, but have disadvantages in dealing with
irregular geometry. Finite element and finite volume methods can deal with irregular geometry and thus
are commonly used and there are quite a few commercial software packages available. Lattice-gas and
Monte-Carlo methods are suitable for many problems in physics, especially for systems with multibodies
or multiparticles. Cellular automaton is a very interesting and powerful method, and it has gradually
become an essential part of the numerical computations owing to its universal computability and the
nature of parallel implementation.

18.3.2 Finite Difference Scheme and Cellular Automata

There is a similarity between finite difference scheme and finite-state cellular automata. Finite difference
scheme is the discretization of a differential equation on a regular grid of points with the evolution of the
states over the discrete time steps. Even the state variable from a differential equation may have continuous

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 276 — #4

Cellular Automata, PDEs, and Pattern Formation 18-277

values; numerical computation on a computer always lead to the discrete values due to the limited bits of
processors or round-off. Similarly, cellular automata are also about the evolution of state variables with
finite number of values on a regular grid of cells at different discrete time steps. If the number of states of
a cellular automaton is comparable with that of the related finite difference equation, then we can expect
the results to be comparable.
To demonstrate this, we choose the 1D heat equation

∂T ∂ 2T
=κ 2,
∂t ∂x

where T is temperature and κ is the thermal diffusivity. This is a mathematical model that is widely used
to simulate many phenomena. The temperature T (x, t ) is a real-valued function, and it is continuous
for any time t > 0 whatever the initial conditions. In reality, it is impossible to measure the temperature
at a mathematical point, and the temperature is always the average temperature in a finite representative
volume over certain short time. Mathematically, one can obtain a closed-form solution with infinite accu-
racy in the domain, but physically the temperature would only be meaningful at certain macroscopic
levels. No matter how accurate the solution may have at very fine scale, it would be meaningless to try
to use the solution at the atomic or subatomic levels where quantum mechanics come into play and, the
solution of temperature is invalid (Toffoli, 1984). Thus, numerical computation would be very useful even
though it has finite discrete values.
The simplest discretization of the above heat equation is the central difference for spatial derivative and
forward scheme for time derivatives, and we have

κt
Tin+1 − Tin = (T n − 2Tin + Ti−1
n
),
(x)2 i+1

where i and n are the spatial and time indices. If we choose the time steps and spatial discretization such
that κt /(x)2 = 1, now we have

Tit +1 = (Ti+1
t
+ Tit + Ti−1
t
) − 2Tit ,

which is something like the “mod 2” cellular automata or Wolfram’s cellular automata with rule 150
(Wolfram, 1984a; Weimar, 1997).
Cellular automata obtained this way are very similar to the finite difference method. If the state variables
are discrete, they are exactly the finite-state cellular automata. However, one can use the continuous-
valued state variable for such simulation, in this case, they are continuous-valued cellular automata from
differential equations as studied by Rucker’s group and used in their the well-known CAPOW program
(Rucker, 2003; Ostrov and Rucker, 1996). In some sense, the continuous-valued cellular automata based
on the differential equations are the same as the finite difference methods, but there are some subtle
differences and advantages of cellular automata over the finite difference simulations due to the CA’s
properties of parallel nature, artificial life-oriented emphasis on experiment and observation and genetic
algorithms for searching large phase space of the rules as proposed by Rucker et al. (1998).

18.3.3 Cellular Automata for Reaction-Diffusion Systems

By extending the discretization procedure from differential equations to derive automaton rules for cel-
lular automata, we now formulate the cellular automata from their corresponding partial differential
equations. First, let us start with the reaction-diffusion equation that may form beautiful patterns in the
2D configuration


∂u ∂ 2u ∂ 2u
=D + 2 + f (u),
∂t ∂x 2 ∂y

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 277 — #5

18-278 Handbook of Bioinspired Algorithms and Applications

where u(x, y, t ) is the state variable that evolves with time in a 2D domain, and the function f (u) can be
either linear or nonlinear. D is a constant depending on the properties of diffusion. This equation can
also be considered as a vector form for a system of reaction-diffusion equations if let D = diag(D1, D2 ),
u = [u1 u2 ]T . The discretization of this equation can be written as

n+1  
ui,j − ui,j
n n
ui+1,j − 2ui,j
n + un
i−1,j
n
ui,j+1 − 2ui,j
n + un
i,j−1 n
=D + + f (ui,j ),
t (x)2 (y)2

by choosing t = x = y = 1, we have

n+1 n n n n n n
ui,j = D[ui+1,j + ui−1,j + ui,j+1 + ui,j−1 ] + f (ui,j ) + (1 − 4D)ui,j ,

which can be written as the generic form


r
t +1 t t
ui,j = ak,l ui+k,j+l + f (ui,j ),
k,l=−r

where the summation is over the 4r + 1 neighborhood. This is a finite-state cellular automaton with the
coefficients ak,l being determined from the discretization of the governing equations, and for this special
case, we have a−1,0 = a+1,0 = a0,−1 = a0,+1 = D, a0,0 = 1 − 4D, r = 1.

18.3.4 Cellular Automata for the Wave Equation

For the 1D linear wave equation,

∂ 2u 2∂ u
2
= c ,
∂t 2 ∂x 2

where c is the wave speed. The simplest central difference scheme leads to

uin+1 − 2uin + uin−1 u n − 2uin + ui−1

n
= c 2 i+1 .
(t ) 2 (x) 2

By choosing t = x = 1, t = n, it becomes

uit +1 = [ui+1
t t
+ ui−1 + 2(1 − c 2 )uit ] − uit −1 .

This can be written in the generic form

uit +1 + uit −1 = g (u t ),

which is reversible under certain conditions. This property comes from the reversibility of the wave
equation because it is invariant under the transformation: t → −t .

18.3.5 Cellular Automata for Burgers Equation with Noise

One of the important equations arising in many processes such as turbulent phenomenon is the noisy
Burgers equation

∂u ∂u ∂ 2u
= 2u + 2 + ∇υ,
∂t ∂x ∂x

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 278 — #6

Cellular Automata, PDEs, and Pattern Formation 18-279

where υ is the noise that is uncorrelated in space and time so that υ(x, t ) = 0 and υ(x, t )υ(x0 , t0 ) =
2Dδ(x − x0 )δ(t − t0 ) (Emmerich and Kahng, 1998). This equation with Gaussian white noise can be
rewritten as

∂u ∂u ∂ 2u
+ ξ = 2u + 2 + η,
∂t ∂x ∂x

where both ξ and η are uncorrelated. By introducing the variables vit = c exp(x uit ), φIt = β ln(vit ),
α = t /(x)2 , (1 − 2α)/cα = exp(−A/β), c 2 = exp(B/β), ξ = exp(), η = exp() and after some
straightforward calculations in the limit of β tends zero, we have the automata rule

φit +1 = φi−1
t
+ max[0, φit − A, φit + φi+1
t
− B, it − φi−1
t
]
t t
− max[0, φi−1 − A, φi−1 + φit − B, ti − φi−1
t
],

where we have used the following identity,

lim ε ln(eA/ε + eB/ε + · · · ) = max[A, B, . . .].

φ→+0

This forms a generalized probabilistic cellular automaton that is referred to as the noisy Burgers cellular
automaton. Burgers equation without noise usually evolves in shock wave, and in the presence of noise,
the states of the probabilistic cellular automata may be taken as discrete reminders of those shock waves
that were disorganized.

18.4 Differential Equations for Cellular Automata

Computer simulations based on cellular automata and partial differential equations work remarkably
well for many different reasons. One possibility is that finite-state cellular automata and finite difference
approximations using discrete time and space with a finite precision can represent physical variables
very well and thus the models are insensitive to very small space and time scales. It is relatively straight-
forward to derive the updating rules for cellular automata from the corresponding partial differential
equations. However, the reverse is usually very difficult. There is no general method available to formulate
the continuum model or differential equations for given rule-based cellular automata despite the obvious
importance. Fortunately, there have been some important progress made in this area (Omohundro, 1984;
Wolfram, 1984b), and we outline some of the procedures of formulating continuum equations for cellular
automata.

18.4.1 Formulation of Differential Equations from Cellular Automata

The mathematical analysis for cellular automata was first pioneered by Wolfram, and has attracted wide
attention since 1980s. Wolfram (1983) gave an extensive analysis of statistical mechanics of cellular
automata. Later on, Omohundro (1984) provided an instructive procedure of formulating the partial
differential equations for cellular automata by using 10 PDE variables in 2D configurations. These ten
variables are the state variable P(x, y, t ) at present, new state N (x, y, t ) and eight variables U1 , . . . , U8 with
eight bump or bell functions. The eight variables have the same format of information as the P(x, y, t ).
On a 2D grid, these eight functions S1 , . . . , S8 are shown in Figure 18.2 together with shifted coordinates.
According to Omohundro’s formulation, we assume the bumps to be α wide and constant outside β
from the transition. If the width of a cell is 1, then α = 1/5 and β = 1/100. The eight functions are

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 279 — #7

18-280 Handbook of Bioinspired Algorithms and Applications

S2 S3 S4
(–1,+1) (0,+1) (+1,+1)

S1 N/P S5

S8 S7 S6
(–1,–1) (0,–1) (+1,–1)

FIGURE 18.2 Diagram of eight neighbors and the positions of Omohundro’s functions.

taken as
  β
S1 (xβ/α) x
S1 = e−1/(β−x) (−β < x < β), 0 (|x| ≥ β),
2
S2 = , S3 = S1 (x)dx S1 (x)dx,
S1 (0) −1 −β
∞

S4 = S3 (x − α/2)S3 (α/2 − x), S5 = S3 (x − α)S3 (α − x), S6 = S4 (x − k),
k=−∞
∞
 ∞

S7 = S5 (x − k), S8 = S2 (x − k).
k=−∞ k=−∞

By using these functions, Omohundro derived the following equation:


∂N dS2 ∂N dS2 ∂N
= −γ + ,
∂t dx ∂x dy ∂y

where γ is a large constant. Differential equations for other variables can be written in a similar manner
although they are more complicated (Omohundro, 1984).

18.4.2 Stochastic Reaction-Diffusion

A stochastic cellular automaton on a 1D ring has N sites with simple rules for the state variable ui (t )
and S = ui (t ) + ui+1 (t ): (1) ui (t + 1) = 0, if S = 0; (2) ui (t + 1) = 1 with probability p1 , if S = 1;
(3) ui (t + 1) = 1 with probability p2 , if S = 2. This Domany–Kinzel cellular automaton is equivalent to
the following stochastic reaction-diffusion equation:

∂v √
= D∇ 2 v + f (v) + ε v,
∂t

where v is the concentration of live sites or ui (t ) = 1, and ε is a zero-mean Gaussian random variable
with unit variance (Ahmed and Elgazzar, 2001). However, there is nonuniqueness associated with the
formulation of differential equations from the cellular automata. Bagnoli et al. (2001) demonstrated that
the above equation can be obtained from the following rules: (1) ui (t + 1) = 0, if  = ui+1 (t ) + ui (t ) +
ui−1 (t ) = 0; (2) ui (t + 1) = 1 with probability p1 , if  = 1; (3) ui (t + 1) = 2 with probability p2 , if
 = 2; (4) ui (t + 1) = 1, if  = 3. This nonuniqueness in the relationship between cellular automata
and PDEs require more research.

18.5 Pattern Formation

The behavior and characteristics of cellular automata are very complicated and there is no general or
universal mathematical analysis available for the description of such complexity. Even for 1D cellular
automata, they are complicated enough as shown by Wolfram classifications. Pattern formation is one
of the typical characteristics in cellular automata. The study of pattern complexity and the conditions

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 280 — #8

Cellular Automata, PDEs, and Pattern Formation 18-281

(a) Space (b) (c)

FIGURE 18.3 Pattern formation in cellular automata: (a) 1D CA with disordered initial conditions; (b) CA for 1D
wave equation; and (c) nonlinear 1D Sine-Gordon equation.

of its formation is essential in many processes such as biological pattern formation, enzyme dynamics,
percolation, and other processes in engineering applications (Turing, 1952; Flake, 1998; Cappuccio et al.,
2001; Boffetta et al., 2002). This section focuses on the pattern formation in cellular automata and
comparison of CA results with the results using differential equations.

18.5.1 Complexity and Pattern in Cellular Automata

Wolfram (1983) pioneered the studies of complexity and patterns in cellular automata. Complex patterns
form in 1D cellular automata even with very simple rules. Figure 18.3(a) shows the typical patterns in 1D
cellular automata with random initial conditions for k = 16 states and r = 1. Figure 18.3(b) is for 1D
wave equation with k = 1024 states and r = 1. Figure 18.3(c) corresponds to the nonlinear wave based
on Sine-Gordon equation utt − uxx = α sin(u) with α = −0.1, k = 1024 states, r = 2 and sinusoidal
initial conditions. The pattern formations are more complicated in higher dimensions. We compare the
results from different methods while we study the pattern formations in the next section.

18.5.2 Comparison of Cellular Automata and PDEs

For a 2D reaction-diffusion system of two partial differential equations:

∂u ∂v
= Du ∇ 2 u + f (u), = Dv ∇ 2 v + g (v), f (u, v) = α(1 − u) − uv 2 ,
∂t ∂t
(α + β)v
g (u, v) = uv 2 − ,
1 + (u + v)

where Du = 0.05. The parameters γ = Dv /Du , α, β can vary so as to produce the complex patterns.
This system can model many systems such as enzyme dynamics and biological pattern formations by slight
modifications (Murray, 1989; Meinhardt, 1982, 1995; Keener and Sneyd, 1998; Yang, 2003). Figure 18.4
shows a snapshot of patterns formed by the simulations of the above reaction-diffusion system at t = 500
for γ = 0.6, α = 0.01, and β = 0.02. The right plot is the comparison of results obtained by three
different methods: cellular automata (marked with CA), finite difference method (FD), and finite element
method (FE). The plot is for the data on the middle line of the pattern shown on the left.

18.5.3 Pattern Formation in Biology and Engineering

Pattern formation occurs in the many processes in biology and engineering. We will give two examples
here to show the complexity and diversity of the beautiful patterns formed. Figure 18.5 shows the
2D and 3D patterns for the reaction diffusion system with f (u, v) = α(1 − u) − uv 2 /(1 + u + v),
g (u, v) = uv 2 − (α + β)v/(1 + u + v).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 281 — #9

18-282 Handbook of Bioinspired Algorithms and Applications

100 0.6 0.8

CA
90 FD
0.55 0.7 FE
80
70 0.5 0.6
60
50 0.45 0.5
40
0.4 0.4
30
20 0.35 0.3
10
0.3 0.2
20 40 60 80 100 20 30 40 50 60 70 80

FIGURE 18.4 A snapshot of pattern formation of the reaction-diffusion system (for α = 0.01, β = 0.02, and
γ = 0.6) and the comparison of results obtained by three different methods (CA-dotted, FD-solid, FE-dashed)
through the middle line of the pattern on the left.

200
0.31
180 0.30
160 0.29
140 0.28
120 0.27
100 0.26
0.25
80
0.24
60
0.23
40
0.22
20 0.21

20 40 60 80 100 120 140 160 180 200

FIGURE 18.5 Pattern formations from 2D random initial conditions with α = 0.05, β = 0.01, and γ = 0.5 (left)
and 3D structures with α = 0.1, β = 0.05, and γ = 0.36 (right).

Another example is the formation of spiral waves studied by Barkley and his colleagues in detail (Barkley
et al., 1990; Margerit and Barkley, 2001)


∂u u v +β ∂v
= ∇ 2 u + 2 (1 − u) u − , = ε∇ 2 v + (u − v),
∂t ε α ∂t

where ε, α, β are parameters. Figure 18.6 shows the formation of spiral waves (2D) and scroll waves (3D)
of this nonlinear system under appropriate conditions.
The patterns formed in terms of diffusion-reaction equations have been observed in many phenomena.
The ring, spots, and stripes exist in animal skin coating, enzymatic reactions, shell structures, and mineral
formation. The spiral and scroll waves and spatiotemporal pattern formations are observed in calcium
transport, Belousov–Zhabotinsky reaction, cardiac tissue, and other excitable systems.
In this chapter, we have discussed some of important development and research results concerning the
connection among the cellular automata and partial differential equations as well as the pattern formation
related to both systems. Cellular automata are rule-based methods with the advantages of local inter-
actions, homogeneity, discrete states and parallelism, and thus they are suitable for simulating systems
with large degrees of freedom and life-related phenomena such as artificial intelligence and ecosystems.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 282 — #10

Cellular Automata, PDEs, and Pattern Formation 18-283

FIGURE 18.6 Formation of spiral wave for α = 1, β = 0.1, and ε = 0.2 (left) and 3D scroll wave for α = 1, β = 0.1,
and ε = 0.15 (right).

PDEs are continuum-based models with the advantages of mathematical methods and closed-form
analytical solutions developed over many years, however, they usually deal with systems with small num-
bers of degree of freedom. There is an interesting connection between cellular automata and PDEs although
this is not always straightforward and sometimes may be very difficult. The derivation of updating rules for
cellular automata from corresponding PDEs are relatively straightforward by using the finite differencing
schemes, while the formulation of differential equations from the cellular automaton is usually difficult
and nonunique. More studies are highly needed in these areas. In addition, either rule-based systems
or equation-based systems can have complex pattern formation under appropriate conditions, and these
spatiotemporal patterns can simulate many phenomena in engineering and biological applications.

References
Ahmed, E. and Elgazzar, A.S. On some applications of cellular automata. Physica A, 296 (2001) 529–538.
Bagnoli, F., Boccara, N., and Rechtman, R. Nature of phase transitions in a probabilistic cellular automaton
with two absorbing states. Physical Review E, 63 (2001) 461161–461169.
Barkley, D., Kness, M., and Tuckerman, L.S. Spiral-wave dynamics in a simple model of excitable media:
The transition from simple to compound rotation. Physical Review A, 42 (1990) 2489–2492.
Boffetta, G., Cencini, M., Falcioni, M., and Vulpiani, A. Predictability: A way to characterize complexity.
Physics Reports, 356 (2002) 367–474.
Cappuccio, R., Cattaneo, G., Erbacci, G., and Jocher, U. A parallel implementation of a cellular automata
based on the model for coffee percolation. Parallel Computing, 27 (2001) 685–717.
Emmerich, H. and Kahng, B.N. A random automata related to the noisy Burgers equation. Physica A, 259
(1998) 81–89.
Flake, G.W. The Computational Beauty of Nature. MIT Press, Cambridge, MA (1998).
Guinot, V. Modelling using stochastic, finite state cellular automata: Rule inference from continuum
models. Applied Mathematical Modelling, 26 (2002) 701–714.
Keener, J. and Sneyd, J. Mathematical Physiology. Springer-Verlag, New York (1998).
Margerit, D. and Barkley, D. Selection of twisted scroll waves in three-dimensional excitable media.
Physical Review Letters, 86 (2001) 175–178.
Margolus, N. Physics-like models of computation. Physica D, 10 (1984) 81–95.
Meinhardt, H. Models of Biological Pattern Formation. Academic Press, London (1982).
Meinhardt, H. The Algorithmic Beauty of Sea Shells. Springer-Verlag, New York (1995).
Murray, J.D. Mathematical Biology. Springer-Verlag, New York (1989).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 283 — #11

18-284 Handbook of Bioinspired Algorithms and Applications

Omohundro, S. Modelling cellular automata with partial differential equations. Physica D, 10 (1984)
128–134.
Ostrov, D. and Rucker, R. Continuous-valued cellular automata for nonlinear wave equations. Complex
Systems, 10 (1996) 91–117.
Rucker, R. Continuous-valued cellular automata in two dimensions. In Griffeath, D. and Moore, C. (Eds.),
New Constructions in Cellular Automata. Oxford University Press, (2003). Website for CAPOW98
Software: http://www.cs.sjsu.edu/faculty/rucker/capow/
Toffoli, T. Cellular automata as an alternative to (rather than an approximate of) differential equations in
modelling physics. Physica D, 10 (1984) 117–127.
Turing, A. The chemical basis of morphogenesis. Philosophical Transactions of the Royal Society of London B,
237 (1952) 37–72.
Vichniac, G.Y. Simulating physics with cellular automata. Physica D, 10 (1984) 96–116.
von Newmann J. Theory of Self-Reproducing Automata (Edited by A.W. Burks). University of Illinois
Press, Urbana, IL (1966).
Weimar, J.R. Cellular automata for reaction-diffusion systems. Parallel Computing, 23 (1997) 1699–1715.
Wolfram, S. Statistical mechanics of cellular automata. Review of Modern Physics, 55 (1983) 601.
Wolfram, S. Cellular automata as models of complexity. Nature, 311 (1984a) 419–424.
Wolfram, S. Universality and complexity in cellular automata. Physica D, 10 (1984b) 1–35.
Wolfram, S. Cellular Automata and Complexity. Addison-Wesley, Reading, MA (1994).
Yang, X.S. Characterization of multispecies living ecosystems with cellular automata. In Standish, Abbass,
and Bedau (Eds.), Artificial Life VIII. MIT Press, Cambridge, MA (2002), pp. 138–141.
Yang, X.S. Turing pattern formation of catalytic reaction-diffusion systems in engineering applications.
Modelling and Simulation in Materials Science and Engineering, 11 (2003) 321–329.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C018” — 2005/8/6 — 14:09 — page 284 — #12

 19
Ant Colonies and the
Mesh-Partitioning
Problem

19.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-285

19.2 Ant-Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-287
19.3 Mesh Partitioning with the Ant-Colonies Algorithm . 19-289
Borut Robič Multiple Ant-Colony Algorithms • Performance Evaluation
Peter Korošec 19.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-301
Jurij Šilc References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-301

The complex social behaviors of ants have been much studied by science, and computer scientists
are now finding that these behavior patterns can provide models for solving difficult combinatorial
optimization problems.

Dorigo and Stützle,

Ant Colony Optimization, MIT Press, 2004.

19.1 Introduction
An ant colony is a “distributed system” that has a highly structured social organization in spite of the
simplicity of its individuals. Because of this organization, an ant colony is capable of accomplishing
complex natural tasks that far exceed the individual capacity of a single ant. It is not surprising, therefore,
that computer scientists have taken inspiration from ants and their behavior in order to design algorithms
for solving computationally demanding problems. This chapter is concerned with applying these ideas to
solve an important, but NP-hard, combinatorial optimization problem: the so-called mesh-partitioning
problem.
Many of the problems that arise in mechanical, civil, automobile, and aerospace engineering can be
expressed in terms of partial differential equations and solved by using the finite-element method. If a
partial differential equation involves a function, f , then the purpose of the finite-element method is to
determine an approximation to f . To do this the domain is discretized into a set of geometrical elements
consisting of nodes: a process known as meshing. The value of f is then computed for each of these nodes,
and the solutions for the other points are interpolated from these values [1].

19-285

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 285 — #1

19-286 Handbook of Bioinspired Algorithms and Applications

(a) (b)

(c) (d)

FIGURE 19.1 Mesh partitioning: (a) sample mesh; (b) mesh with induced graph; (c) after graph partitioning; and
(d) the resulting partitioned mesh.

However, in real-world engineering problems, meshing is a demanding task because meshes usually
have large numbers (e.g., hundreds of thousands) of elements. For this reason, the finite-element method
is usually parallelized, which means the mesh is partitioned and distributed among several processors:
a process known as mesh partitioning. To achieve high computational efficiency, it is important that the
mesh partitioning produces workloads that are well balanced and that the interprocessor communication
is minimized. This is a combinatorial optimization problem and is a special case of the well-known
graph-partitioning problem (see Figure 19.1).
The graph-partitioning problem is defined as follows. Let G(V , E) be an undirected graph consisting
of a nonempty set V of vertices and a set E ⊆ V × V of edges. A k-partition D of G comprises k mutually
disjointed subsets D 1 , D 2 , . . . , D k (called domains) of V , whose union is V . The set of edges that connect
the different domains of a partition D is called an edge-cut, and is denoted by ξ(D). A partition D is
considered to be balanced if the sizes of the domains are roughly the same, that is, if

β(D) = max |D i | − min |D i | ≈ 0.

1≤i≤k 1≤i≤k

The graph-partitioning problem is to find a balanced partition with a minimum edge-cut.

Unfortunately, the graph-partitioning problem is an NP-hard optimization problem. This means that
it is impossible to find optimum solutions in polynomial time (unless P = NP). Consequently, heuristic
approaches are normally used for mesh partitioning. Many mesh-partitioning algorithms are described
in the literature [2]. Some of these just use geometric information about the mesh and employ a fast,
parallel, sorting algorithm in their implementation [3]. Spectral partitioning methods also make use of
the mesh-connectivity information and solve the eigenvalue problem to compute the partition [4].
A promising alternative is to use stochastic heuristics, some of which are based on various fundamental
principles observed in nature. Recently, a number of studies have shown that such techniques have
great potential for solving a wide range of problems [5]. Examples include tabu search [6], simulated

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 286 — #2

Ant Colonies and the Mesh-Partitioning Problem 19-287

TABLE 19.1 Applications of Ant-Colony Optimization Algorithms

Problem name Authors Algorithm name Year

Traveling salesman Dorigo, Maniezzo, and Colorni AS 1991

Gambardella and Dorigo Ant-Q 1995
Dorigo and Gambardella ACS and ACS-3-opt 1996
Stützle and Hoos MMAS 1997
Bullnheimer, Hartl, and Strauss ASrank 1997
Quadratic assignment Maniezzo, Colorni, and Dorigo AS-QAP 1994
Gambardela, Taillard, and Dorigo HAS-QAP 1997
Stützle and Hoos MMAS-QAP 1997
Maniezzo ANTS-QAP 1998
Maniezzo and Colorni AS-QAP 1999
Scheduling problems Colorni, Dorigo, and Maniezzo AS-JSP 1994
Stützle AS-FSP 1997
Bauer et al. ACS-SMTTP 1999
den Besten, Stützle, and Dorigo ACS-SMTWTP 1999
Vehicle routing Bullnheimer, Hartl, and Strauss AS-VRP 1997
Gambardella, Taillard, and Agazzi HAS-VRP 1999
Connection-oriented network routing Schoonderwoerd et al. ABC 1996
White, Pagurek, and Oppacher ASGA 1998
Di Carlo and Dorigo AntNet-FS 1998
Bonabeau et al. ABC-smart ants 1998
Connection-less network routing Di Carlo and Dorigo AntNet and AntNet-FA 1997
Subramanian, Druschel, and Chen Regular ants 1997
Hausse et al. CAF 1998
van der Put and Rothkrantz ABC-backward 1998
Sequential ordering Gambardella and Dorigo HAS-SOP 1997
Graph coloring Costa and Hertz ANTCOL 1997
Shortest common supersequence Michel and Middendorf AS-SCS 1998
Frequency assignment Maniezzo and Carbonaro ANTS-FAP 1998
Generalized assignment Ramalhinho, Lorenço, and Serra MMAS-GAP 1998
Multiple knapsack Leguizamón and Michalewicz AS-MKP 1999
Optical network routing Navarro Varela and Sinclair ACO-VWP 1999
Redundancy allocation Liang and Smith ACO-RAP 1999

Source: Adapted from Dorigo, M., Bonabeau, E., and Therauluz, G. Future Generation Computer Systems, 16,
858, 2000. With permission from Elsevier.

annealing [7], neural networks [8], and genetic algorithms [9]. These methods are widely applicable and
have proven to be very powerful in practice [10].
Ant colonies are yet another such natural phenomenon that has recently given rise to new optimization
methods belonging to the group of “Heuristics from Nature,” which are implemented in some new
bio-inspired algorithms [11,12]. This ant-colony optimization has already been used in solving various
combinatorial optimization problems, such as the traveling salesman [13], quadratic assignment [14],
job-shop scheduling [15], vehicle [16] and network routing [17]. More references to applications of
ant-colony optimization can be found in References 18 and 19 (see Table 19.1). For the case of mesh
partitioning, however, only a few attempts have been made [20–23].

19.2 Ant-Colony Optimization

Ant-Colony Optimization (ACO) is a metaheuristics approach proposed by Dorigo et al. [12]. The
inspiration for ACO is the foraging behavior of real ants. This behavior allows the ants to find the shortest

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 287 — #3

19-288 Handbook of Bioinspired Algorithms and Applications

FIGURE 19.2 Ant-colony optimization algorithm.

paths between food sources and their nest [24]. Ants deploy a chemical trail (or pheromone trail) as they
walk; this trail attracts other ants to take the path that has the most pheromone. This reinforcement process
results in the selection of the shortest path: the first ants coming back to the nest are those that took the
shortest path twice (from the nest to the source and back to the nest), so that more pheromone is present
on the shortest path than on the longer paths immediately after these ants have returned, stimulating nest
mates to choose the shortest path.
Ant-colony optimization algorithms (see Figure 19.2) are based on a parameterized probabilistic model
(pheromone model) that is used to model the chemical pheromone trails. Artificial ants incrementally
construct solutions by adding opportunely defined solution components to a partial solution under
consideration. In order to do this, artificial ants perform randomized walks on a completely connected
graph G(C, L), called a construction graph, whose vertices are the solution components C, and the set
L composed of the connections. When a constrained combinatorial optimization problem is considered,
the problem constraints  are built into the ants’ constructive procedure in such a way that in every step of
the construction process only feasible solution components can be added to the current partial solution.

Ant_activity(): In the construction phase an ant incrementally builds a solution by adding solution
components to the partial solution constructed so far. The probabilistic choice of the next solution
component to be added is done by means of transition probabilities. More specifically, ant n in step t
moves from vertex i ∈ C to vertex j ∈ C with a probability given by:



 τija (t )ηijb
 j ∈ Ni,n ,
pij,n (t ) = m∈Ni,n τim (t )ηim
a b


0 j∈
/ Ni,n ,

where ηij is a priori available heuristic information, a and b are two parameters that determine the
relative influence of the pheromone trail τij (t ) and heuristic information, respectively, and Ni,n is the
feasible neighborhood of vertex i. If a = 0, then only heuristic information is considered. Similarly, if
b = 0, then only pheromone information is at work. Once an ant builds a solution, or while a solution
is being built, the pheromone is being deposited (on nodes or connections) according to the evaluation
of a (partial) solution. This pheromone information will direct the search of the ants in the following
iterations. The solution construction ends when an ant comes to the ending vertex (where the food is
located).
Pheromone_evaporation(): Pheromone-trail evaporation is a procedure that simulates the
reduction of pheromone intensity. It is needed in order to avoid a too quick convergence of the algorithm
to a suboptimal solution.
Daemon_actions(): Daemon actions can be used to implement centralized actions that cannot be
performed by single ants. Examples are the use of a local search procedure applied to the solutions built

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 288 — #4

Ant Colonies and the Mesh-Partitioning Problem 19-289

by the ants, or the collection of global information that can be used to decide whether it is useful or not
to deposit additional pheromone to bias the search process from a nonlocal perspective.
As we can see from the pseudo code, the ScheduleActivities construct does not specify how the three
included activities should be scheduled or synchronized. This means it is up to the programmer to specify
how these procedures will interact (parallel or independent).
Within the ACO metaheuristic framework the currently best-performing versions in practice are
Ant Colony System [13] and MAX–MIN Ant System [25]. Recently, researchers have been dealing
with finding similarities between ACO algorithms and Estimation of Distribution Algorithms [26,27].
Furthermore, connections between ACO algorithms and Stochastic Gradient–Descent algorithms are
shown in Reference 28.

19.3 Mesh Partitioning with the Ant-Colonies Algorithm

The mesh-partitioning problem, like some other problems, for example, data-mining, text-mining, belong
to the wider class of so-called clustering problems, which are concerned with the grouping of objects
into homogeneous subgroups. For these problems, ant-based algorithms have also been proposed [29].
However, ant-based clustering differs from ACO in several fundamental respects:
• It draws its inspiration from the clustering behavior observed in real ants (not the foraging behavior
as in the ACO).
• It is not metaheuristics, in contrast to ACO; it tackles only the specific task of clustering.
• Unlike ACO it does not make use of artificial pheromones.
• It shows no synergetic effect, that is, its performance is mostly independent of population size.
In contrast to the ant-based clustering approach to the partitioning problem, we discuss three
approaches that are based on ACO metaheuristics. Informally, each of these approaches uses multiple
ant colonies instead of only one (as is the case in ACO).

19.3.1 Multiple Ant-Colony Algorithms

19.3.1.1 The Basic Algorithm
The main idea of the Basic Multiple Ant-Colony Algorithm (b-MACA) for k-way partitioning was recently
proposed by Langham and Grant [22]. There are k colonies of ants that are competing for food, which
in this case represents the vertices of the graph (mesh). Eventually, ants gather food to their nests, that
is, they partition the mesh into k submeshes. More precisely, the algorithm b-MACA proceeds as follows
(see Figure 19.3).
Initially, the graph is mapped onto the grid that represents the ants’ habitat (where the ants can move).
There are many possibilities for this mapping, one of which is random mapping, as used in our case. Then
the ants are evenly distributed to k nest loci on the grid, which are chosen according to a certain strategy,
from where they start their foraging and gathering of food. An ant can move in one of three directions
(forward, left, and right). The direction is chosen by a probability rule based on pheromone intensity.
When an ant attempts to move off the grid it is forced to move left or right with equal probability. When
an ant finds food it tries to pick it up. To do this, it checks whether the quantity of the temporarily gathered
food in its nest is at the limit (the capacity of storage is limited by the problem’s constraints). If the limit
has not been reached, then the weight of the food is calculated from the number of cut edges created by
assigning the selected vertex to the partition associated with the nest of the current ant; otherwise the ant
moves in a randomly selected direction. If the food is too heavy for one ant to pick it up (and not too
heavy for a few ants to pick it up) then the ant sends a help signal within a radius of a few cells. So, if other
ants are in the neighborhood, they will help this ant to carry the food to the nest locus. On the way back
to the nest locus an ant deposits pheromones on the trail that it is making, so the other ants can follow
its trail and gather more food from that, or a nearby cell. When an ant reaches the nest locus it drops the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 289 — #5

19-290 Handbook of Bioinspired Algorithms and Applications

FIGURE 19.3 Algorithm b-MACA.

food in the first possible place around the nest (e.g., in a clockwise direction). After an ant has dropped its
food, it starts a new round of foraging. Of course, ants can also gather food from other nest loci. When an
ant tries to pick up food from other nest loci it performs the same procedure as when foraging for food,
with the exception that when the food is too heavy to be picked up, the ant moves on instead of sending a
help signal. In this way the temporary solution is significantly improved.
As was mentioned above, there are some constraints that are imposed on the b-MACA algorithm.
The first is the colony’s storage-capacity constraint, which is implemented so that no single colony can
gather all the food into its nest, and to maintain the appropriate balance between domains. The second
constraint ensures that when the pheromone intensity of a certain cell drops below a fixed value, that
cell’s pheromone intensity is restored to the initial value. In this way we maintain a high exploration level.
Other constraints are as follows: there can only be a limited number of vertices put in a single cell; each
ant can carry only a limited number of pieces of food; the food that is being brought back to the nest is a
kind of a tabu, that is, it is not available to other ants. A short tabu list consisting of the last m pieces of
food that were moved helps the algorithm to escape from local minima.

19.3.1.2 The Multilevel Algorithm

One already-established way to speed up and globally improve the partitioning method is the use of
multilevel techniques. Here, the basic idea is to group vertices together to form clusters that define a
new graph. The next step is to recursively iterate this procedure until the graph size falls below a certain
threshold. This is followed by a successive refinement of these coarser graphs. This procedure is known as
the multilevel paradigm. The multilevel idea (see Figure 19.4) was first proposed by Barnard and Simon
[30] as a method of speeding-up spectral bisection, and improved by Hendrickson and Leland [31] who
generalized it to encompass local refinement algorithms.
The implementation consists of two parts: graph contraction and partition expansion. In the first part,
a coarser graph G
+1 (V
+1 , E
+1 ) is created from G
(V
, E
) by finding the largest independent subset
of graph edges and then collapsing them. Each selected edge is collapsed and the vertices u1 , u2 ∈ V
that
are at either end of it are merged into the new vertex v ∈ V
+1 with weight

|v| = |u1 | + |u2 |.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 290 — #6

Ant Colonies and the Mesh-Partitioning Problem 19-291

phase
Coars
e

ment
ning p

Refine
hase

Initial partitioning phase

s
FIGURE 19.4 The three phases of multilevel k-way mesh partitioning.

The edges that have not been collapsed are inherited by the new graph G
+1 and the edges that have
become duplicated are merged and their weight is summed. Because of inheritance, the total weight
of the graph remains the same and the total edge weight is reduced by an amount equal to the weight
of the collapsed edges, which have no impact on the graph balance or the edge-cut. In the second part,
the already-optimized partition (with the algorithm b-MACA) of graph G
is expanded. The optimized
partition must be interpolated onto its parent graph G
−1 . Because of the simplicity of the coarsening
in the first part, the interpolation itself is trivial. So, if vertex v ∈ V
belongs to domain D i , then
after refinement the matched pair u1 , u2 ∈ V
−1 that represents v, will also be in D i . In this way
the graph is expanded to its original size, and on every level
of our expansion we run our basic
ant-colony algorithm. This is referred to as the Multilevel Multiple Ant-Colony Algorithm (m-MACA)
approach.
Due to the large graphs and the increased number of levels, the number of vertices in a single cell
increases rapidly. To overcome this problem we introduced a method called bucket sort that accelerates
and improves the algorithm’s convergence by choosing the most “promising” vertex from the cell. The
bucket sort, which was first introduced by Fiduccia and Mattheyses [32], has become an essential tool
for the efficient and rapid sorting and adjustment of vertices in terms of their gain. The basic idea
is that all the vertices with a particular gain g are put together in a “bucket” ranked g . In this way
the problem of finding a vertex with maximum gain is converted into finding the nonempty bucket
with the highest rank, and then picking a vertex from it. If a chosen vertex migrates from one domain
to another, only its gain and the gains of all its neighbors have to be recalculated and put back into
appropriate buckets. In our implementation each bucket is represented by a double-linked list of vertices.
Because of the multilevel process, it often happens that the potential gain values are dispersed over a wide
range. For this reason we have introduced the 2–3 tree. With this we avoided large and sparse arrays of
pointers. We store the nonempty buckets in the 2–3 tree, so each leaf in the tree represents a bucket.
For even faster searching we have made one 2–3 tree for each colony on every cell that has vertices on
it (see Figure 19.5). With this we have increased the speed of the search, as well as the add and delete
operations.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 291 — #7

19-292 Handbook of Bioinspired Algorithms and Applications

–1 2

Grid with food (vertices)

–3 0 4 6
Bucket
–7 –3 –1 0 2 4 6 ranked 6

Double linked
list of vertices

FIGURE 19.5 Bucket sorting.

19.3.1.3 The Hybrid Algorithm

What we have done is to merge vector quantization (VQ) and the basic ant-colony algorithm into a single
algorithm called the hybrid ant-colony algorithm (h-MACA) [23]. With the VQ we compute a partition,
which is then used as a starting partition for the b-MACA. With the b-MACA we refine this partition so
that the best possible result is obtained.
The VQ method [33] is a stochastic approximation method that uses the basic structure of the input
vectors to solve a specific problem (e.g., data compression). In other words, the input space is divided
into a finite number of distinct regions (domains) and for each region there is a representative vector.
When a mapping function (device) receives a new input vector it maps it into a region with which this
vector is represented best. This is a simple example of some sort of compression. Of course this is only one
possibility of the use of this method. In our case we used it as a mapping device for our mesh-partitioning
problem. The mesh vertices are usually locally connected to their neighbors. Now we can treat the position
of each mesh vertex as an input vector and each domain in our partition as a region in input space. We try
to divide our “mesh” space into domains, so that the size (the number of vertices) of each domain is
approximately the same, with as few as possible connections between the domains.
A vector quantizer maps
-dimensional vectors in the vector space IR
into a finite set of vectors

Y = {y i : i = 1, 2, . . . , k}.

Each vector y i is called a codeword and the set of all the codewords is called a codebook. Associated with
each codeword y i is a nearest-neighbor region called the Voronoi region, and it is defined by:

υi = {x ∈ IR
: x − y i ≤ x − y j , ∀i  = j}.

The set of Voronoi regions partition the entire space IR

such that
   

k 
k
 υ i = IR
 ∧  υ i = ∅ .
i=1 i=1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 292 — #8

Ant Colonies and the Mesh-Partitioning Problem 19-293

The VQ consists of the following six steps:

Step 1: Determine the number of domains k.
Step 2: Read the input graph and its coordinates. The input vector consists of three elements. The first
two are the vertex coordinates and the third is the vertex density. By density we mean how close
together its connected neighbors are. The closer they are, the higher is the value for the density.
Step 3: Select k codewords. The initial codewords can be selected randomly from the input vectors or
as random points in the input space.
Step 4: Calculate the Euclidian distance between the input vector and the codewords. The input vector is
assigned to the domain of the codeword that returns the minimum value according to the function:

f (x, y i ) = ε(x, y i ) − ξi − βi ,

where ε(x, y i ) represents a function that calculates the Euclidian distance between x and y i , ξi
represents the change in the edge-cut if x belonged to the i-th domain, and βi represents the
difference between the number of vertices in the largest and the i-th domains.
Step 5: Compute the new set of codewords. We add up all the x i vectors in the i-th domain and divide
the summation by the number of input vectors in the domain:

1 
m
yi = xij ,
m
j=1

where m represents the number of input vectors in the i-th domain.

Step 6: Repeat steps 4 and 5 until the values of the codewords converge, usually to a suboptimal solution.
An example of the VQ is shown in Figure 19.6. Here we used a two-dimensional graph (
= 2),
but it can easily be expanded to any other number of dimensions. We can see 45 input vectors that are

X V13 V12 V11

X X
X y13
X

X y12
X
Input vectors

X X
V8 y8 X
y10 y11
X X X X
X X X
y9 V10
X
Codewords

y6 V9 X
X
X
y7 V6 X y4
X X V4
X X X X
y5
X X
V7 X X X
X X V5 X X
y2 y3
Voronoi region

V1 X X
X
X X
X
y1 V2 V3
X X

FIGURE 19.6 Vector quantization.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 293 — #9

19-294 Handbook of Bioinspired Algorithms and Applications

divided into k = 13 domains (Voronoi regions υ 1 , υ 2 , . . . , υ 13 ) and are represented with the codewords
y 1 , y 2 , . . . , y 13 .

19.3.2 Performance Evaluation

19.3.2.1 Experimental Environment
The b-MACA, m-MACA, and h-MACA were implemented in Borland®Delphi™. The experiments were
made on a computer with an AMD Athlon™XP 1800+ processor running the Microsoft®Windows®XP
operating system. The implementation also includes a visualization tool to assist the user in selecting the
appropriate parameters of the algorithm (see Figure 19.7 and Figure 19.8).
Figure 19.7 shows what the output of the MACA-type algorithm looks like. Here we can see how the
algorithm partitioned the graph into four domains and which vertices belong to which partition.
Figure 19.8 consists of two pictures. Figure 19.8(a) represents what the grid looks like after four iterations
of the algorithm. As we can see, ants are foraging for food (one food piece represents a graph vertex, which
in the case of m-MACA can be composed of many vertices of the initial graph), which is scattered all over
the grid. In Figure 19.8(b) we can see that food is laid around the nests and ants are “stealing” food from
other nests.
On the right-hand side of the grid you can see “instruments” with which the algorithm is controlled.
With it, we are able to set the grid size (Grid size), the number of ants per colony (Number of ants),
the desired balance (±diversion), the maximum number of iterations without improvement per level
(Number of iterations), the weight of heuristic information (Weight on prob1), the maximum number
of gathered food pieces (Fetch number), and the maximum length of the tabu list (Tabu number). All the
other data represent the current state of the algorithm.

19.3.2.2 Benchmark Suite

For the purpose of testing, we have randomly generated four graphs. Meanwhile, all the other benchmark
graphs were taken from the Graph Partitioning Archive at the University of Greenwich and from the
Graph Collection Web page at the University of Paderborn. All the graphs are described in Table 19.2.

FIGURE 19.7 Mesh-partitioning visualization.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 294 — #10

Ant Colonies and the Mesh-Partitioning Problem 19-295

(a)

(b)

FIGURE 19.8 User interface: (a) initial stage and (b) final stage.

19.3.2.3 Basic Algorithm

Several experiments were made to evaluate the algorithm b-MACA [20]. We compared b-MACA with
the well-known partitioning programs k-Metis 4.0 [34], Randomized Tabu Search (RTS) [35] and Hybrid
Genetic Algorithm (HGA) [35]. Each test graph was partitioned into two and four domains (k = 2 and
k = 4). The results, which are given in terms of edge-cut ξ(D) and balance β(D), are shown in Table 19.3.
Table 19.3 shows that b-MACA performs well on small graphs (|V | < 500). We did additional exper-
iments [20] on larger graphs (|V | > 1000) and found that b-MACA performs much worse than RTS or
HGA. To remedy this we enhanced the algorithm b-MACA with a multilevel technique.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 295 — #11

19-296 Handbook of Bioinspired Algorithms and Applications

TABLE 19.2 Benchmark Graphs

Number of Number of
Graph vertices |V | edges |E|

graph1a 50 86
graph2a 50 143
grid1c 252 476
grid1_dualc 224 420
graph3a 400 546
graph4a 400 1,006
netz4504_dualc 615 1,171
U1000.5c 1,000 2,394
U1000.10c 1,000 4,696
U1000.20c 1,000 9,339
ukerbe1_dualc 1,866 3,538
netz4504c 1,961 2,578
add20b 2,395 7,462
datab 2,851 15,093
grid2_dualc 3,136 6,112
grid2c 3,296 6,432
airfoilc 4,253 12,289
3eltb,c 4,720 13,722
ukb 4,824 6,837
add32b 4,960 9,462
ukerbe1c 5,981 7,852
airfoil_dualc 8,034 11,813
bcsstk33b 8,738 291,583
3elt_dualc 9,000 13,287
whitaker3b,c 9,800 28,989
crackb,c 10,240 30,380
wing_nodalb 10,937 75,488
fe_4elt2b 11,143 32,818
4eltb 15,606 45,878
bigc 15,606 45,878
fe_sphereb 16,386 49,152
ctib 16,840 48,232
whitaker3_dualc 19,190 8,581
crack_dualc 20,141 30,043
cs4b 22,499 43,858
big_dualc 30,269 44,929

a Randomly generated.
b Graph partitioning archive: www.gre.ac.uk/
∼c.walshaw/partition/.
c Graph collection: www.uni-paderborn.de/cs/ag-monien/
RESEARCH/PART/graphs.html.

19.3.2.4 Multilevel Algorithm

We again partitioned each of the test graphs into two and four domains (k = 2 and k = 4). The results
are again pairs, ξ(D) and β(D). We allowed a 0.2% unbalance, that is, β(D) ≤ 0.002|V |. The algorithm
m-MACA was run 30 times on each graph; and from the solutions obtained we chose the one with
minimum ξ(D) as the final result for that graph.
The results of our experiments are shown in Table 19.4, where m-MACA is compared with the k-Metis
4.0 [34], the Chaco 2.0 [31] with multilevel Kernighan–Lin global partitioning method, and the new
mixed simulated annealing and tabu search algorithm MLSATS [36].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 296 — #12

Ant Colonies and the Mesh-Partitioning Problem 19-297

TABLE 19.3 Experimental Results (b-MACA, k = 2, 4)

Algorithm k-Metis 4.0 RTS HGA b-MACA

Graph k ξ(D) β(D) ξ(D) β(D) ξ(D) β(D) ξ(D) β(D)

graph1 2 15 0 14 0 14 0 13 2
4 32 12 25 1 25 1 25 1
graph2 2 35 0 33 0 33 0 33 0
4 77 9 57 1 57 1 57 1
graph3 2 45 8 41 0 40 0 41 0
4 85 4 82 0 76 0 79 0
graph4 2 208 10 181 0 187 0 182 0
4 331 2 309 0 310 0 316 2

TABLE 19.4 Experimental Results (m-MACA, k = 2, 4)

Algorithm k-Metis 4.0 Chaco 2.0 MLSATS m-MACA

Graph k ξ(D) β(D) ξ(D) β(D) ξ(D) β(D) ξ(D) β(D)

add20 2 773 23 630 1 696 119 601 5

4 1,214 45 1,242 1 1,193 57 1,196 3
data 2 253 67 210 1 196 131 199 1
4 486 42 444 1 395 67 424 1
3elt 2 148 72 124 0 87 108 90 0
4 250 63 258 0 199 45 225 2
uk 2 33 126 23 0 54 186 20 2
4 50 47 60 0 261 74 50 2
add32 2 12 62 11 0 10 2 10 2
4 38 59 53 0 35 66 40 5
bcsstk33 2 13,393 139 10,199 0 10,064 100 10,222 12
4 22,909 129 25,529 1 22,442 219 22,632 14
whitaker3 2 135 96 135 0 130 0 126 16
4 407 97 439 0 448 198 386 6
crack 2 221 248 209 0 184 62 184 0
4 478 106 457 0 401 216 381 8
wing_nodal 2 1,892 315 1,747 1 1,670 543 1,711 23
4 3,898 163 3,817 1 3,596 273 3,619 11
fe_4elt2 2 132 297 130 1 130 1 130 1
4 400 87 378 1 350 113 350 2
4elt 2 149 104 242 0 130 1 130 1
4 385 197 416 1 367 369 389 13
fe_sphere 2 444 418 422 0 384 576 402 4
4 828 46 835 1 786 329 810 15
cti 2 401 300 369 0 318 120 332 24
4 1,104 156 1,000 0 966 360 1,045 14
cs4 2 397 653 418 1 418 907 397 1
4 1,102 312 1,135 1 1,103 403 1,038 21

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 297 — #13

19-298 Handbook of Bioinspired Algorithms and Applications

TABLE 19.5 Results from Graph Partitioning Archive

ξ(D)
Graph k Algorithm(s) with 1% unbalance

add20 2 m-MACA 601

4 m-MACA 1,186
data 2 GrPart 196
4 m-MACA 410
3elt 2 GrPart, m-MACA 89
4 JE, MLSATS 199
uk 2 m-MACA 19
4 JE 44
add32 2 J2.2, MLSATS, m-MACA 10
4 JE 33
bcssth33 2 GrPart 10,109
4 iJ 21,685
whitaker3 2 JE, m-MACA 126
4 JE 380
crack 2 m-MACA 183
4 JE 368
wing_nodal 2 m-MACA 1,697
4 JE 3,572
fe_4elt2 2 MRSB, MLSATS, m-MACA 130
4 JE 349
4elt 2 GrPart 138
4 JE 327
fe_sphere 2 JE 386
4 N/A 768
cti 2 JE, MLSATS, m-MACA 318
4 m-MACA 963
cs4 2 JE 367
4 JE 940

GrPart: A. Kozhushkin’s implementation of iterative multilevel Kernighan–Lin; iJ: Iterated

JOSTLE — iterated multilevel Kernighan–Lin (k-way) [38]; J2.2: JOSTLE — multilevel
Kernighan–Lin (k-way); version 2.2 [38]; JE: JOSTLE Evolutionary — combined evolu-
tionary/multilevel scheme [37]; MLSATS: MultiLevel refinated mixed Simulated Anealing
and Tabu Search [36]; MRSB: Multilevel Recursive Spectral Bisection [30]; and m-MACA:
Multilevel Multiple Ant-Colony Algorithm [21].

Clearly, m-MACA performed very well. Notice that m-MACA is superior to the classical k-Metis and
Chaco algorithms. Notice also that MLSATS produced some results that have better ξ(D) but with much
higher β(D) than m-MACA.
The algorithm m-MACA also returned some solutions that are better than currently available (Winter
2003–2004) solutions in the Graph Partitioning Archive (Table 19.5). Furthermore, m-MACA is even
comparable with the combined evolutionary/multilevel scheme used in the JOSTLE Evolutionary
algorithm [37], which is currently the most promising mesh-partitioning algorithm.

19.3.2.5 Hybrid Algorithm

Now we present and discuss the results of the experimental evaluation of the algorithm h-MACA in com-
parison with the well-known partitioning programs p-Metis 4.0 [34], Chaco 2.0 [31] with the multilevel
Kernighan–Lin global partitioning method and the algorithm m-MACA [21].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 298 — #14

Ant Colonies and the Mesh-Partitioning Problem 19-299

TABLE 19.6 Experimental Results (h-MACA, k = 2, 4)

Algorithm Chaco 2.0 p-Metis 4.0 m-MACA h-MACA

Graph k ξ(D) β(D) ξ(D) β(D) ξ(D) β(D) ξ(D) β(D)

3elt 2 124 0 98 0 90 0 90 0
4 258 0 252 0 225 2 212 4
3elt_dual 2 70 0 70 0 44 6 45 8
4 130 0 120 0 112 4 154 7
airfoil 2 82 1 85 1 74 1 81 1
4 182 1 179 1 176 2 190 3
airfoil_dual 2 60 0 40 0 37 0 40 14
4 111 1 84 1 80 7 110 7
big 2 242 0 165 0 141 0 139 0
4 416 1 405 1 354 11 382 14
big_dual 2 92 1 92 1 78 1 77 11
4 219 1 196 1 215 18 222 25
crack 2 209 0 206 0 184 0 196 2
4 457 0 458 0 377 6 371 9
crack_dual 2 130 1 101 1 80 25 87 1
4 228 1 201 2 169 3 164 9
grid1 2 26 0 20 0 18 0 18 0
4 48 0 40 0 38 0 38 0
grid1_dual 2 16 0 16 0 16 0 16 0
4 37 0 35 0 35 0 35 0
grid2 2 38 0 37 0 34 0 34 2
4 106 0 121 0 94 2 92 2
grid2_dual 2 35 0 32 0 32 0 32 0
4 99 0 91 0 90 2 96 4
netz4504 2 25 1 26 1 22 1 24 1
4 66 1 62 1 50 1 49 3
netz4504_dual 2 21 1 21 1 19 1 19 1
4 54 1 49 1 44 2 44 2
U1000.5 2 10 0 1 0 1 0 2 0
4 20 0 6 0 7 2 12 2
U1000.10 2 115 0 56 0 39 0 39 0
4 200 0 108 2 99 2 107 2
U1000.20 2 294 0 253 0 220 4 221 2
4 554 0 515 2 546 2 497 2
ukerbe1 2 30 1 28 1 27 1 28 1
4 82 1 64 1 63 2 61 1
ukerbe1_dual 2 25 0 25 0 22 0 22 0
4 56 1 51 1 52 3 48 1
whitaker3 2 135 0 128 0 126 16 127 0
4 439 0 424 0 383 0 383 2
whitaker3_dual 2 82 0 74 0 64 18 65 0
4 251 1 210 1 200 6 195 12

We partitioned each of the graphs into two and four domains (k = 2 and k = 4). Each score is described
with the best obtained edge-cut ξ(D). It is important to mention that the balance β(D) was kept inside
0.2% of the |V |. The results of our experiment are shown in Table 19.6.
Table 19.6 shows that in most cases the best partition was obtained with the h-MACA algorithm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 299 — #15

19-300 Handbook of Bioinspired Algorithms and Applications

(a)

(b)

FIGURE 19.9 Mesh-Partitioning Visualization: (a) m-MACA; (b) h-MACA.

Figure 19.9 shows the main drawback of the m-MACA. Here we can see vertices represented as lighter
and darker dots, where lighter belongs to the one domain, and darker to the other. With the solid white
line we have emphasized the border between these two domains. The edges are hidden again. The final
solution obtained by the m-MACA includes “islands” (i.e., a set of connected vertices) that belong to
different domains. The islands are due to a bad initial partition and the inability of the m-MACA to merge
the islands into homogeneous regions (each having only one border). This drawback is eliminated by
using the VQ to obtain the initial partitions. In Figure 19.9(a) (m-MACA) one can see four such islands,
whereas there are only two in Figure 19.9(b) (h-MACA).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 300 — #16

Ant Colonies and the Mesh-Partitioning Problem 19-301

19.4 Conclusions
The graph-partitioning problem is an important component for mesh partitioning in the domain-
decomposition method. The ACO uses a metaheuristic approach for solving hard combinatorial
optimization problems. The purpose of this chapter was to give the reader a basic knowledge of ACO,
to investigate variants of the MACA for mesh partitioning, to suggest modifications to improve this
algorithm, and to evaluate them experimentally.
The b-MACA performed very well on small- or medium-sized graphs (n < 500). With larger graphs,
which are often encountered in mesh partitioning, we had to use a multilevel or hybrid method to
produce results that were competitive with the results given by other algorithms. Both multilevel and
hybrid algorithms performed very well on almost all graphs. Both of them were quite similar in producing
the best results. The only difference was in the standard deviation of the results [39], which was in favor
of the hybrid method. The m-MACA and the h-MACA are very promising algorithms that need to be
thoroughly investigated.
An obvious improvement of our algorithm would be to merge these two methods into one. To do this
one could apply VQ to produce a starting partition, then coarsen the graph to some extent (to a much
smaller extent than in the original multilevel method), and then use the multilevel method to refine the
previously (with VQ) obtained partition.
On the other hand, there are many possibilities for improving multilevel and, to some extent, hybrid
algorithms. One possibility is in the mapping of the graph onto the grid: with a proper mapping conver-
gence the results can be improved. The use of the load-balancing method between levels would also be a
very promising way to go. The next possibility is in determining which and how many vertices from the
cell will be picked and with what probability. Here, the Kernighan–Lin gain method [40] might be used.
We could also add some daemon actions, like the min-cut algorithm, to improve solutions during the
crossing from one level to another. And, finally, we could change the way the pheromone is evaporated,
deposited and restored.
There is a wide range of possibilities to be considered in the future. One of the most appealing is a
merger of the MACA with some other method through daemon actions and parallel implementation of
the MACA.

References
[1] Cook, R.D. et al. Concepts and Applications of Finite Element Analysis, 4th ed. John Wiley & Sons,
New York, 2001.
[2] Farhat, C. and Lesoinne, M. Automatic partitioning of unstructured meshes for the parallel
solution of problems in computational mechanics. International Journal for Numerical Methods in
Engineering, 36, 745, 1993.
[3] Shephard, M.S. et al. Parallel automated adaptive procedures for unstructured meshes. In
Special Course on Parallel Computing in Computational Fluid Dynamics (AGARD-R-807), AGARD,
Neuilly-sur-Seine, France, 1995, p. 6.1.
[4] Pothen, A., Simon, H.D., and Liou, K.P. Partitioning sparse matrices with eigenvectors of graphs.
SIAM Journal on Matrix Analysis and Application, 11, 430, 1990.
[5] Zomaya, A.Y. et al. Non-conventional computing paradigms in the new millennium. IEEE/AIP
Computing in Science and Engineering, 3, 82, 2001.
[6] Kadłuczka, P. and Wala, K. Tabu search and genetic algorithms for the generalized graph
partitioning problem. Control and Cybernetics, 24, 459, 1995.
[7] Tao, L. et al. Simulated annealing and tabu search algorithms for multiway graph partition,
Journal of Circuits, Systems and Computers, 2, 159, 1992.
[8] Bahreininejad, A., Topping, B.H.V., and Khan, A.I. Finite element mesh partitioning using neural
networks. Advances in Engineering Software, 27, 103, 1996.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 301 — #17

19-302 Handbook of Bioinspired Algorithms and Applications

[9] Ż ola, J. and Wyrzykowski, R. Application of genetic algorithm for mesh partitioning. In Proceedings
of the Workshop Parallel Numerics, Bratislava, Slovakia, 2000, p. 209.
[10] Blum, C. and Roli, A. Metaheuristics in combinatorial optimization: Overview and conceptual
comparison. ACM Computing Surveys, 35, 268, 2003.
[11] Colorni, A., Dorigo, M., and Maniezzo, V. Distributed optimization by ant colonies. In Proceedings
of the 1st European Conference on Artificial Life, Paris, France, 1991, p. 134.
[12] Dorigo, M., Maniezzo, V., and Colorni, A. The ant system: Optimization by a colony of cooperating
agents. IEEE Transactions on Systems Man and Cybernetics Part B, 26, 29, 1996.
[13] Dorigo, M. and Gambardella, L.M. Ant colony system: A cooperative learning approach to the
traveling salesman problem. IEEE Transactions on Evolutionary Computation, 1, 53, 1997.
[14] Gambardella, L.M., Taillard, E., and Dorigo, M. Ant colonies for the quadratic assignment problem.
Journal of Operational Research Society, 50, 167, 1999.
[15] Teich, T. et al. A new ant colony algorithm for the job shop scheduling problem. In Beyer, H.,
Canta-Paz, E., GoldBerg, D., Parmee, Spector, L., and Whitley, D., Eds., Proceedings of Genetic
Evolutionary Computation Conference, Morgan Kaufmann Publishers, San Francisco, CA, 2001,
p. 803.
[16] Montemanni, R. et al. A new algorithm for a dynamic vehicle routing problem based on ant
colony system. In Proceedings of the 34th Annual Conference of the Italian Operations Research
Society, Venice, Italy, 2003, p. 140.
[17] Sim, K.M. and Sun, W.H. Multiple ant-colony optimization for network routing. In Proceedings of
the 1st International Symposium Cyber Worlds, Tokyo, Japan, 2002, p. 277.
[18] Dorigo, M. and Stützle, T. Ant Colony Optimization, MIT Press, Cambridge, CA, 2004.
[19] Dorigo, M., Bonabeau, E., and Theraulaz, G. Ant algorithms and stigmergy. Future Generation
Computer Systems, 16, 851, 2000.
[20] Korošec, P., Šilc, J., and Robič, B. An ant-colony-optimization approach to the mesh-partitioning
problem. In Parallel Numerics ’02, Trobec, R. et al., Eds., University of Salzburg and Jožef Stefan
Institute, 2002, p. 123.
[21] Korošec, P., Šilc, J., and Robič, B. A multilevel ant-colony-optimization algorithm for mesh
partitioning. International Journal of Pure and Applied Mathematics, 5, 143, 2003.
[22] Langham, A.E. and Grant, P.W. Using competing ant colonies to solve k-way partitioning problems
with foraging and raiding strategies, Lecture Notes in Computer Science, 1674, 621, 1999.
[23] Šilc, J., Korošec, P., and Robič, B. Combining vector quantization and ant-colony algorithm for
mesh-partitioning. Lecture Notes in Computer Science, 3019, 113, 2004.
[24] Deneubourg, J.-L. et al. The self-organizing exploratory pattern of the argentine ant. Journal of
Insect Behavior, 3, 159, 1990.
[25] Stützle, T. and Hoos, H.H. MAX–MIN ant system. Future Generation Computer Systems, 16, 889,
2000.
[26] Pelikan, M., Goldberg, D.E., and Lobo, F.G. A survey of optimization by building and using
probabilistic models. Computational Optimization and Applications, 21, 5, 2002.
[27] Zlochin, M. et al. Model-based search for combinatorial optimization: A critical survey. Annuals
of Operations Research, 131, 373, 2004.
[28] Meuleau, N. and Dorigo, M. Ant colony optimization and stochastic gradient descent. Artificial
Life, 8, 103, 2002.
[29] Handl, J. and Meyer, B. Improved ant-based clustering and sorting in a document retrival interface.
Lecture Notes in Computer Science, 2439, 913, 2002.
[30] Barnard, S.T. and Simon, H.D. A fast multilevel implementation of recursive spectral bisection for
partitioning unstructured problems. Concurrency — Practice and Experience, 6, 101, 1994.
[31] Hendrickson, B. and Leland, R. A multilevel algorithm for partitioning graphs. In Proceedings of
the ACM/IEEE Conference on Supercomputing, San Diego, CA, 1995, p. 28.
[32] Fiduccia, C.M. and Mattheyses, R.M. A linear time heuristic for improving network partitions,
In Proceedings of 19th IEEE Design Automation Conference, Las Vegas, NV, 1982, p. 175.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 302 — #18

Ant Colonies and the Mesh-Partitioning Problem 19-303

[33] Linde, Y., Buzo, A., and Gray, R.M. An algorithm for vector quantizer design. IEEE Transactions on
Communications, 28, 84, 1980.
[34] Karypis, G. and Kumar, V. Multilevel k-way partioning scheme for irregular graphs. Journal of
Parallel and Distributed Computing, 48, 96, 1998.
[35] Šilc, J., Korošec, P., and Robič, B. An experimental evaluation of modified algorithms for the
graph partitioning problem. In Proceedings of 17th International Symposium on Computer and
Information Science, Orlando, FL, October 28–30, CRC Press, Boca Raton, FL, 2002, p. 120.
[36] Banos, R. et al. Multilevel heuristic algorithm for graph partitioning. Lecture Notes in Computer
Science, 2611, 143, 2003.
[37] Soper, A.J., Walshaw, C., and Cross, M. A combined evolutionary search and multilevel approach
to graph partitioning. In Proceedings of Genetic Evolutonary Computation Conference, Las Vegas,
NV, 2000, p. 674.
[38] Walshaw, C. and Cross, M. Mesh partitioning: A multilevel balancing and refinement algorithm.
SIAM Journal of Scientific Computing, 22, 63, 2000.
[39] Korošec, P., Šilc, J., and Robič, B. Solving the mesh-partitioning problem with an ant-colony
algorithm. Parallel Computing, 30, 785, 2004.
[40] Kernighan, B.W. and Lin, S. An efficient heuristic procedure for partitioning graph. Bell System
Technical Journal, 49, 291, 1970.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C019” — 2005/8/8 — 17:09 — page 303 — #19

 20
Simulating the
Strategic Adaptation
of Organizations
Using OrgSwarm

20.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-305

20.2 Strategic Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-306
20.3 Particle Swarm Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-306
Description of PSA • Particle Swarm as a Metaphor for
Organizational Adaptation
20.4 Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-308
Strategic Landscape • Simulation Model • Outline of Swarm
Anthony Brabazon Algorithm • Simulator
Arlindo Silva 20.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-313
Ernesto Costa Static Landscape • Dynamic Landscape
Tiago Ferra de Sousa 20.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-317
Michael O’Neill References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-318

20.1 Introduction
In an organizational setting, a strategy consists of a choice of what activities the organization will perform,
and choices as to how these activities will be performed [1]. These choices define the strategic config-
uration of the organization. Recent work by Levinthal [2] and Rivkin [3] has recognized that strategic
configurations consist of interlinked individual elements (decisions), and have applied general models of
interconnected systems such as Kauffman’s NK model to examine the implications of this for processes of
organizational adaptation.
Following a long-established metaphor of adaptation as search [4], strategic adaptation is considered
in this study as an attempt to uncover peaks on a high-dimensional strategic landscape. Some strategic
configurations produce high profits, others produce poor results. The search for good strategic configura-
tions is difficult due to the vast number of strategic configurations possible, uncertainty as to the nature of
topology of the strategic landscape faced by an organization, and changes in the topology of this landscape
over time. Despite these uncertainties, the search process for good strategies is not blind. Decision-makers
receive feedback on the success of their current and historic strategies, and can assess the payoffs received
by the strategies of their competitors [5]. Hence, certain areas of the strategic landscape are illuminated.

20-305

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 305 — #1

20-306 Handbook of Bioinspired Algorithms and Applications

Organizations do not exist in isolation but interact with, and receive feedback from their environment.
Their efforts at strategic adaptation are guided by social as well as individual learning. Good ideas dis-
covered by one organization disseminate over time. Particle swarm algorithms (PSAs) also emphasize
the importance of individual and social learning processes. Surprisingly, despite the parallels between
the learning processes in particle swarm algorithms and those in populations of organizations, as yet the
particle swarm metaphor has not been applied to the domain of organizational science. This chapter
describes a novel simulation model based on the particle swarm metaphor, and applies this to examine
the process of organizational adaptation. This study, therefore, extends the particle swarm metaphor into
the domain of organization science.

20.2 Strategic Adaptation

Strategic adaptation and strategic inertia are closely linked. If strategic adaptation is problematic, inertia
is a likely contributing cause. Broadly speaking, the strategic inertia of organizations stems from two
sources, imprinting forces, and as a consequence of market selection forces.
Imprinting forces [6] combine to define and solidify the strategic configuration of a newly formed
organization. These forces include the dominant initial strategy pursued by the organization, the
skills/prior experience of the management team, and the distribution of decision-making influence in
the organization at time of founding [6]. All of these influence the initial choice of organizational strategy.
As consensus concerning the strategy emerges, it is imprinted on the organization through resource alloca-
tion decisions [7]. The imprinting leads to inertia by creating sunk costs, internal political constraints, and
a rigid organizational structure. Over time this inertia intensifies due to the formation of an organiza-
tional history, which creates barriers to industry exit, and legitimacy issues if adaptation is suggested [8].
The resulting inertia serves to circumscribe the organization’s ability to adapt its strategy in the future.
Imprinting also occurs as relationships are built up with suppliers and customers. The creation of a web
of these relationships can serve to constrain the range of strategic alternatives in the future, as strategic
moves that dramatically disrupt the web are less likely to be considered.
The discussion of strategic inertia was extended by Hannan and Freeman [9] who posited that inertia
is also created as a natural consequence of the market selection process, claiming that “selection processes
tend to favor organizations whose structures are difficult to change” (p. 149). The basis of this claim
is that organizations that can produce a good or service reliably (consistently of a minimum quality
standard) are favored for trading purposes by other organizations, and therefore by market selection
processes. The routines required to produce a product or service reliably tend to lead to structural inertia,
as the construction of standardized routines leads to an increase in the complexity of the patterns of
links between organizational subunits [9,10]. It can, therefore, be posited that efficient organizations are
likely to exhibit inertia. As organizations seek better environment-structure congruence, their systems
become increasingly specialized and interlinked, making changes to their activities become costly and
difficult. Tushman and O’Reilly [11] note that structural inertia is rooted in the size, complexity, and
interdependence of the firm’s structures, systems, procedures, and processes. Theoretical support for
these assertions, that increasing organizational complexity can make adaptation difficult, is found in [3]
and [12], as the heightened degree of interconnections between activities within the organization will
increase the “ruggedness” of the strategic landscape on which they are adapting.

20.3 Particle Swarm Algorithm

This section provides an introduction to the canonical Particle Swarm Algorithm (PSA). The term PSA is
used in place of the commonly used PSO (Particle Swarm Optimization) in this chapter, as the objective
is not to develop a tool for “optimizing,” but to adapt and apply the swarm metaphor as a model of
organizational adaptation. The PSA [13,14] has been widely used for function optimization, and is based
on a metaphor of human social interaction [15].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 306 — #2

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-307

Under the swarm metaphor, a swarm of particles (entities) are assumed to move (fly) through an
n-dimensional space, typically looking for a function optimum. Each particle is assumed to have two
associated properties, a current position and a velocity. Each particle also has a memory of the best
location in the search space that it has found so far (pbest), and knows the best location found to date
by all the particles in the population (gbest). At each step of the algorithm, particles are displaced from
their current position by applying a velocity vector to them. The size and direction of this velocity is
influenced by the velocity in the previous iteration of the algorithm (simulates “momentum”), and the
current location of a particle relative to its pbest and gbest. Therefore, at each step, the size and direction
of each particle’s move is a function of its own history (experience), and the social influence of its peer
group. A number of variants of the PSA exist.

20.3.1 Description of PSA

The following paragraphs provide a description of the continuous version of the PSA. The algorithm is
initially described narratively. This is followed by a description of the particle position-update equations.

1. Initialize each particle in the population by randomly selecting values for its location and velocity
vectors.
2. Calculate the fitness value of each particle. If the current fitness value for a particle is greater than
the best fitness value found for the particle so far, then revise pbest.
3. Determine the location of the particle with the highest fitness and revise gbest if necessary.
4. For each particle, calculate its velocity according to Equation (20.1).
5. Update the location of each particle.
6. Repeat steps 2 to 5 until stopping criteria are met.

Each particle i has an associated current position in d-dimensional space x i , a current velocity vi , and a
personal best position yi . During each iteration of the algorithm, the location and velocity of each particle
is updated using Equations (20.1) to (20.4). Assuming a function f is to be maximized, that the swarm
consists of n particles, and that r1 , r2 are drawn from a uniform distribution in the range (0, 1), the velocity
update is described as follows:

vi (t + 1) = W vi (t ) + c1 r1 (yi − xi (t )) + c2 r2 (ŷ − xi (t )), (20.1)

where ŷ is the location of the global-best solution found by all the particles. A variant on the basic
algorithm is to use a local rather than a global version of gbest, and the term gbest is replaced by lbest.
In the local version, lbest is set independently for each particle, based on the best point found thus far
within a neighborhood of that particle’s current location.
In every iteration of the algorithm, each particle’s velocity is stochastically accelerated toward its previous
best position and toward gbest (or lbest). The weight-coefficients c1 and c2 control the relative impact of
pbest and gbest locations on the velocity of a particle. The parameters r1 and r2 ensure that the algorithm
is stochastic. A practical effect of the random coefficients r1 and r2 is that neither the individual nor the
social learning terms are always dominant.
Although the velocity update has a stochastic component, the search process is not random. It is guided
by the memory of past “good” solutions (corresponding to a psychological tendency for individuals to
repeat strategies that have worked for them in the past [15], and by the global best solution found by all
particles thus far. W represents a momentum coefficient that controls the impact of a particle’s prior-
period velocity on its current-period velocity. Each component (dimension) of the velocity vector vi is
restricted to a range [−vmax , vmax ] to ensure that individual particles do not leave the search space. The
implementation of a vmax parameter can also be interpreted as simulating the incremental nature of most
learning processes [15]. The value of vmax is usually chosen to be k × xmax , where 0 < k < 1. Once
the velocity update for particle i is determined, its position is updated and pbest is updated if necessary,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 307 — #3

20-308 Handbook of Bioinspired Algorithms and Applications

as described in Equations (20.2) to (20.4).

xi (t + 1) = xi (t ) + vi (t + 1), (20.2)

yi (t + 1) = yi (t ) if f (xi (t )) ≤ f (yi (t )), (20.3)

yi (t + 1) = xi (t ) if f (xi (t )) > f (yi (t )) (20.4)

After all particles have been updated, a check is made to determine whether gbest needs to be updated.

ŷ ∈ (y0 , y1 , . . . , yn )|f (ŷ) = max(f (y0 ), f (y1 ), . . . , f (yn )). (20.5)

20.3.2 Particle Swarm as a Metaphor for Organizational Adaptation

Although particle swarm algorithms have been used extensively in function optimization (PSO), the
original inspiration for PSAs arose from observations of animal and human social behavior [13]. Kennedy
has published a series of papers that emphasize the social aspects of particle swarm [15–17] and this work
was given prominence in the first major book on particle swarm [14].
The velocity update formula (Equation [20.1]) can be divided into cognitive and social components [15],
with the former relating to the adaptive history of a particle, an individual, or an organization. The
cognitive term can be considered as an interpretation of Thorndike’s Law of Effect [18], which states that a
behavior that is followed by a (positive) reinforcement becomes more likely in the future. The individual
learning component in the velocity update formula (yi (t) − xi (t)) introduces a stochastic tendency to
return to previously rewarded strategies, mimicking a psychological tendency for managers to repeat
strategies that have worked for them in the past [15]. The social learning component of the formula
(ŷi (t) − xi (t)) can be compared with social no-trial learning [19], in which the observation of a peer being
rewarded for a behavior will increase the probability of the observer engaging in the same behavior.
The mechanisms of the canonical PSA bear prima facie similarities to those of the domain of interest,
organizational adaptation. It adopts a populational perspective, and learning in the algorithm, just as in
populations of organizations, is both distributed and parallel. Organizations persist in employing already-
discovered good strategies, and are attracted to, and frequently imitate, good product ideas and business
practices discovered by other organizations. However, the canonical PSA requires modification before
it can employed as a component of a plausible simulation model of organizational adaptation. These
modifications are discussed in the next section.

20.4 Simulation Model

The two key components of the simulation model, the landscape generator (environment), and the
adaptation of the basic PSA to incorporate the activities and interactions of the agents (organizations) are
described next.

20.4.1 Strategic Landscape

In this chapter, the strategic landscape is defined using the NK model [12,20]. It is noted ab initio
that application of the NK model to define a strategic landscape is not atypical and has support from
prior literature in organizational science that has adopted this approach [2,3,21,22], and related work on
technological innovation [23,24]. The NK model considers the behavior of systems that are comprised of
a configuration (string) of N individual elements. Each of these elements are in turn interconnected to K
other of the N elements (K < N ). In a general description of such systems, each of the N elements can

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 308 — #4

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-309

assume a finite number of states. If the number of states for each element is constant (S), the space of all


possible configurations has N dimensions, and contains a total of N i=1 Si possible configurations.
In Kauffman’s operationalization of this general framework [12], the number of states for each element
is restricted to two (0 or 1). Therefore the configuration of N elements can be represented as a binary
string. The parameter K , determines the degree of fitness interconnectedness of each of the N elements
and can vary in value from 0 to N − 1. In one limiting case where K = 0, the contribution of each of
the N elements to the overall fitness value (or worth) of the configuration are independent of each other.
As K increases, this mapping becomes more complex, until at the upper limit when K = N − 1, the fitness
contribution of any of the N elements depends both on its own state, and the simultaneous states of all
the other N − 1 elements, describing a fully connected graph.
If we let si represent the state of an individual element i, the contribution of this element (fi ) to the
overall fitness (F ) of the entire configuration is given by fi (si ) when K = 0. When K > 0, the contribution
of an individual element to overall fitness depends both on its state and the states of K other elements
to which it is linked ( fi (si :si1 , . . . , sik )). A random fitness function (U (0, 1)) is adopted, and the overall
fitness of each configuration is calculated as the average of the fitness values of each of its individual
elements.
Altering the value of K effects the ruggedness of the described landscape, and consequently impacts
on the difficulty of search on this landscape [12,20]. The strength of the NK model in the context of this
study is that by tuning the value of K it can be used to generate strategic landscapes (graphs) of differing
degrees of local-fitness correlation (ruggedness).
The strategy of an organization is characterized as consisting of N attributes [2]. Each of these attributes
represents a strategic decision or policy choice, which an organization faces. Hence a specific strategic
configuration s, is represented as a vector s1 , . . . , sN where each attribute can assume a value of 0 or 1 [3].
The vector of attributes represents an entire organizational form, hence it embeds a choice of markets,
products, method of competing in a chosen market, and method of internally structuring the organiza-
tion [3]. Good consistent sets of strategic decisions — configurations, correspond to peaks on the strategic
landscape.
The definition of an organization as a vector of strategic attributes finds resonance in the work of
Porter [1,25], where organizations are conceptualized as a series of activities forming a value-chain.
The choice of what activities to perform, and subsequent decisions as to how to perform these activities,
defines the strategy of the organization. The individual attributes of an organization’s strategy interact. For
example, the value of an efficient manufacturing process is enhanced when combined with a high-quality
sales force. Differing values for K correspond to varying degrees of payoff-interaction among elements of
the organization’s strategy [3]. As K increases, the difficulty of the task facing strategic decision-makers is
magnified. Local-search attempts to improve an organization’s position on the strategic landscape become
ensnared in a web of conflicting constraints.

20.4.2 Simulation Model

Five characteristics of the problem domain that impact on the design of a simulation model are:

1. The environment is dynamic

2. Organizations are prone to strategic inertia. Their adaptive efforts are anchored by their past
3. Organizations do not knowingly select poorer strategies than the one they already have (election
operator)
4. Organizations make errorful ex-ante and assessments of fitness
5. Organizations coevolve

Although our simulator embeds all of the above, in this chapter we report results that consider the first
three of these factors. We note that this model bears passing resemblance to the eleMentals model of [16],
which combined a swarm algorithm and an NK landscape to investigate the development of culture and
intelligence in a population of hypothetical beings called eleMentals. However, the OrgSwarm simulator is

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 309 — #5

20-310 Handbook of Bioinspired Algorithms and Applications

differentiated from the eleMental model on grounds of application domain, and because it incorporates
the above five characteristics of strategic adaptation.
20.4.2.1 Dynamic Environment
Organizations do not compete in a static environment. The environment may alter as a result of exogenous
events, for example a regime change such as the emergence of a new technology or a change in customer
preferences. This can be mimicked in the simulation by stochastically respecifying the strategic landscape
during the course of a simulation run. These respecifications simulate a dynamic environment, and
a change in the environment may at least partially negate the value of past learning (adaptation) by
organizations. Minor respecifications are simulated by altering the fitness values associated with one of
the N dimensions of the strategic landscape, whereas in major changes, the fitness of the entire strategic
landscape is redefined. The environment faced by organizations can also change as a result of competition
among the population of organizations. The effect of interfirm competition is left for future work.
20.4.2.2 Strategic Anchor
Organizations do not have complete freedom to alter their current strategy. Their adaptive processes are
subject to strategic inertia. This inertia springs from the organization’s culture, history, and the mental
models of its management [6]. In the simulation, strategic inertia is mimicked by implementing a strategic
anchor. The degree of inertia can be varied from zero to high. In the latter case, the organization is highly
constrained from altering its strategic stance. By allowing the weight of this anchor to vary, adaptation
processes corresponding to different industries, each with different levels of inertia, can be simulated.
Inertia could be incorporated into the PSA in a variety of ways. We have chosen to incorporate it into
the velocity update equation, so that the velocity and direction of the particle at each iteration is also a
function of the location of its strategic anchor. Therefore for the simulations, Equation (20.1) is altered
by adding an additional “anchor” term

vi (t + 1) = vi (t ) + R1 (yi − xi (t )) + R2 (ŷ − xi (t )) + R3 (ai − xi (t )), (20.6)

where ai represents the position of the anchor for organization i (a full description of the other terms
such as R1 is provided in the pseudo-code below). The weight attached to the anchor parameter (R3 )
(relative to those attached to pbest and gbest), can be altered by the modeler. The position of the anchor
can be fixed at the initial position of the particle at the start of the simulation, or it can be allowed to
“drag,” thereby being responsive to the adaptive history of the particle. In the latter case, the position of
the anchor for each particle corresponds to the position of that particle “x” iterations ago.
20.4.2.3 Election operator
Real-world organizations do not usually intentionally move to strategies that are poorer (i.e produce a
lower payoff) than the one they already have. Hence, an election operator (also referred to as a conditional
update or ratchet operator) is implemented, which when turned on ensures that position updates that
would worsen an organization’s strategic fitness are discarded. In these cases, an organization remains at
its current location.

20.4.3 Outline of Swarm Algorithm

As the strategic landscape is described using a binary representation (the NK model), the canonical PSA
is adapted for the binary case using the BinPSO version of the algorithm [26]. The binary version of the
PSA is inspired by the idea that an agent’s probability of making a binary decision (yes/no, true/false) is a
function of both personal history and social factors. The probability that an agent chooses a value of for
example 1, for a particular decision in the next time period, is a function of the agent’s history (xi (t), vi (t)
& pbest) and social factors (lbest) (see Equation [20.7])

Prob(xi (t + 1) = 1) = f (xi (t), vi (t), pbest, lbest). (20.7)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 310 — #6

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-311

The vector vi is interpreted as organization i’s predisposition to set each of the N binary strategic choices
j
that it faces to 1. The higher the value of vi for an individual decision j, the more likely that organization
j
i will choose to set decision j = 1, with lower values of vi favoring the choice of decision j = 0.
In order to model the tendency of managers to repeat historically good strategies, values for each
dimension of xi that match those of pbest, should become more probable in the future. Adding the
j j j
difference between pbesti and xi for organization i to vi will increase the likelihood that organization
j j
i will choose to set decision j = 1 if the difference is positive (when pbesti = 1 and xi = 0). If the
j j j j
difference between pbesti and xi for organization i is negative (when pbesti = 0, and xi = 1), adding
j j
the difference to vi will decrease vi .1
In each iteration of the algorithm, the agent adjusts his decision-vector (xi (t)), taking account of his
historical experience (pbest), and the best strategy found by his peer group (lbest). Hence, the velocity
update equation used in the continuous version of the PSA (see Equation [20.6]) can still be used, although
now, vi (t + 1) is interpreted as the updated vector of an agent’s predisposition (or probability thresholds)
to set each of the N binary strategic choices that it faces to one

vi (t + 1) = vi (t) + R1 (pbesti − xi (t)) + R2 (lbesti − xi (t)) + R3 (anchori − xi (t)). (20.8)

To ensure that each element of the vector vi (t + 1) is mapped into (0, 1), a sigmoid transformation is
performed on each element j of vi (t + 1) (see Equation [20.9])

j 1
Sig(vi (t + 1)) = j
. (20.9)
1 + exp(−vi (t + 1))

Finally, the transformed vector of probability thresholds is used to determine the values of each element
of xi (t + 1), by comparing each element of Sig(vi (t)) with a random number drawn from U (0, 1) (see
Equation [20.10]).

j j j
If U (0, 1) < Sig(vi (t + 1)), then xi (t + 1) = 1; else xi (t + 1) = 0 (20.10)

In the binary version of the algorithm, trajectories/velocities are changes in the probability that a coordinate
j j
will take on a 0 or 1 value. Sig(vi ) represents the probability of bit xi taking the value 1 [26]. Therefore,
j j
if Sig(vi ) = 0.3 there is a 30% chance that xi = 1, and a 70% chance it is zero.

20.4.3.1 Pseudocode for Algorithm

The pseudocode for the swarm algorithm in the simulator is as follows:

For each entity in turn

For each dimension (strategic decision) n

v[n] = v[n] + R1*(pbest[n] - x[n]) + R2*(lbest[n]-x[n]) + R3*(a[n]-x[n])

If(v[n] > Max) v[n] = Vmax
If(v[n]<-Vmax) v[n] = -Vmax

j
1 The difference in each case is weighted by a random number drawn from U (0, 1). Therefore, if pbesti = 1,
j j j j
(pbesti − xi ) × U (0, 1) will be nonnegative. Adding this to vi will increase vi , and therefore also increase the
j j j j
probability that xi = 1. On the other hand if pbesti = 0, vi will tend to decrease, and Prob(xi ) = 1 becomes smaller.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 311 — #7

20-312 Handbook of Bioinspired Algorithms and Applications

If(Pr<Sig(v[n]))t[n] = 1 Else t[n] = 0

If(fitness(t)*(1 + e))>fitness(x)) //ratchet operator
For each dimension n x[n] = t[n]
UpdateAnchor(a) //if iteratively update anchor
//option is selected

R1 , R2 , and R3 are random weights drawn from a uniform distribution ranging from 0 to R1max , R2max , and
R3max , respectively, and they weight the importance attached to pbest, lbest and anchor in each iteration
of the algorithm. R1max , R2max , and R3max are constrained to sum up to 4.0 in line with the BinPSO
algorithm of [26]. x is the particle’s actual position, pbest is its past best position, lbest its local best and
a is the position of its anchor. Vmax is set to 4.0 to ensure that Sig(v[n]) does not get too close to either
0 or 1, therefore ensuring that there is a nonzero possibility that a bit will flip state during each iteration.
Pr is a random value drawn from U (0, 1), and Sig is the sigmoid function: Sig(x) = 1/(1 + exp(−x)),
which squashes v into the range 0 → 1 range. t is a temporary record that is used in order to implement
the ratchet operator. If the new strategy is considered better than the organization’s existing strategy, it is
accepted and t is copied into x. Otherwise t is discarded and x remains unchanged. e is the error or noise,
injected in the fitness evaluation, in order to mimic an errorful forecast of strategy fitness.

20.4.4 Simulator
Although the underlying code for the OrgSwarm simulator is written in C + + , the user interacts with the
simulator through a series of easy-to-use screens (Figure 20.1 shows one of the screens in the main control
menu for the simulator). These screens allow the user to select and alter a wide variety of parameters that
determine the nature of the simulation run. In essence, the simulator allows the user to select choices for
four items:

1. The form of NK landscape generated

2. The nature of the search heuristics to be employed by inventors
3. The number of simulations to be run
4. The form of output generated during the simulation run

During the simulation run, a series of graphics (see Figure 20.2 for an example of a graphic that shows
the status of each particle in the population during the simulation run), and a run report (see Figure 20.3)
can be displayed. The report display records the full list of simulation parameters chosen by the modeler,
as well as providing a running record of the best design in the population at the end of each iteration. The
simulator also facilitates the recording of comprehensive run-data to disk during the simulation.

FIGURE 20.1 Main control screen for OrgSwarm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 312 — #8

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-313

FIGURE 20.2 OrgSwarm screendump showing the status of each particle in the population during the simulation
run. Three bars are shown for each of the 20 particles in the population, and these bars represent the fitness of the
anchor location, the fitness of the pbest location, and the fitness of the current location of each particle.

20.5 Results
All simulations were run for 5000 iterations, and all reported fitnesses are the average population fitnesses,
and average environment best fitnesses, across 30 separate simulation runs. On each of the simulation
runs, the NK landscape is specified anew, and the positions and velocities of particles are randomly
initialized at the start of each run. A population of 20 particles is employed, with a neighborhood of size
18. The choice of a high value for the neighborhood, relative to the size of the population, arises from the
observation that real-world organizations know the profitability of their competitors.
Tables 20.1 and 20.2 provide the results for each of 14 distinct PSA variants, at the end of 5000
iterations, across a number of static and dynamic NK landscape scenarios. In each scenario, the same
series of simulations are undertaken. Initially, a basic PSA is employed, without an anchor or a ratchet
(conditional move) heuristic. This simulates a population of organizations searching a strategic landscape,
where the population has no strategic inertia, and where organizations do not utilize a ratchet operator
in deciding whether to alter their position on the strategic landscape.
The basic PSA is then supplemented by inclusion of a series of strategic anchor formulations, ranging
from an anchor that does not change position during the simulation (initial position anchor) to one that
can adapt after a time-lag (moving anchor). Two lag periods are examined, 20 and 50 iterations. Differing
weights can be attached to the anchor term in the velocity Equation (20.6), ranging from 0 (anchor is
“turned off ”) to a maximum of 4. To determine whether the weight factor for the anchor term has a
critical impact on the results, results are reported for weight values of both 1 and 3, corresponding to
low and high inertia weights. Next, to isolate the effect of the ratchet, the conditional move operator is
implemented, and the anchor term is dropped. Finally, to ascertain the combined effect of both ratchet
and anchor, the anchor simulations outlined above are repeated with the ratchet operator “turned on.”

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 313 — #9

20-314 Handbook of Bioinspired Algorithms and Applications

FIGURE 20.3 A typical run report generated by OrgSwarm.

TABLE 20.1 Average (Environment Best) Fitness after 5000 Iterations, Static
Landscape

Fitness
Algorithm (N = 96, K = 0) (N = 96, K = 4) (N = 96, K = 10)
Basic PSA 0.4641 (0.5457) 0.5002 (0.6000) 0.4991 (0.6143)
Initial anchor, w = 1 0.4699 (0.5484) 0.4921 (0.5967) 0.4956 (0.6102)
Initial anchor, w = 3 0.4943 (0.5591) 0.4994 (0.5979) 0.4991 (0.6103)
Mov. anchor (50,1) 0.4688 (0.5500) 0.4960 (0.6003) 0.4983 (0.6145)
Mov. anchor (50,3) 0.4750 (0.5631) 0.4962 (0.6122) 0.5003 (0.6215)
Mov. anchor (20,1) 0.4644 (0.5475) 0.4986 (0.6018) 0.5001 (0.6120)
Mov. anchor (20,3) 0.4677 (0.5492) 0.4994 (0.6156) 0.4994 (0.6229)
Ratchet PSA 0.5756 (0.6021) 0.6896 (0.7143) 0.6789 (0.7035)
Rach-Initial anchor, w = 1 0.6067 (0.6416) 0.6991 (0.7261) 0.6884 (0.7167)
Rach-Initial anchor, w = 3 0.5993 (0.6361) 0.6910 (0.7213) 0.6844 (0.7099)
Rach-Mov. anchor (50,1) 0.6659 (0.6659) 0.7213 (0.7456) 0.6990 (0.7256)
Rach-Mov. anchor (50,3) 0.6586 (0.6601) 0.7211 (0.7469) 0.6992 (0.7270)
Rach-Mov. anchor (20,1) 0.6692 (0.6695) 0.7211 (0.7441) 0.6976 (0.7243)
Rach-Mov. anchor (20,3) 0.6612 (0.6627) 0.7228 (0.7462) 0.6984 (0.7251)

“Real world” strategy vectors consist of a large array of strategic decisions. A value of N = 96 was chosen
in defining the landscapes in this simulation. It is noted that there is no unique value of N that could
have been selected, but the selection of very large values are not feasible due to computational limitations.
However, a binary string of 96 bits provides 296 , or ∼1028 , distinct choices of strategy. It is also noted

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 314 — #10

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-315

TABLE 20.2 Average (Environment Best) Fitness after 5000 Iterations, Entire
Landscape Respecified Stochastically

Fitness
Algorithm (N = 96, K = 0) (N = 96, K = 4) (N = 96, K = 10)

Basic PSA 0.4761 (0.5428) 0.4886 (0.5891) 0.4961 (0.6019)

Initial anchor, w = 1 0.4819 (0.5524) 0.4883 (0.5822) 0.4982 (0.6075)
Initial anchor, w = 3 0.5021 (0.5623) 0.4967 (0.5931) 0.4998 (0.6047)
Mov. anchor (50,1) 0.4705 (0.5450) 0.4894 (0.5863) 0.4974 (0.6008)
Mov. anchor (50,3) 0.4800 (0.5612) 0.4966 (0.6053) 0.5010 (0.6187)
Mov. anchor (20,1) 0.4757 (0.5520) 0.4926 (0.5867) 0.4985 (0.6097)
Mov. anchor (20,3) 0.4824 (0.5632) 0.4986 (0.6041) 0.5004 (0.6163)
Ratchet PSA 0.5877 (0.6131) 0.6802 (0.7092) 0.6754 (0.7015)
Rach-Initial anchor, w = 1 0.6187 (0.6508) 0.6874 (0.7180) 0.6764 (0.7070)
Rach-Initial anchor, w = 3 0.6075 (0.6377) 0.6841 (0.7130) 0.6738 (0.7017)
Rach-Mov. anchor (50,1) 0.6517 (0.6561) 0.7134 (0.7387) 0.6840 (0.7141)
Rach-Mov. anchor (50,3) 0.6597 (0.6637) 0.7049 (0.7304) 0.6925 (0.7225)
Rach-Mov. anchor (20,1) 0.6575 (0.6593) 0.7152 (0.7419) 0.6819 (0.7094)
Rach-Mov. anchor (20,3) 0.6689 (0.6700) 0.7158 (0.7429) 0.6860 (0.7147)

Strategic inertia using particle swarm K = 0

0.7

0.65
Mean average fitness (30 runs)

0.6

Basic PSA
Anchor (1)
Anchor (3)
0.55 Mov. Anchor (50,1)
Mov. Anchor (50,3)
Mov. Anchor (20,1)
Mov. Anchor (20,3)
Ratchet PSA
R-Anchor (1)
R-Anchor (3)
R-Mov. Anchor (50,1)
0.5
R-Mov. Anchor (50,3)
R-Mov. Anchor (20,1)
R-Mov. Anchor (20,3)

0.45
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Iteration

FIGURE 20.4 Plot of the mean average fitness on the static landscape where K = 0.

that we would expect the dimensionality of the strategy vector to exceed the number of organizations in
the population, hence the size of the population is kept below 96, and a value of 20 is chosen. A series of
landscapes of differing K values (0, 4, and 10), representing differing degrees of fitness interconnectivity,
were used in the simulations.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 315 — #11

20-316 Handbook of Bioinspired Algorithms and Applications

Strategic inertia using particle swarm K = 10

Strategic inertia using particle swarm K = 4
0.75 0.7

Basic PSA
0.7 Anchor (1)
0.65
Anchor (3)
Mov. Anchor (50,1)
Mov. Anchor (50,3)
Basic PSA Mov. Anchor (20,1)
Anchor (1)
Mov. Anchor (20,3)

Mean average fitness (30 runs)

0.65 Anchor (3)
Mean average fitness (30 runs)

Mov. Anchor (50,1) Ratchet PSA

0.6 R-Anchor (1)
Mov. Anchor (50,3)
R-Anchor (3)
Mov. Anchor (20,1)
R-Mov. Anchor (50,1)
Mov. Anchor (20,3) R-Mov. Anchor (50,3)
Ratchet PSA R-Mov. Anchor (20,1)
0.6 R-Anchor (1)
R-Mov. Anchor (20,3)
R-Anchor (3)
R-Mov. Anchor (50,1)
R-Mov. Anchor (50,3) 0.55
R-Mov. Anchor (20,1)
R-Mov. Anchor (20,3)
0.55

0.5
0.5

0.45 0.45
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Iteration Iteration

FIGURE 20.5 Plot of the mean average fitness on the static landscape where K = 4 (left) and 10 (right).

20.5.1 Static Landscape

Table 20.1 and Figure 20.4 and Figure 20.5 provide the results for a static NK landscape. Examining these
results suggests that the basic PSA, without anchor or ratchet heuristics, performs poorly, even on a static
landscape. The average populational fitnesses obtained after 5000 iterations (averaged over all 30 runs) is no
better than random search, suggesting that unfettered adaptive efforts, based on “social communication”
between organizations (gbest), and a memory of good past strategies (pbest) is not sufficient to achieve
high levels of populational fitness. When various anchor term mechanisms, simulating strategic inertia,
are added to the basic PSA, the results are not qualitatively altered from those of the basic PSA. This
suggests that social communication and inertia are not sufficient for the attainment of high levels of
populational strategic fitness.
When a ratchet heuristic is added to the basic PSA, a significant improvement (statistically significant
at the 5% level) in both average populational, and average environment best fitness is obtained across
landscapes of all K values, suggesting that the simple decision heuristic of “only abandon your current
strategy for a better one” can lead to notable increases in populational fitness.
Finally, the results of a series of simulations that combine anchor and ratchet mechanisms are reported.
Virtually all of these combinations lead to significantly (at the 5% level) enhanced levels of populational
fitness against the ratchet-only PSA, suggesting that strategic inertia can be beneficial, when organizations
employ a conditional move test before adopting new strategies. Examining the combined ratchet and anchor
results in more detail, the best results are obtained when the anchor is not fixed at the initial location of
each particle on the landscape, but when it is allowed to “drag” or adapt, over time. The results are not
qualitatively sensitive to the weight value (1 or 3).

20.5.2 Dynamic Landscape

The real world is rarely static, and changes in the environment can trigger adaptive behavior by agents
in a system [27]. Table 20.2 and Figure 20.6 and Figure 20.7 provide results for the case where the entire

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 316 — #12

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-317

Strategic inertia using particle swarm K = 0

0.7

0.65

Basic PSA
Anchor (1)
Mean average fitness (30 runs) Anchor (3)
Mov. Anchor (50,1)
0.6

Mov. Anchor (50,3)

Mov. Anchor (20,1)
0.55
Mov. Anchor (20,3)
Ratchet PSA
R-Anchor (1)
R-Anchor (3)
R-Mov. Anchor (50,1)
R-Mov. Anchor (50,3)
R-Mov. Anchor (20,1)
0.5 R-Mov. Anchor (20,3)

0.45
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Iteration

FIGURE 20.6 Plot of the mean average fitness on the dynamic landscape where K = 0.

NK landscape is respecified in any iteration with a p = 0.00025. When the landscape is wholly or partially
respecified, the benefits of past strategic learning by organizations is eroded (see [27–29] for a detailed
discussion of the utility of the PSO in tracking dynamic environments).
Qualitatively, the results in both scenarios are similar to those obtained on the static landscape. The
basic PSA, even if supplemented by an anchor mechanism, does not perform any better than random
search. Supplementing the basic PSA with the ratchet mechanism leads to a significant improvement in
populational fitness, with a further improvement in fitness occurring when the ratchet is combined with
an anchor mechanism. In the latter case, an adaptive or dragging anchor gives better results than a fixed
anchor, but the results between differing forms of dragging anchor do not show a clear dominance for any
particular form. As for the static landscape case, the results for the combined ratchet/anchor, are relatively
insensitive to the choice of weight value (1 or 3).

20.6 Conclusions
In this chapter, a synthesis of a strategic landscape defined using the NK model, and a Particle Swarm
metaphor is used to create a novel simulation model of the process of strategic adaptation of organizations.
The results suggest that a degree of strategic inertia, in the presence of an election operator, can assist rather
than hamper the adaptive efforts of populations of organizations in static and slowly changing strategic
environments. The results also suggest that despite the claim for the importance of social learning in
populations, social learning alone is not always enough, unless learnt lessons can be maintained by means
of an election mechanism.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 317 — #13

20-318 Handbook of Bioinspired Algorithms and Applications

Strategic inertia using particle swarm K = 4 Strategic inertia using particle swarm K = 10
0.75 0.7

0.7 Basic PSA

Anchor (1)
0.65 Anchor (3)
Mov. Anchor (50,1)
Basic PSA Mov. Anchor (50,3)
Anchor (1)
Anchor (3) Mov. Anchor (20,1)
0.65 Mov. Anchor (50,1) Mov. Anchor (20,3)

Mean Average Fitness (30 Runs)

Mean average fitness (30 runs)

Ratchet PSA
Mov. Anchor (50,3)
R-Anchor (1)
Mov. Anchor (20,1) 0.6 R-Anchor (3)
Mov. Anchor (20,3) R-Mov. Anchor (50,1)
Ratchet PSA R-Mov. Anchor (50,3)
R-Anchor (1)
0.6 R-Mov. Anchor (20,1)
R-Anchor (3)
R-Mov. Anchor (20,3)
R-Mov. Anchor (50,1)
R-Mov. Anchor (50,3)
R-Mov. Anchor (20,1) 0.55
R-Mov. Anchor (20,3)

0.55

0.5
0.5

0.45 0.45
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Iteration Iteration

FIGURE 20.7 Plot of the mean average fitness on the dynamic landscape where K = 4 (left) and 10 (right).

It is not possible in a single set of simulation experiments to exhaustively examine every possible com-
bination of settings for each parameter in the simulation model. Future work will extend the range of
settings examined. However, the initial results cast an interesting light on the role of anchoring in organ-
izational adaptation, and the development of the swarm-landscape simulator extends the methodologies
available to researchers to conceptualize and examine organizational adaptation.
Finally, it is noted that the concept of anchoring developed in this chapter is not limited to organizations,
but is plausibly a general feature of social systems. Hence, the extension of the social swarm model to
incorporate inertia may prove useful beyond this study.

References
[1] Porter, M. (1996). What is strategy? Harvard Business Review, Nov.–Dec., pp. 61–78.
[2] Levinthal, D. (1997). Adaptation on rugged landscapes. Management Science, 43: 934–950.
[3] Rivkin, J. (2000). Imitation of complex strategies. Management Science, 46: 824–844.
[4] Wright, S. (1932). The roles of mutation, inbreeding, crossbreeding and selection in evolution.
In Proceedings of the Sixth International Congress on Genetics, 1: 356–366.
[5] Kitts, B., Edvinsson, L., and Beding, T. (2001). Intellectual capital: from intangible assets to fitness
landscapes. Expert Systems with Applications, 20: 35–50.
[6] Boeker, W. (1989). Strategic change: The effects of founding and history. Academy of Management
Journal, 32: 489–515.
[7] Stuart, T. and Podolny, J. (1996). Local search and the evolution of technological capabilities,
Strategic Management Journal, 17: 21–38.
[8] Hannan, M. and Freeman, J. (1977). The populational ecology of organizations. American Journal
of Sociology, 82: 929–964.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 318 — #14

Simulating the Strategic Adaptation of Organizations Using OrgSwarm 20-319

[9] Hannan, M. and Freeman, J. (1984). Structural inertia and organizational change. American
Sociological Review, 49: 149–164.
[10] Levinthal, D. (1991). Random walks and organisational mortality. Administrative Science
Quarterly, 36: 397–420.
[11] Tushman, M. and O’Reilly, C. (1996). Ambidextrous organizations: Managing evolutionary and
revolutionary change. California Management Review, 38: 8–30.
[12] Kauffman, S. (1993). The Origins of Order. Oxford University Press, Oxford, England.
[13] Kennedy, J. and Eberhart, R. (1995). Particle swarm optimization. In Proceedings of the IEEE
International Conference on Neural Networks, December 1995, pp. 1942–1948.
[14] Kennedy, J., Eberhart, R., and Shi, Y. (2001). Swarm Intelligence. Morgan Kauffman, San Mateo, CA.
[15] Kennedy, J. (1997). The particle swarm: Social adaptation of knowledge. In Proceedings of the
International Conference on Evolutionary Computation. IEEE Press, Washington, pp. 303–308.
[16] Kennedy, J. (1999). Minds and cultures: Particle swarm implications for beings in sociocognitive
space. Adaptive Behavior, 7: 269–288.
[17] Kennedy, J. (1999). Small worlds and mega-minds: effects of neighbourhood topology on particle
swarm performance. In Proceedings of the International Conference on Evolutionary Computation.
IEEE Press, Washington, pp. 1931–1938.
[18] Thorndike, E. (1911). Animal Intelligence, Macmillan, New York.
[19] Bandura, A. (1986). Social Foundations of Thought and Action: A Social Cognitive Theory.
Prentice Hall, Englewood Cliffs, NJ.
[20] Kauffman, S. and Levin, S. (1987). Towards a general theory of adaptive walks on rugged
landscapes. Journal of Theoretical Biology, 128: 11–45.
[21] Gavetti, G. and Levinthal, D. (2000). Looking forward and looking backward: Cognitive and
experiential search, Administrative Science Quarterly, 45: 113–137.
[22] Porter, M. and Siggelkow, N. (2001). Contextuality within activity systems. Harvard Business
School Working paper series, No. 01-053.
[23] Lobo, J. and MacReady, W. (1999). Landscapes: A Natural Extension of Search Theory.
Santa Fe Institute Working paper 99-05-037.
[24] Kauffman, S., Lobo, J., and MacReady, W. (1998). Optimal Search on a Technology Landscape.
Santa Fe Institute Working paper 98-10-091.
[25] Porter, M. (1985). Competitive Advantage:Creating and Sustaining Superior Performance. The Free
Press, New York.
[26] Kennedy, J. and Eberhart, R. (1997). A discrete binary version of the particle swarm algorithm.
In Proceedings of the Conference on Systems, Man, and Cybernetics. IEEE Press, Washington,
pp. 4104–4109.
[27] Blackwell, T. (2003). Swarms in dynamic environments. In Proceedings of GECCO 2003, Vol. 2723
of Lecture Notes in Computer Science. Springer-Verlag, Berlin, pp. 1–12.
[28] Eberhart, R. and Shi, Y. (2001). Tracking and optimizing dynamic systems with particle swarms.
In Proceedings of the CEC 2001. IEEE Press, Washington, pp. 94–97.
[29] Hu, X. and Eberhart, R. (2002). Adaptive particle swarm optimization: detection and response to
dynamic systems. In Proceedings of CEC 2002. IEEE Press, Washington, pp. 1666–1670.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C020” — 2005/8/6 — 14:10 — page 319 — #15

 21
BeeHive: New Ideas
for Developing
Routing Algorithms
Inspired by
Honey Bee Behavior

21.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-322

Organization of the Chapter
21.2 An Agent-Based Investigation of Honey Bee
Colony . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-323
Labor Management • The Communication Network of a
Honey Bee Colony • Reinforcement Learning • Distributed
Coordination and Planning • Energy Efficient Foraging •
Stochastic Selection of Flower Sites • Group Organization
21.3 An Engineering Approach to Nature Centered
Routing Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-324
21.4 BeeHive: The Mapping of Concepts from Nature to
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-326
21.5 The Bee Agent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-327
Algorithm Design of BeeHive
21.6 Routing Algorithms Used for Comparison . . . . . . . . . . . . 21-331
AntNet: An Algorithm Inspired from Ant Behavior •
Distributed Genetic Algorithm • Open Shortest
Path First
21.7 Simulation Environment for BeeHive . . . . . . . . . . . . . . . . . . 21-332
21.8 Experimental Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-333
Horst F. Wedde 21.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-337
Muddassar Farooq References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-338

Some parts of the chapter have been taken from H.F. Wedde, M. Farooq, and Y. Zhang. BeeHive: An efficient fault-
tolerant routing algorithm inspired by honey bee behavior. Ant Colony Optimization and swarm intelligence, Lecture
Notes in Computer Science 3172, by kind permission of Springer-Verlag, Heidelberg, Germany. The paper also won
the best paper award at ANTS2004 conference.

21-321

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 321 — #1

21-322 Handbook of Bioinspired Algorithms and Applications

21.1 Introduction
A honey bee colony manages to react to countless changes in the forage pattern outside the hive and
internal changes inside the hive through a decentralized and sophisticated communication and control
system. According to Seeley, a honey bee colony can thoroughly monitor a vast region around the hive for
rich food sources, nimbly redistribute its foragers within an afternoon, fine-tune its nectar processing to
match its nectar collecting, effect cross inhibition between different forager groups to boost its response
differential between food sources, precisely regulate its pollen intake in relation to its ratio of internal
supply and demand, and limit the expensive process of comb building to times of critical need for
additional storage space [1]. A bee colony demonstrates this flexible and adaptive response because it is
organized with morphologically uniform individuals but with different temporary specializations. A bee
takes up four roles during her life span — cleaner, nurse, food-storer, and forager. The foragers could
be further recognized as nectar, pollen, and water collectors [1]. They have two functional roles within
each sub specialty: scouts, who discover new food sources around the hive, and foragers, who transport
nectar from an already discovered flower site by following the dances of other scouts or foragers. The
colony brilliantly allocates, through its communication and control system, its labor force among these
individuals to maintain a balance between collection and processing rate of each commodity, as a result,
an optimum stock of nectar, pollen, and water is piled inside the colony. On an average, a colony extracts
from its environment around 20 kg of pollen, 120 kg of nectar, 25 l of water, and 100 g of resin each
year [1].
Karl von Frisch in 1944 made a revolutionary discovery about the communication paradigm that
foragers use to communicate the information about flower sites around the hive. He experimentally
verified that foragers can inform their fellow foragers, inside the hive on a dance floor, of the direction
and the distance to a food source by means of a dance. He deciphered these communication signals
into a language, in his book Tanzsprache und Orientierung der Bienen [2] (the translation was done by
Chadwick [3]). The foragers, according to von Frisch, used two type of dances: round dances, which
show that a food source is present near the hive (about 100 m), and waggle dances, which further specify
the direction and the distance to a food source (up to few kilometers), by the orientation and duration
portion of the waggle portion of each dance circuit. The food sources located in the direction of the
sun are represented as an upward direction on the vertical comb, and any angle to the right or left of the
sun is encoded as any angle to the right or left of the upward direction. The recruited foragers maintain
this angle from the sun to reach in a small region around the flower and then use the flower fragrance,
which clings to the body of a forager and which they smelt while observing the dances, to identify the food
source unambiguously. Recently, the researchers have also indicated that several components of dance are
correlated with food-source profitability, such as dance duration and probability of sound production;
however, this information is not taken into account by unemployed (dance following) foragers, instead
they randomly choose a dancer before leaving the hive. Nevertheless, the recruited foragers arrive in greater
numbers at more profitable food sources because the dances for richer sources are more conspicuous and
hence likely to be encountered by the unemployed (dance following) foragers [1].

21.1.1 Organization of the Chapter

In Section 21.2 we will briefly provide an overview of the important organizational principles of a
honey bee colony that enabled us to develop a robust routing algorithm. An insight into these principles
will help the reader in understanding the motivation for following an engineering approach, discussed
in Section 21.3, for a systematic transformation of concepts from Nature to Networks, described in
Section 21.4. In Section 21.5 we will first develop the key ideas of the bee agent model underlying the
BeeHive algorithm. Based on this we will present our BeeHive algorithm in Section 21.5.1. In Section 21.6,
we will briefly introduce the important features of three state-of-the-art algorithms, with which we com-
pared the performance of BeeHive. Section 21.7 will describe the simulation and the network environment
that were used to compare the performance of BeeHive with other algorithms. Section 21.8 discusses the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 322 — #2

BeeHive: New Ideas for Developing Routing Algorithms 21-323

results obtained from the extensive simulations. Finally, we conclude our findings, and provide an outlook
to future research.

21.2 An Agent-Based Investigation of Honey Bee Colony

In this section we briefly outline the organizational principles of a honey bee colony that will enable
the computer scientists to develop agent-based algorithms for different optimization and real world
problems. A honey bee colony solves the most interesting multi-objective optimization problem: how to
allocate resources to different tasks under constantly changing operating environment so that the colony
maximizes its profit gains. This is the same problem that many researchers try to solve in the field of
multi-objective optimization under dynamic and time-varying environments. The reader can find details
of these principles in Reference 1.

21.2.1 Labor Management

As discussed in Section 21.1 that a honey bee colony is organized with morphologically uniform individuals
but with different temporary specializations. The benefit of this approach is that it enhances a colony’s
flexibility to adapt its response according to the ever changing environment while at the same time doing
the tasks with acceptable level of efficiency. For example, a nectar forager is able to extensively forage at
a discovered flower site as he/she does not waste time in storing the nectar inside the hive. Moreover, a
bee colony is able to adapt the activity level of its specialists according to the group’s needs. For example,
nectar foragers may become pollen foragers if the amount of protein that they receive from nurse bees
falls below a threshold level, or nurse bees might take the role of food-storer bees if the rate of processing
nectar is slower than the rate of collecting nectar (foragers indicate this by a tremble dance).

21.2.2 The Communication Network of a Honey Bee Colony

A honey bee colony utilizes a hybrid communication network that consists of signals and cues for
information exchange among its members. Signals are information-bearing actions or structures that
have been shaped by the natural selection, specifically to convey information. Cues are variables that
likewise convey information about the state of the colony but have not been modeled by the natural
selection to convey that information [4]. Signals enable direct communication among the members of a
honey bee colony via waggle dance, tremble dance, and shaking signal while cues enable indirect com-
munication among the members through the environment shared by them. Both mechanisms provide an
efficient information exchange protocol that empowers the members to, mostly, communicate indirectly
(group-to-one paradigm), and when required, directly using one-to-one paradigm.
A good example of cue is the search time for finding a food-storer bee that a forager experiences once she
wants to unload her nectar. A nectar forager uses this cue to get an estimate of both rates, nectar collecting
and nectar processing, of the colony; higher search time to find a food-storer bee is an indicator to the
forager that the rate of processing nectar is slower than the rate of collecting nectar. Consequently, she will
decide to perform a tremble dance instead of a waggle dance. The forager by doing a tremble dance will
achieve two objectives: one, she will recruit more food-storer bees to increase the rate of processing nectar
and two, she will make other foragers, on the dance floor, to stop performing waggle dances, as a result,
the rate of collecting nectar will be decreased.

21.2.3 Reinforcement Learning

A colony experiences a strong fluctuation in the external supply or internal demand (or both) for its
commodities: nectar, pollen, and water. The feedback signals, both negative and positive, are important
to regulate their amount so that the colony has sufficient stockpile of each of these commodities. This
is achieved by recruiting unemployed forgers for good supply sites through waggle dances. A forager
decides to dance only if the quality of the food site, visited by her, is above a certain threshold or

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 323 — #3

21-324 Handbook of Bioinspired Algorithms and Applications

it experiences a very small search time to locate a food-storer bee for its commodity (a cue that the colony
needs the commodity). By maintaining the search time within some thresholds, the honey bee colony
reinforces the foraging labor at a site in times of need and vice versa. A stochastic model for the foraging
behavior is presented in Reference 5. Sumpter used this basic model to come up with an agent-based
model in Reference 6. The Sumpter’s model provides a solid foundation for developing an agent-based
reinforcement learning algorithm [7].

21.2.4 Distributed Coordination and Planning

A honey bee colony achieves coordination among its thousands of members without any central authority.
The colony does not have a hierarchy where some individuals require information and then allocate tasks
to different members and monitor them. Each individual decides to do a job depending upon the need of
the colony that it estimates using the above-mentioned communication network.

21.2.5 Energy Efficient Foraging

The foragers tend to optimize the energetic efficiency of foraging at a flower site. For example, if during
an average foraging trip a forager collects G units of energy, expends C units of energy, and spends
time T , then energetic efficiency could be defined as (G − C)/C. Consequently, the neighborhood site
of the hive get preference than the sites far away from it. That is why von Frisch believed that those
foragers (short distance) which return from nearby sites perform round dances while other (long distance)
foragers perform waggle dances. This principle enables a colony to collect the commodities at an optimum
expenditure of energy.

21.2.6 Stochastic Selection of Flower Sites

The unemployed foragers on the dance floor observe, at the maximum, two or three dancers and then
decide to choose one among them at random. They do not broadly survey the dance floor to identify the
best flower site. This concept is contradictory to human society where well informed customers are crucial
to proper functioning of competitive markets. According to Seeley, this stems from the fact that the
individual forager seeks to maximize her colony’s profits, whereas the individual human tries to maximize
her or his own profit gains. This “sacrifice for group principle” enables a colony to distribute its forager
force over different flower sites rather than allocating it to the best site only. Such a policy results in quick
reallocation of foragers, which were foraging at the best site, to other discovered sites, once the best flower
site is about to fade away.

21.2.7 Group Organization

The employed foragers that collect nectar from the same type of flowers recognize one another in the hive
through the flower fragrance that clings to their body. Only group companions respond to the dances and
they show no interest in the dances performed by the foragers of other groups, foraging at other types
of flowers. However, the employed foragers might switch their group if the quality of their flower site
degrades to an extent that is no longer profitable [3].
Section 21.3 introduces the engineering approach that was developed to transform the above-mentioned
principles into a routing algorithm for packet switching networks. The engineering approach, however,
provides a general set of guidelines for developing algorithms/products inspired from natural systems.

21.3 An Engineering Approach to Nature Centered

Routing Protocols
In this section, we will introduce an engineering approach that we followed in the design and development
of a routing protocol inspired from a natural system (a honey bee colony).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 324 — #4

BeeHive: New Ideas for Developing Routing Algorithms 21-325

Definition 21.1 (Natural Engineering). Natural Engineering is an emerging new engineering discipline that
enables scientists/engineers to utilize inspirations and observations from organizational principles of natural
systems, and transform them into structural principles of software organization (algorithms) or industrial
products, in search of efficient/optimal solutions for real world problems under constraints of the resources.

The above-mentioned concept emphasizes six aspects:

1. Understanding the working principles of natural systems.

2. Understanding the operational environment of the target system.
3. Developing an algorithmic model of the organizational principles of natural systems.
4. Mapping the concepts from a natural system to a technical system.
5. Adapting the algorithmic model according to the operational environment of a technical system.
6. Follow test, evaluate, improve feedback loop in search of optimum solutions under the resource
constraints (time, space, computation, money, labor, etc.).

There is, of course, no clear-cut way to achieve one-to-one match of structures/principles in Nature life
organizations with working principles in technical systems. The most important challenge, therefore, is to
identify a natural system whose working principles could be easily abstracted to the one in the technical
system. If one has to add many nonbiological features into the natural system then, we believe, it is more
advisable to look at other natural systems for inspiration. Consequently, we chose honey bees because
their foraging behavior could be easily abstracted into a routing problem in telecommunication networks.
Both systems have to maximize the amount of a commodity (nectar/data delivered to the hive/nodes) as
quickly as possible, under a continuously changing operating environment.
The major focus of our research, however, is to take bio/nature inspired solutions into business and
therefore we decided to follow, for developing the BeeHive algorithm, a feedback oriented engineering
approach as displayed schematically in Figure 21.1 that incorporates most of the features discussed here.
First, we considered the ensemble of constraints under which the envisioned routing protocol is
supposed to operate:

• Nonavailability of a global clock for trip time calculation.

• Routers and links could crash.
• Routers have limited queue capacity.
• Links have a BER (Bit Error Rate) associated with them.
• The requirements from the Linux kernel routing framework needed to support the protocol.
• The requirements of the IP protocol which is currently used at the network layer in Internet.

At the same time we decided that the bee agents should explore the network, collect important parameters,
and make the routing decisions in a decentralized fashion (in the style as real scouts/foragers do decision
making during collecting nectar from flowers). Bee agents should measure the quality of a route and then
communicate to other bee agents like foragers do that in Nature. The structure of the routing table should
provide the functionality of a dance floor for exchange of information among bee agents, and among
bee agents and data packets. Moreover, later on we should be able to utilize it in a real kernel of the Linux
operating system.
We implemented our ideas in a simulation environment and then refined our algorithmic mapping
through the feedback channel 1 (see Figure 21.1). During this phase we did not use any simulation
specific features that were not available inside the Linux kernel, for example, using vector, stack, or similar
data structures. Once we reached a relative optimum of the BeeHive concept, we started to develop
an engineering model of the algorithm. The engineering model could be easily transported to the Linux
kernel routing framework. We tested it on the real network of Linux machines and refined our engineering
model through the feedback channel 2 (see Figure 21.1). At the moment we are evaluating our conceptual
approach in two prototype projects: BeeHive [8], which deals with the design and development of a
routing algorithm for fixed networks, and BeeAdHoc, whose goal is to design and develop an energy
efficient routing algorithm for mobile ad hoc networks (MANETS) [9].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 325 — #5

21-326 Handbook of Bioinspired Algorithms and Applications

Natural
system

Algorithmic
Working model
environment

Natural algorithm to
Model system algorithm
mapping

Testing and 1
evaluation

Engineering
model

Testing and evaluation

on real system

FIGURE 21.1 Nature centered protocol engineering.

In Section 21.4 we will elaborate the most important step of our engineering approach: the mapping of
concepts in a honey bee colony, discussed in Section 21.2, to operating environment of real packet switch-
ing networks. This step will help the reader in understanding our algorithm described in Section 21.5.
We do not discuss the implementation of the BeeHive inside the network stack of Linux kernel because it is
beyond the scope of the chapter, however, interested reader will find the technical details in Reference 10.

21.4 BeeHive: The Mapping of Concepts from

Nature to Networks
This section will briefly illustrate the mapping of concepts from nature to networks, one of the most
important step in Natural Engineering, that will simplify for the reader to trace back the origin of important
features of our BeeHive algorithm into the principles of a honey bee colony:
1. We could consider each node in the network as a hive that consists of bee agents. Each node
periodically launches the bee agents that explore the network and collect the routing information
that provides the nodes, which they visit, the partial information on the state of the network. These
bee agents could be considered as scouts that explore and evaluate the quality of multiple paths
between their launching node and the nodes that they visit.
2. Bee agents provide to the nodes, which they visit, the information on the propagation and queuing
delay of the paths, which they explored, leading to their launching nodes from the visited nodes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 326 — #6

BeeHive: New Ideas for Developing Routing Algorithms 21-327

One could consider propagation delay as a distance information and queuing delay as a direction
information (please remember bee scouts also provide these parameters in their dances).
3. A bee agent decides to provide its path information only, if the quality of the path it traversed,
is above a threshold. The threshold is controlled by the number of hops that a bee agent is allowed
to take. Moreover, the agents model the quality of a path as a function of the propagation and the
queuing delay of the path; lower values of the parameters result in higher values for the quality
parameter.
4. Majority of bee agents in BeeHive algorithm explore the network near their launching node and
very few explore the complete network. The idea is borrowed from a honey bee colony as this
reduces not only the overhead of collecting the routing information but also helps in maintaining
smaller/local routing tables.
5. We consider routing table as a dance floor where the bee agents provide the information about
the quality of the paths they traversed. The routing table is used for information exchange among
bee agents, launched from the same node but arrived at a node via its different neighbors. This
information exchange helps in evaluating the overall quality of a node (that has multiple pathways
to a destination) for reaching a certain destination.
6. The nectar foragers exploit the flower sites according to their quality while the distance and direction
to the sites is communicated through dances made by their fellow foragers, on the dance floor. In our
algorithm, we have to map the quality of paths onto the quality of nodes to utilize the bee principle.
Consequently, we formulate the quality of a node, for reaching a destination, as a function of
proportional quality of only those neighbors, which possibly lie in the path toward the destination.
7. We consider data packets as foragers. Once they arrive at a node, they access the information in the
routing tables, stored by bee agents, about the quality of different neighbors of the node for reaching
their destinations. They select the next neighbor toward the destination in a stochastic manner
depending upon its goodness, as a result, not all packets follow the best paths. This will help in
maximizing the system performance though a data packet may not follow the best path, a concept
directly borrowed from a principle of bee behavior: a bee could only maximize her colony’s profit
if she refrains from broadly monitoring the dance floor to identify the single most desirable food [1]
(see Section 21.2).
Now we are in a position to introduce our bee agent model and BeeHive algorithm in the following section.

21.5 The Bee Agent Model

Our bee agent model consists of two types of agents: short distance bee agents and long distance bee agents.
Both agents undertake the same responsibility: exploring the network and evaluating the quality of the
paths that they traverse. However, they only differ in the distance (hops) that they are allowed to take
starting from their launching node. Short distance bee agents collect and disseminate routing information
in the neighborhood (up to a specific number of hops) of their source node while long distance bee
agents collect and disseminate routing information to all nodes of a network. This helps in collecting
the routing information with minimum overhead, both processing and the bandwidth, and as quickly as
possible.
The bee agents that are launched from the same node form an affinity group in which they show
interest in each others information. Once the bee gents of the same group arrive at the same node, but
via different neighbors of the node, they access the routing information, collected by their fellow bee
agents in the group, in the routing table. They will decide to discontinue their exploration of the network
after storing their information in the routing table, if one of their members has already arrived at the
node. The communication model among bee agents is known as a blackboard system [11]. Note that,
in comparison, Ant Colony Optimization (ACO) [12] algorithms utilize the principle of stigmergy [13]
for communication among agents. The agents, using this principle, communicate indirectly through the
shared environment.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 327 — #7

21-328 Handbook of Bioinspired Algorithms and Applications

Informally, the BeeHive algorithm and its main characteristics could be summarized as follows:

1. The network is organized into fixed partitions called foraging regions. A partition results from
particularities of the network topology. Each foraging region has one representative node. Currently
the lowest IP address node in a foraging region is elected as the representative node. If this node
crashes then the next higher IP address node takes over the job.
2. Each node also has a node specific foraging zone that consists of all nodes from whom short distance
bee agents can reach this node.
3. Each nonrepresentative node periodically sends a short distance bee agent, by broadcasting replicas
of it to each neighbor site.
4. When a replica of a particular bee agent arrives at a site it updates the routing information there,
and will be flooded again; however, it will not be sent to the neighbor from where it arrived. This
process continues until the life span of the agent has expired, or if a replica of this bee agent had
been received already at a site, the new replica will be killed there.
5. Representative nodes only launch long distance bee agents that would be received by the neighbors
and propagated as in 4. However, their life span (number of hops) is limited by the long distance
limit.
6. The idea is that each agent while traveling, collects and carries path information, and it leaves,
at each node visited, the trip time estimate for reaching its source node from this node over the
incoming link. Bee agents use priority queues for quick dissemination of routing information.
7. Thus each node maintains current routing information for reaching nodes within its foraging zone
and the representative nodes of foraging regions. This mechanism enables a node to route a data
packet (whose destination is beyond the foraging zone of the given node) along a path toward the
representative node of the foraging region containing the destination node.
8. The next hop for a data packet is selected in a stochastic fashion according to the quality measure
of the neighbors. The motivation for this routing policy is explained in Section 21.4. Note that
the currently employed routing algorithms in Internet always choose a next hop on the shortest
path [14].

Figure 21.2 provides an exemplary working of the flooding algorithm. Short distance bee agents can
travel up to 3 hops in this example. Each replica of the shown bee agent (launched by node 10) is specified
with a different trail to identify its path unambiguously. The numbers on the paths show their costs.
The flooding algorithm is a variant of breadth first search algorithm. Nodes 2, 3, 4, 5, 6, 7, 8, 9, 11
constitute the foraging zone of node 10.
Now we will briefly discuss the estimation model that bee agents utilize to approximate the trip time tis
that a packet will take in reaching its source node s from current node i (ignoring the protocol processing
delays for a packet at node i and s).

qlin
tis ≈ + txin + pdin + tns , (21.1)
bin

where qlin is the size of the queue (in bits) for neighbor n at node i, bin is the bandwidth of the link
between node i and neighbor n, txin and pdin are transmission and propagation delays, respectively, of the
link between node i and neighbor n, and tns is trip time from n to s. Bandwidth and propagation delays
of all links of a node are approximated by transmitting hello packets.

21.5.1 Algorithm Design of BeeHive

In BeeHive, each node i maintains three types of routing tables: Intra-Foraging Zone (IFZ), Inter-Foraging
Region (IFR), and Foraging Region Membership (FRM). IFZ routing table Ri is organized as a vector of
size |D(i)| × (|N (i)|), where D(i) is the set of destinations in foraging zone of node i and N (i) is the set of
neighbors of i. Each entry Pjk is a pair of queuing delay and propagation delay (qjk , pjk ) that a packet will

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 328 — #8

BeeHive: New Ideas for Developing Routing Algorithms 21-329

Rep node
11
1

1 8
7
1 1

10
1
5
1
4 9
1
2
6
1

1 Foraging region 1

0
3
Representative 1
node
1
1
1 2

Foraging region 0

FIGURE 21.2 Bee agents flooding algorithm.

TABLE 21.1 IFZ Routing Table

Ri D1 (i) D2 (i) ... Dd (i)

N1 (i) ( p11 , q11 ) ( p12 , q12 ) ··· ( p1d , q1d )

.. .. .. .. ..
. . . . .
Nn (i) ( pn1 , qn1 ) ( pn2 , qn2 ) ··· ( pnd , qnd )

experience in reaching destination k via neighbor j. Table 21.1 shows an example of Ri . In the IFR routing
table, the queuing delay and propagation delay values for reaching the representative node of each foraging
region through the neighbors of a node are stored. The structure of the IFR routing table is similar to
the one shown in Table 21.1 where destination is replaced by a pair of (representative, region). The FRM
routing table provides the mapping of known destinations onto a foraging region. In this way we eliminate
the need to maintain O(N × D) (where D is total number of nodes in a network) entries in a routing table
as done by AntNet and save a considerable amount of router memory needed to store this routing table.

Goodness of a Neighbor: The goodness of a neighbor j of node l (l has n neighbors) for reaching
a destination d is gjd and defined as follows:

(1/pjd )(e−qjd /pjd ) + (1/qjd )(1 − e−qjd /pjd )

gjd =
n −qkd /pkd ) + (1/q )(1 − eqkd /pkd ))
. (21.2)
k=1 ((1/pkd )(e kd

The fundamental motivation behind Definition 21.2 is to approximate the behavior of the real network.
When the network is experiencing a heavy traffic load then the queuing delay plays the primary role in
the delay of a link. In this case it is trivial to say that qjd pjd and we could see from Equation (21.2)


that gjd ≈ (1/qjd )/ nk=1 1/qkd . When the network is experiencing low traffic then the propagation
delay plays an important role in defining the delay of a link. As qjd
pjd , from Equation (21.2) we get

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 329 — #9

21-330 Handbook of Bioinspired Algorithms and Applications

gjd ≈ (1/pjd )/ nk=1 1/pkd . We use stochastic sampling with replacement [15] for selecting a neighbor.
This principle ensures that a neighbor j with goodness gjd will be selected as the next hop with at least the


probability gjd / nk=1 gkd . Algorithm 21.1 provides the pseudo-code of BeeHive.

Algorithm 21.1 (BeeHive)

t := current_time, tend := time to end simulation

// Short_Limit := 7, Long_Limit := 40, Bee_Generation_Interval := 1
// i = current node, d=destination node, s=source node
// n = successor node of i, p=predecessor node of i
// z = Representative node of the foraging region containing s
// w = Representative node of the foraging region containing d
// q is queuing delay estimate of a bee agent from node p to s
// p is propagation delay estimate of a bee agent from node p to s
t := Bee_Generation_Interval, h := hello packet generation interval
bip := estimated_link_band_width_to_neighbor_p
pip := estimated_propagation_delay_to_neighbor_p
hi := hop limit for bees of i, lip := size_normal_queue_i_to_p (bits)
foreach Node // concurrent activity over the network
while (t ≤ tend )
if (t mod t = 0)
if (i is representative node of the foraging region)
set hi := Long_Limit, bi is long distance bee agent
else
set hi := Short_Limit, bi is short distance bee agent
endif
launch a bee bi to all neighbors of i
endif
foreach bee bs received at i from p
if (bs was launched by i or its hop limit reached)
kill bee bs
elseif (bs is inside foraging zone of node s)
l
q := q + bipip and p := p + pip
update IFZ routing
table entries qps = q and pps

=p
update q (q := (qks × gks )) and p (p := (pks × gks ))
k∈N (i) k∈N (i)
else
lip
q := q + and p := p + pip
bip

table entries qpz = q and ppz

update IFR routing =p
update q (q := (qkz × gkz )) and p (p := (pkz × gkz ))
k∈N (i) k∈N (i)
endif
if (bs already visited node i)
kill bee bs
else
use priority queues to forward bs to all neighbors of i except p
endif
endfor
foreach data packet dsd received at i from p
if (node d is within foraging zone of node i)
consult IFZ routing table of node i to find delays to node d
calculate goodness of all neighbors for reaching d using equation 21.2

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 330 — #10

BeeHive: New Ideas for Developing Routing Algorithms 21-331

else
consult FRM routing table of node i to find node w
consult IFR routing table of node i to find delays to node w
calculate goodness of all neighbors for reaching w using equation 21.2
endif
probabilistically select a neighbor n (n = p) as per goodness
enqueue data packet dsd in normal queue for neighbor n
endfor
if (t mod h = 0)
send a hello packet to all neighbors
if (time out before a response from neighbor) (4th time)
neighbor is down
update the routing table and launch bees to inform other nodes
endif
endif
endwhile
endfor

21.6 Routing Algorithms Used for Comparison

The focus of our research is on dynamic algorithms, therefore we now provide an overview of two
such algorithms, AntNet, which is inspired from the pheromone laying behavior of ants, and Distributed
Genetic Algorithm (DGA), which uses an evolutionary algorithm for routing. However, we will also briefly
introduce Open Shortest Path First (OSPF), a classic static routing algorithm. The reader can find a
detailed description of the algorithm in Reference 16 and its complete implementation in Reference 17.
We use the algorithm in our comparative simulation for the sake of comprehensiveness.

21.6.1 AntNet: An Algorithm Inspired from Ant Behavior

AntNet was proposed by Di Caro and Dorigo [18]. In AntNet the network state is monitored through
two ant agents: Forward_Ant and Backward_Ant. A Forward_Ant agent is launched at regular intervals
from a source to a certain destination. The authors proposed a model in Reference 19 that enables a
Forward_Ant agent to estimate queuing delay without waiting inside data packet queues. Forward_Ant
agent is equipped with a stack memory on which the address and entrance time of each node on its path are
pushed. Once the Forward_Ant agent reaches its destination it creates a Backward_Ant agent and transfers
all information to it. Backward_Ant visits the same nodes as Forward_Ant in reverse order and modifies
the entries in the routing tables based on the trip time from the nodes to the destination. At each node the
average trip time, the best trip time, and the variance of the trip times for each destination are saved. The
trip time values are calculated by taking the difference of entrance times of two subsequent nodes pushed
onto the stack. Backward_Ant agent uses the system priority queues so that it quickly disseminates the
information to the nodes. The interested reader may find more details in References 18 and 19. Later on
the authors of Reference 20 made significant improvements in the routing table initialization algorithm
of AntNet, bounded the number of Forward_Ant agents during congestion, and proposed a mechanism
to handle routing table entries at the neighbors of crashed routers.

21.6.2 Distributed Genetic Algorithm

The authors of Reference 21 showed that the information needed by AntNet for each destination is difficult
to obtain in real networks. Their idea of global information is that there is an entry in the routing table
for each destination. This shortcoming motivated the authors to propose in Reference 22, an evolutionary
routing algorithm, DGA, that eliminates the need for having an entry for each destination node in the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 331 — #11

21-332 Handbook of Bioinspired Algorithms and Applications

routing table. In this algorithm ants are asked to traverse a set of n nodes in a particular order, known as a
chromosome. Once an agent visits the nth node it is then converted into a backward agent that returns to
its source node. In contrast to AntNet, the backward agents only modify the routing tables at the source
node. The source node also measures the fitness of this agent based on the trip time value, and then it
generates a new population using single point cross over. New agents enter the network and evaluate the
assigned paths. The routing table stores the agents’ IDs, their fitness values and trip times to the visited
nodes. Routing of a data packet is done through the path that has the shortest trip time to the destination.
If no entry for a particular destination is found then a data packet is routed with the help of an agent that
has the maximum fitness value. DGA was designed assuming that the routers could crash during network
operations. The interested reader will find more details in Reference 22.
Note that in contrast to above-mentioned algorithms, bee agents need not be equipped with a stack to
perform their duties. Moreover, our agent model requires only forward moving agents and they utilize
an estimation model to calculate the trip time from their source to a given node. This model eliminates
the need for global clock synchronization among routers, and it is expected that for very large networks
routing information could be disseminated quickly with a small overhead as compared with AntNet.
Our agent model does not require to store the average trip time, the variance of trip times, and the
best trip time for each destination at a node to determine the goodness of a neighbor for a particular
destination. Last but not the least, BeeHive works with a significantly smaller routing table as compared
with AntNet.

21.6.3 Open Shortest Path First

Open Shortest Path First is currently the state-of-the-art routing algorithm employed as an Interior
Gateway Protocol (IGP) in Internet [16]. OSPF stores the entire topology of a network in a weighted
directed graph, in which each edge corresponds to a link and each node corresponds to a router. The
cost of a link is a function of the propagation delay of the link. However, the network administrators are
also allowed to change these costs on the basis of their on-field knowledge about network traffic loads. In
our implementation we simply take the fixed propagation costs and do not allow the costs to be changed
by the network administrators. Such approach was also taken by the authors of AntNet. Each node, in
OSPF algorithm, measures the costs of the links to its neighbors. It then encapsulates the addresses of
its neighbors and the costs of the links to the neighbors in a link state packet and then broadcasts it to
all the neighbors. The neighbors in-turn send the link state packet to their neighbors and so on until all
the nodes in the network get the packet. Each node builds the network topology in a distributed fashion.
Finally, each node builds a shortest path tree to all destinations by considering itself as a root. The shortest
paths from the root to all nodes are calculated using the deterministic Dijkstra Algorithm [23]. The next
hop on the shortest path to each destination is stored in a routing table.

21.7 Simulation Environment for BeeHive

In order to evaluate our algorithm BeeHive in comparison with AntNet, DGA, and OSPF, we implemented
all of them in the OMNeT++ simulator [24]. For AntNet and DGA we used the same parameters that
were reported by the authors in References 18 and 22, respectively. The network instance that we used
in our simulation framework is the Japanese Internet Backbone (NTTNet) (see Figure 21.3). It is a
57 node, 162 bidirectional links network. The link bandwidth is 6 Mbits/sec and propagation delay is
from 2 to 5 msec. Traffic is defined in terms of open sessions between two different nodes. Each session
is characterized completely by sessionSize (2 Mbits), inter-arrival time between two sessions, source,
destination, and packet inter-arrival time during a session. The size of data packet is 512 bytes, the size of
a bee agent is 48 bytes. The queue size is limited to 500 Kbytes in all experiments. To inject dynamically
changing data traffic patterns we have defined two states: uniform and weighted. Each state lasts 10 sec
and then a state transition to another state occurs. In Uniform state (U) a destination is selected from a

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 332 — #12

BeeHive: New Ideas for Developing Routing Algorithms 21-333

54
52 55

56
Node 40 51
53
48
47 50

44 49
45

26 34 40 46
13 18 19 32
10 43
5 14 22 21 20 25 273036 35 39 41
9 11 29
6 15 42
3 24 38
7
2 1 4 12
17 16 23
28 31 33 37

0 8
Node 21
Node 4 (hot spot)

FIGURE 21.3 Japanese Backbone NTTNet.

uniform distribution. While, in Weighted state (W), a destination selected in the previous Uniform state
is favored over other destinations. This approach provides a more challenging experimental environment
than the one in which AntNet was evaluated.

21.8 Experimental Findings

Now we will report our results obtained from the extensive simulations. MSIA is the mean of session inter-
arrival times and MPIA is the mean of packet inter-arrival times during a session. The session inter-arrival
and packet inter-arrival times are taken from negative exponential distributions with MSIA and MPIA,
respectively. All reported values are an average of the values obtained from ten independent runs. The
percentage of deliverable packets that were actually delivered is reported by % Packets Delivered (on top
of bars). By deliverable packet we mean a packet whose destination router is up.

Saturating loads: The purpose of the experiments was to study the behavior of the algorithms by
gradually increasing the traffic load, through decreasing MSIA from 4.7 to 1.7 sec. MPIA is 0.005 sec
during these experiments. Figure 21.4 shows the average throughput and 90th percentile of the
packet delay distribution. It is obvious from Figure 21.4 that BeeHive delivered approximately the
same number of data packets as that of AntNet but with lesser packet delay. Both OSFP and DGA
are unable to cope with a saturated load yet the performance of DGA is the poorest.
Size of foraging zones: Next, we analyzed the effect of the size of a foraging zone in which a bee agent
updates the routing table. We report the results for sizes of 7, 10, 15, and 20 hops in Figure 21.5.
Figure 21.5 shows that increasing the size of foraging zone after 10 does not bring significant
performance gains. This shows the power of BeeHive that it converges to an optimum solution
with a size of just 7 hops.
Size of the routing table: The fundamental motivation of the foraging zone concept was not only to elim-
inate the requirement for global knowledge but also to reduce memory needed to store a routing
table. BeeHive requires 88, 94, and 104 entries, on the average, in the routing tables for foraging
zone sizes of 7, 10, and 20 hops, respectively. OSFP needs just 57 entries while AntNet needs 162
entries on the average. Hence BeeHive achieves similar performance as that of AntNet but size of
the routing table is of the order of OSPF.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 333 — #13

21-334 Handbook of Bioinspired Algorithms and Applications

(a) 70 (b) 8
OSPF
OSPF 58
85 AntNet

90th percentile of packet delays (sec)

82 AntNet 7
60 DGA
DGA
6 BeeHive 48
BeeHive
50 68 39
Throughput (Mbps)

97 99
5
40 81
4
30 39 85
99 3
92 97 68 82
48
20 81
58 2
97
10 1 92 97
99
99
0 0
1.7 2.7 4.7 1.7 2.7 4.7
MSIA MSIA

FIGURE 21.4 Behavior under saturating traffic loads.

(a) 70 (b) 8
BeeHive(7) BeeHive(7)
90th percentile of packet delays (sec)

60 83 82 81 82 BeeHive(10) 7 BeeHive(10)
BeeHive(15) BeeHive(15)
BeeHive(20) 6 BeeHive(20)
50
Throughput (Mbps)

99 99 99 99
5
40
4
30 99 99 99 99
3 83 82 81 82
20
2
10 1
99 99 99 99
99 99 99 99
0 0
1.7 2.7 4.7 1.7 2.7 4.7
MSIA MSIA

FIGURE 21.5 Effect of foraging zones sizes.

Hot spot : The purpose of this experiment is to study the effect of transient overloads in a network. We
selected node 4 (see Figure 21.3) as a hot spot. The hot spot was active from 500 to 1000 sec and
all nodes sent data to node 4 with MPIA = 0.05 sec. This transient overload was superimposed
on a normal load of MSIA = 2.7 sec and MPIA = 0.3 sec. Figure 21.6 shows that both BeeHive
and AntNet are able to cope with the transient overload, however the average packet delay for
BeeHive is less than 100 msec as compared with 500 msec for AntNet. Again DGA shows the
poorest performance.
Router crash: The purpose of this experiment was to analyze the fault-tolerant behavior of BeeHive so
we took MSIA = 4.7 sec and MPIA = 0.005 sec to ensure that no packets are dropped because
of the congestion. We simulated a scenario in which Router 21 crashed at 300 sec, and Router 40
crashed at 500 sec and then both were repaired at 800 sec. Figure 21.7 shows the results. BeeHive
is able to deliver 97% of deliverable packets as compared with 89% by AntNet. Observe that from
300 to 500 sec (just Router 21 is down), BeeHive has a superior throughput and lesser packet
delay but once Router 40 crashes, the packet delay of BeeHive increases because of higher load at
Router 43. From Figure 21.3 it is obvious that the only path to the upper part of the network is via
Router 43 once Router 40 crashed. Since BeeHive is able to deliver more packets the queue length
at Router 43 increased and this led to relatively poorer packet delay as compared with AntNet.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 334 — #14

BeeHive: New Ideas for Developing Routing Algorithms 21-335

(a) 70
OSPF (81.39%)
AntNet (96.85%)
DGA (48.08%)
60 BeeHive (99.9%)
BeeHive
50

OSPF AntNet
40

30
DGA

20

10

0
0 100 200 300 400 500 600 700 800 900 1000

(b) 2
OSPF (81.39%)
1.8 AntNet (96.85%)
DGA DGA (48.08%)
1.6 BeeHive (99.9%)

1.4

1.2

1
OSPF
0.8

0.6

0.4

0.2 BeeHive
AntNet
0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 21.6 Node 4 acted as hot spot from 500 to 1000 sec.

In addition, observe that Router 21 is critical but in case of its crash still multiple paths exist to the
middle and upper part of the topology via 15, 18, 19, 20, 24.
Overhead of BeeHive: We first define the terms routing overhead and suboptimal overhead.

Definition 21.2 (Routing overhead). Routing overhead is defined as the ratio between the bandwidth
occupied by the routing packets and the total available network bandwidth [18].

The metric suboptimal overhead was introduced in Reference 25 in the context of MANETS but we believe
that it is equally relevant in fixed networks as well.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 335 — #15

21-336 Handbook of Bioinspired Algorithms and Applications

(a) 40
AntNet (89.43%)
DGA (46.34%)
BeeHive (97.45%)

30
BeeHive

20

AntNet

DGA
10

0
0 100 200 300 400 500 600 700 800 900 1000

(b)
AntNet (89.43%)
1.4
DGA (46.34%)
BeeHive (97.45%)
1.2

1 DGA

0.8

0.6

0.4
AntNet

0.2
BeeHive
0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 21.7 Router 21 is down at 300 and Router 40 at 500 and both repaired at 800.

Definition 21.3 (Suboptimal overhead). The difference between the bandwidth consumed when trans-
mitting data packets from all the sources to destinations and the bandwidth that would have been
consumed, should the data packets have followed the shortest hop count path. Formally, we could define the
parameter as

n
n
k
i=1 (hi − hosd ) × Lisd
sd
d=1 s=1
So = , s = d, (21.3)
Bt

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 336 — #16

BeeHive: New Ideas for Developing Routing Algorithms 21-337

(a) 2 (b) OSPF

OSPF
20 AntNet

Suboptimal routing ovehead (%)

AntNet
DGA DGA
Routing overhead (%)

1.5 BeeHive
BeeHive 15

1
10

0.5 5

0 0
4.7 2.7 1.7 4.7 2.7 1.7
MSIA MSIA

FIGURE 21.8 Saturated loads.

where n is total number of nodes in the network, k is total number of packets generated, hisd number of hops
that packet i took to reach its destination d from its source s, hosd minimum hops between node s and node d,
Lisd is length of packet i from source s to destination d, and Bt is the total bandwidth of the network.

We report this parameter because it implicitly includes the overhead of loops.

Figure 21.8 shows the control overhead and suboptimal overhead of the algorithms. It is quite interesting
to note that the suboptimal overhead is much higher than the control overhead (note that the x- and y-
axes have different scales on the two figures). OSPF has the smallest suboptimal overhead but then it
delivers less packets as well. The routing overhead of DGA decreases with an increase in the load and
vice versa. This happens due to the genetic algorithm. Remember that the next generation of the agents
are launched once four agents from the previous generation are received. Under low load, the return times
for the agents are smaller, as a result, the agents are launched at a higher rate and vice versa. Since in
AntNet, Forward_Ant agents use the same queues that data packets also use, therefore more ants were
dropped under increased network load and this explains the decrease in routing overhead behavior of
the algorithm. It is evident from Figure 21.8 that BeeHive has significantly smaller routing overhead and
suboptimal overhead as compared to AntNet and DGA.

21.9 Conclusion
A honey bee colony is able to optimize its stockpiles of nectar, pollen, and water through an intelligent
allocation of labor among different specialists, which communicate with each other using a sophisticated
communication protocol that consists of signals and cues, in continuously changing internal and external
environments. The dance language and foraging behavior of honey bees inspired us to develop a fault-
tolerant, adaptive, and robust routing protocol. The algorithm does not need any global information
such as the structure of the topology and cost of links among routers, rather it works with the local
information that a short distance bee agent collects in a foraging zone. It works without the need of global
clock synchronization, which not only simplifies its installation on real routers but also enhances fault
tolerance. In contrast to AntNet our algorithm utilizes only forward moving bee agents that help in
disseminating the state of the network to the routers in real time. The bee agents take less than 1% of the
available bandwidth but provide significant enhancements in throughput and packet delay.
We implemented two state-of-the-art adaptive algorithms (AntNet and DGA) for the OMNeT++
simulator and then compared our BeeHive algorithm with them. Through extensive simulations represent-
ing dynamically changing operating network environments we have demonstrated that BeeHive achieves
a better or similar performance as compared with AntNet. However, this enhancement in performance is
achieved with a routing table whose size is of the order of OSPF.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 337 — #17

21-338 Handbook of Bioinspired Algorithms and Applications

We should emphasize that the foraging model described in References 6 and 7 could be used
for developing any multi-objective optimization algorithm for dynamic environments. The impor-
tant concepts of BeeHive, such as bee agent propagation algorithm and bee agent communication
paradigm, could be applied to any optimization problem that could be represented in a graph. We
believe that BeeHive concept will motivate the researchers to develop Bee Colony Optimization (BCO)
algorithms.

References
[1] T.D. Seeley. The Wisdom of the Hive. Harvard University Press, London, 1995.
[2] K. von Frisch. Tanzsprache und Orientierung der Bienen. Springer-Verlag, Heidelberg, 1965.
[3] K. von Frisch. The Dance Language and Orientation of Bees. Harvard University Press, Cambridge,
1967.
[4] H.A. Simon. Administrative Behavior: A Study of Decision-Making Processes in Administrative
Organization. Free Press, New York, 1976.
[5] T.D. Seeley and W.F. Towne. Collective decision making in honey bees: How colonies choose
among nectar sources. Behavior Ecology and Sociobiology, 12: 277–290, 1991.
[6] D.J.T. Sumpter. From bee to society: An agent-based investigation of honey bee colonies.
Ph.D. thesis, The University of Manchester, UK, 2000.
[7] L.P. Kaelbling, M.L. Littman, and A.W. Moore. Reinforcement learning: A survey. Journal of
Artificial Intelligence, 4: 237–285, 1996.
[8] H.F. Wedde, M. Farooq, and Y. Zhang. Beehive: An efficient fault-tolerant routing algorithm
inspired by honey bee behavior. In Proceedings of ANTS Workshop, Lecture Notes in Computer
Science 3172, Springer-Verlag, September 2004.
[9] H.F. Wedde and M. Farooq et al. BeeHive — An Energy-Aware Scheduling and Routing Framework.
Technical report, Project Group 439, LSIII, School of Computer Science, University of Dortmund,
2004.
[10] Y. Zhang. Design and implementation of bee agents based algorithm for routing in high speed,
adaptive and fault-tolerant networks. Master thesis, LSIII, The University of Dortmund, Germany,
2003.
[11] P. Nii. The blackboard model of problem solving. AI Magazine, 7: 38–53, 1986.
[12] E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence from Natural to Artificial Systems.
Oxford University Press, Oxford, 1999.
[13] P.P. Grassé. La reconstruction du nid et les coordinations interindividuelles chez bellicositermes
natalensis et cubitermes sp. la théorie de la stigmergie: essai d’interprétation du comportement
des termites constructeurs. Insectes Sociaux, 6: 41–81, 1959.
[14] L.L. Peterson and B.S. Davie. Computer Networks a Systems Approach. Morgan Kaufmann
Publishers, San Francisco, CA, 2000.
[15] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
[16] J.T. Moy. OSPF Anatomy of an Internet Routing Protcol. Addison-Wesley, Reading, MA, 1998.
[17] J.T. Moy. OSPF Complete Implementation. Addison-Wesley, Reading, MA, 2000.
[18] G. Di Caro and M. Dorigo. AntNet: Distributed stigmergetic control for communication networks.
Journal of Artificial Intelligence, 9: 317–365, 1998.
[19] G. Di Caro and M. Dorigo. Two ant colony algorithms for best-effort routing in datagram networks.
In Proceedings of the Tenth IASTED International Conference on Parallel and Distributed Computing
and Systems (PDCS’98), IASTED/ACTA Press, 1998, pp. 541–546.
[20] B. Barán and R. Sosa. A new approach for antnet routing. In Proceedings of the Ninth International
Conference on Computer, Communications and Networks, 2000.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 338 — #18

BeeHive: New Ideas for Developing Routing Algorithms 21-339

[21] S. Liang, A.N. Zincir-Heywood, and M.I. Heywood. The effect of routing under local informa-
tion using a social insect metaphor. In Proceedings of IEEE Congress on Evolutionary Computing,
May 2002.
[22] S. Liang, A.N. Zincir-Heywood, and M.I. Heywood. Intelligent packets for dynamic network
routing using distributed genetic algorithm. In Proceedings of Genetic and Evolutionary Computa-
tion Conference, GECCO, July 2002.
[23] E.W. Dijkstra. A note on two problems in connection with graphs. Numerical Mathematics,
1: 269–271, 1959.
[24] A. Varga. OMNeT++: Discrete Event Simulation System: User Manual. http://www.omnetpp.org.
[25] C. Santivanez, B. McDonald, I. Stavrakakis, and R. Ramanathan. On the scalability of ad hoc
routing protocols. In Proceedings of IEEE INFOCOM 2002, IEEE, June 2002.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C021” — 2005/8/6 — 14:10 — page 339 — #19

 22
Swarming Agents for
Decentralized
Clustering in
Spatial Data

22.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-341

22.2 A Multiagent Spatial Clustering Algorithm . . . . . . . . . . . . 22-343
The DBSCAN Algorithm • The Reynolds’ Flock Model •
SPARROW: A Flocking Algorithm for Spatial Clustering
22.3 The SPARROW-SNN Algorithm . . . . . . . . . . . . . . . . . . . . . . . 22-346
The SNN Clustering Algorithm • The SPARROW-SNN
Implementation
22.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-349
SPARROW Results • SPARROW-SNN Results
22.5 Entropy Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-354
Gianluigi Folino 22.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-356
Agostino Forestiero Acknowledgment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-357
Giandomenico Spezzano References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-357

22.1 Introduction
In recent years several approaches to knowledge discovery and data mining, and in particular, clustering,
have been developed, but only a few of them are designed using a decentralized approach. Clustering
data is the process of grouping similar objects according to their distance, connectivity, or relative density
in space [1]. There are a large number of algorithms for discovering natural clusters, if they exist, in a
dataset, but they are usually studied as a centralized problem. These algorithms can be classified into
partitioning methods [2], hierarchical methods [3], density-based methods [4], and grid-based methods
[5]. Han et al.’s paper [6] is a good introduction to this subject. Many of these algorithms work on
data contained in a file or database. In general, clustering algorithms focus on creating good compact
representation of clusters and appropriate distance functions between data points. For this purpose, they
generally need to be given one or two parameters by a user that indicate the types of clusters expected.
Since they use a central representation where each point can be compared with each other point or cluster
representation, points are never placed in a cluster with largely different members. Centralized clustering

22-341

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 341 — #1

22-342 Handbook of Bioinspired Algorithms and Applications

is problematic if we have large data to explore or data is widely distributed. Parallel and distributed
computing is expected to relieve current mining methods from the sequential bottleneck, providing the
ability to scale to massive datasets, and improving the response time. Achieving good performance on
today’s high performance systems is a nontrivial task. The main challenges include synchronization and
communication minimization, workload balancing, finding good data decomposition, etc. Some existing
centralized clustering algorithms have been parallelized and the results have been encouraging. Centralized
schemes require high level of connectivity, impose a substantial computational burden, are typically more
sensitive to failures than decentralized schemes, and are not scalable, which is a property that distributed
computing systems are required to have.
Recently, other algorithms based on biological models [7–10] have been introduced to solve the clus-
tering problem in a decentralized fashion. These algorithms are characterized by the interaction of a large
number of simple agents that sense and change their environment locally. Furthermore, they exhibit
complex, emergent behavior that is robust with respect to the failure of individual agents. Ant colonies,
flocks of birds, termites, swarms of bees, etc., are agent-based insect models that exhibit a collective intel-
ligent behavior (SWARM intelligence, SI) [11] that may be used to define new algorithms of clustering.
The use of new SI-based techniques for data clustering is an emerging new area of research that has
attracted many investigators in the last years. SI models have many features in common with evolutionary
algorithms (EAs). Like EA, SI models are population-based. The system is initialized with a population of
individuals (i.e., potential solutions). These individuals are then manipulated over many iteration steps
by mimicking the social behavior of insects or animals, in an effort to find the optima in the problem
space. Unlike EAs, SI models do not explicitly use evolutionary operators, such as crossover and muta-
tion. A potential solution simply “flies” through the search space by modifying itself according to its past
experience and its relationship with other individuals in the population and the environment. In these
models, the emergent collective behavior is the outcome of a process of self-organization, in which insects
are engaged through their repeated actions and interaction with their evolving environment. Intelligent
behavior frequently arises through indirect communication between the agents using the principle of
stigmergy [12]. This mechanism is a powerful principle of cooperation in insect societies. According to
this principle, an agent deposits something in the environment that makes no direct contribution to the
task being undertaken but is used to influence the subsequent behavior that is task related. The advantages
of SI are twofold. First, it offers intrinsically distributed algorithms that can use parallel computation
quite easily. Second, these algorithms show a high level of robustness to change by allowing the solution to
dynamically adapt itself to global changes by letting the agents self-adapt to the associated local changes.
Social insects provide a new paradigm for developing decentralized clustering algorithms.
In this chapter, we present a method for decentralized clustering based on an adaptive flocking algorithm
proposed by Macgill [13]. We consider clustering as a search problem in a multiagent system in which
individual agents have the goal of finding specific elements in the search space, represented by a large
dataset of tuples, by walking efficiently through this space. The method takes advantage of the parallel
search mechanism a flock implies, by which if a member of a flock finds an area of interest; the mechanics
of the flock will draw other members to scan that area in more detail. The algorithm selects interesting
subsets of tuples without inspecting the whole search space guaranteeing a fast placing of points correctly in
the clusters. We have applied this strategy as a data reduction technique to perform efficiently approximate
clustering [14]. In the algorithm, each agent can use hierarchical, partitioned, density-based, and grid-
based clustering methods to discover if a tuple belongs to a cluster.
To illustrate this method we present two algorithms: SPARROW (SPAtial clusteRing algoRithm thrOugh
sWarm intelligence) and SPARROW-SNN. SPARROW combines the flocking algorithm with the density-
based DBSCAN algorithm [15]. SPARROW-SNN combines the flocking algorithm with a shared nearest-
neighbor (SNN) cluster algorithm [16] to discover clusters with differing sizes, shapes, and densities in
noise, high dimensional data. We have built a SWARM [17] simulation of both algorithms to investigate
the interaction of the parameters that characterize them. First, experiments showed encouraging results
and a better performance of both algorithms in comparison with the linear randomized search using an
entropy-based model.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 342 — #2

Swarming Agents for Decentralized Clustering in Spatial Data 22-343

The rest of this chapter is organized as follows: Section 22.2 presents the SPARROW clustering algorithm,
first, introducing the heuristics of the DBSCAN algorithm, used by each agent to discover clusters, and
the flocking algorithm for the interaction among the agents. Section 22.3 describes the centralized SNN
clustering algorithm and how it is combined with the flocking algorithm to produce the SPARROW-SNN
algorithm. Section 22.4 discusses the results obtained. Section 22.5 presents an entropy-based model to
theoretically explain the behavior of the algorithm and Section 22.6 draws some conclusions.

22.2 A Multiagent Spatial Clustering Algorithm

In this section, we will present the SPARROW algorithm that combines the stochastic search of an adaptive
flocking with the DBSCAN heuristics for discovering clusters in parallel. SPARROW replaces the DBSCAN
serial procedure for clusters identification with a multiagent stochastic search that has the advantage of
being easily implementable on parallel computers and that is robust compared with the failure of individual
agents. We will first introduce the DBSCAN algorithm and then will present the Reynolds’ flock model
that describes the standard movement rules of birds. Finally, we will illustrate the details of the SPARROW
behavioral rules of the Swarming agents (SAs) that move through the spatial data looking for clusters and
communicating their findings to each other.

22.2.1 The DBSCAN Algorithm

One of the most popular spatial clustering algorithms is DBSCAN, which is a density-based spatial
clustering algorithm. A complete description of the algorithm and its theoretical basis is presented in
the paper by Ester et al. [15]. In the following, we briefly present the main principles of DBSCAN. The
algorithm is based on the idea that all points of a dataset can be regrouped into two classes: clusters and
noise. Clusters are defined as a set of dense connected regions with a given radius (Eps) and containing at
least a minimum number (MinPts) of points. Data are regarded as noise if the number of points contained
in a region falls below a specified threshold. The two parameters, Eps and MinPts, must be specified by the
user and allowed to control the density of the cluster that must be retrieved. The algorithm defines two
different kinds of points in a clustering: core point s and non-core points. A core point is a point with at least
MinPts number of points in an Eps neighborhood of the point. The non-core points in turn are either
border points if not core points but are density reachable from another core point or noise points if not core
points and are not density reachable from other points. To find the clusters in a dataset, DBSCAN starts
from an arbitrary point and retrieves all points with the same density reachable from that point using Eps
and MinPts as controlling parameters. A point p is density reachable from a point q if the two points are
connected by a chain of points such that each point has a minimal number of data points, including the
next point in the chain, within a fixed radius. If the point is a core point, then the procedure yields a cluster.
If the point is on the border, then DBSCAN goes on to the next point in the database and the point is
assigned to the noise. DBSCAN builds clusters in sequence (i.e., one at a time), in the order in which they
are encountered during space traversal. The retrieval of the density of a cluster is performed by successive
spatial queries. Such queries are supported efficiently by spatial access methods such as R ∗ -trees.

22.2.2 The Reynolds’ Flock Model

The flocking algorithm was originally proposed by Reynolds [18] as a method for mimicking the flocking
behavior of birds on a computer both for animation and as a way to study emergent behavior. Flocking is
an example of emergent collective behavior: there is no leader, that is, no global control. Flocking behavior
emerges from the local interactions. In the flock algorithm, each agent has direct access to the geometric
description of the whole scene, but reacts only to flock mates within a certain small radius. The basic
flocking model consists of three simple steering behaviors:

Separation gives an agent the ability to maintain a certain distance from others. This prevents agents
from crowding too closely together, allowing them to scan a wider area.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 343 — #3

22-344 Handbook of Bioinspired Algorithms and Applications

Cohesion gives an agent the ability to cohere (approach and form a group) with other nearby agents.
Steering for cohesion can be computed by finding all agents in the local neighborhood and com-
puting the average position of the nearby agents. The steering force is then applied in the direction
of that average position.
Alignment gives an agent the ability to align with other nearby characters. Steering for alignment can
be computed by finding all agents in the local neighborhood and averaging together the “heading”
vectors of the nearby agents.

22.2.3 SPARROW: A Flocking Algorithm for Spatial Clustering

SPARROW is a multiagent adaptive algorithm able to discover clusters in parallel. It uses a modified
version of standard flocking algorithm that incorporates the capacity for learning that can be found in
many social insects. In the algorithm, the agents are transformed into hunters with a foraging behavior
that allow them to explore the spatial data while searching for clusters.
SPARROW starts with a fixed number of agents that occupy a randomly generated position. Each
agent moves around the spatial data testing the neighborhood of each location in order to verify if the
point can be identified as a core point. In such a case, a temporary label is assigned to all the points of
the neighborhood of the core point. The labels are updated concurrently as multiple clusters take shape.
Contiguous points belonging to the same cluster take the label corresponding to the smallest label in the
group of contiguous points. Each agent follows the rules of movement described in Reynolds’ model. In
addition, our model considers four different kinds of agents, classified based on the density of data in their
neighborhood. These different kinds are characterized by a different color: red, revealing a high density
of interesting patterns in the data, green, a medium one, yellow, a low one, and white, indicating a total
absence of patterns. The main idea behind our approach is to take advantage of the colored agent in order
to explore more accurately the most interesting regions (signaled by the red agents) and avoid the ones
without clusters (signaled by the white agents). Red and white agents stop moving in order to signal this
type of regions to the others, while green and yellow ones fly to find more dense clusters. Indeed, each
flying agent computes its heading by taking the weighted average of alignment, separation, and cohesion
(as illustrated in Figure 22.1). The following are the main features that make our model different from
Reynolds’ model:

• Alignment and cohesion do not consider yellow agents, since they move in a not very attractive
zone.
• Cohesion is the resultant of the heading toward the average position of the green flockmates
(centroid), of the attraction toward reds, and of the repulsion from whites, as illustrated in
Figure 22.1.
• A separation distance is maintained from all the boids, apart from their color.

Ignore it White
Yellow

Centroid Green
Red

Resultant

FIGURE 22.1 Computing the direction of an agent in SPARROW.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 344 — #4

Swarming Agents for Decentralized Clustering in Spatial Data 22-345

for i =1..MaxGenerations
foreach agent (yellow, green)
age =age +1;
if (age >Max_Life)
generate_new_agent( ); die( );
endif
if (not visited (current_point))
density = compute_local_density( );
mycolor = color_agent( );
endif
end foreach

foreach agent (yellow, green)

dir =compute_dir( );
end foreach
foreach agent (all)
switch (mycolor){
case yellow, green:
move(dir, speed(mycolor)); break;
case white:
stop( );generate_new_agent( ); break;
case red:
stop( ); merge( ); generate_new_close_agent( );
break;
}
end foreach

FIGURE 22.2 Pseudo code of the SPARROW algorithm.

The SPARROW algorithm consists of a setup phase and a running phase shown in Figure 22.2. During
the setup phase agents are created, data are loaded, and some general settings are made. In the running
phase each agent repeats four distinct procedures for a fixed number of times (MaxGenerations). We
use the foreach statement to indicate that the rules are executed in parallel by the agents whose color is
specified in the argument. The mycolor procedure chooses the color and the speed of the agents with
regard to the local density of the points of clusters in the data. It is based on the same parameters used in
the DBSCAN algorithm: MinPts, the minimum number of points to form a cluster and Eps, the maximum
distance that the agents can look at. In practice, the agent computes the local density (density) in a circular
neighborhood (with a radius determined by its limited sight, i.e., Eps) and then it chooses the color
according to the following simple rules:

density > MinPts → mycolor = red (speed=0)

MinPts/4 < density <= MinPts → mycolor = green (speed=1)

0 < density <= MinPts/4 → mycolor = yellow (speed=2)

density = 0 → mycolor = white (speed=0)

In the running phase, yellow and green agents will compute their direction, according to the rules
previously described, and will move following this direction with the speed corresponding to their color.
In case the agent falls in the same position of that of an older one it will die and will be regenerated in
another place.
Agents will move toward the computed destination with a speed depending on their color: green agents
more slower than yellow agents since they will explore denser zones of clusters. Green and yellow agents
have a variable speed, with a common minimum and maximum for all agents. An agent will speedup to
leave an empty or uninteresting region whereas will slowdown to investigate an interesting region more
carefully.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 345 — #5

22-346 Handbook of Bioinspired Algorithms and Applications

FIGURE 22.3 The cage effect.

The variable speed introduces an adaptive behavior in the algorithm. In fact, agents adapt their move-
ment and change their behavior (speed) based on their previous experience represented from the red and
white agents. Red and white agents will stop signaling to the others the interesting and desert regions.
Note that for any agent that has become red or white, a new agent will be generated in order to maintain
a constant number of agents exploring the data. In the first case, the new agent will be generated in a
close random point, since the zone is considered interesting, while in the latter it will be generated in a
random point over all the space. Next, red agents will run the merge procedure, which will merge the
neighboring clusters. The merging phase considers two different cases: when we have never visited points
in the circular neighborhood and when we have points belonging to different clusters. In the first case, the
points will be labeled and will constitute a new cluster; in the second case, all the points will be merged
into the same cluster, that is, they will get the label of the cluster discovered first. The first part of code
executed by yellow and green agents was added to avoid a “cage effect ” (see Figure 22.3), which occurred
during the first simulations; in fact, some agents could remain trapped inside regions surrounded by red
or white agents and would have no way to go out, wasting useful resources for the exploration. Therefore,
a limit was imposed on their life. When their age exceeded a determined value (Max_Life) they were made
to die and were regenerated in a new randomly chosen position of the space.
SPARROW suffers the same limitation as DBSCAN, that is, cannot cope with clusters of different
densities. The new algorithm SPARROW-SNN, introduced in Section 22.3, is more general and overcomes
these drawbacks. It can be used to discover clusters with differing sizes, shapes, and densities in noise data.

22.3 The SPARROW-SNN Algorithm

In this section, we present the SPARROW-SNN multiagent clustering algorithm, which combines the
stochastic search of the SPARROW algorithm with the SNN heuristics for discovering clusters in spatial
data. This approach has a number of nice properties. It can be easily implemented on parallel computers
and is robust compared with the failure of individual agents. It can also be applied to perform efficiently
approximate clustering since the points, that are visited and analyzed by the agents, represent a significant
(in ergodic sense) subset of the entire dataset. The subset reduces the size of the dataset while keeping
the accuracy loss as small as possible. Moreover, SPARROW-SNN can discover clusters with differing
sizes, shapes, and densities in noise data in parallel. The behavior requires the individual members to
first explore the environment, searching for a goal whose position was not known a priori, and then,
after the goal is located, all the flock members should move toward the goal. The same strategy is used
by SPARROW to transform the agents into hunters with a foraging behavior in order to explore the
spatial data while searching for clusters. Clusters are discovered using the heuristics principles of the SNN
clustering algorithm.

22.3.1 The SNN Clustering Algorithm

The SNN is a clustering algorithm developed by Ertöz et al. [16] to discover clusters with differing
sizes, shapes, and densities in noise, high dimensional data. The algorithm extends the nearest-neighbor
nonhierarchical clustering technique by Jarvis and Patrick [19] redefining the similarity between pairs

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 346 — #6

Swarming Agents for Decentralized Clustering in Spatial Data 22-347

of points in terms of how many nearest neighbors the two points share. Using this new definition of
similarity, the algorithm eliminates noise and outliers, identifies representative points, and then builds
clusters around the representative points. These clusters do not contain all the points, rather they represent
relatively uniform group of points.
The SNN algorithm starts performing the Jarvis–Patrick scheme. In the Jarvis–Patrick algorithm, a set
of objects is partitioned into clusters based on the number of shared nearest neighbors. The standard
implementation is constituted by two phases. The first, a preprocessing stage identifies the K nearest
neighbors of each object in the dataset. In the subsequent clustering stage, two objects i and j join the
same cluster if:
• i is one of the K nearest neighbors of j
• j is one of the K nearest neighbors of i
• i and j have at least Kmin of their K nearest neighbors in common
where K and Kmin are used-defined parameters. For each pair of points i and j is defined as a link with an
associate weight. The strength of the link between i and j is defined as:

strength(i, j) = (k + 1 − m)(k + 1 − n) where im = jn .

In this equation, k is the nearest-neighbor list size, m and n are the positions of a shared nearest
neighbor in i and j’s lists. At this point, clusters can be obtained by removing all edges with weights less
than a user-specified threshold and taking all the connected components as clusters. A major drawback
of the Jarvis–Patrick algorithm is that, the threshold needs to be set high enough since two distinct set of
points can be merged into same cluster even if there is only a link across them. On the other hand, if a
high threshold is applied, then a natural cluster will be split into many small clusters due to the variations
in the similarity in the cluster.
Shared nearest-neighbor addresses these problems introducing the following steps:

1. For every node (data point) calculates the total strength of links coming out of the point.
2. Identify representative points by choosing the point that have high density (>core_threshold).
3. Identify noise points by choosing the points that have low density (<noise_threshold) and
remove them.
4. Remove all links between points that have weight smaller than a threshold (merge_threshold).
5. Take connected components of points to form clusters, where every point in a cluster is either a
representative point or is connected to a representative point.

The number of clusters is not given to the algorithm as a parameter. Also, note that not all the points are
clustered.

22.3.2 The SPARROW-SNN Implementation

As the first step, SPARROW-SNN computes for each data element the nearest-neighbor list using a
similarity threshold that reduces the number of data elements to take into consideration. The intro-
duction of the similarity threshold produces variable-length nearest-neighbor lists and therefore i and
j must have at least Pmin of the shorter nearest-neighbor list in common; where Pmin is a user-defined
percentage.
After the nearest-neighbor list is computed, SPARROW-SNN starts a fixed number of agents that will
occupy a randomly generated position. The agents have an attribute that defines their color. From their
initial position, each agent moves around the spatial data testing the neighborhood of each location in
order to verify if the point can be identified as a representative (or core) point.
Figure 22.4 illustrates the behavioral rules of each agent. All agents execute the same set of rules for
a fixed number of times (MaxGenerations). When an agent falls on a data point A, not yet visited, it

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 347 — #7

22-348 Handbook of Bioinspired Algorithms and Applications

for i=1..MaxGenerations
foreach agent (yellow, green)
if (not visited (current_point))
conn =compute_conn( );
if (conn <noise_threshold)
consider the point for the removal from the clustering
endif
endif
mycolor = color_agent( );
end foreach

foreach agent (yellow, green)

dir =compute_dir( );
end foreach
foreach agent (all)
switch (mycolor){
case yellow, green:
move(dir, speed(mycolor)); break;
case white:
stop( ); generate_new_agent( ); break;
case red:
stop( ); merge( ); generate_new_close_agent( );
break;
}
end foreach
end for

FIGURE 22.4 Pseudo code of the SPARROW-SNN algorithm.

computes the connectivity, conn[A], of the point, that is, computes the total number of strong links the
points has, according to the rules of the SNN algorithm. Points having connectivity smaller than a fixed
threshold (noise_threshold) are classified as noise and are considered for removal from clustering. A color
is then assigned to each agent, based on the value of the connectivity computed in the visited point, using
the following procedure (called color_agent() in the pseudo code):

conn > core_threshold → mycolor = red

noise_threshold < conn <= core_threshold → mycolor = green

0 < strength < noise_threshold → mycolor = yellow

strength = 0 → mycolor = white

The colors assigned to the agents are: red, revealing representative points, green, border points, yellow,
noise points, and white, indicating an obstacle (uninteresting region). After the coloration step, the green
and yellow agents, compute their movement observing the positions of all other agents that are at some
fixed distance (dist_max) from their and applying the same rules of SPARROW.
In any case, each red agent (placed on a representative point) will run the merge procedure so that it
will include, in the final cluster, the representative point discovered together with the points that share
a significant (greater than Pmin ) number of neighbors and that are not noise points. The merging phase
considers two different cases: when we have visited none of these points in the neighborhood and when
we have points belonging to different clusters. In the first case, the points will be assigned the same
temporary label and will constitute a new cluster; in the second case, all the points will be merged into
the same cluster, that is, they will get the label corresponding to the smallest one. Therefore, clusters will
be built incrementally.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 348 — #8

Swarming Agents for Decentralized Clustering in Spatial Data 22-349

(a) (b)

GEORGE: 5463 points DS4: 8843 points

FIGURE 22.5 The two datasets used in our experiments.

In addition, in this case, during simulations a “cage effect,” was observed. This part is not reported in
the pseudo code, as it is the same used in the SPARROW algorithm.

22.4 Experimental Results

In Sections 22.4.1 and 22.4.2, we present the experimental results obtained for the two algorithms. Both
algorithms have been implemented using SWARM, a multiagent software platform for the simulation
of complex adaptive systems. In the SWARM system, the basic unit of simulation is the SWARM, a
collection of agents executing a schedule of actions. SWARM provides object-oriented libraries of reusable
components for building models and analyzing, displaying, and controlling experiments on those models.
More information about SWARM can be obtained from Reference 17.

22.4.1 SPARROW Results

We evaluated the accuracy of the solution supplied by SPARROW in comparison with the one of DBSCAN
and the performance of the search strategy of SPARROW in comparison with the standard flocking search
strategy and with the linear randomized search. Furthermore, we evaluated the impact of the number of
agents on foraging for clusters performance. Results are compared with the ones obtained using a publicly
available version of DBSCAN.
For the experiments, we used two synthetic datasets, shown in Figure 22.5(a) and (b). The first dataset,
called GEORGE, consists of 5463 points. The second dataset, called DS4, contains 8843 points. Each point
of the two datasets has two attributes that define the x and y coordinates. Furthermore, both datasets have
a considerable quantity of noise.
Although DBSCAN and SPARROW produce the same results if we examine all points of the dataset, our
experiments show that SPARROW can obtain, with an average accuracy about 93% on GEORGE dataset
and about 78% on DS4, the same number of clusters with a slightly smaller percentage of points for each
cluster using only 22% of the spatial queries used by DBSCAN. The same results cannot be obtained by
DBSCAN because of the different strategy of attribution of the points to the clusters. In fact, if we stop
DBSCAN before which it has performed the spatial queries on all the points, we should obtain a correct
number of points for the clusters already individuated and probably a smaller number of points for the
cluster that we were building but of course, we will not discover all the clusters.
Table 22.1 and Table 22.2 show, for the two datasets, the number of clusters and the percentage of points
for each cluster found by DBSCAN and SPARROW.
To verify the effectiveness of the search strategy we have compared SPARROW with the random walk
search (RWS) strategy of the Reynolds’ flock algorithm and with the linear randomized search (LRS)
strategy.
Figure 22.6 and Figure 22.7 give the number of clusters found through the three different strategies
versus number of visited points for the DS4 and GEORGE dataset.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 349 — #9

22-350 Handbook of Bioinspired Algorithms and Applications

TABLE 22.1 Number of Clusters and Number of

Points for Clusters for GEORGE Dataset (Percentage
in Comparison to the Total Point for Cluster found
by DBSCAN) when SPARROW Analyzes 7, 12, and
22% Points

Percentage of
Clustering data points for
using the cluster found by SPARROW
GEORGE
dataset 7% 12% 22%

G 57.6 83.2 92.4

E 58.2 71.1 91.3
O 61.3 85.8 94.2
R 50.6 72.7 93.3
G 48.8 77.2 89.7
E 61.1 81.2 94.5

TABLE 22.2 Number of Clusters and Number of

Points for Clusters for DS4 Dataset (Percentage in
Comparison to the Total Point for Cluster found by
DBSCAN) when SPARROW Analyzes 7, 12, and 22%
Points

Percentage of
data points for
Clustering cluster found by SPARROW
using the
DS4 dataset 7% 12% 22%
1 51.16 70.99 78.86
2 45.91 64.74 74.40
3 40.68 59.36 81.95
4 44.21 60.66 81.67
5 54.65 58.72 71.54
6 48.77 59.91 78.10
7 54.29 66.43 79.18
8 51.16 70.99 78.86
9 45.91 64.74 74.40

SPARROW
400 Random
Flock
300
Core points

200

100

0
0 100 200 300 400 500
Visited points

FIGURE 22.6 Number of core points found for SPARROW, random, and flock strategy versus total number of visited
points for the DS4 dataset.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 350 — #10

Swarming Agents for Decentralized Clustering in Spatial Data 22-351

SPARROW
400 Random
Flock

Core points
300

200

100

0
0 100 200 300 400 500
Visited points

FIGURE 22.7 Number of core points found for SPARROW, random, and flock strategy versus total number of
visited points for the GEORGE dataset.

2000
100 boids
1800
50 boids
1600
25 boids
1400
Core points

1200
1000
800
600
400
200
0
0 20 40 60 80 100
Generations

FIGURE 22.8 The impact of the number of agents on the foraging for clusters strategy (DS4).

2000
1800 100 boids
1600 50 boids
25 boids
1400
Core points

1200
1000
800
600
400
200
0
0 20 40 60 80 100
Generations

FIGURE 22.9 The impact of the number of agents on the foraging for clusters strategy (GEORGE).

At the beginning, the random strategy, and also (to a minor extent) the flock, overcomes SPARROW;
however, after 200 to 250 visited points SPARROW presents a superior behavior on both the search
strategies because of the adaptive behavior of the algorithm that allows agents to learn on their previous
experience. Finally, we present the impact of the number of agents on the foraging for clusters performance.
Figure 22.8 and Figure 22.9 give the number of clusters found in 100 time steps (generations) for 25, 50,
and 100 agents. A comparative analysis reveals that a 100-agents population discovers a larger number of
clusters than the other two populations with a smaller number of agents (the scalability is almost linear).
This scalable behavior of the algorithm determines a faster completion time because a smaller number of
iterations are necessary to produce the solution.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 351 — #11

22-352 Handbook of Bioinspired Algorithms and Applications

(a) (b)

Boston

New York

Philadelphia

DS1: 8000 points North – East: 123593 points

FIGURE 22.10 The two datasets used in our experiments (circles surround the three towns of North–East dataset).

22.4.2 SPARROW-SNN Results

In this section, we present an experimental evaluation of SPARROW-SNN algorithm. We intent to explore
the following issues:

1. Determine the accuracy of the approximate solution that we obtain if we run our cluster algorithm
on only a percent of points, as opposed to running the SNN clustering algorithm on the entire
dataset.
2. Determine the effects of using SPARROW-SNN searching strategy as opposed to random-walk
searching strategy in order to identify clusters.
3. Determine the impact of the number agents on the foraging for clusters performance.

For the experiments, we used synthetic datasets and a real life dataset extracted from a spatial database.
The structures of these datasets are shown in Figure 22.10(a) and 22.10(b).
The first dataset, called DS1, contains 8,000 points and presents different densities in the clusters.
The second dataset, called North–East, contains 123,593 points representing postal addresses of three
metropolitan areas (New York, Boston, and Philadelphia) in the North East States.
We first illustrate the accuracy loss of our SPARROW-SNN algorithm in comparison with SNN
algorithm when SPARROW-SNN is used as a technique for approximate clustering. For this purpose,
we implemented a version of SNN and we computed the number of clusters and the number of points for
cluster for the two datasets. Table 22.3 presents a comparison of these results with respect to ones obtained
from SPARROW-SNN when a population of 50 agents visited, respectively, 7, 12, and 22% of the entire
dataset.
Note that with only 7% of points we can have a clear vision of the found clusters and with a few more
points we can obtain the near totality of the points. This trend is well marked in the North–East dataset.
For the DS1 dataset, the results are not so well defined because the clusters called numbers 3 and 4 have
very few points, so they are very hard to discover. For the real dataset, we only reported the results for the
three main clusters representing the towns of Boston, New York, and Philadelphia.
We can explain the good results through the adaptive search strategy of SPARROW-SNN that requires
to the individual agents to first explore the data searching for representative points whose position is not
known a priori, and then, after the representative points are located, all the flock member are steered
to move toward the representative points, that represent the interesting regions, in order to help them,
avoiding the uninteresting areas that are instead marked as obstacles and adaptively changing their speed.
To verify the effectiveness of the search strategy we have compared SPARROW-SNN with the RWS
strategy and with the standard flocking search strategy. Figure 22.11 and Figure 22.12 show the number
of representative points found with SPARROW-SNN and those found with the random search and with
the standard flock versus the total number of visited points. Both figures reveal that the number of
representative points discovered at the beginning (and until about 170 visited points for DS1 and 150 for

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 352 — #12

Swarming Agents for Decentralized Clustering in Spatial Data 22-353

TABLE 22.3 Number of Clusters and Number of Points for

Clusters for DS1 and North–East (Percentage in Comparison to
the Total Point for Cluster Found by SNN) when
SPARROW-SNN Analyzes 7, 12, and 22% Points

Percentage of data points for

cluster
Clustering using
North–East dataset 7% 12% 22%

Philadelphia 42.5 65.2 79.4

New York 38.7 52.3 67.6
Boston 46.5 68.6 82.3

Percentage of data points for

cluster found by SPARROW-SNN
Clustering using
the DSI dataset 7% 12% 22%

1 41.35 58.31 70.37

2 30.08 53.58 60.72
3 29.28 40.99 53.02
4 20.9 30.5 51.41
5 53.38 65.36 76.56
6 56.89 69.87 73.63
7 33.89 43.5 61.58

SPARROW-SNN
400 Random
Flock
300
Core points

200

100

0
0 100 200 300 400 500
Visited points

FIGURE 22.11 Number of core points found for SPARROW-SNN, random, and flock versus total number of visited
points for DS1 dataset.

North–East dataset) from the random strategy, and (to a minor extent) for the flock, is slightly greater
than of SPARROW-SNN.
Next, our strategy presents a superior behavior on both the strategies because of the adaptive behavior
of the algorithm that allows to the agents to learn on their previous experience.
Finally, in order to study the scalability of our approach, we present the impact of the number of agents
on the foraging for clusters performance. Figure 22.13 and Figure 22.14 give, respectively, the DS1 and
North–East dataset, the number of clusters found in 100 time steps (generations) for 25, 50, and 100
agents. A comparative analysis reveals that a 100-agents population discovers a larger number of clusters
(almost linear) than the other two populations with a smaller number of agents. This scalable behavior of
the algorithm should determine a faster completion time because a smaller number of iterations should
be necessary to produce the solution.
We try to partially give a theoretical explanation of the behavior observed in Figure 22.11 and
Figure 22.12, that is, the improvement in accuracy of the SPARROW-SNN strategy, introducing an
entropy-based model in Section 22.5.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 353 — #13

22-354 Handbook of Bioinspired Algorithms and Applications

400
SPARROW-SNN
Random
300 Flock

Core points
200

100

0
0 100 200 300 400
Visited points

FIGURE 22.12 Number of core points found for SPARROW-SNN, random, and flock versus total number of visited
points for North–East dataset.

2000
1800 100 boids
1600 50 boids
25 boids
Core points

1400
1200
1000
800
600
400
200
0
0 20 40 60 80 100
Generations

FIGURE 22.13 The impact of the number of agents on foraging strategy for the DS1 dataset.

2000
100 boids
1800
50 boids
1600
25 boids
1400
Core points

1200
1000
800
600
400
200
0
0 20 40 60 80 100
Generations

FIGURE 22.14 The impact of the number of agents on foraging strategy for the North–East dataset.

22.5 Entropy Model

In order to verify and explain the behavior of our system, we used a model based on the entropy as
introduced in Reference 20. The authors use a measure of entropy to analyze emergence in multiagent
systems. The fundamental claim of this chapter is that the relation between self-organization based on

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 354 — #14

Swarming Agents for Decentralized Clustering in Spatial Data 22-355

emergence in multiagent systems and concepts as entropy is not just a loose metaphor, but can provide
quantitative, analytical guidelines for designing and operating agent systems. These concepts can be applied
in measuring the behavior of multiagent systems. The main result suggested here concerns the principle
that the key to reducing disorder in a multiagent system to achieve a coherent global behavior is coupling
that system to another in which disorder increases. This corresponds to a macrolevel where the order
increases, that is, a coherent behavior arises, and the microlevel where an increase in disorder is the cause
for this coherent behavior at the macrolevel.
A multiagent system should follow the second law of thermodynamics “energy spontaneously disperses
from being localized to becoming spread out if it is not hindered,” if agents move without any constriction.
However, if we add information in an intelligent way the agents natural tendency to maximum entropy
will be contrasted and system goes toward self-organization.
Here, we will apply this model to the SPARROW-SNN algorithm. As stated in Reference 20, we can
observe two levels of entropy: a macrolevel in which organization takes place, balanced by the micro in
that we have an increase of entropy. For the sake of clarity, in SPARROW-SNN, microlevel is represented
by red and white agents’ positions, signaling, respectively, interesting and desert zones, and the macrolevel
is computed considering all the agents positions. Therefore, we expect to observe an increase of micro
entropy by the birth of new red and white agents and, on the contrary, a decrease in macro entropy
indicating organization in the coordination model of the agents.
In the following, we give a more exact description of the entropy-based model. In information theory,
entropy can be defined as

S=− pi log pi .
i

Now, to adapt this formula to our need, we use the location-based entropy. Consider an agent moving
in the space of data divided in a grid N × M = K , where each cell has the same dimension. Hence, if N
and M are quite large, each random agent has the same probability to be in one of the K cells of the grid.
We measured the entropy running a simulation for 100 times, for 2000 time steps using 50 agents and
counting how many times the agent falls in the same cell i for each time step. Dividing this number by T
we obtain the probability pi that the agent will be in this cell.
Then, the locational entropy will be:

a
i=1 pi log pi
S=− . (22.1)
log a

In case of random distribution, every state has probability 1/a, so the overall entropy will be
(log a)/(log a) = 1; this explains the factor of normalization a.
This equation can be generalized for P agents, summing the over all agents and dividing the average
by P. Equation (22.1) represents the macro entropy; if we consider only red and white points, it represents
the micro entropy.
We computed the micro and macro locational entropy for the real dataset North–East with the
SPARROW-SNN algorithm using the parameters cited earlier. The results are showed in Figure 22.15
and Figure 22.16.
As expected, we can observe an increase in micro entropy and a decrease in macro entropy of SPARROW-
SNN due to the organization introduced in the coordination model of the agents by the attraction toward
red agents and the repulsion of white agents. On the contrary, in random and standard flock model, the
curve of macro entropy is almost constant, confirming the absence of organization.
These experiments partially explain the validity of our approach, but further studies are necessary and
will be conducted to better demonstrate the positive effects of organization and the correlation between
macro entropy and validity of searching strategy.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 355 — #15

22-356 Handbook of Bioinspired Algorithms and Applications

0.35

0.3

0.25
Micro entropy
0.2

0.15

0.1

0.05

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Iterations

FIGURE 22.15 Micro entropy (red and white agents) for the North–East dataset using SPARROW-SNN.

1
Random
0.995
Flock
0.99

0.985
Macro entropy

0.98

0.975

0.97 SPARROW-SNN

0.965

0.96

0.955
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Iterations

FIGURE 22.16 Macro entropy (all the agents) for the North–East dataset using SPARROW-SNN.

22.6 Conclusions
In this chapter, we have described a fully decentralized approach to clustering spatial data by a multi-
agent system. The approach is based on the use of SI techniques. Two novel algorithms that combine
density-based and shared nearest-neighbor clustering strategy with a flocking algorithm have been presen-
ted. The algorithms have been implemented in SWARM and evaluated using synthetic datasets and one
real word dataset. Measures of accuracy of the results show that the flocking algorithm can be efficiently
applied as a data reduction strategy to perform approximate clustering. Moreover, by an entropy-based
model we have theoretically demonstrated that the adaptive search strategy of SPARROW is more efficient
than that of the RWS strategy. Finally, the algorithms show a good scalable behavior.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 356 — #16

Swarming Agents for Decentralized Clustering in Spatial Data 22-357

Acknowledgment
This work was supported by the CNR/MIUR project — legge 449/97-DM 30/10/2000 and by Project
“FIRB Grid.it” (RBNE01KNFP).

References
[1] Han, J. and Kamber, M. Data Mining: Concepts and Techniques. Morgan Kaufmann, San Mateo,
CA, 2000.
[2] Kaufman, L. and Rousseeuw, P.J. Finding Groups in Data: An Introduction to Cluster Analysis.
John Wiley & Sons, New York, 1990.
[3] Karypis, G., Han, E., and Kumar, V. CHAMELEON: A hierarchical clustering algorithm using
dynamic modeling. IEEE Computer, 32, 68–75, 1999.
[4] Sander, J., Ester, M., Kriegel, H.-P., and Xu, X. Density-based clustering in spatial databases: The
algorithm GDBSCAN and its applications, Data Mining and Knowledge Discovery, 2, 169–194,
1998.
[5] Wang, W., Yang, J., and Muntz, R. STING: A statistical information grid approach to spatial data
mining. In Proceedings of International Conference on Very Large Data Bases (VLDB’97), 1997,
pp. 186–195.
[6] Han, J., Kamber, M., and Tung, A.K.H. Spatial clustering methods in data mining: A survey. In
Geographic Data Mining and Knowledge Discovery, H. Miller and J. Han (Eds.), Taylor & Francis,
London, 2001.
[7] Deneubourg, J.L., Goss, S., Franks, N., Sendova-Franks, A., Detrain, C., and Chretien, L. The
dynamic of collective sorting robot-like ants and ant-like robots. In Proceedings of the first Confer-
ence on Simulation of Adaptive Behavior, J.A. Meyer and S.W. Wilson (Eds.), MIT Press/Bradford
Books, Cambridge, MA, 1990, pp. 356–363.
[8] Lumer, E.D. and Faieta, B. Diversity and adaptation in populations of clustering ants. In Proceedings
of the Third International Conference on Simulation of Adaptive Behavior: From Animals to Animats
(SAB94), D. Cliff, P. Husbands, J.A. Meyer, and S.W. Wilson (Eds.), MIT Press, Cambridge, MA,
1994, pp. 501–508.
[9] Kuntz, P. and Snyers, D., Emergent colonization and graph partitioning, In Proceedings of the Third
International Conference on Simulation of Adaptive Behavior: From Animals to Animats (SAB94),
D. Cliff, P. Husbands, J.A. Meyer, and S.W. Wilson (Eds.), MIT Press, Cambridge, MA, 1994,
pp. 494–500.
[10] Monmarché, N., Slimane, M., and Venturini, G. On improving clustering in numerical databases
with artificial ants. In Advances in Artificial Life: 5th European Conference, ECAL 99, Vol. 1674 of
Lecture Notes in Computer Science, Springer-Verlag, Berlin, 1999, pp. 626–635.
[11] Bonabeau, E., Dorigo, M., and Theraulaz, G. Swarm Intelligence: From Natural to Artificial Systems.
Oxford University Press, Oxford, 1999.
[12] Grassé, P.P., La Reconstruction du nid et les Coordinations Inter-Individuelles chez Beellicositermes
Natalensis et Cubitermes sp. La Théorie de la Stigmergie: Essai d’interprétation du Comportement
des Termites Constructeurs, Insect. Soc. 6, pp. 41–80, 1959.
[13] Macgill, J. Using flocks to drive a geographical analysis engine. In Artificial Life VII: Proceedings of
the Seventh International Conference on Artificial Life. MIT Press, Reed College, Portland, Oregon,
2000, pp. 1–6.
[14] Kollios, G., Gunopoulos, D., Koudas, N., and Berchtold, S. Efficient biased sampling for
approximate clustering and outlier detection in large datasets. IEEE Transactions on Knowledge
and Data Engineering, 15, 1170–1187, 2003.
[15] Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. A density-based algorithm for discovering clusters in
large spatial databases with noise. In Proceedings of the 2nd International Conference on Knowledge
Discovery and Data Mining (KDD-96), Portland, OR, 1996, pp. 226–231.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 357 — #17

22-358 Handbook of Bioinspired Algorithms and Applications

[16] Ertöz, L., Steinbach, M., and Kumar, V. A new shared nearest neighbor clustering algorithm and its
applications. In Workshop on Clustering High Dimensional Data and its Applications at 2nd SIAM
International Conference on Data Mining, Washington, USA, April 2002, pp. 105–115.
[17] Minar, N., Burkhart, R., Langton, C., and Askenazi, M. The Swarm Development Group, 1996,
http://www.santafe.edu/projects/swarm.
[18] Reynolds, C.W. Flocks, herds, and schools: A distributed behavioral model. Computer Graphics,
21, 25–34, 1987.
[19] Jarvis, R.A. and Patrick, E.A. Clustering using a similarity measure based on shared nearest
neighbors. IEEE Transactions on Computers, C-22(11), 1025–1034, 1973.
[20] Van Dyke Parunak, H. and Brueckner, S. Entropy and self-organization in multi-agent systems.
In Proceedings of the Fifth International Conference on Autonomous Agents, ACM Press, 2001,
pp. 124–130.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C022” — 2005/8/6 — 14:10 — page 358 — #18

 23
Biological Inspired
Based Intrusion
Detection Models
for Mobile
Telecommunication
Systems

23.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-359

Related Works • Distributed Management Security System
23.2 Specification and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-361
SSTCC Service Specification • SSTCC Protocol
Specification • Formal SSTCC Validation
23.3 Intrusion Detection by Neural Networks . . . . . . . . . . . . . . 23-365
Neural Network Construction • Neural Network
Implementation
Azzedine Boukerche 23.4 Toward an Artificial Immune Human System for
Mobile Phone Intrusion Detection . . . . . . . . . . . . . . . . . . . . . 23-368
Kathia Regina Lemos Jucá Anamnestic Response (Memory) • Artificial Immune
João Bosco Mangueira Sobral System • Artificial Telecom Immune System
Mirela Sechi Moretti 23.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-370
Annoni Notare References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-371

23.1 Introduction
Security management against malicious intrusions and frauds in mobile telecommunication systems
represents one of the most challenging problems for the wireless networking and mobile phone com-
munities. Currently, mobile telephone carriers had been plagued by fraudulent use of cloned and stolen
phone numbers. These fraudulent acts are costing them dearly in many countries, especially in countries
where mobile phones can be rented freely and exchanged freely from some telecom carriers [1]. They are
also witnessing, every year, a significant increase of their profit losses that are passed, unfortunately, to the
mobile phone users. As a consequence, many of these mobile phone carriers are investing a large amount

23-359

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 359 — #1

23-360 Handbook of Bioinspired Algorithms and Applications

of money in the R&D to develop efficient security management solutions to deal with their security
concerns. In this chapter, we focus mainly on the cloning and subscription frauds. In a subscription fraud,
an impostor may subscribe to the service provided by the mobile phone telecommunication company,
with a prior intention of not paying for the service. The impostor can also steal, subscribe, or use, each
month, a different mobile phone (e.g., using a different name/number). However, in all of these cases,
these mobile phones will most likely have a common usage pattern. The cloning fraud is characterized
by calls that were made by cloned phones where the calls will appear in a monthly billing statement of
a legitimate phone. The cloning may occur via a simple radio that can capture the two identification
numbers: ESN, Electronic Serial Number and MIN, Mobile Identification Number) of a legitimate phone.
Here again, calls from cloned phones and calls from legitimate phones will most likely have different usage
patterns, as usage pattern is specific to each mobile phone user. While it is true that cloning frauds can
be significantly reduced using new hardware technologies, subscription frauds are hardware independent.
Therefore, new software-based technologies must be developed and deployed to identify these malicious
intruders. To the best of our knowledge little work has been done at the software level to deal with the
cloning and subscription frauds problem [2–4].

23.1.1 Related Works

Several strategies might be envisaged to treat the cloning and fraud subscriptions problem in mobile
phone operation systems. In what follows, we briefly describe each of the available strategies:

1. Cryptography schemes will increase the difficulty for malicious intruders to capture the ESN/MIN
numbers and perform clandestine hearings. While one might argue that it is simple to use such
schemes in digital phones, it is not quite obvious as to how they can be used in analog phones.
Thereby, new and efficient schemes are needed to deal with the security problem in mobile phone
operations. Furthermore, GSM digital phones using encryption model have already been cloned
and broken recently by malicious intruders.
2. Denial of service schemes are basically based upon denying international calls, for instance. While
this approach might help the mobile phone carriers to reduce their profit losses, it is clear that this
solution is an alternative solution and not an efficient security management solution.
3. User verification schemes are based upon the use of passwords by a mobile phone user before using
his phone set. Note that this approach is widely used by many mobile phone carriers. While this
approach provides a certain level of security, it is our belief that a good security solution should
not complicate the mobile phone usage by requiring the users to memorize all of these passwords
in order to use their mobile phones.
4. Traffic analysis schemes are based upon the profile and the behavior of the mobile phone users.
Nowadays, the main problem of this approach is its’ simplistic and rule based analysis where only
the monthly statement bill and the price of the calls are being considered.

In this chapter, we focus upon the last approach. We show how nature and biological inspired techniques
can be used efficiently for traffic analysis purposes and the identification of malicious intruders in mobile
phone operation systems.

23.1.2 Distributed Management Security System

This chapter proposes a security management system for intrusion detection in wireless telecommunica-
tion networks through traffic analysis, where the users are classified into clusters according to their behavior
patterns and phone usages — through neural networks and immune human-based system employment.
The classification of the users helps the security system to identify when the phone calls do not corre-
spond to their users’ usage pattern, which may constitute a possible fraudulent use of the mobile phone.
The fraud could take the form of a cloned phone, a fraud subscription, or a frequent delinquent mobile
phone user, known through his past, to subscribe to a service with a prior intention of not paying for the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 360 — #2

Biological Inspired Based Intrusion Detection Models 23-361

service (e.g., user has history of a bad debt). An immediate and automatic message (thereby, minimizing
the size of the personal staff to do that) is sent to the mobile phone user in order to confirm the cloning
or fraud subscription. As opposed to waiting until the end of the monthly billing statement cycle, this
immediate notification will help the telecom carriers to reduce their profit losses, thereby reducing the
damage that would have been passed on, somehow, to the mobile phone users, such as increased phone
call fees, surcharges, and so on. In the case of detection of fraudulent subscription, personalized reports
are quickly generated and used by the system and the telecom carrier to investigate who actually made the
suspected calls.
In addition to the security management, Web/WAP application tools could be developed where users
can observe online their phone bills, which will allow them to track down their mobile phone statement
bill and identify malicious intruders using their phone number, thereby reducing the fraudulent usage of
the mobile phone. The system may encompass services of access control, authentication, confidentiality,
integrity, availability and nonrepudiation of communication. In our previous work [2,4,5], we have
incorporated and implemented all of these security services using Java with the support of CORBA.
Moreover, as an important characteristic, our system was formally validated according to ISO 8807
standard, see References 3 and 6.
This chapter is organized as follows: Section 23.2 describes the use of ISO 8807 standard that we have
used to specify and validate our security management system. Section 23.3 demonstrates how neural
network techniques can be used to detect intrusion by pattern recognition of telecom users. Section 23.4
comments about how the immune human system can be used as an alternative solution to identify
malicious intruders and fraudulent usage of mobile phone users. Section 23.5 concludes the chapter.

23.2 Specification and Validation

This section presents the formal specification and validation of the security management system according
to ISO 8807 (International Organization for Standardization), FDT (Formal Description Technique) and
LOTOS (Language of Temporal Ordering Specifications) [7]. One of the major issues of the use of FDTs, is
to demonstrate the correctness of the system. In this work, the LOTOS specifications are validated through
the use of the Eucalyptus ToolSet/CADP (Caesar/Aldebaran Development Package) [8], on Sun Solaris
environment, that includes the following two main components: (1) Caesar and (2) Aldebaran.

Caesar is a compiler that translates a LOTOS specification into a C program (to be executed or simulated)
or into an LTS, Labeled Transition Systems (to be verified using bi-simulation and temporal logic
tools). As will be demonstrated in Section 23.2.3, it is possible to compare the protocol LTS with
the related LTS’ service. Both LTSs are generated using the Caesar compiler and are compared using
the Aldebaran tool.
Aldebaran is a tool for communication systems verification. It is represented by LTSs, where the state
transitions are labeled by action names. Aldebaran allows the LTSs reduction using several equiv-
alence relations (e.g., strong bi-simulation, observational equivalence, delay bi-simulation and
secure equivalence).

The stepwise refinements approach used in our work to build the LOTOS specifications, allows the
system to be validated along the system development.

23.2.1 SSTCC Service Specification

At the most abstract level, the SSTCC (Security System for Telecommunications against Cellular Cloning)
and Impostors [2,3,9], can be seen as a black box, containing four communication gates (mail_alarm,
phone_alarm, online_bill, and check_owner), to send/receive messages to and from the
telecommunication users. As illustrated in Figure 23.1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 361 — #3

23-362 Handbook of Bioinspired Algorithms and Applications

mail_alarm

phone_alarm
SSTCC Telecom
Service online_bill Users
check_owner

FIGURE 23.1 SSTCCService graphical representation.

specification SstccService[mail_alarm,phone_alarm,online_bill,check_owner]:noexit
behaviour SstccService[mail_alarm,phone_alarm,online_bill,check_owner] where
process SstccService[mail_alarm,phone_alarm,online_bill,check_owner]:noexit:=
(i;mail_alarm;
(phone_alarm;SstccService[mail_alarm,phone_alarm,online_bill,check_owner]
[] SstccService[mail_alarm,phone_alarm,online_bill,check_owner]))
[](online_bill;SstccService[mail_alarm,phone_alarm,online_bill,check_owner])
[](check_owner;SstccService[mail_alarm,phone_alarm,online_bill,check_owner])
endproc
endspec

FIGURE 23.2 SSTCCService LOTOS specification.

The mail_alarm gate is used by SSTCC to send alarm signals by snail mail to the users about a
malicious intruder. The phone_alarm gate allows the SSTCC system to send these alarms using the
mobile phone (e.g., short message or interactive voice message). The advantage of using snail mail is its
delivery guarantee, whereas short time has the benefit of using mobile phone. The online_bill gate
is used by the users to observe their online phone bill via the Web/WAP tool. Finally, the check_owner
gate is used by the telecom carrier to investigate probable impostors (with potential future bad debits). The
SSTCC system remains permanently active, thereby characterizing an infinite behavior of this system —
This is illustrated in Figure 23.2 using the noexit functionality.
The behavior of our SSTCC system is defined by the process SSCCService, that can execute an action
in the gate mail in order to send an alarm by a regular surface mail (this action is always possible); the
sequence is followed by a nondeterministic choice with two options. The first option is related to the alarm
sent by mobile phone, in the gate phone (this action is not always possible) and, following, the process
SSTCCService is called recursively in order to deal with another case. The first option may not be
successful, since the phones may not work properly (e.g., the mobile phone users may be out of their area
after a certain period of time). Thus, an internal action i occurs (which has not been observed) and the
process SSTCCService is executed recursively. Similarity, the online_bill and check_owner
are the two gates used to observe the phone bills and verify/confirm the user identity, respectively.
This abstract level of SSTCC specification corresponds to a formal specification of the user requirements
of the system, which will represent the basis for future refinements (i.e., for the protocol specification).

23.2.2 SSTCC Protocol Specification

In a more detailed level, the behavior of our SSTCC system is defined by the SstccProtocol specifi-
cation, where it is possible to observe its three main components: (1) SSCC, Security System against
Cellular Cloning, represented by the SsccClone process; (2) SETWeb/Wap, Online Phone Bill by
Web/WAP System, represented by the SetwebBill process; and (3) SIPI, System to Identify Probable
Impostors, represented by the SipiImpostor process. Figure 23.3 illustrates the general architecture
of the refined SSTCC system, including the three main processes that compose it.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 362 — #4

Biological Inspired Based Intrusion Detection Models 23-363

SSTCCProtocol

C
CloneCorbaAgent1 clone_notif O
R
user_pattern online_call B mail
A alarm
UserPatternFile1 OnlineCallsFile1
M
A
CloneCorbaAgentn clone_notif N phone
A alarm
user_pattern online_call G
UserPatternFilen OnlineCallsFilen E
R
SSCC

C U
BillCorbaAgent1 O O S
bill_notif R E
bill_db online_call B R
A S
BillDB1 UserOnlineCallsFile1
M online
A bill
BillCorbaAgentn bill_notif N
A
bill_db online_call G
E
BillDB1 UserOnlineCallsFile1 R
SETWeb/SETWap

impostor _notif C
ImpostorCorbaAgent2 O
R
user_1 impostor_1 online_1 B
A
Patterns1 Impostors1 OnLine1
M check
impostor_notif A owner
ImpostorCorbaAgentn N
A
user_n impostor_n online_n G
E
Patternsn Impostorsn OnLinen R

SIPI

FIGURE 23.3 SSTCCProtocol graphical representation.

Figure 23.3 displays each of the three components that are identified by a manager and several
distributed agents. The SSCC/SIPI agents compare the online calls database with the users’ pattern data-
base identified by the neural networks technology. The SETWeb/Wap agents access both the online and the
existing databases to build the phone bills. The procedural definition of the SSTCCProtocol behavior
is shown in Figure 23.4.
Figure 23.4 presents the three components composed in parallel using the parallel independent com-
position operator (|||). Based upon the use of the two specifications (SstccService.lotos and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 363 — #5

23-364 Handbook of Bioinspired Algorithms and Applications

specification SstccProtocol[mail_alarm,phone_alarm,online_bill,check_owner]:noexit
behaviour
hide clone_notif,web_notif,impostor_notif in
SsccClone[mail_alarm,phone_alarm]
|||SetwebBill[online_bill]
|||SipiImpostor[check_owner]
where
process SsccClone[mail_alarm,phone_alarm]:noexit:= ... endproc
process SetwebBill[online_bill]:noexit:= ... endproc
process SipiImpostor[check_owner]:noexit:= ... endproc
endspec

FIGURE 23.4 SSTCCProtocol LOTOS specification.

FIGURE 23.5 SSTCCProtocol LTS.

SstccProtocol.lotos, as presented in Sections 23.2.1 and 23.2.2, respectively) it is possible to

perform the validation of the security system, as presented in Section 23.2.3.

23.2.3 Formal SSTCC Validation

System’s validation can be done via simulation, test and verification. While simulation and testing are
used (only) to find errors, verification provides the formal (mathematical) proof and the correctness of
the system. In order to obtain the system’s correctness proof, initially, the LTS (SstccProtocol.bcg)
file is generated. This file is related to the protocol specification (file SstccProtocol.lotos) using
the Caesar tool. Figure 23.5 shows the SSTCCProtocol LTS containing 194401 states and 1645926
transitions.
Using a similar approach, the LTS (SstccService.bcg) file is generated. This file is related to
the service specification (file SstccService.lotos) using the Caesar tool. Figure 23.6 shows the
SSTCCService LTS containing three states and six transitions.
Then, it is possible to verify, by observational equivalence, the protocol LTS with the service LTS using
the Aldebaran tool. The obtained result, TRUE, signifies that the protocol executes the desired service,
that is, the formal (mathematical) proof of correctness of the SSTCC system has been obtained. See
Figure 23.7.
This proof aims at satisfying the requirements in order to obtain the ClassA1 of the Security Level
Class_A1. Table 23.1 provides the security levels approved by the U.S. – Department of Defense (DoD) [10].
The formal analysis and the mathematical proof presented in this section, aims at guaranteeing the
verification of the design based upon the ClassA1 security level.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 364 — #6

Biological Inspired Based Intrusion Detection Models 23-365

FIGURE 23.6 SSTCCService LTS.

-----------------------------------------------------------
aldebaran –bddsize 4 –oequ –std SstccProtocol_bmin.bcg ./SstccService.bcg | tee aldebaran.seq
TRUE
-----------------------------------------------------------

FIGURE 23.7 SstccProtocol and SstccService verification.

TABLE 23.1 Table of Security Levels (DoD of USA)

Security level Major feature introduced

Division D Minimal protection
Division C Discretionary protection
Division B Mandatory protection
Division A Verified protection
Class A1 Verified design. Formal analysis and mathematical
proof that the computer system matches the system’s
security policy and its design specifications

23.3 Intrusion Detection by Neural Networks

In this section, we will show how neural networks technology can be used to design intrusion detection
systems based upon the traffic analysis and the classification [11] of the mobile phone users. The main
reasons behind using neural networks techniques, among others, are: (1) they have the intrinsic capacity of
learning input data and capturing the user’s behavioral patterns, (2) they are based upon nonparametric-
based models and have the potential to provide better understanding of the distribution of input data than
the traditional statistical methods (e.g., Bayesian methods), and (3) they have the capabilities of creating
decision boundaries that are highly nonlinear in the space of characteristics.
Neural networks and neural brain share certain similarities, among others: (1) the knowledge is acquired
via a learning process; and (2) the weights of the connections between neurons, known as synapses, are
used to store the knowledge. The procedure used to represent the learning process, commonly called
learning algorithm, has the capabilities of modifying the weights of the connections of the network in
order to get the targeted design. The algorithms could easily be implemented using the MatLab tool [2].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 365 — #7

23-366 Handbook of Bioinspired Algorithms and Applications

Comada de entrada
u0 C
Camada escondida
W Camada de saida
u1 y1

u2 y2

uN y0

Junçâo de somatório
Funçao Gaussina
Funça–o linear

FIGURE 23.8 RBF architecture.

23.3.1 Neural Network Construction

In order to conduct the classification of mobile phone users based upon their phone usage patterns, an
artificial neural network using the a Radial Base Function (RBF) can be used [2]. The architecture of
the RBF network consists mainly of an input layer, a hidden layer, and an output layer. As illustrated in
Figure 23.8.
The input layer is composed of the source nodes (i.e., sensorial units) and the output nodes that
constitute a linear combination of the RBF calculated by the nodes of the hidden layer. The RBF in the
hidden layer produces a response for the input (stimulus pattern), that is, it produces responses that are
different from zero when the input pattern is within a small region located in the input space. The learning
component in the hidden (intermediate) layer is executed using the “non-supervised” method, such as
a simple heuristic clustering algorithm, or a “supervised” algorithm to find the centers (i.e., the C nodes
in the hidden layer). In our work [2], we made use of the K-means algorithm in order to determine the
centers that represent the connections between the input layer and the intermediate layer. Gauss function
was used in order to determine which of the various centers are similar to the input vector. The Gauss
function used to obtaining the output of a hidden layer is defined as follows:

n   
1 1  xj − c j 2
f = (x − c) = exp − . (23.1)
(2π )n/2 σ1 , σ2 , . . . , σn 2 σj
j−1

To increase the RBF’s functionality, the Mahalanobis distance could be used within the Gauss function, as
follows:


1 1 T −1
f = (x − c) = exp − (x − c) K (x − c) , (23.2)
(2π )n/2 |K |1/2 2

where K −1 is the inverse of the X covariance matrix, associated with the node of the hidden C layer.
Given n-vectors (input data) of p-samples, representing p-classes, the network can be initiated with
the knowledge of the centers (i.e., locations of the samples). If the J-th vector sample is represented, then
the weight matrix C can be defined as: C = [c1 , c2 , . . . , c3 ]T so that the weights of the hidden layer in

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 366 — #8

Biological Inspired Based Intrusion Detection Models 23-367

the j node are composed of the center vector. The output layer is a weight sum of the outputs of the hidden
layer. When presenting an input vector for the network, the network implements Equation (23.3)

y = W · f (x − c), (23.3)

where f represents the functional output vector of the hidden layer, and C the corresponding center vector.
After supplying some data with the desired results, the weights (W ) can be determined using the training
algorithm either interactively or noninteractively, based upon the descendant or pseudo-inverse gradient
scheme, respectively. In order to determine the σ 2 variation parameter for the Gauss function, one has,
first, to compute the median distance between all the training data, see Equation (23.4).

1 
σj2 = (x − c)T (x − c), (23.4)
Mj
x∈j

where j is the group of training patterns grouped in the center of the cluster Cj , and Mj is the number
of patterns in .
Another way of choosing the parameter σ 2 is to calculate the distances between the centers in each
dimension and use a scaling factor. This approach will allow the p-nearest neighboring algorithm to be
used and obtain the variance related to each center.

23.3.2 Neural Network Implementation

In order to implement the neural network algorithm that is used to classify the mobile phone users into
groups based upon their phone usage pattern, the following steps must be performed. Step 1: identify
the number of neurons of the hidden layer. Step 2: find the centers cj (j = 1, . . . , M ) which makeup the
basis of an M -dimensional space. Step 3: for each input pattern, the xi − cj  is sent as a parameter to
the RBF, which describes the level of classification of the input patterns. The output of the hidden layer
makes up a G matrix, which serves as a basis for the calculation of the weigh W (i.e., connections for the
output layer), following the formula: W = G + t , where G + is the pseudo-inverse of the G matrix given
by G + = (G T G + λG0 )−1 · G T ; and t is the matrix that contains the group of training data. In the last
step, the output is calculated as a sum of the activated neurons of a hidden layer.
The obtained output, that is, the classified users based upon their phone usage (see Figure 23.9) makes
up the database used in the implementation of the SSTCC (see Section 23.4). In our implementation, the
online calls are compared to the classified data in order to identify calls that do not match with that of the
behavioral pattern of the client, which imply fraud detection.
The results we have obtained are summarized in Figure 23.9. They illustrate the efficiency of neural
network and pattern recognition techniques to design efficient nature based intrusion detection system that
can be used to identify the cloning and subscription frauds in mobile phone operations [12]. Considering
the error rate obtained in the classification (2.5% using 80 neurons in the hidden layer), the currently
losses of US$ 500,000 daily (data from IDC, Sept 99) could be reduced to US$ 12.500 daily. This error

FIGURE 23.9 Users (before and after classification).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 367 — #9

23-368 Handbook of Bioinspired Algorithms and Applications

rate is quite encouraging, when compared to previous results obtained by other researchers [13], where a
5.4% error rate was obtained using the Back Propagation algorithm.

23.4 Toward an Artificial Immune Human System for

Mobile Phone Intrusion Detection
In this section, we show how an immune human based model can be used to design mobile phone intrusion
detection system. First, let us review the basic understanding of the immune human systems. The Natural
(Human) Immune System [14] protects our bodies from infected agents such as viruses, bacteria, and
fungi. It can distinguish between self and nonself genes. It is able to remember previous experiences and
react accordingly. Therefore, once a human being has a flu, the immune human system will prevent him
from getting it again. The immune system displays both enormous diversity and extraordinary specificity.
Indeed, it is not only able to recognize the many millions of distinctive nonself genes, but it can also produce
molecules and cells to match up with nonself genes, and have them counteract each other. An antigen
is defined as a substance that reacts with antibody molecules and antigen receptors on lymphocytes. An
immunogen is an antigen that is recognized by the body as a nonself gene and stimulates an adaptive
immune response [15].
There are two types of immunity, innate and adaptive. Innate immunity primarily consists of the
endocytic and phagocytic systems, which involve motile scavenger cells such as macrophages that ingest
extracellular molecules and materials, clearing the system of both remaining items and pathogens.
An Adaptive (or Acquired) immunity, allows the immune system to launch an attack against any foreign
intruders that the innate system cannot remove. The adaptive system is directed against specific intruders,
it learns how to recognize specific kinds of pathogens, and keep track of them, so that it can identify them
later if they reappear again, thereby speeding up the response to deal with them. The learning process
occurs during a primary response to a kind of pathogen that was not encountered before by the immune
system. The primary response is slow, often only becoming apparent, at first, many hours after the initial
infection. However, the immune system will retain a memory of the kind of pathogens that has caused
the infection for future intrusions [16].

23.4.1 Anamnestic Response (Memory)

As a result of the T-dependent B-lymphocytes being sensitized to a specific antigen, numerous circulating
B-memory cells developed primary anamnestic response or memory. A subsequent exposure to the same
antigen may result not only in a more rapid production but also of a higher number of antibodies and
for a longer period of time. The primary response to a new antigen generally peaks at 10 to 17 days;
however, because of the numerous circulating B-memory cells, the secondary anamnestic response peaks
in just 2 to 7 days. Due to the clonal expansion and affinity maturation, there is now a pool of B-memory
cells having the “fine-tuned” B-cell receptors on their surface. The pool of B-memory cells circulates
throughout the body waiting to encounter the original antigen when it again enters the human body.
B-memory cells have high longevity. They replicate and produce antibodies periodically when they are
exposed to a persisting epitope that remains on the surface of the follicular dendritic cells in the lymphoid
organs. This secondary immune response is said to be specific to the antigen that has initiated the immune
response the first time and has been the factor behind the faster response that is attributed to the memory
cells remaining in the immune system, so that when an antigen, or a similar antigen, is encountered,
a new immunity does not need to be built up, since it is already there [17,18].

23.4.2 Artificial Immune System

Artificial Immune Systems are known to use adaptive search algorithms based on the biological immune
system with the central task of pattern matching between antigens and antibodies. Two distinct

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 368 — #10

Biological Inspired Based Intrusion Detection Models 23-369

groups of algorithms have emerged as successful implementations of artificial immune systems: (1) the
negative selection algorithm developed by Forrest et al. [19] is used to differentiate between normal system
operation (i.e., self) and abnormal operation (i.e., nonself), and (2) the immune network based model as
discussed by Jerne [20], then formally described by Perelson [21] and Farmer et al. [16]. Our intrusion
detection model for mobile phone operations is based upon the Jerne model. Both approaches are partic-
ularly well suited to data-mining related tasks that involve searching through large databases and finding
the matching patterns [22–25].
The Idiotypic network theory, introduced by Jerne, maintains that interactions in the immune system
do not occur only among antibodies and antigens since antibodies may interact with each others as well.
Hence, an antibody may be matched to other antibodies, which in turn may be matched to yet other
antibodies, and this process can spread throughout the entire population. According to the immune
network theory, immunological memory is achieved by the B cells supporting each other through an
Idiotypic network. B cells are not only stimulated by antigens, but also stimulated and, to a degree,
suppressed by neighboring B cells. The interaction between B cells takes place via the idiotopes on each B
cell which acts as a recognizer for other B cells. Therefore, the more neighboring cells a particular B cell
has that are similar to itself, the more stimulated that particular B cell becomes. This allows the clustering
of a population of identical B cells, and produces a self-supporting structure to aid the immune responses.
The network self-organizes and stabilizes within a reasonable time since its survival is achieved by mutual
reinforcement among the B cells via a feedback mechanism. Survival of a new B cell produced, as part
of the immune response, either by the bone marrow or by hypermutation, depends on its affinity to the
antigens that are present, and to its neighbors in the immune network. The new B cells may have an
improved match for an antigen and, thus, will proliferate and survive longer than the existing B cells. By
repeating the mutation and the selection procedures several times, the immune system learns to produce
better antigen matches. This theory could help in explaining how the memory of past infections is retained.
Furthermore, it could result in the suppression of similar antibodies, thereby, encouraging a good diversity
in the antibody pool [26]. See Figure 23.10.

Stimulation
Supression

Antigen
Epitope
B cell 2

Idiotope 3
Paratope 2
Antibody 2
B cell 1

Idiotope 1
Paratope 1
Antibody 1

Paratope 4
B cell 3 Antibody 4
Idiotope 4

Idiotope 3
B cell 4
Antibody 3 Paratope 3

FIGURE 23.10 Jernes Idiotypic Network Hypothesis. (From Jerne, N.K. Annals of Immunology, 125, 373–389, 1973.
With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 369 — #11

23-370 Handbook of Bioinspired Algorithms and Applications

Figure 23.10 shows the basic idea of the immune network hypothesis. There are many antibodies on
the surface of the B cells that act as antigen detectors. The specialized portion of the antibody used for
identifying other molecules (of antigens or antibodies) is called paratopes. The region on any molecules
that can be recognized by the paratopes is called epitope. The binding between idiotopes and paratopes has
the effect of stimulating the B cells. This is mainly due to the fact that the paratopes on B cell2 reacts to the
idiotope on B cell1 as it would on an antigen. However, to counter the reaction, there is a certain amount
of pressure among the B cells to act as a regulatory mechanism. The authors [16,21] have formulized the
theory, and they have suggested that the B cells are stimulated by an interaction not only between B cells
and antigens, but also the surrounding of the B cells that are connected to form a network of cells. As
illustrated in Figure 23.10, B cell1 stimulates three cells, B cell2, B cell3, and B cell4, and it also receives a
certain amount of pressure from each one. This creates a network type structure that provides a regulatory
effect on the neighboring B cells. The immune network acts as a self-organizing and self-regulatory system
that captures antigen information, and is ready to launch an attack against any similar antigens.
In the immune network, a node represents a type of antigens or antibodies, and the links between
nodes represent the affinity between them. The affinity is determined by the matching degree between
the paratope and the epitope, and the population of each type of antibodies (or antigens). An antibody
type is thought to be stimulated when its paratope recognizes the epitope of antigens or other types of
antibodies, so that the corresponding lymphocyte is stimulated to reproduce more lymphocytes, and the
lymphocytes secrete more antibodies (with a certain mutation rate). This process is called clonal selection.
On the contrary, an antibody or antigen may be suppressed if its epitopes are recognized by others [25].
The essential aspect of the immune network model is that the list of antigens and antibody types is
dynamically changing as some types that are added or removed, change with time.

23.4.3 Artificial Telecom Immune System

Using similar analogies between the immune human system and the mobile telecommunication systems,
the artificial Telecom immune system will consist of a set of B cells and the links between the B cells and
their cloned ones, and the data items represent the antigens, as well as the pathogens. These data are,
first, converted from their raw state (numerical data) to a normalized form into values within the interval
[0, 1]. In what follows, we discuss briefly the main components that shall be performed by the immune
human-based system for mobile phone operations, as outlined in Figure 23.11.
Figure 23.11 shows a simple diagram of the learning algorithm that is described hereafter:

1. Data training : The initial network B cell population is made up by sampling the raw data set and
creating the B cell objects (for instance, phone calls on a Sunday between 12 and 14 h).
2. Data items: The remainder of the data items are taken to create the antigens set. These data are
introduced during the training process and will be repeatedly presented to the network in an
attempt to capture the trends of the mobile phone users using the data set.
3. Expose: This level computes the B cell stimulation level, which takes into account the match scores
with the antigens.
4. Cloning : New clones may be produced when the cells are sufficiently stimulated.
5. Feedback: New items are introduced in the system.
6. Results: Adopting the primary immune response’s analogy, the components of the network
repeatedly present these new items for training purposes to the network, in an attempt to pro-
voke a reply from B cell, clone and mutate, thereby creating a diverse set of the raw data to keep a
network of B cells.

23.5 Conclusion
Biological inspired techniques have received a great deal of interests due to their promises to resolve
challenging combinatorial and real-world problems. In this chapter, we have reviewed two promising

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 370 — #12

Biological Inspired Based Intrusion Detection Models 23-371

Data
items

Start

Training
set
Expose data
training

Stimulate
B cell

Clone B cell

Add new
clones

Training No
finish ?

Results

FIGURE 23.11 Immune human-based intrusion detection model for mobile phone operations.

nature-based models, neural networks and immune human-based systems, and show how these two
approaches could be used to design efficient distributed security management systems for mobile phone
operations.
In the future, we plan to investigate further the immune human-based intrusion detection system and
study its performance evaluation for mobile phone operations. Next, we wish to investigate further the
design of efficient security models using biological inspired techniques for mobile ad hoc networks.

References
[1] Ottawa Citizen, October 11, 2004.
[2] Boukerche, A. and Notare, M.S.M.A. Behavior-based intrusion detection in obile phone systems.
Journal of Parallel and Distributed Computing, 62, 1476–1490, 2002.
[3] Notare, M.S.M.A., Boukerche, A., Cruz, F.A., Riso, B.G., and Westphall, C.B. Security management
against cloning mobile telephones. In IEEE GLOBECOM’99, 1999, Rio de Janeiro, Brazil. 1999,
pp. 1969–1973.
[4] Boukerche, A., Sobral, J.B.M., Juca, K., and Notare, M.S.M.A. Anomaly and misuse intrusion detec-
tion based in the immune human system. In Proceedings of the 17th IEEE IPDPS/NIDISC’2003 —
International Parallel and Distributed Processing Symposium, 2003, Nice, France. IEEE Press, 2003,
p. 146.
[5] Boukerche, A., Juca, K., Notare, M.S.M.A., and Sobral, J.B.M. Intrusion detection based on the
immune human system. In IEEE 16th IPDPS’2002/BioSP3 — International Parallel and Distributed

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 371 — #13

23-372 Handbook of Bioinspired Algorithms and Applications

Processing Symposium/Workshop on Bio-Inspired Solutions to Parallel Processing Problems,

Fort Lauderdale, FL, IEEE, 2002.
[6] Riso, B., Maciel, C., Souza, I., Rossi, L., Notare, M.S.M.A., and Tonin, N.A. Protocols Engineering
with LOTOS/ISO (in Portuguese). Florianópolis : EdUFSC, 2004. p. 215.
[7] Brinksma, E. ISO 8807: International Organization for Standardization, LOTOS — Language of
Temporal Ordering Specifications, 1988.
[8] Garavel, H. CADP: Eucalyptus Manual. Grenoble, France: INRIA/VASY, 1997. http://www.inrialpes.
fr/vasy/cadp/
[9] Notare, M.S.M.A. Conception, analysis and development of a security management system for tele-
communication networks (in Portuguese). Florianopolis, Brazil, 2000. Thesis (Computer Science
Program) — Department of Informatic and Statistic, Technological Center, Federal University of
Santa Catarina.
[10] Simon, E. Distributed Informations Systems: From Client/Server to Distributed Multimedia.
McGraw-Hill, Maidenhead, Berkshire, England, 1996, p. 414.
[11] Duda, R.O. and Hart, P.E. Pattern Classification and Scene Analysis. John Wiley and Sons, New York,
1973.
[12] Boukerche, A. and Notare, M.S.M.A. (Wiley and Sons) Applications of neural networks to mobile
communication systems. In Solutions to Parallel and Distributed Computing Problems: Lessons from
Biological Sciences. John Wiley and Sons, New York, 2001, pp. 255–268.
[13] Todesco, J.L. Pattern Recognition using Artificial Neuronal Networks with a Radial Basis Function:
an Application for a Human Chromosome Classification (in Portuguese). Florianopolis, Brazil,
1995. Thesis (Production Eng. and Systems Program) — Department of Production Engineering,
Technological Center, Federal University of Santa Catarina.
[14] NIH/NCI. Understanding the Immune System. National Institutes of Health. National
Cancer Institute. U.S. Department of Health and Human Services. Public Health Service.
http://press2.nci.nih.gov/sciencebehind/immune/immune00.htm. access 08-02-2003.
[15] Alberts, B. Molecular Biology of the Cell, 4th ed. Garland Science, Taylor and Francis Group,
New York, 2002.
[16] Farmer, J.D., Packard, N.H., and Perelson, A.S. The immune system, adaptation, and machine
learning. Physica D, 22, 187–204, 1986.
[17] Male, David. Immunology — An Illustrated Outline. Gower Medical Publishing Ltd., 1986, p. 45.
[18] Roitt, I.M. Essential Immunology, 5th ed. Blackwell Scientific Publications, 1995.
[19] Forrest, S., Perelson, A.S., Allen, L., and Cherukuri, R. Self–nonself discrimination in a computer.
In Proceedings of the 1994 IEEE Symposium on Research in Security and Privacy, 1994, pp. 202–212.
[20] Jerne, N.K. Towards a network theory of the immune system. Annals of Immunology, 125, 373–389,
1973.
[21] Perelson, A.S. Immune Network Theory. Immunological. Theoretical Division, Los Alamos
National Laboratory, NM 87545.1989. Review 110, pp. 5–36.
[22] Cayzer, S. and Aickelin, U. A recommended system based on the immune network. In Proceedings
of CEC 2002.
[23] Hofmeyer, S. An Immunological Model of Distributed Detection and Its Applications to Computer
Security. Albuquerque, 1999. Ph.D. thesis. Computer Science, University of New Mexico, p. 23
[24] Morrison, T. and Aickelin, U. An artificial immune system as a recommender for web sites.
In Proceedings of the 1st International Conference on Artificial Immune Systems (ICARIS-2002),
Canterbury, UK, 2002, pp. 161–169.
[25] Yixin, D. and Passino, K.M. Immunity-based hybrid learning methods for approximate structure
and parameter adjustment. Engineering Applications of Artificial Intelligence, 15, 587–600, 2002.
[26] Timmis, J.I. Artificial Immune Systems: A Novel Data Analysis Technique Inspired by the Immune
Network Theory. Aberystwyth. Ph.D. thesis. Department of Computer Science, University of
Wales, 2000.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C023” — 2005/8/17 — 18:14 — page 372 — #14

 24
Synthesis of
Multiple-Valued
Circuits by
Neural Networks

24.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-374

Artificial Neural Networks • Multiple-Valued Logic •
Multiple-Valued Logic Neural Networks • Problem
Statement • Related Studies • Motivations • Organization
of this Chapter
24.2 Multiple-Valued Multiple-Threshold Perceptrons . . . . 24-378
The (n, k, s)-Perceptrons • Decomposition of
(n, k, s)-Perceptrons • Multilinear Separability and
Multilinear Partition
24.3 Learning with Permutably Homogeneous
Multiple-Valued Multiple-Threshold Perceptrons . . . . 24-383
Permutably Homogeneous (n, k, s)-Perceptron Learning
Algorithm • Permutably Homogeneous (n, k, s)-Perceptrons
Convergence Properties • Determining the Output Vectors of
Permutably Homogeneous (n, k, s)-Perceptrons • Experiments
24.4 A Strip-Based Neural Network Growth Algorithm
for Learning Multiple-Valued Functions . . . . . . . . . . . . . . . 24-394
Problem Statement • Background and Motivations • Known
Partitioning Methods • Longest Strip-Based Growth
Algorithm • Determining Longest Strips by GA •
Constructing the Neural Network • Experimental Results
24.5 ES for Learning Multiple-Valued Logic Functions . . . . 24-412
Determining Longest Strips by Evolutionary Strategy • Weight
Neighborhood • Evolution Strategy • Experimental Results
24.6 Minimization of Multiple-Valued Multiple-Threshold
Perceptrons using GAs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-416
Computing Optimal s-Representations with GAs •
Experiments and Discussions
24.7 Conclusion and Open Problems . . . . . . . . . . . . . . . . . . . . . . . . 24-420
Alioune Ngom Conclusions • Open Problems
Ivan Stojmenović References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-423

24-373

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 373 — #1

24-374 Handbook of Bioinspired Algorithms and Applications

24.1 Introduction
24.1.1 Artificial Neural Networks
Several novel methods of computation that are collectively known as soft computing have recently
emerged. The raison d’être of these modes is to exploit the tolerance for imprecision and uncertainty
in real-world problems to achieve tractability, robustness, and low cost. Soft computing is usually used to
find an approximate solution to a precisely or imprecisely formulated problem. Neural computing, fuzzy
computing, and evolutionary computing are the major components of this approach.
Artificial neural networks are an attempt to mimic some or all of the characteristics of biological
neural networks. This soft computational paradigm differs from a programmed instruction sequence, in
that information is stored in form of weights. Each neuron is an elementary processor with primitive
operations, such as summing the weighted inputs coming to it and then amplifying or thresholding the
sum. Assembly of such neurons can, in principle, perform universal computations for suitably chosen
weights.
A well-known model of neuron studied extensively in Reference 1 is called the perceptron
(see Figure 24.1). The perceptron computes a weighted sum of its input signals and generates an output
of 1 if this sum is above a certain threshold t ∈ R, otherwise, an output of 0 results. In General, given a
weight vector w  = (w1 , . . . , wn ) ∈ R n and an input vector x = (x1 , . . . , xn ) ∈ R n , such neuron computes
a simple function of the form fw : R n  → S, where n ≥ 1, S ⊆ R and

fw (x ) = g (w
 x ) (24.1)

for some transfer function g : R  → S. There are two choices for the set S currently popular in literature.
The first is the discrete model with S = {0, 1}. In this case, if a neuron’s threshold is t , then its transfer
function g is a linear threshold function (see Figure 24.2[a]) defined by


0 if y < t
g (y) = (24.2)
1 if y ≥ t

and f is called a weighted linear threshold function. The second is the continuous model with S = [0, 1]. In
this case, g is typically a monotone increasing function such as the sigmoid function (see Figure 24.2[c])
given by

1
g (y) = (24.3)
1 + a −(by)

for a, b ∈ R + . The continuous model is popular because it is easier to construct. The discrete model is
popular because its behavior is easier to analyze (however, it uses more hardware). A neural network is
characterized by the network topology, the connection strength (i.e., weights) between pairs of neurons,
the neurons properties (i.e., transfer functions), and the learning algorithms.

x1
. w1
.
y
. f (x) = g (y)
.
. wn
xn

FIGURE 24.1 Perceptron.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 374 — #2

Synthesis of Multiple-Valued Circuits by Neural Networks 24-375

(a) (b) (c) (d)

FIGURE 24.2 Examples of transfer functions: (a) linear threshold function, (b) piecewise linear function, (c) sigmoid
function, and (d) Gaussian function.

(a) x1 (b) x1

x2 x2

x3 x3

FIGURE 24.3 Examples of neural networks architectures: (a) feed-forward network and (b) feed-back network.

Artificial neural networks can be viewed as weighted directed graphs in which artificial neurons are
nodes and directed edges (with weights) are connections between the neuron’s outputs and inputs. Based
on the interconnection pattern (architecture), artificial neural networks are grouped in two categories:
feed-forward networks (in which graphs have no loops and cycles) and feed-back (or recurrent ) networks
(in which loops or cycles occur because of feed-back connections). Different network topologies yield
different network behaviors and require appropriate learning algorithms. Figure 24.3 illustrates the two
types of network topologies.
A learning process in artificial neural networks is the problem of updating the connection weights so
that the network can efficiently perform a specific task. Learning can also be viewed as the problem of
minimizing an objective function which is the error between the network output and the desired output.
Efficient learning algorithms for specific topologies have been proposed in the literature.

24.1.2 Multiple-Valued Logic

Let k be a fixed positive integer and K = {0, . . . , k − 1}. A k-valued logic function f maps the Cartesian
power K n (of all ordered n-tuples of elements of K ) onto K . Denote by Pkn the set of all such functions
n
f : K n  → K . Thus, Pkn consists of k k multiple-valued logic functions. The set Pkn is called the set of n-ary
operations on K in universal algebras and the set of n-ary functions of k-valued logic in multiple-valued
logic algebras. The set Pk defined by Pk = ∪n≥1 Pkn is the set of all k-valued logic functions.
Multiple-valued logic has been the subject of research over many years [2]. Besides reduction in
chip area and interconnections, multiple-valued logic offers other benefits such as potential for image
processing and speech recognition. The development of multiple-valued logic algebras has also proceeded
since the pioneering works of Lukasiewicz [3] and Post [4]. Reference 4 presented the first functionally
complete set as basic operations in multiple-valued logic. Since then, there have been many multiple-valued
logic algebras developed [5] and, applications for image processing network and knowledge information
processing systems presented.

24.1.3 Multiple-Valued Logic Neural Networks

A discrete neuron is a processing unit whose transfer function outputs a discrete value. An example of such
transfer function is the linear threshold function. A discrete n-input multiple-valued neuron has a discrete

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 375 — #3

24-376 Handbook of Bioinspired Algorithms and Applications

transfer function and realizes a function of n variables ranging in the set S ⊆ R with values in K , that is,
computes a function f : S n ⊆ R n  → K . For S = K we refer to the processing unit as a multiple-valued
logic neuron since it simulates a multiple-valued logic function f : K n → K . Multiple-valued logic neural
networks are thus neural networks composed of multiple-valued logic neurons as processing units. The
first model of multiple-valued logic neural networks were introduced in Reference 6 and since then various
other models have been described [7–11].

24.1.4 Problem Statement

The problem we address is that of implementing multiple-valued logic functions by neural networks.
In particular, we discuss interesting models of multiple-valued logic neurons (and neural networks) and
how they might be used to learn and compute multiple-valued logic functions. In most models, we will be
interested in optimizing some identified resource of space or time such as number of thresholds, number
of neurons or size of strip. We will also be interested to study the learning abilities of our main model of
neuron: the multiple-valued multiple-threshold perceptron.

24.1.5 Related Studies

Chan et al. [6] first developed a model of multiple-valued logic neural network. In that model, neurons
simulate Kleenean functions — a special class of three-valued logic functions — using the well-known
back-propagation learning algorithm. The model can be extended to make probabilistic predictions. It can
also be used to make decisions with fuzzy input.
A design of multiple-valued logic adaptive linear neuron — adaline — is described in Reference 12.
Such adaline contains multiple-valued logic operations in place of the usual linear combiner and, thus,
has a faster operation speed and a more robust separation ability. The model uses a modified adaline
learning algorithm to learn two-valued logic functions with noisy inputs.
References 13 and 14 defined a continuous multiple-valued neural network containing piecewise linear
units as processing elements. A multiple-valued logic algebra with only three basic arithmetic operations
is introduced and is proven to be functionally complete. Several logic functions can thus be synthesized
using the algebraic system. The authors also described a learning technique, which is largely borrowed
from back-propagation training algorithm, to learn any multiple-valued logic function.
Tang et al. [10] described a learning multiple-valued logic network based on back-propagation. The
learning network is derived directly from a canonical realization of multiple-valued logic functions and
therefore its functional completeness is guaranteed.
It has been shown in Reference 15 that continuous neurons with limited precision are essentially
equivalent to discrete multiple-valued neurons. Such model of neurons uses a class of multiple-valued
multiple-threshold functions as transfer functions and hence generalizes the discrete binary perceptron
model which uses the linear threshold function. The perceptrons discussed in this chapter belong to
the more general class of multiple-valued multiple-threshold perceptrons and in Reference 16 it is given
the first instance of learning algorithm for multiple-valued multiple-threshold perceptrons.
Some results and methods are known in literature concerning the synthesis of two-valued logic functions
by feed-forward multilayer neural networks composed of discrete binary perceptrons [1, 17–19]. Minsky
and Papert [1] are the first who gave necessary and sufficient conditions for two-valued logic functions to
be linearly separable, that is, to be learned and hence simulated by discrete two-valued perceptrons.
Two-valued logic perceptrons are very closely related to threshold logic elements which have long been
investigated by many authors [20]. Perceptrons can be thought of as threshold logic elements with
learning capability. A threshold logic element forms a functionally complete set by itself and thus is used
as a building block to construct complex logic circuits implementing logic functions. Therefore, comes
the important problem of minimizing the size or depth of such threshold circuits.
Extensions of the two-valued threshold logic elements to the two-valued multiple-threshold logic elements
were also introduced and analyzed in the literature [21–23]. It is known that any two-valued logic function

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 376 — #4

Synthesis of Multiple-Valued Circuits by Neural Networks 24-377

can be implemented by such a single element. More general circuit elements called multiple-valued
multiple-threshold elements have also been studied [24–26].
Interesting methods for the synthesis of multiple-valued logic functions have been described in the
literature. These new approaches to synthesis, based on natural or physical laws, were introduced
mostly to search for a minimal circuit representation (or equivalently, a minimal logic expression) of
a given multiple-valued logic function. The only known algorithm for finding minimal multiple-valued
logic expressions is exhaustive search. The excessive computation time makes this approach impractical.
Especially, multiple-valued sum-of-products expressions are interesting because of the ease with which
they can be implemented by programmable logic arrays [26,27]. Because of the computational com-
plexity associated with minimal sum-of-products solutions, there is considerable interest in heuristics.
Typical heuristics is that, first a minterm is selected and then an implicant is chosen that covers the min-
term [26, 28]. This process is repeated until the given expression is covered. Yildirim et al. [29] proposed
multiple-valued logic design methods which employ simulated annealing [30]. Kaczmarek et al. [31]
proposed neural network techniques. Hata et al. [32] proposed solutions using genetic algorithms.
Lloris-Ruiz et al. [33] used information theoretic (entropy) approaches for the minimization of logic
expressions.

24.1.6 Motivations
References 6, 10, 12, and 14 do not discuss how to construct minimal neural networks from their respective
models. In all of these papers, the size and depth of the networks is fixed in advance. This is a major
drawback since many logic functions can be synthesized with minimal size or minimal depth networks.
The homogeneous multiple-valued perceptron learning algorithm presented in Reference 15 has the
weakness of being able to learn only a very tiny portion of the set of separable functions, so it has a
low capacity. In addtion, no learning algorithm is known, in general, for the multiple-valued multiple-
threshold perceptrons.
Multiple-valued multiple-threshold perceptrons are multiple-valued multiple-threshold logic elements
with learning abilities. A minimal multiple-valued logic perceptron which computes a given logic function
is a perceptron containing the least number of thresholds. The problem of finding such minimal perceptron
for a function is difficult and is still left open.
Finding a minimal logic expression for a given function is a very difficult problem. The simulated
annealing [29], genetic algorithms [32], and neural networks [31] based-approaches to express minimiza-
tion introduced in the literature have the tendency to produce local optimum solutions. These are still very
good solutions; however, it seems to us that better solutions can be obtained by improving these methods.

24.1.7 Organization of this Chapter

In Section 24.2, we introduce the multiple-valued multiple-threshold perceptron (or (n, k, s)-perceptrons,
for short), and define many (new) concepts (such as multilinear partition and multilinear separability,
for instance) that are used throughout this chapter. All subsequent discussions will be about this model.
The rest of the chapter is subdivided into four main sections. In Section 24.4, the learning ability of
(n, k, s)-perceptrons is examined. The previously studied homogeneous (n, k, k − 1)-perceptron learning
algorithm is generalized to the permutably homogeneous (n, k, s)-perceptron learning algorithm with
guaranteed convergence property. We also introduce a high capacity learning method that learns any
permutably homogeneously separable k-valued function given as input. In Section 24.4, we consider the
problem of synthesizing multiple-valued logic functions by neural networks. A genetic algorithm (GA)
that finds the longest strip in V ⊆ K n is described. A strip contains points located between two parallel
hyperplanes. Repeated application of GA partitions the space V into certain number of strips, each of them
corresponding to a hidden unit. We construct two neural networks based on these hidden units and show
that they correctly compute the given but arbitrary multiple-valued function. Preliminary experimental
results are presented and discussed. In Section 24.5, an evolutionary strategy (ES) that finds the longest strip

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 377 — #5

24-378 Handbook of Bioinspired Algorithms and Applications

in V ⊆ K n is described and the results obtained are compared with those of Section 24.4. In Section 24.6,
we address the problem of minimizing the size of single multiple-valued multiple-threshold perceptrons.
Every n-input k-valued logic function can be implemented using a (n, k, s)-perceptron, for some number
of thresholds s. We propose a GA to search for an optimal (n, k, s)-perceptron that efficiently realizes a
given multiple-valued logic function, that is, to minimize the number of thresholds. Experimental results
show that the GA finds optimal solutions in most cases.

24.2 Multiple-Valued Multiple-Threshold Perceptrons

Our main problem is to synthesize multiple-valued logic functions by neural networks. Most of our
research focuses on the study of the computational and learning abilities of multiple-valued multiple-
threshold perceptrons. As far as we know, this is a new model for neural networks. Some special cases of
this model are known in the literature References [9,15,16]describe learning algorithms for these cases.
Our main contribution in the domain is the development of learning algorithms for multiple-valued
multiple-threshold perceptrons and the constructions of neural networks composed of such perceptrons.
Concepts that are known for the binary perceptron such as linear partition, linear separability, and capacity
are extended to the multiple-valued perceptron.

24.2.1 The (n, k, s)-Perceptrons

Let R be the set of real numbers. Let V ⊆ R n and K = {0, . . . , k − 1}, with k ≥ 2. An n-input k-valued
real function f : V → K is a function with real-valued inputs and k-valued outputs. When V ⊆ K n , we
will refer to f as an n-input k-valued logic function. We denote by Pkn the set of all n-input k-valued real
(respective logic) functions f : V  → K and by Pk = ∪n≥1 Pkn the set of all k-valued real (logic) functions.
n
Thus, Pkn consists of k k k-valued logic functions. The set Pkn is called the set of n-ary operations on K in
universal algebras and the set of n-ary functions of k-valued logic in multiple-valued logic algebras. For
instance, for k = 2, P2 is the set of all two-valued logic functions.
A discrete neuron is a processing unit whose transfer function outputs a discrete value. An example
of such a transfer function is the linear-threshold function. A discrete n-input multiple-valued neuron
has a discrete transfer function and realizes a function of n variables ranging in a set V ⊆ R n with
values in K , that is, computes a function f : V → K . For V ⊆ K n we refer to the processing unit as a
multiple-valued logic neuron since it simulates a multiple-valued logic function f : K n → K . Our model
of multiple-valued logic neuron is the n-input k-valued s-threshold perceptron defined in the following
paragraph.
In the theory of multiple-valued logic functions there is an important class of functions called multiple-
valued multiple-threshold functions [25]. Such functions are used in the design of classes of multiple-valued
logic circuits called programmable logic arrays [26]. A k-valued s-threshold function of one variable y ∈ R
is defined as



 o0 if y < t1

t ,o
gk,s (y) = oi if ti ≤ y < ti+1 for 1 ≤ i ≤ s − 1 (24.4)



os if ts ≤ y,

where o = (o0 , . . . , os ) ∈ K s+1 is the output vector, t = (t1 , . . . , ts ) ∈ R s is the threshold vector — with
ti ≤ ti+1 (1 ≤ i ≤ s − 1) — and s (1 ≤ s ≤ k n − 1) is the number of threshold values.
Multiple-threshold devices [21] are threshold elements containing multiple levels of excitation
(thresholds). Among their qualities, given enough thresholds, a single multiple-threshold element can
realize any given function operating on a finite domain [25].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 378 — #6

Synthesis of Multiple-Valued Circuits by Neural Networks 24-379

An n-input k-valued s-threshold perceptron [8,34,35], abbreviated as (n, k, s)-perceptron, computes an

 t , o) given by
n (w,
n-input k-valued weighted s-threshold function Fk,s



 o0  x < t1 ,
if w

t ,o
n
 t , o)(x ) = gk,s (w
Fk,s (w,  x ) = oi if ti ≤ w
 x < ti+1 , (24.5)



os if ts ≤ w
 x ,

where input vector x = (x1 , . . . , xn ) ∈ V , weight vector w  = (w1 , . . . , wn ) ∈ V , threshold vector

t = (t1 , . . . , ts ) ∈ R s , and ti ≤ ti+1 (1 ≤ i ≤ s − 1 and 1 ≤ s ≤ k n − 1, the number of threshold values),
and output vector o = (o0 , . . . , os ) ∈ K s+1 .
t ,o
The perceptron’s transfer function is a k-valued s-threshold function gk,s : R  → K . An
(n, k, s)-perceptron simulates a k-valued function f : V → K . Depending on V we will refer to either real
or logic (n, k, s)-perceptrons. It is well known that any n-input k-valued logic function can be transformed
into a k-valued s-threshold function [25,36] for some s.
An (n, k, s)-perceptron is, monotone if o is monotone, that is, o0 ≤ · · · ≤ os or o0 ≥ · · · ≥ os , otherwise
it is nonmonotone. For example, the well-known binary perceptron (which is an (n, 2, 1)-perceptron)
studied in Reference 1 is monotone. An (n, k, s)-perceptron is homogeneous if s = k − 1 and o is the
identity permutation of K (i.e., oi = i for 0 ≤ i ≤ s), it is permutably homogeneous if s ≤ k − 1 and
o is any permutation of s + 1 elements out of K . Functions computed by such single neurons are called
permutably homogeneous functions.
The (n, k, k − 1)-perceptrons were introduced by Obradović and Parberry [15] as a model for sim-
ulating continuous perceptrons (i.e., perceptrons containing analog transfer functions) with limited
precision on their inputs. Their computational abilities were extensively studied in Reference 9 whereas
Reference 16 examined their learning power. Learning algorithms for some classes of (n, k, s)-perceptrons
are investigated in References 7 and 34. Takiyama [23] introduced that the (n, 2, s)-perceptrons and
their computational power have been intensively studied in References 22 and 37. The neulonets’
units discussed in Reference 6 are a class of (n, 3, 2)-perceptrons with t = (−d, +d), d ∈ R, and
o = (1, 0, 2).
Multiple-valued logic neural networks, in our definition, are thus neural networks composed of
(n, k, s)-perceptrons as processing units. It should be noted that the units are not necessarily the same.
For example, one layer of the network may contain (2, 4, 8)-perceptrons while another layer may contain
(5, 4, 3)-perceptrons.
Analog computers are inherently inaccurate due to imperfections in fabrication and fluctuations in
operating temperatures. The classical solution to this problem uses extra hardware to enforce discrete
behavior. However, the brain appears to compute reliably well with inaccurate components without
necessarily resorting to discrete techniques. The continuous neural network is a computational model
based on certain observed features of the brain. Experimental evidence has shown continuous neural
network to be extremely fault tolerant; in particular, their performance does not appear to be significantly
impaired when precision is limited. It has been shown by Obradović and Parberry [15] that analog neurons
of limited precision are essentially multiple-valued logic neurons. The first model of multiple-valued logic
neuron (and neural network) was introduced by Chan et al. [6] and since then various other models have
been described (see, for instance, References 11, 15, and 38).

24.2.2 Decomposition of (n, k, s)-Perceptrons

An (n, k, s)-perceptron can be decomposed into a two-hidden-layer network composed of n-input nodes,
that is, one linear combiner in the first hidden layer, s (n, 2, 1)-perceptrons (usual linear threshold units)
in the second hidden layer and one linear combiner in the output layer [25,36]. The transfer function
t ,o
gk,s (w
 x ) of the (n, k, s)-perceptron is expressed as the linear summation of s linear threshold functions

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 379 — #7

24-380 Handbook of Bioinspired Algorithms and Applications

(a)
os-2
os
os-1
o3

o1
o2
o0

0 y

(b)
1 g(y)1

0 y
t1

1 g(y)2

0 y
t2

1 g (y )3

0 y
t3

1 g (y)s–1

0 y
ts–1

1 g (y)s

0 y
ts

t ,o 
FIGURE 24.4 Decomposition of (a) gk,s (w
 x ) into (b) s linear threshold functions. (From A. Ngom, I. Stojmenović,
and V. Milutinović. IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

g (w
 x )1 , . . . , g (w
 x )s , that is,



s
0  x < ti ,
if w
gst ,o (w
 x ) = o0 + ai g (w
 x )i and g (w
 x )i = (24.6)
i=1
1  x ≥ ti ,
if w

where ai = oi − oi−1 is the weight associated with g (w

 x )i and 1 ≤ i ≤ s. Figure 24.4 shows an example
of such decomposition. Figure 24.5 shows the depth-two network corresponding to Equation (24.6).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 380 — #8

Synthesis of Multiple-Valued Circuits by Neural Networks 24-381

s linear threshold units

xn
t1
wn o0
a1

t2
a2
w2

x2
as

w1
ts
x1
Linear summers
t ,o 
FIGURE 24.5 Two-hidden-layers network for gk,s (w
 x ). (From A. Ngom, I. Stojmenović, and V. Milutinović. IEEE
Transactions on Neural Networks, 12, 212–227, 2002. With permission.)

24.2.3 Multilinear Separability and Multilinear Partition

The problem of computing (or simulating) a given function f by a (n, k, s)-perceptron is to determine s
and a vector r = (w,  t , o) ∈ R n+s × K s+1 , such that Fk,s n ( r )(x ) = f (x ) (∀x ∈ V ), that is, f = Fk,s n (r ).
We will refer to r as an s-representation of Fk,s for f .
n

Let V = {x1 , . . . , xv } ⊆ R n be a set of v vectors (v ≥ 1). A k-valued function f with domain V and
specified by the input–output pairs {(x1 , f (x1 )), . . . , (xv , f (xv ))}, where xi ∈ R n , f (xi ) ∈ K , is said to be
s-separable if there exist vectors w  ∈ R n , t ∈ R s , and o ∈ K s+1 , such that


 o0  xi < t1
if w

f (xi ) = oj if tj ≤ w
 xi < tj+1 for 1 ≤ j ≤ s − 1 (24.7)



os if ts ≤ w
 xi

for 1 ≤ i ≤ v. Equivalently, f is s-separable if and only if it has a s-representation defined by (w, t , o).
A k-valued function over V is said to be s-nonseparable if it is not s-separable.
In other words, an (n, k, s)-perceptron partitions the space V ⊂ R n into s + 1 distinct classes
[oj ]
H0 , . . . , Hs[os ] , using s parallel hyperplanes, where Hj = {x ∈ V |f (x ) = oj and tj ≤ w
[o0 ]
 x < tj+1 }.
We assume that, t0 = −∞ and ts+1 = +∞. Each hyperplane equation denoted by Hj (1 ≤ j ≤ s) is of
the form

 x = tj .
Hj : w (24.8)

Multilinear separability (s-separability) extends the concept of linear separability (1-separability of the
common binary one-threshold perceptron) to the (n, k, s)-perceptron. Linear separability in two-valued
case tells us that an (n, 2, 1)-perceptron can only learn from a space V ⊆ [0, 1]n in which there is a single
hyperplane which separates it into two disjoint halfspaces: H0[0] = {x |f (x ) = 0} and H1[1] = {x |f (x ) = 1}.
From the (n, 2, 1)-perceptron convergence theorem [1], concepts which are linearly nonseparable cannot
be learned by an (n, 2, 1)-perceptron. One example of linearly nonseparable two-valued logic function is
the n-input parity function. Likewise, the (n, k, s)-perceptron convergence theorems [16,34] state that an
(n, k, s)-perceptron computes a given function f ∈ Pkn if and only if f is s-separable. Figure 24.6 shows an
example of two-separable four-valued logic function of P52 .

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 381 — #9

24-382 Handbook of Bioinspired Algorithms and Applications

y
4 2 2 2 2 2

3 2 2 2 0 0

2 0 0 0 0 0

1 0 0 4 4 4

0 4 4 4 4 4
0 1 2 3 4x

FIGURE 24.6 A two-separable function of P52 . (From A. Ngom, I. Stojmenović, and J. Žunić. IEEE Transactions on
Neural Networks. 14, 469–477, 2004, Proceedings of the 29th IEEE International Symposium on Multiple-Valued Logic.
IEEE Computer Society Technical Committee on Multiple-Valued Logic, May 1999, pp. 208–213, IEEE Computer
Society. With permission.)

FIGURE 24.7 A (2, 52 , 2)-partition. (From A. Ngom, I. Stojmenović, and J. Žunić. IEEE Transactions on Neural
Networks, 14, 469–477, 2004, Proceedings of the 29th IEEE International Symposium on Multiple-Valued Logic, IEEE
Computer Society Technical Committee on Multiple-Valued Logic, May 1999, pp. 208–213, IEEE Computer Society.
With permission.)

Let V ⊂ R n with |V | = v ≥ 2. An (n, v, s)-partition is a partition of V by s ≤ v −1 parallel hyperplanes

(namely (n − 1)-planes), which do not pass through any of the v points. For instance, Figure 24.7 shows
an example of (2, 52 , 2)-partition.
An (n, v, s)-partition determines s + 1 distinct classes S0 , . . . , Ss ⊂ V separated by s parallel (n − 1)

planes such that si=0 Si = V and si=0 Si = ∅. An (n, v, s)-partition corresponds to an s-separable
k-valued function f ∈ Pkn if and only if all points in the set Si , for 0 ≤ i ≤ s + 1, have the same value
taken out of K . In addition, we assume that any two neighboring classes have distinct values. If for a given
(n, v, s)-partition we have Si  = ∅ (0 ≤ i ≤ s + 1) then, clearly, the number of associated functions is
k(k − 1)s .
A linear partition of a point set V is an (n, v, 1)-partition; therefore, only a single (n − 1) plane is
required to separate an n-dimensional space V ⊆ R n into two half spaces. The enumeration problem for
linear partitions is closely related to the efficiency measurement problem for linear discriminant functions
in pattern recognition [17] and to many other algorithmic problems [39].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 382 — #10

Synthesis of Multiple-Valued Circuits by Neural Networks 24-383

24.3 Learning with Permutably Homogeneous Multiple-Valued

Multiple-Threshold Perceptrons
The (n, k, s)-perceptrons partition the input space V ⊂ R n into s + 1 regions using s parallel hyperplanes.
Their learning abilities are examined in this section. The previously studied homogeneous (n, k, k − 1)-
perceptron learning algorithm is generalized to the permutably homogeneous (n, k, s)-perceptron learning
algorithm with guaranteed convergence property. We also introduce a high capacity learning method that
learns any permutably homogeneously separable k-valued function given as input.
A function implementable by a homogeneous (n, k, s)-perceptron is said to be homogeneously sepa-
rable (or homogeneous, for short). A function computable by an (n, k, s)-perceptron with given output
vector o is said to be o-separable. A function implementable by an (n, k, s)-perceptron whose output vector
is monotone is said to be monotoneously separable. A function computable by a permutably homogeneous
(n, k, s)-perceptron is said to be permutably homogeneously separable (or simply, permutably homoge-
neous). For instance, the functions f1 , f2 , and f3 shown in Figure 24.8 are all three-separable and, moreover,
are respectively (0, 2, 1, 3)-separable, (0, 1, 2, 3)-separable, and (3, 2, 1, 3)-separable; f1 is nonmonoto-
neously separable and permutably homogeneous; f2 is (permutably) homogeneous and monotoneoulsy
separable; f3 is nonpermutably homogeneous and nonmonotoneously separable.
Note that since o ∈ K s+1 then every o-separable function (for some o) is also s-separable. However,
the converse is not true, that is s-separability does not implies o-separability (for some o). The only case
where s-separability is equivalent to o-separability is the two-valued one-threshold case, that is, when
k = 2, s = 1, and o = (0, 1) or (1, 0). Every one-separable two-valued logic function is (0, 1)-separable,
and also, every (0, 1)-separable two-valued logic function is one-separable.

24.3.1 Permutably Homogeneous (n, k, s)-Perceptron Learning Algorithm

In this section we propose a learning algorithm for permutably homogeneous (n, k, s)-perceptrons.
An (n, k, s)-perceptron is permutably homogeneous if its output vector is a (s + 1, k)-permutation. To
the best of our knowledge there are no known learning algorithms for (n, k, s)-perceptrons in general.
Our algorithms, when applied to the special case s = k − 1, that is, the (n, k, k − 1)-perceptrons, are more
powerful than the homogeneous (n, k, k − 1)-perceptron learning algorithm described in Reference 16,
in that they can learn a larger class of multiple-valued logic functions.
A permutation does not need to contain all values of K . There are k-valued logic functions whose set
of output values is a subset S ⊆ K , that is, functions of the form f : K n → S ⊆ K . Examples of such
functions are, for instance, functions of the form f : K n → {0, 1}, that is, functions with k-valued inputs
and two-valued outputs.

y y y
f1(x, y) f2(x, y) f3(x, y)
3 1 1 3 3 3 2 2 3 3 3 2 2 3 3

2 2 1 1 3 2 1 2 2 3 2 1 2 2 3

1 0 2 1 1 1 0 1 2 2 1 3 1 2 2

0 0 0 2 1 0 0 0 1 2 0 3 3 1 2
x x x
0 1 2 3 0 1 2 3 0 1 2 3

FIGURE 24.8 Examples of three-separable two-input four-valued logic functions. (From A. Ngom, C. Reischer,
D.A. Simovici, and I. Stojmenović. Neural Processing Letters, 12, 2000, Proceedings of the 28th IEEE International
Symposium on Multiple-Valued Logic, May 1998, pp. 161–166. With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 383 — #11

24-384 Handbook of Bioinspired Algorithms and Applications

FIGURE 24.9 Permutably homogeneous (n, k, s)-perceptron learning algorithm. (From A. Ngom, C. Reischer,
D.A. Simovici, and I. Stojmenović. Neural Processing Letters, 12, 2000, Proceedings of the 28th IEEE International
Symposium on Multiple-Valued Logic, May 1998, pp. 161–166. With permission.)

Let o ∈ K s+1 be the output vector of an (n, k, s)-perceptron. When o is a (s + 1, k)-permutation, that is,
there are no i and j (i  = j) such that oi = oj , we propose the permutably homogeneous (n, k, s)-perceptron
learning algorithm (for a fixed (s + 1, k)-permutation o) as shown in Figure 24.9.
In Figure 24.9, the constant 0 < η ≤ 1 is the learning rate. The initial weights can be set to any
(random) values. The initial thresholds can also be set to any (random) values, however, empirical tests
show that the algorithm converges faster when the initial thresholds are set in such a way that ti+1 − ti = c
(e.g., c = kn) and that tv−1 ≤ tv ≤ tv+1 each time we update tv or tv+1 . We can also generate a new
random η before each call to MultiPerceptronUpdate. Poso [z] is the position (or the index) of z in o. For
example, if o = (3, 0, 2, 1) then Poso [3] = 0, Poso [0] = 1, Poso [2] = 2, and Poso [1] = 3.
In the algorithm, the weight and the threshold vectors are always updated in opposite directions
using the error value δ = Poso [f (x )] − Poso [v]. If Poso [f (x )] < Poso [v] then δ < 0 means that the
weights are too large or tPoso [v] is too small. Therefore, we decrease the weights and increase tPoso [v] .
If Poso [f (x )] > Poso [v] then δ > 0 means that the weights are too small or tPoso [v]+1 is too large.
Thus, we increase the weights and decrease tPoso [v]+1 . When Poso [f (x )] = Poso [v] no modification is
done and the algorithm goes to the next step. So, w  and t are always updated in opposite directions
given by the position of f (x ) relative to that of v in o. Note that o is known and given as input to the
algorithm.
Our algorithm can learn any o-separable logic function as long as o is a fixed (s + 1, k)-permutation.
In other words, the algorithm can learn any function whose input vectors can be separated by a set of paral-
lel hyperplanes (i.e., s-separable, for some s) and whose classes — separated by these hyperplanes — have
distinct values (i.e., o-separable, for some (s + 1, k)-permutation o). In fact, the permutably homogeneous
algorithm generalizes the homogeneous algorithm in which o is any permutation (such permutation need
to be the identity permutation or a k-permutation).
When o is not a permutation, that is, there are i and j (i  = j) such that oi = oj , then it becomes
difficult to obtain a learning algorithm with guaranteed convergence. This problem is left open for further
research.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 384 — #12

Synthesis of Multiple-Valued Circuits by Neural Networks 24-385

24.3.2 Permutably Homogeneous (n, k, s)-Perceptrons

Convergence Properties
Given a (s + 1, k)-permutation o, experiments show that the permutably homogeneous (k, s)-perceptron
learning algorithm always converges for o-separable functions. That is, given the appropriate output vector
any permutably homogeneous k-valued logic function will be learned. In this section we give a formal
proof of convergence of the algorithm given in Figure 24.9.
The latency (or delay) of a learning algorithm is the worst case running time between the output of one
set of assignments and the next. We will assume unit-cost latency; that is, we will assume that the algorithm
is implemented on a digital computer with word-size large enough that each elementary arithmetic and
logic operation can be implemented in constant time. The mistake bound is the worst case total number
of distinct assignments outputs. The latency and the mistake bound of the permutably homogeneous
(n, k, s)-perceptron learning algorithm are, respectively, O(n) and (k n ).
Let π = (π(0), . . . , π(k − 1)) be a k-permutation. The identity permutation is denoted by σ . Thus,
a homogeneous (n, k, s)-perceptron is a neuron whose output vector is σ and whose s = k − 1. It has
been proven in Reference 16 that the homogeneous (n, k, k − 1)-perceptron learning algorithm always
terminates in learning homogeneous k-valued logic functions.
Let f ∈ Pkn be a homogeneous function, then we will denote by fπ ∈ Pkn the function obtained by
permuting the output values of f with respect to π . That is, we define the homogeneous transformation
fπ (x ) = π( f (x )) for all x ∈ K n . Clearly, fσ = f , and the function fπ is a permutably homogeneous
function. In fact, the set of all permutably homogeneous (k − 1)-separable functions can be constructed
in this way from the set of all homogeneous functions.
Lemma 24.1 Let f be homogeneous and π be a k-permutation. Then rσ = (w,  t , σ ) is a
 t , π ) is a (k − 1)-representation for fπ .
(k − 1)-representation for f if and only if rπ = (w,
Proof: ⇒) If rσ is a (k − 1)-representation for f then, for x ∈ K n , we have fπ (x ) = π( f (x )) =
t ,
σ t ,π
π(Fk,k−1
n (rσ )(x )) = π(gk,k−1 (w
 x )) = π(σ (i)) = π(i) = gk,k−1 (w
 x ) = Fk,k−1
n (rπ )(x ) (if ti ≤ w
 x ≤ ti+1
with t0 = −∞ and tk = +∞). So rπ is a (k − 1)-representation for fπ . ⇐) Similar proof using the fact
that, for z ∈ K , we have σ (z) = π −1 (π(z)) = z where the permutation π −1 is the inverse of π (π −1 is
guaranteed to exist and is unique since the set of all permutations on K forms a group).

Lemma 24.1 tells us that, given a (k − 1)-representation (w,  t , σ ) for a homogeneous function f , the
homogeneous transformation of f into a permutably homogeneous function fπ leaves the weights and
the thresholds invariant. Therefore, the positions of the k − 1 separating parallel hyperplanes do not
change after transformation of f . Clearly, in Figure 24.8, the three hyperplanes remain invariant even after
transformation of f2 into f1 = f2(0,2,1,3) and vice versa.
Theorem 24.1 Given the output vector π,  the permutably homogeneous (n, k, k − 1)-perceptron learning
algorithm for learning a function f ∈ Pkn terminates if and only if f is π -separable.
Proof: ⇒) If the permutably homogeneous (n, k, k − 1)-perceptron algorithm with output vector π
terminates on learning f then a (k −1)-representation (w,  t , π ) exists for f . Therefore, f is π-separable
 and
thus permutably homogeneous. ⇐) Let f be π -separable, we want to show that the algorithm terminates
for f . The algorithm, instead of learning f , learns fπ −1 using the homogeneous (n, k, k − 1)-perceptron
learning algorithm with output vector π −1 . Since f is permutably homogeneous and π -separable then
from the ⇐ part of Lemma 24.1 we have that fπ −1 is homogeneous and thus σ -separable. Once fπ −1 has
been learned, f can be reconstructed using the ⇒ part of Lemma 24.1. The algorithm is guaranteed to
terminate by the homogeneous (n, k, k − 1)-perceptron convergence theorem of Reference 16.

 then we clearly have Poso = π −1 and therefore the algorithm truly

If in Figure 24.9 we replace o by π,
learns fπ −1 and reconstruct f using the parameters found for fπ −1 .
These results concern output vectors which are k-permutations of K , that is, full permutations. Next,
we consider the case of (s +1, k)-permutations for fixed s ≤ k −1. In the sequel we let ν = (ν(0), . . . , ν(s))

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 385 — #13

24-386 Handbook of Bioinspired Algorithms and Applications

be a (s + 1, k)-permutation and ν c = (ν c (0) = ν(0), . . . , ν c (s) = ν(s), ν c (s + 1), . . . , ν c (k − 1)) be the

k-permutation (full permutation) obtained by adding k − s coordinates to ν .

Lemma 24.2 f is ν -separable if and only if it is ν c -separable.

Proof: Clearly, if r = (w,  t = (t1 , . . . , ts ), ν ) is a s-representation for f , then it is easy to see that
 t c = (t1c = t1 , . . . , tsc = ts , ts+1
r c = (w, c , . . . , t c ), ν
k−1  ) is a (k − 1)-representation for f and vice versa
c

(where ts+1 ≤ · · · ≤ tk−1 are arbitrary and their corresponding hyperplanes contain no points between
c c

them).

Theorem 24.2 (Permutably homogeneous perceptron convergence theorem) Given the output
vector ν , the permutably homogeneous (n, k, s)-perceptron learning algorithm for learning a function f ∈ Pkn
terminates if and only if f is ν -separable.

Proof: ⇒) Same as in Theorem 24.2 ⇐) The algorithm learns f using the output vector ν c instead
of ν . Since from Lemma 24.2 f is ν c -separable, then by Theorem 24.2 the algorithm is guaranteed to
terminate.

24.3.3 Determining the Output Vectors of Permutably Homogeneous

(n, k, s)-Perceptrons
A more difficult learning problem is when the output vector o is not known and that it should be
determined along with w  and t . In this section we let o be a (s + 1, k)-permutation. An obvious solution
to this problem is to generate each (s + 1, k)-permutation p and apply the learning algorithm with o = p .
This method takes O(enk n (k!/(k − s − 1)!)) time complexity (where e is the number of learning epochs)
and thus is nonrealistic for even small values of k. A better method is, for a given function f ∈ Pkn , to
search for a partial order relation defined over f ’s values and, if such relation exists then we search for
a good linear extension of the corresponding partially ordered set and use it as the output vector of an
(n, k, s)-perceptron. Before we describe the algorithm we need a few definitions.
A partially ordered set (P, <P ) or (poset ) is a set P = {a0 , . . . , ap−1 } equipped with an irreflexive,
antisymmetric, and transitive relation <P . A linear extension L of P is a linear (or total) ordering <L of
the elements of P such that ai <L aj whenever ai <P aj . For example, L = {1 <L 0 <L 3 <L 2 <L 4}
and L = {4 <L 1 <L 3 <L 0 <L 2} are two linear extensions of the poset shown in Figure 24.10. We also
define a covering relation ≺P between two elements of P: x ≺P y if and only if x <P y and there is no z
such that x <P z <P y.

2 4 3

0 01

FIGURE 24.10 Example of partially ordered set. (From A. Ngom, C. Reischer, D.A. Simovici, and I. Stojmenović.
Neural Processing Letters, 12, 2000, Proceedings of the 28th IEEE International Symposium on Multiple-Valued Logic,
May 1998, pp. 161–166. With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 386 — #14

Synthesis of Multiple-Valued Circuits by Neural Networks 24-387

FIGURE 24.11 Extended (n, k, s)-perceptron learning algorithm. (From A. Ngom, C. Reischer, D.A. Simovici,
and I. Stojmenović. Neural Processing Letters, 12, 2000, Proceedings of the 28th IEEE International Symposium on
Multiple-Valued Logic, May 1998, pp. 161–166. With permission.)

Our extended permutably homogeneous (n, k, s)-perceptron learning algorithm, for searching an out-
put vector o ∈ K s+1 and then learning a given function f using o, is shown in Figure 24.11. For f ∈ Pkn ,
let Kf be its set of values. Denote by <dKf an order relation over Kf with respect to the d-th variable, that
is xd , where 1 ≤ d ≤ n. We will refer to d as direction since, as we will see later, it selects the dimension of
the n-cube K n along which we construct a poset.
The extended learning algorithm goes as follows. Using the partial order construction algorithm of
Figure 24.12 we attempt to construct a poset (Kf , <dKf ) with respect to some variable xd . If such (Kf , <dKf )
exists and is a chain then o is the concatenation of the unique linear extension of (Kf , <dKf ) and a
(s − |Kf | + 1, |K − Kf |)-permutation of K − Kf and it will be used to learn f . If such (Kf , <dKf )
exists but is not a chain then we attempt to obtain (Kf , <d+1 Kf ) and so on until either we construct a
chain poset (Kf , <dKf ) for some d, or there is some d such that a poset (Kf , <dKf ) cannot be obtained, or
d = n. When n nonchain posets (Kf , <1Kf ), . . . , (Kf , <nKf ) are constructed then, using the partial orders
combination algorithm of Figure 24.13, we attempt to combine these n nonchain posets into a chain poset
(Kf , <Kf ).
The partial order construction algorithm goes as follows. For a given direction 1 ≤ d ≤ n, we start with
an antichain (Kf , <dKf ). Then, for every x = (x1 , . . . , xn ), we construct poset (Kf , <dKf ) by adding new
comparable pairs f1 = f (x1 , , . . . , xd , . . . , xn ) <dKf f (x1 , . . . , xd + 1, . . . , xn ) = f2 whenever f1 ≤dKf f2 ,
and also, we add new comparable pairs y <dKf f2 whenever y <dKf f1 and comparable pairs f1 <dKf y
whenever f2 <dKf y, for some y. We exit the loops as soon as there is some new comparable pair y <dKf z
(for some y and z) that cannot be added to (Kf , <dKf ). That is, z <dKf y is already in (Kf , <dKf ) and,
therefore, adding its inverse leads to inconsistency. In this case, the construction of (Kf , <dKf ) cannot be
completed along the direction d (meaning that (Kf , <dKf ) simply do not exist). In case (Kf , <dKf ) exists —
its construction can be completed along d — then it has a unique linear extension if and only if it is a
chain. In other words, (Kf , <dKf ) is always constructed in the positive direction along the d-th dimension
of the n-cube K n , or equivalently, starting from any point y in the hyperplane xd = 0 we move toward the
hyperplane xd = k − 1 by following the line segment orthogonal to both hyperplanes and whose origin
is y . Figure 24.14 shows examples of constructed posets. For illustration purpose, we have also shown the
graphs obtained from function g (x, y) (Figure 24.6[b]) when attempting to complete the construction

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 387 — #15

24-388 Handbook of Bioinspired Algorithms and Applications

FIGURE 24.12 Partial order construction algorithm.

with possible contradictory pairs. As one can see, such graph cannot be embedded into a poset. So posets
(Kg , <1Kg ) and (Kg , <2Kg ) do not exist.
For given permutably homogeneous and s-separable function f ∈ Pkn , not any linear extension e of a
nonchain (Kf , <dKf ) is good for learning. The (n, k, s)-perceptron learning algorithm may not terminate
when e is used. For instance, from Figure 24.6(c) the linear extension (3, 0, 1, 2) of (Kh , <1Kh ) is not good for
learning h since h is (2, 0, 1, 3)-separable (assuming s = 3). Some directions may give more information
on order than others. For example, f (x1 , x2 ) = (x1 + 1) mod k is irrelevant on x2 (or direction 2) and
so (Kf , <2Kf ) is an antichain whereas (Kf , <1Kf ) is a chain. In general, given a direction d there are two
possibilities for failure. Either (Kf , <dKf ) cannot be constructed, or (Kf , <dKf ) exists but is a nonchain poset,
such that a selected linear extension (among its many linear extensions) does not yield a convergence of the
(n, k, s)-perceptron learning algorithm (but some other will do so). Because of this fact, when a nonchain
poset (Kf , <dKf ) is obtained for any direction 1 ≤ d ≤ n, we must combine these n posets in some way in
order to obtain a unique linear extension. Next we describe how to combine them.
The partial orders combination algorithm goes as follows. Let (Kf , <Kf ) be a combination poset of
d consistent nonchains posets (Kf , <1Kf ), . . . , (Kf , <dKf ). Initially (Kf , <Kf ) is set to (Kf , <1Kf ) and is
constructed according to some binary string c = [c1 , . . . , cd ] ∈ {0, 1}d , where 1 ≤ d ≤ n. When
ci = 1 then the inverse of (Kf , <iKf ), that is poset (Kf , >iKf ), is in (Kf , <Kf ), otherwise (Kf , <iKf ) itself

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 388 — #16

Synthesis of Multiple-Valued Circuits by Neural Networks 24-389

FIGURE 24.13 Partial orders combination algorithm.

FIGURE 24.14 Constructed partial orders for some f ∈ P42 .

is in (Kf , <Kf ) (obviously, (Kf , <iKf ) and (Kf , >iKf ) cannot both be in (Kf , <Kf ) at the same time). The
combination poset (Kf , <Kf ) is constructed using an algorithm for generating binary strings of lengths ≤n
in lexicographic order. For example, for n = 4, the lexicographic generation of binary strings of lengths ≤4
goes in the following manner:

0 00 000 0000
0001
001 0010
0011
01 010 0100
0101
011 0110
0111
1 ...
..
.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 389 — #17

24-390 Handbook of Bioinspired Algorithms and Applications

FIGURE 24.15 Extension algorithm.

(a) y f (x, y) Direction = 1 Direction = 2

3 2 2 1 1 1 1

2 0 2 2 1 2 2

1 3 0 0 2 0 0

0 3 3 0 0 3 3
x
0 1 2 3

(b) y
g (x, y) 1 1

3 2 2 1 1 3
2
2
2 0 2 3 1
0 1
0
1 3 1 0 2 3
1

0 3 3 0 0 3 0
x
0 1 2 3
(c) y
h (x, y)

3 3 1 1 1
2 1 3

2 1 1 1 1

1
1 1 1 1 1

0 3 0 3
0 0 0 0 2
x
0 1 2 3

FIGURE 24.16 Cutting algorithm. (From A. Ngom, C. Reischer, D.A. Simovici and I. Stojmenović. Neural Pro-
cessing Letters, 12, 2000, Proceedings of the 28th IEEE International Symposium on Multiple-Valued Logic, May 1998,
pp. 161–166. With permission.)

The algorithm is in ExtendPoset phase when it goes from left to right staying in a row (Figure 24.15).
It is in CutPoset phase when the algorithm shifts to some row (possibly far) below (Figure 24.16). The
algorithm is used to construct poset (Kf , <Kf ) as follows. If with the string [c1 , . . . , cd ] poset (Kf , <Kf )
exists but is a nonchain for d < n then we extend to next string [c1 , . . . , cd cd+1 = 0] in the row and
add poset (Kf , <d+1
Kf ) into (Kf , <Kf ). If with the string [c1 , . . . , cd ] poset (Kf , <Kf ) cannot be constructed
then we do not need to extend since poset (Kf , <Kf ) simply do not exist and that we cannot add a poset

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 390 — #18

Synthesis of Multiple-Valued Circuits by Neural Networks 24-391

Direction = 1 Direction = 2

2 1 3

y h (x, y)
1
3 3 1 1 1

0 3 0 3
2 1 1 1 1

1 1 1 1 1 Combination = 00 Combination = 01
1 2
0 0 0 0 2
0 1 2 3 x
3 0 2

1
1

0 2 3

FIGURE 24.17 Examples of combinations posets for Figure 24.14c. (From A. Ngom, C. Reischer, D.A. Simovici,
and I. Stojmenović. Neural Processing Letters, 12, 2000, Proceedings of the 28th IEEE International Symposium on
Multiple-Valued Logic, May 1998, pp. 161–166. With permission.)

to an undefined poset. Hence, in this case we can bypass the lexicographic generation of binary strings
to an appropriate point: we say that the algorithm is in CutPoset phase. We cut in the following manner.
Starting from position d of c we search for the first position r ≤ d such that cr = 0. We remove posets
(Kf , <rKf ) and (Kf , >r<i≤d
Kf ) from (Kf , <Kf ) and add poset (Kf , >rKf ) into (Kf , <Kf ), then finally, we set
d to r and cd to 1.
To summarize, with a given string [c1 , . . . , cd ] we have three possibilities. We generate the next
string whenever d < n and (Kf , <Kf ) is a nonchain poset (the extension phase) and update (Kf , <Kf )
accordingly. We bypass the lexicographic generation to some row below whenever (Kf , <Kf ) cannot be
constructed (the cutting phase) and compute (Kf , <Kf ) appropriately. We exit the algorithm as soon as
(Kf , <Kf ) is a chain or c1 = 1 or, (Kf , <Kf ) is a nonchain and d = n. In the first case, we learn f using
the unique linear extension of (Kf , <Kf ). In the second case, when no chain poset (Kf , <Kf ) is found, we
may either randomly select one poset (Kf , <iKf ) and look for a good linear extension of it to learn f , or
we select among all posets constructed so far the one which has the smallest width (since it will have the
smallest number of linear extensions to search); the selection can be done by computing the width of the
currently constructed consistent poset and keeping track of the smallest width (and storing the associated
poset). The width of a poset is the size of its longest antichain.
In addition, we do not need to continue generating new binary strings when c1 becomes 1. Because
they are symmetric to (i.e., complement of) those generated already (the poset constructed according to a
string c is dual to the poset constructed according to the complement of c, and hence, both posets behave
exactly the same way). See Figure 24.17 for examples of combination posets. Next, we explain how to add
or remove from a combination poset (Kf , <Kf ).
Given (Kf , <dKf ) to be added to (Kf , <Kf ) the addition (Kf , <Kf ) ⊕ (Kf , <dKf ) is defined by
(Kf , <Kf ) ⊕ (Kf , <dKf ) = (Kf , <Kf ∪γ (<dKf )), where γ (<dKf ) is the transitive closure of every compa-
rable pair of relation <dKf in relation <Kf . That is, O(|Kf |2 ) comparabilities from (Kf , <dKf ) are added to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 391 — #19

24-392 Handbook of Bioinspired Algorithms and Applications

(Kf , <Kf ) during addition. In addition, for every such comparability from (Kf , <dKf ) its transitive closure
in (Kf , <Kf ) is also added, that is, O(|Kf |) more comparabilities. In sum, operation ⊕ take O(|Kf |3 ) steps.
Given a poset (Kf , <dKf ) to be removed from poset (Kf , <Kf ), the substraction (Kf , <Kf )  (Kf , <dKf )

is defined by (Kf , <Kf )  (Kf , <dKf ) = [c1 ,...,ci ,...,cd−1 ] (Kf , <iKf ) (where (Kf , <iKf ) is reversed when
necessary). That is, to remove (Kf , <dKf ) from (Kf , <Kf ) is equivalent to restore (Kf , <Kf ) in the state it
was before (Kf , <dKf ) was added to it. To achieve efficiency, substraction operation is done in the following
way. Whenever we add a poset (Kf , <dKf ) in (Kf , <Kf ) we store into a separate data structure Rd all
comparabilities of (Kf , <dKf ) that are not in (Kf , <Kf ). So that when we later remove (Kf , <dKf ) from
(Kf , <Kf ) we will eliminate only comparabilities of Rd from (Kf , <Kf ). The ⊕ operation modified in this
way still operates in O(|Kf |3 ). O(|Kf |2 ) comparabilities of Rd are removed from (Kf , <Kf ). Therefore,
substraction takes O(|Kf |2 ) steps hence faster than addition. Inconsistency and unicity can be tested,
respectively, in O(1) and O(|Kf |) during addition.
In Figure 24.14 we show examples of constructed posets (Kf , <1Kf ) and (Kf , <2Kf ) for some f ∈ P42 .
Suppose s = 3. Poset (Kf , <1Kf ) and (Kf , <2Kf ) are chains (Figure 24.14[a]), so they have unique linear
extensions and thus f is permutably homogeneous and three-separable and can be learned. In an attempt
to construct (Kg , <1Kg ) and (Kg , <2Kg ) we obtain graphs (Figure 24.14[b]) that cannot be embedded into
posets because of inconsistencies, so g is not permutably homogeneous and three-separable and thus g
cannot be learned. Poset (Kh , <1Kh ) and (Kh , <2Kh ) are both nonchains (Figure 24.14[c]), so they have
many linear extensions and h is permutably homogeneous and three-separable; however, we do not know
which linear extensions are good for learning f , so we must combine (Kh , <1Kh ) and (Kh , <2Kh ) to search
for a unique linear extension. In Figure 24.17 we show two combination posets (Kh , <Kh ) according to
binary strings 00 and 01. As we can see, with string 00 poset (Kh , <Kh ) cannot be obtained because of
inconsistencies whereas with string 01 it is a chain.
A thick s-separable function is a function f ∈ Pkn for which the distance between any two neighboring
separating hyperplanes, in any direction, is strictly greater than one.

Theorem 24.3 If a permutably homogeneous function f ∈ Pkn is thick s-separable then (Kf , <dKf ) is a chain
for any 1 ≤ d ≤ n.

Proof: Let a and b be two neighboring distinct values connected by an edge. There is at least one separating
hyperplane between them. However, there is at most one separating hyperplane since, otherwise, two such
separating hyperplanes will be at distance strictly less than one along the dimension of that edge. Thus,
a ≺dKf b or b ≺dKf a, that is a and b are neighbors in the poset. All such neighboring pairs are detected in
at least one dimension. Therefore (Kf , <dKf ) has a unique linear extension.

For some nonthick s-separable functions, all combination posets (Kf , <Kf ) may have many linear
extensions and the last repeat loop of the algorithm in Figure 24.13 can be modified as follows to make
it more efficient. In parallel using several processors, we generate each linear extension of (Kf , <dKf )
and test it for learning, until one processor succeeds. This can be simulated on one processor by
time sharing, that is, generate linear extensions and test each of them for the same time in succes-
sion, until one successfully terminates. Next, we discuss the time complexity of the extended learning
algorithm.
The worst case scenario, in terms of time complexity, for the partial order construction algorithm is
when there is no contradiction for a given direction d. Therefore, the while loop associated with the selected
variable xd will be iterated k − 1 times and the n − 1 remaining for loops associated with nonselected
variables will be iterated each k times. Also, each of the two inner for loops will be iterated |Kf | times
and it takes O(|Kf |) steps to test whether (Kf , <dKf ) is a chain. Therefore the partial order construction
algorithm has a time complexity of O(k n |Kf |).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 392 — #20

Synthesis of Multiple-Valued Circuits by Neural Networks 24-393

The worst case scenario for the partial orders combination algorithm is when there is no contradiction
for d < n but always contradiction for d = n. Therefore, 2n − 2 combination posets are construc-
ted each by either extension or cutting and then O(|Kf |!) linear extensions are checked for learning f .
Extension and cutting involve ⊕ operations and tests for unicity and inconsistency. Cutting is slower
than extension since it also involves  operations and a search for the first bit equals to 0 (starting
from the end of the current string). Therefore, extension and cutting take, respectively, O(|Kf |3 ) and
O(n|Kf |2 ) steps. The (n, k, s)-perceptron learning algorithm takes O(enk n ) steps (e is the number of
learning epochs) and thus the partial orders combination algorithm has O(2n n|Kf |2 + (s + enk n )|Kf |!)
time complexity. Since in practice e is large and that 2n ≤ k n and |Kf |2 ≤ |Kf |! then the complexity
becomes O(enk n |Kf |!).
The worst case scenario for the extended learning algorithm is when poset (Kf , <dKf ) is a nonchain
for any direction d. So, n posets are constructed and combined. Consequently, the extended learning
algorithm has O(nk n |Kf | + enk n |Kf |!), that is, O(enk n |Kf |!) time complexity.
Recall the first method: generate each (s + 1, k)-permutation p and apply the (k, s)-perceptron learning
algorithm with output vector o = p for learning f until the learning terminates for some permutation p .
This method takes O(enk n (k!/(k −s −1)!)) time complexity. Let us refer to it as the permutation generating
learning algorithm. Next, we compare the extended algorithm to the permutation algorithm.
First, note that the time complexity of the permutation generating algorithm is always the same for any
function. That is not true for the extended algorithm. For instance, for any nonpermutably homogeneous
function f ∈ Pkn the extended algorithm takes O(nk n |Kf |) steps; the algorithm takes O(enk n ) steps for
any permutably homogeneous thick s-separable function. The worst time complexity is achieved only for
permutably homogeneous nonthick s-separable functions f whose any combination poset (Kf , <Kf ) is
a nonchain or cannot be constructed. We believe that the probability to obtain such function f is very
close to zero (if not equal to zero), so that in practice, the extended learning algorithm runs in O(enk n )
for permutably homogeneous s-separable functions. This proves its superiority over the permutation
generating learning algorithm.

24.3.4 Experiments
We tested our extended learning algorithm on nonpermutably homogeneous functions and on permutably
homogeneous thick or nonthick functions. However, we could not obtain nonthick functions whose
combination posets are all nonchains. This suggests that such functions are very rare if not inexistent.
The nonthick functions we used have at least one chain combination poset. In our test we set the learning
rate η to 0.5 and the maximum number of learning epochs e to 5000. We experimented with different
values of n and k. Also, the number of threshold s was not given to the learning algorithm, it was to be
found by the algorithm itself. The initial weigth vector is set to 0 and the initial threshold vector is set to
(k n , 2k n , . . . , sk n ) after s was found.
For nonpermutably homogeneous functions, the algorithm behaved as expected, that is, no learning is
effected on these functions. For permutably homogeneous (thick or nonthick) functions the algorithm
always terminated after learning the function with its unique linear extensions.
Next, we discuss an example of nonthick function which we have used in our experiment for k = 4 and
n = 3. Consider the two-place function h shown in Figure 24.17. To obtain a three-place function f we
project the values h (which will correspond to points in plane x3 = 0) in the three planes x3 = 1, x3 = 2,
and x3 = 3. So, f is also a nonthick function as h. Now, to make it more difficult, to find one of its good
linear extensions, we replace the value 3 that lies in plane x3 = 0 by value 1, and also change the value 2
that lies in plane x3 = 3 to 0. Here, the function f has no unique linear extension at all in any direction
and thus the extended learning algorithm must combine the three constructed posets to search for a good
linear extension. The algorithm did indeed, as we expected, find a chain poset which has the unique linear
extension (2, 0, 1, 3). The function has been learned successfully in 61 learning epochs. We also obtain
same results when extending f to an (n ≥ 3)-place functions.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 393 — #21

24-394 Handbook of Bioinspired Algorithms and Applications

Examples of permutably homogeneous thick functions are given by the following formula:

 

n
1
f (x ) = xi + n mod k,
ai
i=1

where ai = 2i + 1. For example, we tested with the four-place four-valued logic function f (x ) =
x1 /3 + x2 /5 + x3 /7 + x4 /9 + 4 mod f 4. Clearly, such function is permutably homogeneous since it
defines itself its separating hyperplanes and their number. It is easy to see that the function has three
possible values, namely 0, 1, and 2 and thus there must be 2 separating hyperplanes, also, the three classes
of input are separated in the order (0, 1, 2). Therefore, we expect that our extended learning algorithm
will find two separating hyperplanes and the output vector (0, 1, 2). Indeed, the function was learned
successfully in 946 learning epochs after the algorithm has found the output vector.

24.4 A Strip-Based Neural Network Growth Algorithm for

Learning Multiple-Valued Functions
We consider the problem of synthesizing multiple-valued logic functions by neural networks. A GA which
finds the longest strip in V ⊆ K n is described. A strip contains points located between two parallel
hyperplanes. Repeated application of GA partitions the space V into certain number of strips, each of
them corresponding to a hidden unit. We construct two neural networks based on these hidden units
and show that they correctly compute the given but arbitrary multiple-valued function. Preliminary
experimental results are presented and discussed.

24.4.1 Problem Statement

The problem we address in this section is that of learning multiple-valued logic functions using minimal
neural networks composed of (n, k, s)-perceptrons. Multilayer feed-forward neural networks are in prin-
ciple able to learn any arbitrary mapping, provided that enough hidden units are present [1]. For these
networks, learning algorithms such as back-propagation [40] have been found to be computationally
prohibitive. Also, the topology of the networks must be fixed before learning.
Till recently, architecture design is still very much a human expert’s job. It depends heavily on the
expert experience and a tedious trial and error process. There is no systematic way to design a near
optimal architecture for a given task automatically. Research on destructive and constructive algorithms
represents an effort toward the automatic design of architectures [41]. Design of the optimal architecture
for a neural network can be formulated as a search problem in the architecture space where each point
represents an architecture. Given some performance criteria about architectures, the performance level
of all architectures forms a discrete surface in the space. The optimal architecture design is equivalent to
finding the highest point on this surface.
A way to improve the performance of a neural network is to match its topology to a specific task
(i.e., a set of input–output pairs) as closely as possible. However, the problem of deciding whether or not
a given task can be performed by a given architecture is known to be NP-complete [42]. In addition, it
has been shown in Reference 27 that the problem of finding the absolute minimal architecture for a given
task is NP-hard. A third problem known to be NP-complete [19] and which also concerns us here is the
following. Given two subsets S1 , S2 ⊆ R n such that |S1 ∪ S2 | = c, determine subsets E1 ⊆ S1 and E2 ⊆ S2
such that E1 ∪ E2 is of maximum cardinality, and for some w  ∈ R n and t ∈ R, w
 x − t > 0 for all x ∈ E1
and w  x − t ≤ 0 for all x ∈ E2 . This problem is a generalized version of the maximum cardinality problem
(i.e., find a linearly separable subset of maximum cardinality). When applied to multiple-valued logic,
these three problems are also NP-complete since they are known to be NP-complete in the synthesis of
two-valued logic functions, which are special case of multiple-valued logic functions.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 394 — #22

Synthesis of Multiple-Valued Circuits by Neural Networks 24-395

Our approach to the problem’s solution is discussed in Section 24.4.4. The learning method is based
on the general principle of partitioning algorithms discussed in Section 24.4.3. A partitioning algorithm
seeks to construct a minimal network by partitioning the input space into classes that are as large as
possible. Each class of partition is then assigned to a new hidden unit. The connections and weights of
the new units are determined appropriately in such a way that the constructed network will always give
the correct answer for any input. Distinct partitioning algorithms differ in the way the input space is
partitioned. Also, network topologies obtained from different partitioning algorithms may differ in the
way new hidden units are connected.

24.4.2 Background and Motivations

Chan et al. [6] described three-valued logic networks called neural logic networks, which combine the
strengths of neural networks and expert systems. Such networks are used to represent a set of nonre-
cursive propositional rules, and to infer the truth values of the unknown propositions by applying the
common back-propagation training method. Tang et al. [10] proposed a multiple-valued logic network
with functional completeness properties and learning capabilities. The multiple-valued logic network
consists of layered arithmetic piecewise linear units. Since the arithmetic operations of the network are
basically wired sums and piecewise linear operations, their implementation should be rather simple and
straightforward. In Wang and shi [11], neural networks composed of a novel model of multiple-valued
logic neuron suitable for representing arbitrary multiple-valued logic expressions are described. Obradović
and Parberry [16] described algorithms for learning multiple-valued logic functions on either a single
homogeneous (n, k, k − 1)-perceptron or a depth-two network composed of k (n, 2, 1)-perceptrons in
the hidden layer and one homogeneous (k, k, k − 1)-perceptron in the output layer. Ngom et al. [34]
introduced learning algorithms for permutably homogeneous (n, k, s)-perceptrons and proved them to
be more powerful than the algorithms mentioned in Reference 16. These are a few mentioned methods
for learning multiple-valued logic functions; the reader can refer to Reference 8 for a survey on multiple-
valued logic neural networks. The minimal size of a neural network for learning a given but arbitrary
multiple-valued logic function has not been studied in the literature. That is, the network is fixed in
advance before learning and its size does not change during and after learning. As stated earlier, such
networks may not be powerful enough for learning or may overfit the training data. In particular, there
is no evidence of any function for which the size of the learning network is significantly smaller than the
size of its sum-of-products or other standard gates implementations.
Techniques are known in the literature for learning multiclass functions (multiple-valued logic func-
tions are special cases of multiclass functions). The most powerful one is the error-correcting output
codes (or ECOC) approach which is a robust method for solving multiclass learning problems. Dieterich
and Bakiri [43] has shown that ECOC learning provides a general purpose method for improving the
performance of inductive learning programs on multiclass problems. Other simpler and less powerful
methods use back-propagation (or any appropriate learning algorithm such as madaline or radial basis
function, etc.) networks with multiple output units assigned to distinct classes. All these approaches use
fixed-architecture networks for learning arbitrary multiclass functions and thus a given network may not
be the optimal one for such functions.

24.4.3 Known Partitioning Methods

In Marchand et al. [44], given a function f ∈ P2n , the cube {0, 1}n is partitioned into regions by hyperplanes
in such a way that f is constant in each region containing input vectors. The halfspaces defined by these
hyperplanes correspond to threshold units in the hidden layer. The construction of the halfspaces implies
that one can add an output unit in such a way that the resulting neural network indeed computes f . The
hyperplanes are determined sequentially. First, an hyperplane is found such that f is constant on one of
its sides. The points in the corresponding halfspace are removed and they continue with the remaining
points in a similar manner, until the set of remaining points becomes empty. Thus, each halfspace is

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 395 — #23

24-396 Handbook of Bioinspired Algorithms and Applications

used to cut off a set of points with identical function values from the remaining set of points. Let the
halfspaces constructed by their algorithm be H1 , . . . , Hr and define ui to be 0 (or 1) if f is 0 (or 1) on the
region cut off by Hi for 1 ≤ i ≤ r. Then adding an output unit with threshold 0 and weight ui 2r−i for
the edge leaving the hidden unit corresponding to Hi , they get a neural network that computes f . They
assume that linear threshold units produce an output 0 or 1. In a restricted version of their approach
called regular partitioning, Ruján and Marchand [45], the hyperplanes do not intersect. This description
gives the general principle of partitioning algorithms. Particular implementations depend on the way the
next hyperplane is selected and the way new hidden units are connected (in term of network weights
and topology). Experiments indicate that partitioning algorithms are successful in the sense that they
efficiently construct (near) minimal neural networks [44–47].
Another way to view this process, without any reference to neural networks, is to consider a sequence
of halfspaces H1 , . . . , Hr . For each halfspace Hi we specify the value ui of f for those points in Hi that
are not contained in any of the previous halfspaces. Thus, for every input vector x ∈ {0, 1}n it holds that
f (x ) = ui , where i is the smallest index such that x ∈ Hi . It is assumed that Hr contains {0, 1}n , thus i is
always defined. This model is called a linear decision list [47, 48].
More formally, Rivest [48] defines a linear decision list in the following way. A linear test L over the

variable x = (x1 , . . . , xn ) ∈ {0, 1}n is of the form ni=1 wi xi ≥ t , where w1 , . . . , wn ∈ R n are the weights
and t ∈ R is a threshold. A linear decision list D over x is a sequence (L1 , u1 ), . . . , (Lr , ur ) where Li is
a linear test and ui is 0 or 1 for 1 ≤ i ≤ r. It is assumed that Lr , the last linear test is true for all input
vectors x . The length of D is r. The two-valued function fD computed by D assigns to every x the value
ui , where Li is the first linear test in the list that is satisfied by x . Every two-valued function f ∈ P2n can
be computed by some linear decision list. For instance, a disjunctive normal form with m terms can be
represented by a linear decision list of length m + 1.
Many heuristics such as Fahlman’s cascade correlation algorithm [49], Frean’s upstart algorithm [50],
Sethi’s entropy nets [51], Sirat’s neural trees [52], Mezard’s tiling algorithm [53], Barkema’s patch algorithm
[54], Frattale-Mascioli’s oil-spot algorithm [55], Young’s carve algorithm [56], regular partitioning [45], and
other partitioning algorithms [44,47,57–60], are proposed as approaches for building networks which are
(near) minimal for a given arbitrary task. These heuristics are known as constructive or growth algorithms
since they all construct a network starting from a fixed small number of units.
Partitioning techniques such as cascade correlation [49], upstart algorithm [50], and entropy nets [51]
apply only for functions with Boolean-valued outputs. Moreover, most of the growth algoritms described
in literature (see, for instance, References 44, 45, 53–55, 58 and 59) are only applicable to problems with
Boolean-valued inputs. Very few constructive methods deal with multiclass problems (k-valued functions
are multiclass functions). For instance, the carve algorithm [56] and Marchand’s neural decision list
Reference [47] both apply to functions f : R n → K .
The techniques in References 47 and 56 seek to identify the largest subset of input vectors of same class
that is separable from vectors of other classes. The hyperplane determined by the maximum separable
subset is then assigned to a newly created (n, 2, 1)-perceptron. A drawback of Marchand’s neural decision
list (NDL) [47] is that, since it employs linear programming to determine whether specific subsets of the
input vectors are linearly separable, the number of possible subsets of points that could be considered
for linear separability is exponential in the size of the inputs set. In order to circumvent this problem,
Marchand and Golea [47] worked only with a specific class of functions called halfspace intersections.
Young’s CARVE [56], which is an extension of Marchand et al. [44] to multiclass functions with real-
valued inputs, avoids the issue of testing for linear separability of subsets of points by directly searching
for hyperplanes that separate sets of points of one class only. Let Si be a set of points of class i and Hi (the
hull set for Si ) be the set of all points in the training set except those in Si , i ∈ K . Clearly, points outside
the convex hull formed by Hi are all of the same class i and thus, only points outside the convex hull
can actually be separated by a hyperplane boundary from the hull set. So to find the maximum separable
subset, CARVE considers only hyperplanes that touch the boundary (i.e., (n − 1)-dimensional faces or
vertices) of the convex hull. The hyperplane with the largest set of points in its open halfspace outside
the convex hull is the maximum separable subset. The algorithm is a simple hill-climbing technique that

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 396 — #24

Synthesis of Multiple-Valued Circuits by Neural Networks 24-397

searches for a (near) optimal hyperplane by partially traversing (or covering) the convex hull (i.e., the
boundary of the hull set) starting from a randomly selected convex hull vertex and a hyperplane that passes
through that vertex. This is repeated a fixed number of times nv and each time, the initial hyperplane
is randomly rotated a fixed number of times nr around the boundary of the hull set. The set of points
encountered during each rotation is determined; these are potential solutions. The main difficulty with
CARVE is that the likelihood to find the maximum separable subset depends on the parameters nv and nr .
For larger values of the product nv nr , the more likely is one to obtain a (near) optimal hyperplane since
large section of the hull set boundary will be traversed (it should be noted that the number of facets and
vertices of the convex hull may be exponential in the dimension of the input space). The optimal values
for nv and nr depend on the actual problem and the user must find such values by trial and error. In the
end, the size of the network obtained by CARVE depends on nv and nr .
In this section we introduce a method of partitioning a set V ⊆ K n using GA [61] to grow a
multiple-valued logic neural network for learning a function. In our own approach, the maximum
separable subset is treated as a special case of the longest strip (as will be discussed later). We introduce
(n, k + 1, 2)-perceptrons which will freely reduce the network size for given arbitrary k-valued functions.
Both CARVE and NDL use local search to search for good solutions. They extensively search some
parts of the space (around faces of a convex hull as in CARVE, or, around neighborhood of separating
hyperplanes as in NDL), while leaving some other parts of the search space untouched. On the other hand,
GA (in our implementation) performs a global search; it treats all parts of the space equally, due to its
implicit parallelism.

24.4.4 Longest Strip-Based Growth Algorithm

In this section, we describe a novel neural network growth algorithm based on the search for the longest
strip rather than the maximum separable subset. Our technique can be applied to various kinds of
functions. That is, it is capable of handling (1) real-valued and k-valued inputs (k ≥ 2) and (2) two-class
problems as well as multiclass problems. In particular, we apply our growth algorithm to the synthesis of
multiple-valued logic functions, that is, functions of the form f : K n → K .
A strip is a set of points between two parallel hyperplanes which have the same value. The longest strip
is the strip with the maximum possible cardinality. A maximum separable subset is a set of points having
equal values with the maximum possible cardinality that can be separated from all other points by exactly
one hyperplane. Examples of longest strip and maximum separable subset are shown, respectively, in
Figure 24.18 and Figure 24.19.
We describe a search method for finding the longest strip as well as the maximum separable subset
of a currently given set of training examples, namely GA. The maximum separable subset problem is
a special case of the longest strip problem. Indeed, they both share the same problem representation
used by GA, but differ only by their respective objective functions and constructed network architectures.
In the next paragraph we briefly describe our main growth algorithm which constructs a network by

3 1 1 3 3

2 0 1 1 2

1 0 1 1 1

0 3 0 0 1
0 1 2 3 x

FIGURE 24.18 Example of longest strip for k = 4 and n = 2. (From A. Ngom, I. Stojmenović, and V. Milutinović.
IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 397 — #25

24-398 Handbook of Bioinspired Algorithms and Applications

3 1 1 3 3

2 0 1 1 2

1 0 1 1 1

0 3 0 0 1
x
0 1 2 3

FIGURE 24.19 Example of maximum separable subset for k = 4 and n = 2. (From A. Ngom, I. Stojmenović, and
V. Milutinović. IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

FIGURE 24.20 A-based synthesis algorithm.

removal of points in a predefined objective subset A ⊆ V . A is either the longest strip or the maximum
separable subset in V . Our main objective is, using GA, to obtain a subset G ⊆ V such that |G| is
as close as possible to |A| if not equal to |A| (of course we have |G| ≤ |A|). Our maximum separable
subset problem approach is an extension of the sequential learning algorithm described in Reference 44
to multiclass functions f : R n  → K . The growth algorithm of Reference 44 is based on the perceptron
learning algorithm (more specifically, on the pocket algorithm) and its performance was hampered by the
fact that the pocket algorithm does not converge in the case where the points are not linearly separable.
In our own approach, however, we use an evolutionary approach to find optimal subsets.
In our particular implementation of partitioning algorithm (see Figure 24.20), GA is used to obtain
subsequent halfspaces delimited by either one or two hyperplanes (depending on the predefined objective
subset A). To each halfspace we assign a hidden unit that correctly classifies all elements of it. Let the
objective subset A be the longest strip in the given training set. Our growth algorithm begins with an
empty first hidden layer into which new (n, k + 1, 2)-perceptrons are inserted one after another until no
more insertion is possible. An (n, k + 1, 2)-perceptron implements two parallel hyperplanes in the input
domain and the aim is to find two parallel hyperplanes that define a strip G such that |G| is as close as
possible to |A|. The strip G is then removed from the training set. The next (n, k + 1, 2)-perceptron to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 398 — #26

Synthesis of Multiple-Valued Circuits by Neural Networks 24-399

be added to the network aims to separate another (near) longest strip G, but now only from the reduced
training set. Once a (near) longest strip for this unit is found, the unit is added to the layer and the strip is
removed from the training set. The construction of the first hidden layer continues with each subsequent
unit separating a strip from the remaining training examples. The first hidden layer is complete when
only points of one class remains in the training set. Once the first hidden layer is complete, the remaining
weights, layers and units of the networks are determined to complete the network construction (the details
of the network architecture are described in Section 24.4.6).
Our principal objective in this section is to synthesize any arbitrary k-valued logic function by a
neural network constructed by our growth algorithms when (portion of) the function is given. We obtain
networks that either exactly or approximately implement k-valued logic functions.

24.4.5 Determining Longest Strips by GA

We use an evolutionary method to find the set of partitioning hyperplanes. Holland [62] first proposed GAs
in the early 1970s as computer program to mimic the evolutionary processes in nature. GAs manipulate
a population of potential solutions to an optimization (or search) problem. Specifically, they operate on
encoded representations of the solutions, equivalent to the genetic material of individuals in nature, and
not directly on the solutions themselves. Holland’s GA encodes the solutions as binary chromosome (strings
of bits). As in nature, selection provides the necessary driving mechanism for better solutions to survive.
Each solution is associated with a fitness value that reflects how good or bad it is, compared with other
solutions in the population. The higher the fitness value of an individual, the higher its chances of survival
and reproduction and the larger its representation in the subsequent generations. Recombination of
genetic material in GAs is simulated through a crossover mechanism that exchanges portions between two
chromosomes. Another operation, mutation, causes sporadic and random alterations of the chromosomes.
Mutation too has a direct analogy from nature and plays the role of regenerating lost genetic material and
thus reopening the search.

24.4.5.1 Problem Representation

Fundamental to the GA structure is the encoding mechanism for representing the problem’s variables.
For the problem of determining A, the search space is the power set of the training set (i.e., the set of all
subsets of K n ). Each element of the search space is a potential solution and must be represented in such a
way that meaningful and suitable genetic operators can be designed for (and applied to) it.
More formally, our population consists of chromosomes which encode the potential solutions of the
problem. A potential solution is a subset G ⊆ K n and the best solution is one whose size is closest to |A|.
Given a weight vector w  = (w1 , . . . , wn ) ∈ R n we can find the strip (or separable subset) of maximum
cardinality that is associated with w  (see Section 24.4.5.2). For instance, the strip shown in Figure 24.18
can be obtained given the vector w  = (1, 1) and the function’s table (moreover in this example, the
obtained strip is the absolute solution). Given a training set, then searching for a subset whose size is
as close to |A| as possible is equivalent to searching the weight vectors space (i.e., R n ) for a vector that
generates such subset. In our problem representation, a chromosome will be a weight vector w  ∈ R n (as in
Reference 63) since such a vector clearly encodes (i.e., represents) a potential solution to the problem. Each
chromosome will uniquely determine a partition of V ⊆ K n into s + 1 classes with s parallel hyperplanes
(for some s) and the best chromosome is the one that maximizes the number of points between a pair of
parallel hyperplanes. To determine how good a solution the GA needs is an objective function to evaluate
each chromosome w. 
We initialize the population with random real-coded chromosomes w  ∈ R n whose coordinates are
random real numbers taken from the interval [−1, 1]. Each initial chromosome is then normalized to
a unit vector. Another method we used for the initialization of the population is to set wi = cos αi
(for 1 ≤ i ≤ n) for each vector w,  where αi is a random number in the interval [−π/2, π/2]. Initial
population should consist of random unit hyperplanes w.  Starting from the initial population, we apply
genetic operators such as selection, crossover, and mutation to create the next generation of chromosomes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 399 — #27

24-400 Handbook of Bioinspired Algorithms and Applications

This generational cycle is repeated until a predefined maximum number of generations is reached. Subset
G will be the best solution generated so far in the population.
Our main objective here is, for a given function f , to obtain a chromosome w  which generates G such
that |G| ≈ |A|. Once such w  is found we create a hidden unit to be inserted in the neural network. We then
eliminate all points x in G and again apply the GA on the remaining points. The algorithm terminates as
soon as there are no points left. The created hidden units will then be collected to construct a feed-forward
network. The parameters (weight, threshold, and output vectors) of the hidden units and the topology of
the network will be discussed later.
24.4.5.2 Fitness Function
The objective function, the function to optimize, provides the mechanism for evaluating each
chromosome.
Let S ⊆ K n be the set of remaining points. Initially, S = K n . To compute the longest strip generated by
 we calculate for every x ∈ S the value w
w,  x and construct a sorted list of records of the form (w  x , f (x )).
The list is sorted using w  x as primary key and f (x ) as secondary key. Let these records be sorted as follows:
x1 , . . . , x|S| , or more precisely, Pi = (w xi , f (xi )), 1 ≤ i ≤ |S|, where w
 xi ≤ · · · ≤ w
 x|S| . A strip in S is a
(f (xi ))
sequence Tw = Pi Pi+1 . . . Pi+j such that
1. f (xi ) = f (xi+1 ) = · · · = f (xi+j )
 xi−1  = w
2. w  xi and w
 xi+j  = w
 xi+j+1
with 1 ≤ i ≤ |S| and 0 ≤ j ≤ |S| − i. The length of the strip is j − i + 1 and f (xi ) is the value of the strip.
For example, in Figure 24.18 we have w = (1, 1) and

P1 = (0, 3) P2 = (1, 0) P3 = (1, 0) P4 = (2, 0)

P5 = (2, 0) P6 = (2, 1) P7 = (3, 1) P8 = (3, 1)
P9 = (3, 1) P10 = (3, 1) P11 = (4, 1) P12 = (4, 1)
P13 = (4, 1) P14 = (5, 2) P15 = (5, 3) P16 = (6, 3)

(1)
which gives the longest strip T(1,1) = P7 P8 P9 P10 P11 P12 P13 generated by w.

Given a set of points S ⊆ K n and a function f over S, let P1 · · · Pj1 and Pj2 · · · P|S| (1 ≤ j1 < j2 ≤ |S|)
be, respectively, the leftmost and rightmost strips generated by w, with strip values c1 and c2 . We denote by
 the length of the longest strip generated by w
L(S, w)  and denote by M (S, w)
 the length of the maximum
between the leftmost and rightmost strips, on set S and function f . To evaluate how good is w  we propose
the following fitness function with respect to the definition of A:
• A = longest strip


L(S, w)
Fitness1L (w)
 = (24.9)
|S|

• A = maximum separable subset

M (S, w)

Fitness1M (w)
 = (24.10)
|S|

Let Sc = {x ∈ S|f (x ) = c, c ∈ K }, that is, the set of points of value c. An alternative objective is to
(c) (c) (c) (c)
select a strip of value c, that is, Tw , which maximizes |Tw |/|Sc |, where |Tw | denotes the length of Tw .
That is, as in References 44 and 56, the selection criterion chooses the strip that constitutes the largest
proportion of a class of points that can be separated. We denote by L(Sc , w)  the length of the largest strip

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 400 — #28

Synthesis of Multiple-Valued Circuits by Neural Networks 24-401

proportion of value c and denote by L(Sc1 , w)  and L(Sc2 , w)

 the lengths of the leftmost and rightmost
strips, respectively, on set S and function f . Our alternative fitness function with respect to A is

• A = largest strip proportion

 

L(Sc , w)
Fitness2L (w)
 = max (24.11)
|Sc |

• A = largest separable proportion

 
 L(Sc2 , w)
L(Sc1 , w) 
Fitness2M (w)
 = max , , (24.12)
|Sc1 | |Sc2 |

where the maximum is over S for every class c presents in the training set. As it was stated in Reference 56,
choosing the largest proportion rather than the largest set do have some advantage for some functions
(v )
f : S → K . For, if the number of points of a class v1 is small and a strip Tw 11 constituting the whole class
(v )

Sc1 is found, it may be that a longer strip Tw 22 with a smaller proportion L(Sv2 , w)/|S v2 | can be found.
(v )
However, it is preferable to select Tw 11 because this removes the entire set Sv1 from S and brings the neural
network construction closer to the hidden layer termination criteria of having only points of one class
remaining in the training set. This is illustrated in Figure 24.21 where f ∈ P32 is a random function to
which both fitness selection criteria were applied.
As seen in Figure 24.21, it is impossible to obtain a network with exactly three hidden units (which is
the absolute minimum here) when using Fitness1L . The number in circles indicates the order in which
a generated strip is assigned to hidden units. So in Figure 24.21[b]) a fourth hidden unit is needed for
the last remaining point of value 2. In Figure 24.21(a), Fitness2L removes the set S1 immediately after
removing S0 (unlike in Figure 24.2[b] where a proper subset of S2 is removed after S0 ), hence one more
unit is needed to remove S2 .
A note on the time complexity of the evaluation function. For a given w,  both fitness functions take
n|S| steps to compute the w  x ’s, n|S| log |S| steps to sort them and at most |S| steps to compute L(S, w)
 or
 Therefore, the evaluation of Fitness(1 or 2)(L or M ) (w)
M (S, w).  has a time complexity of O(n|S| log |S|).
In addition, crossover and mutation operations below take O(n) steps each and the initialization of the
population takes O(pnk n log k n ) steps (p is the number of chromosomes and all initial chromosomes are
evaluated for their fitness). Thus, the evaluation of Fitness(w)  is the most expensive operation in our GA.
Let g be the number of generations, then at each new generation p/2 new chromosomes are evaluated for
their fitnesses and hence, our GA has a time complexity of O(gpn|S| log |S|) ≈ O(gpn 2 k n log k).

(a) 2 (b) 2 3

2 2 1 2 2 1

1 0 0 0 0 0 0 1

2 1 2 2 1 2

FIGURE 24.21 Behaviors of (a) Fitness2L and (b) Fitness1L on some f ∈ P32 . (From A. Ngom, I. Stojmenović, and
V. Milutinović. IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 401 — #29

24-402 Handbook of Bioinspired Algorithms and Applications

24.4.5.3 Crossover
Crossover is the GA’s crucial operation. Pairs of randomly selected chromosomes are subjected to
crossover. For our problem representation we propose the following mixed crossover method for
real-coded chromosomes as described in Reference 64. Let p1 and p2 be two unit vectors to be crossed over
and let c1 and c2 be the result of their crossing. Vectors c1 and c2 are obtained using, with equal probability,
two of the following three crossovers operations:

ci = p1 + p2 , s (24.13)

ci = p1 − p2 , (24.14)


p1j if random() ≤ 0.5,
cij = (24.15)
p2j otherwise.

Crossover in Equation (24.13) is simply the addition of two parents and the child is assured to be their
exact middle vector since the parents are unit vectors. Crossover in Equation (24.14) is the substraction of
two parents and the child is the vector orthogonal to the sum of its parents. Crossover in Equation (24.15)
is a uniform crossover of two parents, that is, at coordinate i each parent has 50% chances to be selected
as cij (1 ≤ j ≤ n). For a more efficient search, c1 and c2 must not be obtained from the same cross-
over operator. That is, if c1 is obtained using Equation (24.13) then c2 must be generated from either
Equations (24.14) or (24.15); this helps to maintain a certain level of diversity among chromosomes in the
population. Also, crossover is applied only if a randomly generated number in the range 0 to 1 is less than
or equal to the crossover probability pcros (in large population, pcros gives the fraction of chromosomes
actually crossed).
We must emphasize that each chromosome is a unit vector at any moment in the population. Thus,
the initial random vectors are all normalized and the childs are also normalized to unit vectors after any
crossover or mutation operation.
24.4.5.4 Mutation
After crossover, chromosomes are subjected to random mutations. We propose three methods of
coordinate-wise mutations as described in Reference 64. They correspond to bitwise mutation for binary
chromosomes. Let p be a unit vector to be mutated to a child c .
Random replacement. With some probability of mutation, each coordinate pi (1 ≤ i ≤ n) of a parent p
may be replaced in the following way:

ci = random[−1, 1], (24.16)

where random[−1, 1] returns a random real number in the interval [−1, 1] with uniform
probability.
Orthogonal replacement. With some probability of mutation, each coordinate pi (1 ≤ i ≤ n) of a parent
p may be replaced in the following way:

ci = ± 1 − pi2 . (24.17)

Neighborhood replacement. With some probability of mutation, each coordinate pi (1 ≤ i ≤ n) of a

parent p may be replaced in the following way:

m
ci = pi ± , (24.18)
kn

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 402 — #30

Synthesis of Multiple-Valued Circuits by Neural Networks 24-403

where m ≤ k is a random constant. Unlike the two previous methods of mutation, this method
 to a neighboring one.
slightly rotates the current hyperplane w
Just as pcros controls the probability of crossover, the mutation rate pmuta gives the probability for a
given coordinate to be mutated. For a vector to be mutated, one of the three mutation operators is selected
with probability 13 .
Here, we treat mutation only as a secondary operator with the role of restoring lost genetic material
or generating completely new genetic material which may be probably (near) optimal. Mutation is not a
conservative operator, it is highly disruptive. Therefore, we must set pmuta ≤ 0.1.

24.4.6 Constructing the Neural Network

In the rth iteration of the A-based synthesis algorithm, GA will find a chromosome w  r which generates
a subset Gr . Subset Gr is the longest strip (or maximum separable subset) found in the population. The
optimization ability of GA makes it possible to attain a solution as close (in size) to |A| as possible, if not
equal. The ability of GA to produce |A| depends on its many control parameters and the complexity of
the tasks to learn.
1. Longest strip-based network. At every iteration r of the A-based synthesis algorithm, GA finds a
chromosome w  r which produces the longest strip Gr = Pi · · · Pi+j , where 1 ≤ i ≤ |Sr |, 0 ≤ j ≤ |Sr | − i
and strip value vr = f (xi ). Let ur = vr + 1. Then, we create an (n, k + 1, 2)-perceptron (hidden unit Ur )
whose weight vector is w  r , threshold vector is tr = (w
 r xi , w
 r xi+j+1 ), and output vector is or = (0, ur , 0).
(w
 r xi ,w
 r xi+j+1 ),(0,ur ,0)
In other words, the perceptron has a transfer function of the form gk+1,2 : R → {0, ur } (i.e.,
a (k + 1)-valued two-threshold function). The (n, k + 1, 2)-perceptron will output the value ur for all
points x ∈ Gr and will output the value 0 for all points x ∈ Sr − Gr .
In order to achieve a good accuracy on the testing set, that is, a good generalization ability of our
algorithm when approximating a function, we set the threshold vector to tr = (w  r xi − τ1 , w
 r xi+j+1 + τ2 ).
Thus, test points of value vr = f (xi ) which are outside but close to the strip — that is, test points that lie
between w  r xi−1 and w
 r xi and between w
 r xi+j+1 and w
 r xi+j+2 — will be correctly classified by unit Ur ,
since they are now spanned by Ur . The offsets τ1 and τ2 are given by

|(w
 r xi − w r xi−1 )|  r xi+j+2 − w
|(w  r xi+j+1 )|
τ1 = and τ2 = . (24.19)
2 2

2. Maximum separable subset-based network. At every iteration r of the A-based synthesis algorithm,
GA finds a chromosome w  r which produces a maximum separable subset Gr = P1 · · · Pj1 , or Gr =
Pj2 · · · P|Sr | (i.e., the maximum between the leftmost and the rightmost strips), where 1 ≤ j1 < j2 ≤ |Sr |
and strip value vr = f (x1 ) or f (xj2 ). Let ur = vr + 1. We then create an (n, k + 1, 1)-perceptron (hidden
unit Ur ) whose weight vector is w  r , threshold vector is tr = (w
 r xj1 +1 + τ2 ) if Gr is the leftmost strip, or
tr = (w  r xj2 −τ1 ) if Gr is the rightmost strip, and output vector is or = (ur , 0) or or = (0, ur ) depending on
(w
 r xj +1 +τ2 ),(ur ,0)
Gr . In other words, the perceptron has a transfer function of the form gk+1,11 : R → {0, ur } or
(w
 r xj −τ1 ),(0,ur )
gk+1,12 → {0, ur } (i.e., a (k + 1)-valued one-threshold function). The (n, k + 1, 1)-perceptron
:R
will output the value ur for all points x ∈ Gr and will output the value 0 for all points x ∈ Sr − Gr . Offsets
τ1 and τ2 are determined as earlier.
After defining all units U1 , . . . , Ur (where r is the number of runs of the A-based synthesis algorithm),
the next step is to construct a feed-forward multilayer neural network. We propose two network topologies.
24.4.6.1 Three Hidden Layers and r + k + 2 Units Architecture
The network in Figure 24.22 (which shows the case for strip-based method) has three hidden layers
and r + k + 2 neurons. Hidden layer 1 contains the units (the Ui s) obtained by the GA. Each unit
is connected to the inputs and their parameters (weight, threshold, and output vectors) are defined as

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 403 — #31

24-404 Handbook of Bioinspired Algorithms and Applications

tj = j + 1 1 ⱕ j ⱕ k–1
oj = j 0 ⱕ j ⱕ k–1

1 1 1

t=0
0 1 1 1 o = 0 or 1
0 0
–1 –j –k

ti = (k + 1)i 1 ⱕ i ⱕ r –1
oi = (k + 1)i 0 ⱕ i ⱕ r –1

(k +1)r–1 (k +1)r–1 (k +1)0 (k +1)0

(k +1)r–i
(k +1)0
(k +1)r–1 (k +1)r –i (k +1)r –i

(k +1)0

(k +1)r –1
(k +1)r –i
v1 vi vr 1 ⱕ vi ⱕ k
U1 0 0 Ui 0 0 Ur 0 0 1ⱕi ⱕr

FIGURE 24.22 Three hidden layers and r + k + 2 units network. (From A. Ngom, I. Stojmenović, and V. Milutinović.
IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

described previously. So the units in this layer are all either (n, k + 1, 2)-perceptrons or (n, k + 1, 1)-
perceptrons, depending on the definition of A, and there are r such units (Figure 24.22 shows the case
A = longest strip).
Hidden layer 2 has only one unit which is a (r, (k + 1)r−1 + 1, r − 1)-perceptron. Its weight vector
w = ((k + 1)r−1 , (k + 1)r−2 , . . . , (k + 1)0 ), that is, wi = (k + 1)r−i for 1 ≤ i ≤ r; its threshold vector
t = ((k + 1)1 , (k + 1)2 , . . . , (k + 1)r−1 ), that is, ti = (k + 1)i for 1 ≤ i ≤ r − 1; and its output
vector o = ((k + 1)0 , (k + 1)1 , . . . , (k + 1)r−1 ), that is, oi = (k + 1)i for 0 ≤ i ≤ r − 1. All units of layer 1
are connected to this unit and the connection weight vector is w. 
Hidden layer 3 contains k units. Each unit of layer 1 and 2 is connected to every unit in this layer. Each
unit is an ordinary linear threshold element (thus o = (0, 1)) and the connection weight vector from
layer 1 to that unit is the same as the connection weight vector from layer 1 to the unit at layer 2. The
connection weight wi,r+1 (1 ≤ i ≤ k) from layer 2 to the ith unit in layer 3 is −i. The threshold of units
in layer 3 are all set to 0.
The output layer has one unit which is a (k, k, k − 1)-perceptron whose threshold vector t = (2, . . . , k),
that is, ti = i + 1 for 1 ≤ i ≤ k − 1, and output vector o = (0, . . . , k − 1), that is, oi = i for 0 ≤ i ≤ k − 1
(or equivalently, oi = ti − 1). The connection weight from a unit in layer 3 to the output unit is 1.

24.4.6.2 One Hidden Layer and r + 1 Units Architecture

The network in Figure 24.23 is equivalent to Marchand’s construction [44] but generalized to multiclass
functions. It has one hidden layer and r +1 neurons. Hidden layer 1 is same as in Figure 24.22 (Figure 24.23
shows the case of maximum separable subsets).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 404 — #32

Synthesis of Multiple-Valued Circuits by Neural Networks 24-405

t = 1(k +1) ,0..., k(k +1) 0, ..., 1(k +1)r –1 , ..., k(k +1)r –1

o = 0, ..., k –1, 0, ..., k –1, 1..., 0, ..., k –1

(k +1)0
(k +1)r –1

(k +1)r–i

v1 vi vr
U1 0 Ui 0 Ur 0

FIGURE 24.23 One hidden layer and r + 1 units networks. (From A. Ngom, I. Stojmenović, and V. Milutinović.
IEEE Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

TABLE 24.1 Network Complexities

Networks #Layers #Nodes #Thresholds #Connections

SEPAR d2 2 r +1 (k + 1)r (n + 1)r

STRIPd2 2 r +1 (k + 2)r (n + 1)r
CARVE [56] 2 r +k r +k (n + k)r
SEPAR d4 4 r +k +2 2r + 2k − 2 (n + k + 1)r + 2k
STRIPd4 4 r +k +2 3r + 2k − 2 (n + k + 1)r + 2k
NDL [47] r +1 r +k r +k 1 r 2 + (n + k − 3 )r
2 2

Source: From A. Ngom, I. Stojmenović, and V. Milutinović. IEEE Transactions on Neural

Networks, 12, 212–227, 2001. With permission.

The output layer has only one unit which is a (r, k(k + 1)r−1 + 1, kr)-perceptron. Its weight vector
 = ((k + 1)r−1 , (k + 1)r−2 , . . . , (k + 1)0 ), that is wi = (k + 1)r−i for 1 ≤ i ≤ r; its threshold vector
w
t = (1(k + 1)0 , . . . , k(k + 1)0 , 1(k + 1)1 , . . . , k(k + 1)1 , . . . , 1(k + 1)r−1 , . . . , k(k + 1)r−1 ); and output
vector o = (0, . . . , k − 1, 0, . . . , k − 1, . . . , 0, . . . , k − 1). All units of layer 1 are connected to this unit and
the connection weight vector is w. 

24.4.6.3 Networks Complexities

Table 24.1 shows the complexity of our constructed networks (given f : R n → K ) along with other
networks that deal with multiclass functions, namely the NDL network of Reference [47] and the CARVE
network of Reference [56]. In the table, r is the number of created hidden units and #Layers does not
include the input layer. Also, STRIP (respectively, SEPAR) are our networks based on strips (respectively,
maximum separable subsets) and the superscript d4 (or d2) specifies the architecture of Figure 24.22
(or Figure 24.23).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 405 — #33

24-406 Handbook of Bioinspired Algorithms and Applications

As seen from the table, STRIPd2 , CARVE, and STRIPd4 networks are all within the same order of
complexities with respect to r, that is, O(1) number of layers, O(r) number of nodes, thresholds, and
weight connections. The total number of parameters, that is, the number of thresholds plus the number
of weight connections, is an important complexity measure of neural networks because of their hardware
cost and also their influence on generalization performance. With respect to this measure, the NDL
network has the worst overall complexity, which is O(r) number of layers and nodes and O(r 2 ) number of
parameters. CARVE network achieves the best overall complexity but is only slightly better than STRIPd2
network (which has 2r − k more parameters but k − 1 less nodes). STRIPd4 network is more complex
than CARVE and STRIPd2 networks because of its asymptotic constant.
An important observation from the experiments is that, as n and k increase in many classes of functions,
STRIPd2 and STRIPd4 networks have significantly much smaller values for r than CARVE, NDL, and
SEPAR networks. If GA is used to construct a SEPAR network for a given function, then such SEPAR
network should be at least smaller than a CARVE network (for the same function) for reasons explained
in the last paragraph of Section 24.4.3. That is, higher values of r in a CARVE network are due to the less
efficient search of CARVE algorithm compared with SEPAR algorithm (using GA), given the same task.
In practical applications the relation between the number of new hidden units r and the target function
is important. CARVE networks have the simplest operation of the basic units among them, and therefore
require more new hidden units than STRIP networks. Thus, we should discuss the overall complexity
with the number of new hidden units r for some function realizations. SEPAR algorithm already improves
CARVE and NDL for the given reasons. For most classes of functions, STRIP network implementations are
significantly smaller in r than maximum separable-based network implementations of the same functions
such as CARVE, NDL, SEPAR, and other networks. The reason is that removing a strip generated by a w 
may create a completely new strip which is the union of two strips that enclosed the removed strip. This
can happen only when the removed strip is not an end strip. For instance, consider three strips where
one has value 1 and the other two have values 0 and suppose strip of value 1 is between strips of values 0,
that is, we have the sequence 010. Then removing strip 1 creates a new strip 00, which may be longer than
the strips of values 1 and 0 together. If the longer strips are created then the smaller strips will be the
number of new added nodes. This situation cannot happen when maximum separable subset techniques
are used. That is why the maximum separable method creates more nodes than the longest strip method.
To illustrate this fact, consider the example function of P92 (i.e., k = 9 and n = 2) in Figure 24.24.
This function has a mirror-symmetric table, that is, all rows, columns, and the two main diagonals are
symmetric about their center; the second half of a row, column, or diagonal is a mirror reflection of
the first half. Moreover, at row y (1 ≤ y ≤ (k − 1)/2) the first y entries in that row are equal, and
at column x (0 ≤ x ≤ (k − 1)/2) the first x + 1 entries in that column are equal. Such two-input
mirror-symmetric-table function can be constructed for any odd k and the analysis is similar. This class
of functions is very interesting. First, the smallest possible size for a strip-based network or a maximum
separable subset-based network that realizes a function in this class can be obtained analytically. Second,
like random functions, these functions have small separations between inputs and, therefore, seem least
difficult to realize. Third, as described later, they clearly demonstrate the power of STRIP compared with

8 6 4 2 0 2 4 6 8
7 6 4 2 0 2 4 6 7
5 5 4 2 0 2 4 5 5
3 3 3 2 0 2 3 3 3
1 1 1 1 0 1 1 1 1
3 3 3 2 0 2 3 3 3
5 5 4 2 0 2 4 5 5
7 6 4 2 0 2 4 6 7
8 6 4 2 0 2 4 6 8

FIGURE 24.24 Mirror-symmetric-table function of P92 . (From A. Ngom, I. Stojmenović, and V. Milutinović. IEEE
Transactions on Neural Networks, 12, 212–227, 2001. With permission.)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 406 — #34

Synthesis of Multiple-Valued Circuits by Neural Networks 24-407

CARVE, NDL, and SEPAR algorithms; the difference in size between a smallest STRIP network and a
smallest CARVE network for a given function in this class is O(k 2 ).
Clearly, a function in this class has a minimal representation of exactly k units in the first hidden layer
of STRIP. In our example function, STRIP algorithm will extract nine strips out of it and these will have
values 0, 1, 2, 3, 4, 5, 6, 7, 8 in that order. A weight vector that generates the strip of value 0 (the longest
 = (0, 1) — also note that there are four strips of values 2 (and lengths four) in that
strip initially) is w
direction. After removal of the strip of value 0 the algorithm must change direction, that is w  = (1, 0), in
order to remove the strip of value 1. The next longest strip is the strip of value 2 in direction w  = (0, 1);
the four short strips of length 4 are now joined together into a single strip of length 16, since strip 0 and
strips 1 which were between them are removed.
What CARVE, NDL, and SEPAR algorithms can do at best? It can be shown that the minimum number
of hyperplanes needed to partition a mirror-symmetric-table function is (k 2 + 2k − 3)/2. For instance,
the smallest CARVE network associated with our example function contains 48 units in its first hidden
layer. Clearly, the ratio between the sizes of a smallest STRIP network and a smallest CARVE network for
a function in this class → 0 as k → +∞. A smallest CARVE, NDL, or SEPAR network for such functions
is O(k) times larger than a corresponding smallest STRIP network.
Suppose for some function realization, a maximum separable subset-based algorithm, say CARVE,
achieves the smallest network size rm . Suppose, also for the same function realization that STRIP achieves
its smallest network size rs . For STRIP to be fundamentally better than CARVE we must have rs < 12 rm
(this is the case for mirror-symmetric-table functions realizations). Simply put one STRIP neuron (a two-
threshold perceptron) is equal in complexity to two CARVE neurons (one-threshold perceptrons). For
functions where the inputs are separated by a number of parallel hyperplanes (such as linear, permutably
homogeneous, some monotone functions), STRIP is not fundamentally better than CARVE. For these
functions, a smallest STRIP network has a value rs ≥ 12 rm . We will discuss more about this fact in
Section 24.5.7.
STRIP networks are fundamentally better than CARVE, NDL, SEPAR (and other maximum separable
subset based) networks for many classes of functions other than those cited in the previous paragraph.
More complex examples than the mirror-symmetric-table functions can be made, for some n and k,
where STRIP selects optimal directions for partitioning whereas SEPAR, for instance, cannot know such
information and therefore will most likely do poor job. Preliminary experiments (see Table 24.2) seem
to indicate that such functions (including mirror-symmetric-table functions) are very hard to realize by
CARVE and similar algorithms; indeed, they look at least harder than random functions. For example,
given the function in Figure 24.24, if CARVE algorithm removes a single point of value 6 in its first few
iterations, then the corresponding CARVE network will never be minimal. CARVE algorithm proceeds
by corner (and border) separation and it is very likely that one of the points (1, 0), (1, 9), (8, 0), (8, 9) of
value 6 will be in a singleton class that may be separated.
We can also compare the complexity and latency of a minimal STRIP network implementation of
an arbitrary multiple-valued logic function to the complexity and latency of a minimal direct circuit
implementation of the same function. For some functions realizations, STRIP network implementations
are sometimes better and sometimes worse (in depth or size) than their corresponding direct circuit
implementations. The circuit implementations can be from any basis of gates, such as {XOR} or {AND,
OR, NOT } bases for example. See Section 24.5.7 for a comparison with n-bit parity circuits.

24.4.6.4 Proof of Correctness

In this section we prove that both networks, for both definitions of A, effectively compute any given but
arbitrary function f : V ⊂ R n  → K .
Let x be an input vector applied to the network (we refer to both networks and both definitions of A).
Let u = (u1 , . . . , ur ) be the set of output values of the units of layer 1 when x is applied, that is unit Ui
outputs the value ui ≤ k on input x (1 ≤ i ≤ r). Let p ≤ (k + 1)r−1 be the output of the unit (we call
it P) of layer 2. Let q = (q1 , . . . , qk ) be the outputs of layer 3, that is unit Qj has value qj ≤ 1 (1 ≤ j ≤ k)
on input x . Let the value of the output unit (we call it Z ) be z ∈ K .

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 407 — #35

24-408 Handbook of Bioinspired Algorithms and Applications

TABLE 24.2 Results of 10 runs for k = 4 and n = 4

Using 100% of K n Using 60% of K n

#Nodes Min Time #Nodes Time Accuracy (%)

pmuta = 10% and pcros = 75%

Random STRIPF 2 28.0 ± 0.89 27 (3) 33 17.6 ± 0.66 16 25.92 ± 2.42
STRIPF 1 28.9 ± 0.94 27 (1) 29 19.2 ± 0.75 16 23.11 ± 1.43
SEPARF 2 65.8 ± 1.66 62 (1) 58 38.4 ± 1.28 29 25.53 ± 3.39
SEPARF 1 66.3 ± 2.37 63 (1) 58 40.1 ± 2.12 29 24.95 ± 3.42
Linear STRIPF 2 10.0 ± 0.00 10 (10) 10 09.7 ± 0.46 6 87.28 ± 5.98
STRIPF 1 10.0 ± 0.00 10 (10) 10 09.9 ± 0.30 7 87.09 ± 3.79
SEPARF 2 20.3 ± 3.58 13 (1) 19 17.8 ± 2.56 11 85.24 ± 6.78
SEPARF 1 20.2 ± 3.60 13 (1) 19 17.0 ± 2.83 11 83.88 ± 4.21
Monotone STRIPF 2 09.4 ± 0.49 9 (6) 7 08.0 ± 0.77 3 80.68 ± 5.31
STRIPF 1 10.1 ± 0.70 9 (2) 7 08.3 ± 0.46 5 80.97 ± 6.41
SEPARF 2 10.6 ± 1.28 10 (4) 7 09.0 ± 1.34 4 85.05 ± 6.46
SEPARF 1 12.7 ± 1.00 11 (2) 7 10.8 ± 1.08 5 80.58 ± 6.22
Permutably STRIPF 2 3.0 ± 0.00 3 (10) 1 3.1 ± 0.30 1 96.99 ± 1.33
STRIPF 1 3.1 ± 0.30 3 (9) 4 3.0 ± 0.00 3 98.16 ± 2.15
SEPARF 2 3.1 ± 0.30 3 (9) 1 3.1 ± 0.30 1 97.09 ± 2.46
SEPARF 1 3.3 ± 0.90 3 (9) 5 3.3 ± 0.90 3 96.99 ± 1.76

pmuta = 0% and pcros = 75%

Random STRIPF 2 27.9 ± 0.94 26 (1) 28 17.8 ± 0.87 11 24.85 ± 3.17
Linear STRIPF 2 10.0 ± 0.00 10 (10) 13 09.9 ± 0.54 9 86.12 ± 3.28
Monotone STRIPF 2 09.8 ± 0.60 9 (3) 4 08.2 ± 0.40 3 82.82 ± 4.64
Permutably STRIPF 2 03.5 ± 0.50 3 (5) 1 03.2 ± 0.40 1 94.85 ± 2.64

pmuta = 75% and pcros = 0%

Random STRIPF 2 27.5 ± 1.20 26 (3) 36 18.6 ± 0.80 13 24.95 ± 3.16
Linear STRIPF 2 10.0 ± 0.00 10 (10) 15 10.1 ± 0.30 12 87.57 ± 4.27
Monotone STRIPF 2 09.8 ± 0.40 9 (2) 12 07.7 ± 0.64 7 82.14 ± 2.61
Permutably STRIPF 2 03.0 ± 0.00 3 (10) 1 03.0 ± 0.00 1 94.76 ± 2.46

Source: From A. Ngom, I. Stojmenović, and V. Milutinović. IEEE Transactions on Neural Networks, 12, 212–227, 2001. With
permission.

Clearly, on input x each unit Ui has either value ui = 0 or ui = vi = 0, where vi is the maximum
amplitude of the unit (see Figures 24.22 and Figure 24.23). Recall that each Ui corresponds to a subset
Gi found and removed by our A-based synthesis algorithm during its ith run. The collection of the Gi s

is a partition of K n , that is ri=1 Gi = K n and ri=1 Gi = ∅, and therefore, the subset Gi such that
input x ∈ Gi corresponds to the first unit in layer 1 (starting from the left) which outputs a nonzero
value on input x . That is u = (0, . . . , 0, ui = vi , ui+1 , . . . , ur ), where uj for i + 1 ≤ j ≤ r is either 0 or
vj . Recall that, according to our definition of Ui , vi − 1 is the function value of all points in Gi , that is
f (x ) = vi − 1.
By definition, unit P always outputs the value p = (k + 1)r−i whenever i is the least index in u such that
ui > 0 on input x . In the next paragraph we let i be the least such index, that is, such that ui = vi = 0
 
and u1 = · · · = ui−1 = 0. Let a = rl=1 (k + 1)r−l ul = rl=i (k + 1)r−l ul be the dot product of the
weights and inputs (i.e., the outputs of layer 1) of P.
Each unit Qj (1 ≤ j ≤ k) in layer 3 performs the sum bj = a − j(k + 1)r−i (recall that the weight
connection from P to Qj is −j and that the output of P is (k + 1)r−i ). We have bj = (k + 1)r−i ui −

j(k + 1)r−i + rl=i+1 (k + 1)r−l ul . From our definition of the weight connection vector between layer 1

and layer 3 it is easy to see that 0 ≤ rl=i+1 (k + 1)r−l ul < (k + 1)r−i . Therefore, we obtain bj ≥ 0 for
1 ≤ j ≤ ui and bj < 0 for ui+1 ≤ j ≤ k. That is, exactly ui units in layer 3 will output the value 1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 408 — #36

Synthesis of Multiple-Valued Circuits by Neural Networks 24-409


By definition, the output unit Z which has only unit weights, will perform the sum c = kj=1 1bj =
ui
j=1 1 = ui . Since 1 ≤ ui ≤ k and the threshold vector of Z is such that tui = ui + 1, therefore, we
obtain tui −1 ≤ ui < tui . From the definition of the output vector of Z we know that the output of the
neural network is z = tui −1 − 1 = ui − 1 = vi − 1 = f (x ). Thus, the networks have effectively classified
input x correctly. This completes our proof for the network of Figure 24.22. The proof for Figure 24.23 is
straightforward and therefore omitted.

24.4.7 Experimental Results

In this section we present experimental results from the applications of STRIP and SEPAR algorithms
to a number of multiple-valued logic functions, namely: random functions, linear functions, monotone
functions, permutably homogeneous functions, n-bit parity functions, mirror-symmetric functions, and
mirror-symmetric-table functions. In analyzing the performance of STRIP on these functions we looked
at two important aspects, that is, the size of the networks it generates and its generalization ability. We
constructed STRIP networks and compared their performances with constructed SEPAR networks along
with other well-known construction techniques.
Throughout the experiments, we used the following parameters: 1000 generations for GA, 75% cross-
over rate, and 10% mutation rate. We also used an elitist strategy in the GA, which is the best individual
of the current generation is always reproduced to the next generation. This elitist strategy helps counter
balance the disruptive effect of our high mutation rate. On the other hand, we may need this high mutation
rate in order to efficiently search the infinite space of weight vectors. These parameters were obtained by
trial and error and seemed optimal.
For each given class of functions and method of network construction, we ran our algorithms ten times.
All ten runs of the algorithms used the same functions (generated randomly) but with different random
seeds (and different random training sets for approximation).
Table 24.2 shows, respectively for both objective functions (the superscripts F 1 and F 2 stand for
Fitness1 and Fitness2), the results of ten runs of each method on randomly generated functions from each
of the four test classes. We display the average number of created hidden nodes, r (in the first layer of the
constructed networks), with its standard deviation, the minimum value found by the method (the number
in parenthesis is the number of times it was found), the smallest running time (in minutes) over ten runs,
and the average generalization accuracy (with its standard deviation) on the test set over ten runs. One
can see from the table that, in general, Fitness2 yields smaller but less accurate networks than Fitness1.
For functions approximations (see column Using 60% of K n ) we simply used random portions of K n
as training sets. Each of these training sets were generated using a different random seed, also no two
elements in the sets are equal. The network obtained with a training set is then tested on the test set, which
is the remaining 40% of K n , and its accuracy on the test set is computed.
To see if there is any advantage to set pmuta > 0 or pcros > 0 we did two experiments with STRIP using
Fitness2; however, with either pmuta = 0% and pcros = 75% or pmuta = 75% and pcros = 0% (see Table 24.2).
Under the column Using 100% of K n , results in the third part of the table (pcros = 0%) are, in general,
slightly better than those in the first part of the table (pcros , pmuta > 0) which in turn are slightly better
than those in the second part of the table (pmuta = 0%). Thus, there is advantage in setting pmuta > 0.
Clearly, mutation is the operator that helps GA explore the search space. Crossover helps to focus the
search on interesting parts of the space, in parallel. Mutation probability must be set small in order to
avoid a random walk in space and to better exploit good parts of the space. Under the column Using 60%
of K n , the situation is somewhat reversed.
Random functions are more difficult to learn because of the smaller separation among their inputs.
Therefore, it is not surprising that they give the highest number of nodes in the tables and the lowest
generalization accuracy. On the other hand, permutably homogeneous functions have larger separation
of inputs which makes them easier and faster to learn. So they naturally produce smaller networks of
O(k) nodes on average and also give a very high generalization accuracy. Other classes of functions such
as linear, monotone, and other functions lie between these two extremes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 409 — #37

24-410 Handbook of Bioinspired Algorithms and Applications

1. Permutably homogeneous functions. Functions in this class are partitioned by at most k −1 separating
parallel hyperplanes such that no two distinct classes have equal values. Therefore, the minimal number of
new hidden units for both STRIPand SEPAR is ≤k. We have experimented with the following permutably

homogeneous functions: f (x ) = ( ni=1 (1/ai )xi )+n mod k, where ai = 2i+1. We randomly generated
the four-input four-valued logic function f (x ) = x1 /3 + x2 /5 + x3 /7 + x4 /9 + 4 mod 4 and tested our
algorithms with it. The function has three possible values, namely 0, 1, 2. The three classes are separated by
two parallel hyperplanes (classes 1 being in the middle) and the distance between two adjacent hyperplanes
is ≥2.
The results of STRIP and SEPAR are consistent: they gave a value of about the minimal solution r = 3.
Clearly, there is no benefit of using STRIP for these functions; there is no reduction at all and thus
the STRIP network is twice more complex than its corresponding SEPAR network. In general, STRIP is
not better than SEPAR for functions where the inputs are separated by nonintersecting hyperplanes with
distinct values in distinct regions.
2. Monotone functions. We experimented with a (random) function of P44 which is monotonic under
the natural non-decreasing order on K , that is 0 ≤ 1 ≤ · · · ≤ k − 2 ≤ k − 1. For such functions, STRIP
and SEPAR produced closed results. This suggests that, as for permutably homogeneous functions, SEPAR
is better than STRIP for these monotone functions since its achieves smaller network complexity even
though STRIP has smaller values for r.
3. Linear functions. These functions are partitioned by a number of parallel hyperplanes where, many
separated distinct classes of inputs have equal values and any two adjacent classes have distinct values. The
minimal STRIP network for such functions will be exactly twice smaller than the corresponding minimal
SEPAR network, for reasons explained in Section 24.4.6.3. Therefore, even though the STRIP network is
twice smaller in r, it will have exactly the same complexity as its corresponding SEPAR network for linear
functions realizations. Thus STRIP is no better than SEPAR and CARVE for such functions.
The random four-valued linear function generated was f (x ) = (3x1 + x2 + 3x3 + x4 ) mod k. Both
STRIP and SEPAR algorithms, have difficulties realizing this function. First, the minimum r produced by
SEPAR, 13, is very far from the average result, 20.3. Second, STRIP should give a result which is about
two times smaller than the minimum obtained by SEPAR. The reason for this is that, for linear functions,
the distance between two adjacent separating parallel hyperplanes is very small (≤1). Therefore, the
algorithms are very sensitive to rotations, that is, small rotations from separating hyperplanes could cause
the algorithms to produce large networks.
The n-bit parity functions are linear functions and it is well-known that a single layer minimal solution
exists with n hidden (n, 2, 1)-perceptrons. We carried out experiments with STRIP and SEPAR algorithms
for n = 0, . . . , 12 using 2500 generations of GA to learn such functions. The results obtained were
consistently #n/2$ + 1 hidden units for STRIP and n + 1 hidden units for SEPAR, using both fitness
measures, respectively. For SEPAR we cannot obtain n hidden units like the other maximum separable-
based methods because the last training set (of points of same class) is always assigned to a new hidden
unit. Thus, for instance, we will obtain exactly two hidden units for the binary AND function whereas the
other methods would obtain one hidden unit. For SEPAR, the results obtained for n = 10, 11, 12 were
in reality 13, 15, 18 for 2500 generations of GA; however, we obtained the correct results, 11, 12, 13 when
we increased the number of generations to 4000. This suggests that our algorithms are able to find the
minimal value if they are given enough time.
n-bit parity functions can be implemented by {XOR} circuits of depth one and size O(n), {XOR} circuits
of depth O(log2 n) and size n, {AND, OR} circuits of depth d and size O(21/(d−1) n (d−2)/(d−1) ) or {AND,
OR, NOT } circuits of depth two and size 2n−1 . For such functions, a STRIPd2 network has size #n/2$ + 2
and a STRIPd4 network has size #n/2$ + k + 3. As one can see, STRIP networks are not always better
than their corresponding direct circuit implementations. The difference in complexity between a STRIP
network and a direct circuit depends on factors such as the basis set of gates and the fan-in or fan-out of
the gates in the circuit.
4. Random functions. The experiments clearly show that STRIP is fundamentally better than SEPAR
and CARVE for random functions realizations. The average and the minimum obtained STRIP networks

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 410 — #38

Synthesis of Multiple-Valued Circuits by Neural Networks 24-411

are at least twice smaller than the corresponding results for SEPAR networks. As already stated, STRIP
is able to change (and select good) directions for separation and to create larger classes, while SEPAR
cannot do any of these two things. For random functions, the separation between classes is very small.
STRIP can see inside a random function to decide the best directions to choose and, by doing that, it
maximizes the size of all classes (they will get bigger and bigger until they can be removed). SEPAR only
maximizes the size of the classes that are at the boundaries of the inputs space.
Table 24.3 contains results reported for the best three constructive algorithms (so far in the literature)
that have been applied to learning random two-valued logic functions. The value in parenthesis at the
top of each column is the number of trials over which the network is averaged. In each trial of STRIP, we
generated a different random function and used 2500 iterations of GA. STRIP produces smaller networks
for this classification task than any other growth method. Significantly, STRIP gives much smaller networks
than CARVE as n and k increase. Also, SEPAR performs better than Upstart as n and k increases.
5. Mirror-symmetric functions. We have carried out experiments with mirror-symmetric functions.
A mirror-symmetric function has value 1 if the second half of an input vector is a mirror reflection of the
first half, that is the input vector is symmetric about its center. This function is known to have a minimal
representation of two hidden units. For n = 2, . . . , 7 we have always found the optimal number of hidden
units in STRIP and SEPAR, two, using 1000 generations of GA. For n = 8, . . . , 12 we have always found the
optimum when using 3000 generations of GA. Here, as for permutably homogeneous functions, STRIP is
no better than SEPAR.
6. Mirror-symmetric-table (mst) functions. Table 24.4 shows the comparisons of performances between
our example mst function in Figure 24.24 and ten random functions. For each algorithm (STRIPF 2 ,
STRIPF 1 , SEPARF 2 , and SEPARF 1 ), we did ten runs on our mst function and averaged the results. Also,
each algorithm was applied on ten distinct random functions (different from the random functions used
for the other three algorithms) and the results were then averaged. STRIPF 2 is a clear winner (except its
performance in generalization for random functions); it produced the minimal size for mst. The table also

TABLE 24.3 Comparison of Network Size for Random Functions of P2n

STRIPF 2 CARVE [56] Sequential [44] Upstart [50] SEPARF 2

n (100) (100) (100) (100) (100)

4 2.81 ± 0.39 2.40 ± 0.69 3 3.62 ± 0.61

5 3.23 ± 0.42 3.73 ± 0.58 4.5 4.93 ± 0.50
6 4.56 ± 0.50 5.88 ± 0.67 7.28 ± 0.82 8 6.99 ± 0.71
7 6.95 ± 0.30 9.47 ± 0.74 16 11.14 ± 0.77
8 11.56 ± 0.50 16.23 ± 0.86 18.3 ± 0.69 33 19.51 ± 1.03

Source: From A. Ngom, I. Stojmenović, and V. Milutinović. IEEE Transactions on Neural

Networks, 12, 212–227, 2001. With permission.

TABLE 24.4 Performances Comparisons Between Mirror-Symmetric-Table and Random

Functions of P92

STRIPF 2 STRIPF 1 SEPARF 2 SEPARF 1

Using 100% of K n
Mst 09.0 ± 0.00 09.8 ± 0.40 60.2 ± 1.08 69.0 ± 3.69
Random functions 29.2 ± 1.99 31.8 ± 1.60 58.8 ± 2.89 64.7 ± 4.61
Using 60% of K n
Mst 09.0 ± 0.00 11.7 ± 1.95 31.0 ± 3.13 36.0 ± 2.49
Random functions 19.1 ± 1.37 21.6 ± 1.91 36.1 ± 2.98 36.7 ± 2.90
Accuracies
Mst 65.15% ± 12.44 57.27% ± 09.53 18.18% ± 06.91 13.33% ± 05.62
Random functions 09.70% ± 04.85 16.06% ± 06.36 10.91% ± 07.20 12.42% ± 05.33

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 411 — #39

24-412 Handbook of Bioinspired Algorithms and Applications

shows that the single mst function is harder to realize by SEPAR than random functions (see the last two
entries of the first two rows): the average result of SEPAR for mst is very far from the absolute minimum,
48, and is significantly larger than the averaged result of ten random functions. For reasons explained in
Section 24.4.5.2, Fitness2 helps to remove the classes faster than Fitness 1, and therefore, it yields better
results in the table for both algorithms (except on generalization of random functions).

24.5 ES for Learning Multiple-Valued Logic Functions

We consider again the problem of synthesizing multiple-valued logic functions by neural networks.
However, unlike in Section 24.4, an ES is used (in place of GA) to optimize a network.
In this section we apply STRIP algoritm using evolutionary strategy [64] to grow a multiple-valued
logic neural network for learning a function. As in Section 24.4, the maximum separable subset is treated
as a special case of the longest strip.

24.5.1 Determining Longest Strips by Evolutionary Strategy

In this section we use ES [65], instead of GA, to search for the longest strip or maximum separable subset
of training examples. The A-based synthesis algorithm shown in Figure 24.20 is used to grow the network.
The ESs have been proposed by Schwefel [66] as optimization methods for real-valued parameters. ES
manipulates a single potential solution to a problem. Specifically, it operates on an encoded representation
of the solution, and not directly on the solution itself. Schwefel’s ES encodes a solution as a real-valued
vector. We will refer to such a vector as chromosome. The solution is associated with an objective value
that reflects how good or bad it is, compared with other potential solutions in the space. ES is a random
guided HC technique in which a good candidate solution is obtained by applying a certain number of
small mutations to a given parent solution. The best result of mutation is used again to generate the next
best solution, and so on until some convergence criterion is satisfied.
Our ES method uses the same solution representation, the same objective functions and the same
mutation operators as in Section 24.4. That is, a potential solution G (a subset of K n ) is represented as a
weight vector w  (such a vector can be decoded to obtain G). For the problem of determining A, the search
space is the power set of the training set (i.e., the set of all subsets of K n ). Each element of the search space
is a potential solution and must be represented in such a way that meaningful mutation operators can be
designed for (and applied to) it.
We initially start with a unique random unit weight w  0 = (w0,1 , . . . , w0,n ) ∈ [−1, 1]n (Another method
we used for the initialization is to set wi = cos αi (for 1 ≤ i ≤ n) for each vector w,  where αi is a random
number in the interval [−π/2, π/2]. Initial chromosome should consist of random unit hyperplane w). 
Then the next weight vector w  1 = (w1,1 , . . . , w1,n ) is chosen to be a neighbor of w  0 which has the
best fitness value higher than the fitness of w  0 and which yields the least change in the ordering of
the function values according to w  0 (see Section 24.4.5.2). Then, we set w  2 to be the best neighbor of
w 1 , . . . , and continue this process until there is no more best neighbor or a certain number of iterations
is reached.

24.5.2 Weight Neighborhood

(c)
Recall that to find the subset Tw for a current w  in a set of points S with |S| = v we sort the v points
with respect to w  (first key) and f (x ) (second key) and find the longest strip or the maximum separable
subset from the sorted list (w  x1 , f (x1 )), . . . , (w
 xv , f (xv )). The neighbors of w
 are precisely those weight
vectors that yield the least change in the ordering of the f (x )’s.
For instance, if v = 4 and w  gives the order (f (x1 ), f (x2 ), f (x3 ), f (x4 )), then a near neighbor of
 may produce the order (f (x2 ), f (x1 ), f (x3 ), f (x4 )) and a far neighbor of w
w  may produce the order
(f (x4 ), f (x3 ), f (x1 ), f (x2 )).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 412 — #40

Synthesis of Multiple-Valued Circuits by Neural Networks 24-413

Consider the jth coordinate, wj , of w,  1 ≤ j ≤ n. Let w  x and w

 y be two consecutive elements in the

order produced by w  and let d = w y − w  y − x ) = wj (yj − xj ) + i=j wi (yi − xi ). Likewise, for
 x = w(

an arbitrary vector π, we will have d % = πj (yj − xj ) + i=j πi (yi − xi ).
 with respect to coordinate j if πi = wi for i = j and |πj −wj | is minimal such
Vector π is a neighbor of w
 
that the sorted order is changed. Then only wj is affected and hence i=j wi (yi − xi ) = i=j πi (yi − xi ).

Since i=j wi (yi − xi ) = d − wj (yj − xj ) therefore we have d % = πj (yj − xj ) + d − wj (yj − xj ) =
d + wj (yj − xj ), where wj = (πj − wj ). Taking d % = 0 and solving for wj we obtain

d
wj = − . (24.20)
yj − x j

Only differences di = w(

 xi+1 − xi ), for 1 ≤ i ≤ v − 1, between distinct consecutive elements in sorted
order are considered. For each coordinate j (1 ≤ j ≤ n) there should be one neighbor only, chosen such
that |wji | = | − di /(xi+1,j − xi,j )| is minimized and is not equal to zero. Let wj = min |wji | and
denote the nearest neighbor of w with respect to j by πj . Then πj differs from w
 only in the jth coordinate,
by wj . That is πj,j = wj + wj and πj,c = wc for c = j. In other words, the Euclidean distance E(w,  πj )
between w  and πj is |wj |. Thus, there are n neighbors of w,  that is one nearest neigbhor per each
 We denote the set of all n such neighbors of w
coordinate j of w.  by w = {π1 , . . . , πn }. This set is called
the neighborhood of w.

24.5.3 Evolution Strategy

Our ES is shown in Figure 24.25. In the algorithm, w  i is the current solution at generation i, b is the best
solution generated so far, µ  w i is the best solution in w i and w  i+1 is the next solution to generate. We
initially start with a random unit vector w 
 0 and set b to w 0.
The algorithm works as follows. At generation i, we compute the best neighbor of w  i , that is, a vector
µ w i in w i that has the highest fitness value. If µ  w i is better than b then we are done for this generation;
w i+1 and b are both set to µ  w i and we move to the next generation. If µ  w i is worse than b then we
must decide how to set w  i+1 . We first attempt to find a solution better than b in the neighborhood of
the neighbors of w  i . That is we search successively in each πj (1 ≤ j ≤ n) until such solution is found
or all n neighborhoods are tried without success. If a solution is found then we set both b and w  i+1 to
that solution, µ
 πj for some j, and move on to the next generation. If, otherwise, no such solution is
found then b is possibly a local optimum. We have two choices, with probability 12 we either (i) set w  i+1
to the best between µ  w i and µ
 πj (for all j’s) and move on to generation i + 1, or (ii) to jump out of a
possible local optimum, we set w  i+1 to a random unit vector or to a random alteration of b or µ w or i
µ
 πj for some random j. In the second choice, one of the four possibilities is selected with probability
1
4 and the alteration of a vector is done by choosing any of our three mutation techniques described in
Section 24.4.5.4 (with probability 13 ).
Let g be the number of generations, then at each new generation n new chromosomes are evaluated
for their fitness and hence, our ES has a time complexity of O(gn 2 |S| log |S|) ≈ O(gn 3 k n log k). The time
complexity in Section 24.4 is O(gpn 2 k n log k) where p is the population size of the GA. Since p & n then
our ES is much faster than the GA in Reference 28.
Using ES in the A-based algorithm, we construct neural networks in the same way as in Section 24.4.

24.5.4 Experimental Results

In this section we present experimental results from the applications of STRIP and SEPAR algorithms,
using ES, to the same multiple-valued logic functions as in Section 24.4. We constructed STRIP and
SEPAR networks by ES (STRIP_ES and SEPAR_ES) and compared their performances with constructed

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 413 — #41

24-414 Handbook of Bioinspired Algorithms and Applications

FIGURE 24.25 Evolution strategy to find G such that |G| ≈ |A|.

STRIP_GA and SEPAR_GA networks reported in Section 24.4 along with other well-known construction
techniques.
Throughout the experiments, we used the following parameters: 2000 generations for ES (since ES
is at least twice faster than GA and that in Section 24.4 uses 1000 generations for GA) and 10%
mutation rate. Section 24.4 used 10% mutation rate, 75% crossover rate and an elitist strategy in the
GA (i.e., the best individual of the current generation is always reproduced to the next generation).
In some experiments such as random binary functions (and other functions) learning we increased
the number of generation of GA (see Section 24.4), and accordingly use twice as much iterations for
the ES.
Tables 24.5 and 24.6 show, respectively, for both objective functions, the results of ten runs of each
method on randomly generated functions from each of the four test classes. We display the average
number of created hidden nodes, r (in the first layer of the constructed networks), with its stand-
ard deviation, the minimum value found by the method (the number in parenthesis is the number
of times it was found), the smallest running time (in minutes) over ten runs, and the average gener-
alization accuracy (with its standard deviation) on the test set over ten runs. From these two tables
we can see that, in general, Fitness2 yields (slightly) smaller but (slightly) less accurate networks than
Fitness1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 414 — #42

Synthesis of Multiple-Valued Circuits by Neural Networks 24-415

TABLE 24.5 Results of 10 Runs for k = 4 and n = 4 (Using Fitness1)

Using 100% of K n Using 60% of K n

Fitness1 #Nodes Min Time #Nodes Time Accuracy (%)

Random STRIP_GA 28.9 ± 0.94 27 (1) 29 19.2 ± 0.75 16 23.11 ± 1.43

STRIP_ES 32.5 ± 0.67 32 (6) 29 21.4 ± 1.02 14 23.69 ± 3.28
SEPAR_GA 66.3 ± 2.37 63 (1) 58 40.1 ± 2.12 29 24.95 ± 3.42
SEPAR_ES 69.2 ± 2.79 64 (1) 53 40.3 ± 1.35 23 25.92 ± 3.86
Linear STRIP_GA 10.0 ± 0.00 10 (10) 10 09.9 ± 0.30 7 87.09 ± 3.79
STRIP_ES 10.0 ± 0.00 10 (10) 9 09.8 ± 0.60 6 92.23 ± 4.14
SEPAR_GA 20.2 ± 3.60 13 (1) 19 17.0 ± 2.83 11 83.88 ± 4.21
SEPAR_ES 27.9 ± 10.97 13 (2) 17 22.3 ± 7.69 9 73.20 ± 15.84
Monotone STRIP_GA 10.1 ± 0.70 9 (2) 7 08.3 ± 0.46 5 80.97 ± 6.41
STRIP_ES 09.6 ± 0.92 9 (6) 6 09.0 ± 0.77 4 83.50 ± 4.62
SEPAR_GA 12.7 ± 1.00 11 (2) 7 10.8 ± 1.08 5 80.58 ± 6.22
SEPAR_ES 12.2 ± 0.60 11 (1) 7 10.6 ± 1.11 4 86.89 ± 3.77
Permutably STRIP_GA 3.1 ± 0.30 3 (9) 4 3.0 ± 0.00 3 98.16 ± 2.15
STRIP_ES 5.4 ± 0.80 4 (2) 4 4.4 ± 0.66 3 94.08 ± 1.65
SEPAR_GA 3.3 ± 0.90 3 (9) 5 3.3 ± 0.90 3 96.99 ± 1.76
SEPAR_ES 7.8 ± 1.17 6 (2) 5 6.9 ± 1.14 3 90.29 ± 3.22

TABLE 24.6 Results of 10 Runs for k = 4 and n = 4 (Using Fitness2)

Using 100% of K n Using 60% of K n

Fitness2 #Nodes Min Time #Nodes Time Accuracy (%)

Random STRIP_GA 28.0 ± 0.89 27 (3) 33 17.6 ± 0.66 16 25.92 ± 2.42

STRIP_ES 31.6 ± 1.11 30 (2) 34 20.7 ± 1.00 18 24.56 ± 3.88
SEPAR_GA 65.8 ± 1.66 62 (1) 58 38.4 ± 1.28 29 25.53 ± 3.39
SEPAR_ES 69.1 ± 3.11 64 (1) 52 42.7 ± 3.13 25 25.63 ± 3.48
Linear STRIP_GA 10.0 ± 0.00 10 (10) 10 09.7 ± 0.46 6 87.28 ± 5.98
STRIP_ES 10.0 ± 0.00 10 (10) 11 10.1 ± 0.54 8 88.45 ± 2.55
SEPAR_GA 20.3 ± 3.58 13 (1) 19 17.8 ± 2.56 11 85.24 ± 6.78
SEPAR_ES 25.3 ± 7.79 13 (2) 18 23.5 ± 8.27 10 73.88 ± 20.28
Monotone STRIP_GA 09.4 ± 0.49 9 (6) 7 8.0 ± 0.77 3 80.68 ± 5.31
STRIP_ES 09.4 ± 0.49 9 (6) 10 8.1 ± 0.54 6 84.37 ± 4.26
SEPAR_GA 10.6 ± 1.28 10 (4) 7 9.0 ± 1.34 4 85.05 ± 6.46
SEPAR_ES 10.6 ± 1.43 10 (3) 10 9.1 ± 1.22 6 88.06 ± 5.00
Permutably STRIP_GA 3.0 ± 0.00 3 (10) 1 3.1 ± 0.30 1 96.99 ± 1.33
STRIP_ES 5.9 ± 0.94 4 (1) 3 4.4 ± 0.66 2 93.69 ± 2.00
SEPAR_GA 3.1 ± 0.30 3 (9) 1 3.1 ± 0.30 1 97.09 ± 2.46
SEPAR_ES 6.1 ± 0.83 5 (2) 4 4.9 ± 0.94 3 93.69 ± 1.58

GA performed slightly better than ES in most experiments, due to the GA manipulation of population
and also its implicit parallelism. However, ES is at least twice faster than GA.

Permutably homogeneous functions. GA performed much better than ES.

Monotone functions. GA and ES produced very close results.
Linear functions. GA and ES produced very close results.
Random functions. STRIP_GA and STRIP_ES performed much better than any other constructive
technique. Also, ES gave worse results than GA (Table 24.7).
Mirror-symmetric functions. For n = 2, . . . , 7 we have always found the optimal number of hidden
units in STRIP and SEPAR, two, using 2000 generations of ES (1000 for GA in Reference 7). For
n = 8, . . . , 12 we have always found the optimum when using 6000 generations of ES.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 415 — #43

24-416 Handbook of Bioinspired Algorithms and Applications

TABLE 24.7 Comparison of Network Size for Random Functions of P2n

STRIPFGA
2 [7] STRIPFES2 [7] CARVE [56] Sequential [44] Upstart [50] SEPARFGA
2 [7]
n (100) (100) (100) (100) (100) (100)

4 2.81 ± 0.39 3.17 ± 0.49 2.40 ± 0.69 3 3.62 ± 0.61

5 3.23 ± 0.42 3.64 ± 0.52 3.73 ± 0.58 4.5 4.93 ± 0.50
6 4.56 ± 0.50 5.14 ± 0.62 5.88 ± 0.67 7.28 ± 0.82 8 6.99 ± 0.71
7 6.95 ± 0.30 7.84 ± 0.37 9.47 ± 0.74 16 11.14 ± 0.77
8 11.56 ± 0.50 13.04 ± 0.62 16.23 ± 0.86 18.3 ± 0.69 33 19.51 ± 1.03

24.6 Minimization of Multiple-Valued Multiple-Threshold

Perceptrons using GAs
We address the problem of computing and learning multiple-valued multiple-threshold perceptrons. Every
n-input k-valued logic function can be implemented using an (n, k, s)-perceptron, for some number of
thresholds s. We propose a GA to search for an optimal (n, k, s)-perceptron that efficiently realizes a given
multiple-valued logic function, that is to minimize the number of thresholds. Experimental results show
that the GA find optimal solutions in most cases.
A problem still left open in the domain of multiple-valued multiple-threshold functions is how to min-
imize the number of thresholds in order to construct the most efficient multiple-valued multiple-threshold
networks or units. To minimize the number of thresholds, traditional techniques of multiple-valued
multiple-threshold circuit synthesis use either trial and errors, or allow to synthesize only classes of func-
tions for which an optimal number of thresholds can be obtained (multiple-valued multiple-threshold
synthesis of k-valued symmetric functions, for instance). The multiple-valued multiple-threshold networks
considered in literature have no learning capabilities, that is, their parameters are set by the designers
once and for all using some traditional techniques of multiple-valued multiple-threshold networks syn-
thesis. In addition only some small classes of k-valued logic functions are considered for multiple-valued
multiple-threshold synthesis techniques.
Given f ∈ Pkn we are looking for a vector r = (w,  t , o) ∈ R n+s ×K s+1 such that Fk,s
n (
r )(x ) = f (x ) (∀x ∈
K ), that is, f = Fk,s (r ). We will refer to r as a s-representation of Fk,s . In this section, we will be mainly
n n n

interested in finding minimal s for which there exist a s-representation for a given f ∈ Pkn . In other
words, given f ∈ Pkn , we want to find a s-representation r with the least possible number of thresholds s
n (
such that Fk,s r ) = f . We propose GAs as techniques for minimizing multiple-valued multiple-threshold
perceptrons.

24.6.1 Computing Optimal s-Representations with GAs

24.6.1.1 Problem Representation
Fundamental to the GA structure is the encoding mechanism for representing the problem’s variables. For
 and the representation is
the s-representation problem, the search space is the space of weight vectors w
more complex. Unfortunately, there is no practical way to encode s-representation problem as a binary
chromosome to which the classical genetic operators discussed in Reference 61 can be applied in a
meaningful fashion. Therefore it is natural to represent the possible solutions as vectors w ∈ R n and
design appropriate genetic operators which are suitable for the s-representation problem. Each weight
vector will uniquely determine a s-representation. To determine how good the solution is, the GA needs
a fitness function to evaluate the chromosomes.
A note on the initial population. We initialize the population with random real-coded chromosomes
whose coordinates are random real numbers taken from the interval [−1, 1]. Each initial chromosome is
then normalized to a unit vector. Another method we used for the initialization of the population is to set
wi = cos αi (for 1 ≤ i ≤ n) for each vector w,
 where αi is a random number in the interval [−π/2, π/2].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 416 — #44

Synthesis of Multiple-Valued Circuits by Neural Networks 24-417

What we are trying to do in both methods of initialization is to generate random hyperplanes (since each
 represent a hyperplane).
w

24.6.1.2 Fitness Function

The objective function, the function to be optimized, provides the mechanism for evaluating each chro-
mosome. To describe our fitness function we will need the concept of valid and invalid thresholds
(hyperplanes).
To compute the thresholds for a given chromosome w,  we calculate for every x ∈ K n the value w  x
and construct a sorted array (or list) of records of the form (w  x , f (x )). The array is sorted using w  x
as primary key and f (x ) as secondary key. Let these records be sorted as follows: x1 , . . . , xk n , or more
 xi , f (xi )), 1 ≤ i ≤ k n , where w
precisely, (w  xi ≤ · · · ≤ w  xk n . Then w  xj is a threshold if f (xj−1 ) = f (xj ).
We collect all thresholds in a list t . Some thresholds in t may be duplicated (i.e. ti−1 = ti for some i).
Let T (w)
 = V (w)  + I (w),
 where T (w)  is the total number of thresholds generated by w,  V (w)  and
I (w)
 are, respectively, the number of valid thresholds and invalid thresholds generated by w.  A threshold
ti (1 ≤ i ≤ T (w)  ≤ k n − 1) is valid if all points x ∈ K n lying in its corresponding hyperplane Hi
(given by w  x = ti ) are in the same class (i.e., f (x ) has the same value for all points in hyperplane Hi ),
otherwise it is invalid. In other words, invalid thresholds are those for which there exist at least two
points x1 and x2 ∈ K n such that w  x1 = w  x2 but f (x1 )  = f (x2 ). A hyperplane is valid (invalid) if it
corresponds to valid (invalid) threshold. With these definitions then duplicated thresholds in t are invalid
while nonduplicated thresholds are valids.
 is the total number of thresholds in t and can be used to evaluate how good or bad is a chro-
T (w)
mosome. The best chromosomes are those that have the least T (w).  We can therefore define our fitness
function as follows:

T (w)

 =1−
Fitness1(w) . (24.21)
k −1
n

Note that a GA always maximizes an objective function and since 1 ≤ T (w)  ≤ k n − 1, then Fitness1(w)

is maximal when T (w)  is minimal.
However, invalid thresholds must need severe penalty. For instance, assume a n-input k-valued
logic function f : K n  → {0, 1} chosen at random. Then one may take hyperplanes x1 = 0,
x1 = 1, . . . , x1 = k − 1 as invalid thresholds. These k hyperplanes (or k 2 thresholds) will separate in
our sense but are not really separating as such random function needs actually an exponential number of
thresholds. Because of this fact, instead of using Formula (24.21) we can alternatively use Formula (24.22).

2 − (T w)/(k
 n − 1) − I (w)/T
 (w)
 T (w)
 I (w)

 =
Fitness2(w) =1− − . (24.22)
2 2 · (k − 1) 2 · T (w)
n 

Here, we not only minimize T (w)  (in second term) but we also punish a chromosome that generates
a large number of invalid hyperplanes (in last term). That is we are minimizing T (w)  and I (w)  at the
same time. Note that 0 ≤ I (w)  ≤ T (w)  and thus Fitness2(w)
 will be maximal if both T (w)  and I (w)
 are
minimal.
In all our experiments, both formulae of fitness yield the same results for I (w)  = 0. We do not know
for now how they do behave for I (w)   = 0 since the w’s
 generated valid thresholds only. The probability
to generate invalid thresholds seems to be very close to zero.
A note on the time complexity of the evaluation function. For a given w,  it takes n · k n steps to compute
 · x ’s, k · log k steps to sort them and at most k steps to compute T (w).
all the w n n n  Therefore the evaluation
of Fitness(w)  has a time complexity of O(n · k n · log k).
Also, crossover and mutation operations take O(n) steps each and the initialization of the population
takes O(n · p · k n · log k) steps (p is the number of chromosomes and all initial chromosomes are evaluated
for their fitness). Thus the evaluation of Fitness(w)  is the most expensive operation in our GA (and is

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 417 — #45

24-418 Handbook of Bioinspired Algorithms and Applications

true in general for any GA). Let g be the number of generations, then at each new generation p/2 new
chromosomes are evaluated for their fitness and hence, our GA has a time complexity of O(n·g ·p·k n ·log k).

24.6.1.3 Crossover and Mutation

They are same as those discussed in Section 24.4.

24.6.2 Experiments and Discussions

In our experiments, the control parameters’ setting for the GA were: population size p = 100; number of
generations g = 1000; crossover probability pcros = 0.75; and mutation probability pmuta = 0.005. The
most important parameters here are p, pcros , and pmuta and the values used for them seem to be optimal
in that they yield better results (than other possible values) in all experiments we have done. The high
crossover rate is necessary to widen the search while the low mutation rate is necessary to avoid too much
chromosome disruptions. Because we use an elitist strategy some best chromosome in a current generation
is always reproduced to the next generation in order to avoid lost of good genetic material (and to reduce
the disruptive effect of a high crossover rate). We use a large population size to preserve the diversity of the
population, that is to avoid premature convergence. The fact that we used a mixed crossover technique also
helps maintain the diversity. In all experiments, we used Fitness2 as our evaluation function (for reasons
explained in Section 24.6.1.2). Also, we used stochastic universal selection shceme as our reproduction
method.
It is interesting that the proposed population (chromosomes) representation does not depend on k. It
make us wonder how the number of invalid thresholds vary with k (or n). For a fixed n (or k), larger k
(or n) means smaller separation among classes and these problems are typically more difficult to learn.
We did some experiments with small k versus large k and small n versus large n in order to see how the
number of invalid thresholds changes. Such experiment were performed on random functions and their
results are reported in Tables 24.8 and 24.9.

TABLE 24.8 Results for Some

Two-Input k-Valued Random Functions

k #Invalids #Seconds
2 0 3.25
4 0 8.92
8 0 36.72
16 0 153.76
32 0 700.12
64 0 3,289.59
128 0 14,511.80
256 0 99,434.36

TABLE 24.9 Results for Some n-Input

Two-Valued Random Functions

n #Invalids #Seconds
2 0 3.25
4 0 10.40
8 0 217.47
9 0 496.61
10 0 1,047.64
11 0 2,155.61
12 0 4,797.34
13 0 9,918.57
14 0 21,959.14

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 418 — #46

Synthesis of Multiple-Valued Circuits by Neural Networks 24-419

TABLE 24.10 Some Results for 10 Runs

Average number
n Optimal s Number of runs of generations

k=4
2 1 10 0
3 2 10 24.6
4 2 9 430.9

k=3
5 1 3 669

k=2
6 1 9 283.78
7 1 3 660.33

As we can see in both tables (and also in Table 24.10), the number of invalid thresholds obtained is
always zero. The last column in both tables shows the running time for s = 100 and g = 100. Although our
method is slow, it is no surprise that the algorithm is slower as n grows than as k grows. Such results agree
with the complexity analysis given in Section 24.6.1.2. From neural networks applications perspective,
results on the number of invalid thresholds for k ≤ 32 and k ≥ 64 given in Table 24.8 are interesting,
since these values of k correspond to discretizations of real-valued neurons by at most 5 bits or at least 6
bits, for fixed n. A more theoretical, not well understood problem would address n/ log n, n, or constant
number of bits since we know that n/ log n bits is sufficient, n bits is the best known lower bound, while
constant number of bits appears to be sufficient in practice for nonmalicious threshold functions.
In Table 24.10 we show results of ten runs of the GA on examples of n-place k-valued logic functions
(for 2 ≤ k ≤ 4 and 2 ≤ n ≤ 7) given by

 

n
1
f (x ) = xi + n mod k (24.23)
ai
i=1

where ai = 2i + 1. We can easily guess the minimum number of thresholds needed for a perceptron
to simulate these functions. Indeed, each of these functions defines itself its separating hyperplanes and
their number. The number of hyperplanes is simply the number of distinct values of such function

minus one, and each hyperplane Hj is defined by the equation ni=1 (1/ai )xi = tj for some threshold tj
(1 ≤ t ≤ number of thresholds). The output vector can also be obtained by computing the value of f for
x1 = · · · = xn = 0 and x1 = · · · = xn = k − 1 = 3 and listing in increasing order modulo k the sequence
of other distinct values of f in between. In Table 24.11 we show examples of optimal solutions obtained
by the GA.
In Table 24.10, the second column indicates the optimal number of thresholds that the GA must find,
the third columns contains the number of runs where the GA reached the optimum, and the fourth
column shows the average number of generations over all successful runs needed to obtain the optimal
solution. All solutions found by the GA, optimal or not, were valid solutions in that they do not contains
invalids thresholds.
As seen from the table, the difficulty for the GA to find an optimal solution within 1000 generations
depends mostly on n rather than k. This is not surprising since the search space is exponential on n and
thus the GA needs more and more generations (meaning more genetic operations) to successfully obtain
an optimum. This is indicated by the fourth column. For k = 4 and n = 5, for example, the GA could
not find an optimum within five runs of 1000 generations each; however, it was successful within one run
with 2000 generations. This suggest that given enough time (which depends on n) the GA will always find

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 419 — #47

24-420 Handbook of Bioinspired Algorithms and Applications

TABLE 24.11 Some Optimal Solutions Found

n 
w t o

k=4
2 0.953823 0.300371 2.508386 2 3
3 0.830144 0.473697 0.294061 2.303273 4.793706 3 0 1
4 0.785003 0.441867 0.347428 0.260417 2.351256 4.714851 0 1 2
k=3
5 −0.754707 −0.465486 −0.348600 −0.228958 −0.199491 −2.285579 0 2
k=2
6 0.487796 0.827506 0.205418 0.075150 0.111406 0.130513 1.595870 0 1
7 0.746777 0.459637 0.299486 0.148247 0.255323 0.191369 0.132577 1.654147 1 0

the minimal s-representation for a logic function f . We do not have rows for higher values of n because
of the fact that the algorithm is slow as n grows.
It is interesting to note that the functions we used in our experiments are the most difficult for
the GA since the their s-representations are very small (e.g., s ∈ {1, 2}). This indicates that for most
(random) functions the GA will perform much better than for our test functions because s is larger on
average.
We compared our technique with the extended permutably homogeneous (n, k, s)-perceptron learning
algorithm (EPHPLA) described in Section 24.3. A permutably homogeneous perceptron has a (s + 1, k)-
permutation as its output vector, that is a permutations of s + 1 elements out of K with s ≤ k − 1.
The EPHPLA generalizes the homogeneous (k, k − 1)-perceptron learning algorithm of Reference 16
and has a time complexity of O(e · n · k n ), where e is the number of learning epochs. The EPHPLA
can only learn permutably homogeneous functions and an example of such class of functions are our
test functions given by Equation (24.23). It is proven in Reference 34 that the EPHPLA always converges
for permutably homogeneous functions, and that also, it always finds a minimal s-representation for a
learned function f . The EPHPLA is faster and outperform the GA on learning these same test functions
within one run of 1000 learning epochs. The GA converged better only for n = 2 and any k. The main
advantage of the GA method over the EPHPLA is that it can learn any logic function provided enough
time is given.

24.7 Conclusion and Open Problems

Neural net computing is parallel distributed computing with an ensemble of elemental processors
interconnected in ways reminiscent of biological neural nets.
Although the idea of computing with arrays of interconnected elemental processors is certainly not
new, there is currently a resurgence of interest in this area. This resurgence was initiated through the
introduction of a few new algorithms which made it possible, perhaps for the first time, to implement
and experiment with some simple realizations of such interconnected nets of elemental processors. Some
imaginative demonstrations of such nets helped to activate further interest and fascination with these new
developments.
At present there is widespread activity in this area with much interest in the potential use of neural
networks in signal and image processing and so on. There is also interest in the possibility of using such
nets as computer models for studying the functioning of neuro-biological nets.
This thesis addresses an intellectual issue which is neither of these categories but is more related to the
former than to the latter.
The point is that since we can now implement neural nets which can actually perform certain basic
information processing functions, it is of interest to see if we can fashion nets or systems of nets which can
reproduce (i.e., mimic) certain trains of actions regularly performed by humans. The challenge is to be

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 420 — #48

Synthesis of Multiple-Valued Circuits by Neural Networks 24-421

able to do this without an oracle within the net, that is without devine revelation or guidance, so to speak,
at critical junctures. Positive results gained in activities of such nature would not be interpreted to indicate
that that is indeed how human biological nets do function but might serve to suggest preferences among
various ways of thinking of processing in biological nets.
The issue we have addressed in this chapter is that of implementing multiple-valued logic systems in
neural networks. There are many kinds of multiple-valued logic algebras such as fuzzy logic, probabilistic
logic, logical calculus with rough sets, etc. However, for the present we have concentrated on one such logic
system, that is the classical multiple-valued logic as defined in this chapter.
In particular, we have discussed original models of multiple-valued neurons and multiple-valued neural
networks and studied their learning and computing powers.

24.7.1 Conclusions
We have addressed the issues of synthesizing multiple-valued logic circuits with (n, k, s)-perceptrons.
The (n, k, s)-perceptron learning problem is the problem of determining an (optimal) s-representation
 t , o) required to compute a given function. Another problem related to learning with (n, k, s)-
r = (w,
perceptron networks is the search for an optimal size of a network during learning.

1. Learning abilities of (n, k, s)-perceptrons are examined. The previously studied homogeneous
(n, k, k − 1)-perceptron learning algorithm have been generalized to the permutably homoge-
neous (n, k, s)-perceptron learning algorithm with guaranteed convergence property. A permutably
homogeneous perceptron is a neuron whose output vector o is a permutation on K . We have
obtained a powerful learning method that learns any permutably homogeneous separable k-valued
function given as input. When the number of thresholds is not fixed, the algorithm will always find
the minimal one to be used for learning a separable function.
2. We have discussed a particular implementation of partitioning algorithm to construct (near)
minimal (n, k, s)-perceptron networks for learning given but arbitrary multiple-valued functions.
We used GA or ES to find a (near) minimal set of hidden units that partition the space V ⊆ K n
into strips or maximum separable subsets. A strip contains points located between two parallel
hyperplanes. We have constructed two neural networks based on these hidden units and show that
they correctly compute the given but arbitrary multiple-valued function. STRIP and SEPAR can
be used for functions with real-valued inputs, k-valued inputs, two-valued outputs, or k-valued
outputs. More research is needed in order to increase the speed of the A-based synthesis algorithm.
3. Every n-input k-valued logic function can be implemented using a (n, k, s)-perceptron, for some
number of thresholds s. We proposed a GA to search for a minimal (n, k, s)-perceptron that
efficiently realizes a given but arbitray function, that is to minimize its number of thresholds.
Experimental evidence show that the genetic search can be very effective however slow it
may be.

24.7.2 Open Problems

The following are some interesting open problems related to research in this area:

1. Another complexity measure of neural networks is the Vapnik-Chervonenkis (VC) dimension. It is

defined as the maximum size of a training set T ⊆ V such that a given network realizes all functions
defined on T . If the (n, k, s)-perceptrons have a finite VC-dimension then they may be PAC-learnable.
That is, there may exist an efficient (n, k, s)-perceptron learning algorithm which uses only a polynomial
number of examples and that, with high probability (which can be made as high as desired), the algorithm
outputs a good approximation of a given function within a desired degree of accuracy (which can be
made as high as desired). The VC-dimension of (n, k, s)-perceptrons gives a valid lower bound on the
VC-dimension of linear decision lists and other related learning architectures.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 421 — #49

24-422 Handbook of Bioinspired Algorithms and Applications

2. The computing capacity of linear decision lists or any architecture obtained by a partitioning
algorithm is not known and is still an open problem. The technique we have applied for deriving capacity
results on (n, k, s)-perceptrons may be extended or improved to give results for partitioning architectures.
The main question is: In how many ways can we partition a finite set V ⊆ R n using s hyperplanes? The
hyperplanes are not necessarily parallel. The answer to this question gives (bounds on) the capacity of
neural networks constructed by partitioning algorithms. Such neural network architectures include neural
trees and neural decision lists. This question also motivates for the investigation of the VC-dimension and
PAC-learnability of such structures.
3. The combinatorial arguments used to derive our results for the number of linear (or multilinear)
partitions and the capacity of (n, k, s)-perceptrons may possibly be extended for the general case of
n-dimensional set. For instance, enumerate partitions for higher dimension grids.
4. Suppose the following generalization of a threshold function: g (P(x)) = 0 if P(x) < t and,
g (P(x)) = 1 if t ≤ P(x), where P(x) is a polynomial of degree d ≥ 0 and t is a threshold level. We
say that g is a threshold function of order d. For instance, the well-known linear threshold function
is a threshold function of order 1 with P(x) = w  x . Geometrically, an order d threshold function is a
separating hypersurface of degree d. For example, a linear threshold is a separating hyperplane that can be
expressed by a polynomial of degree 1. Very little results are known in neural networks literature on using
hypersurfaces as discriminant functions, such as quadratic surfaces (parabola, hyperbola, circle, ellipses,
spheres, etc.), cubic surfaces, or surfaces of degree d ≥ 4. It may be that hypersurfaces are better separators
than linear surfaces (i.e., hyperplanes). More studies need to be done.
5. An interesting research is to design efficient learning algorithms for multilayer (n, k, s)-perceptron
networks. Such network would have the ability to learn any multiple-valued function. Learning and
computing abilities of discrete Hopfield networks (or any other type of discrete associative neural net-
work) composed of (n, k, s)-perceptrons can be investigated. Support vector machines (SVMs) with
(n, k, s)-perceptrons as processing units can be studied. The capacity of single (n, k, s)-perceptrons may
be increased if learning algorithms are designed for output vectors which are (s + 1, k)-permutations with
possible repetitions of their distinct elements.
6. The neural networks we have considered in this thesis are heteregeneous networks, meaning that
the (n, k, s)-perceptron elements are not all identical (e.g., the number of thresholds or inputs are not
all the same). Designing learning algorithms for homogeneous networks seems (to us) much easier than
for heteregeneous networks. However for homogeneous networks, the problem we must solve first is to
define a good differentiable error function in order to do learning with gradient descent similar to the
back-propagation learning method.
7. The generalization properties of STRIP and SEPAR need to be studied. We believe that our methods
generalize better than CARVE since CARVE always overfits the weights. We avoided this overfitting by
translating the obtained hyperplanes away from their original positions so to include some test points in
the closed space delimited by these hyperplanes. In addition, in higher dimension and for many classes
of functions, our STRIP networks are much less complexe (in terms of total number of parameters) than
CARVE and others maximum separable based networks, and therefore by the Occam razor principle STRIP
would generalize better. One can further improve the generalizability by transforming an obtained STRIP
network into a radial basis functions network and apply back-propagation to the resulting network. Young
and Downs [56] have done the same thing with CARVE using sigmoidal units and proved it efficient.
In our case, we will use radial basis functions because Gaussian units are continuous version of our
(n, k + 1, 2)-perceptrons.
8. In minimizing the number of thresholds of an (n, k, s)-perceptron, the generalization properties of
the GA can be studied when the fitness function is modified to work with small subsets of K n . Another
aspect that would be of interest to neural networks community is related to our proposed GA optimization.
If it generates small number of invalid thresholds when working with K n , it should also be able to discover
a near optimal representation for smaller subsets of K n . If so, and if we believe in Occam’s razor principle
(that the simplest satisfactory explanation of a phenomenon is most likely to be a correct one) then this is
interesting from generalization perspective. In particular, instead of calculating w  x for every x ∈ K n , the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 422 — #50

Synthesis of Multiple-Valued Circuits by Neural Networks 24-423

same fitness estimation algorithm can be applied to a small enough subset S ⊂ K n (e.g., (k − 1)n randomly
selected points from K n ). Once a faithful representation for S (called the training set) is learned, it is easy
(but interesting) to measure classification error on out of sample data (e.g., total number of mistakes, and
maybe sum of squared errors on K n − S). If applied for generalization and if needed, the fitness function
can be further modified to allow some error if it results in smaller representation. One major problem with
our method is that GA is very slow for even small values of n (n ≥ 10) and k (k ≥ 8). Better techniques
are needed for efficient computation of (n, k, s)-perceptrons. For example, one possible solution could
be to use a population of s-representations r (instead of weights w)  and then design a fitness function
which minimizes s and minimizes the error between the (n, k, s)-perceptron and a given multiple-valued
logic function. Sorting is not needed here since we select only chromosomes whose t is sorted. Those with
unsorted t will have severe penalty.

References
[1] M. Minsky and S. Papert, Perceptrons: An Introduction to Computational Geometry, MIT Press,
Cambridge, MA, 1969, Expanded edition, 1988.
[2] K.C. Smith, A multiple-valued logic: A tutorial and appreciation. Computer, 21, 17–27, 1988.
[3] J. Lukasiewicz, O logice trojwartosciowej. Ruch Filozoficzny, 15, 169–171, 1920.
[4] E.L. Post, Introduction to a general theory of elementary propositions. American Journal of
Mathematics, 43, 163–185, 1921.
[5] A. Ngom, C. Reischer, D.A. Simovici, and I. Stojmenović, Set-valued logic algebra: A carrier
computing foundation. Multiple-Valued Logic — An International Journal, 2, 183–216, 1997.
[6] S.C. Chan, L.S. Hsu, and H.H. Teh, On neural logic networks. Neural Networks, 1 (Suppl. I.), 428,
1988.
[7] V. Milutinović, A. Ngom, and I. Stojmenović, Strip — a strip-based neural network growth
algorithm for learning multiple-valued functions. IEEE Transactions on Neural Networks, 12,
212–227, 2001.
[8] A. Ngom, Synthesis of Multiple-Valued Logic Functions by Neural Networks, Ph.D. thesis,
Computer Science Department, University of Ottawa, Ottawa, Ontario, Canada, October 1998.
[9] Z. Obradović, Computing with nonmonotone multivalued neurons. Multiple-Valued Logic — An
International Journal, 1, 271–284, 1996.
[10] Z. Tang, Q. Cao, and O. Ishizuka, A learning multiple-valued logic networks: Algebra, algorithm,
and application. IEEE Transactions on Computers, 47, 247–251, 1998.
[11] G. Wang and H. Shi, Tmlnn: Triple-valued or multiple-valued logic neural network. IEEE
Transactions on Neural Networks, 9, 1099–1117, 1998.
[12] T. Watanabe, M. Matsumoto, M. Enokida, and T. Hasegawa, A design of multi-valued logic
neuron. In Proceedings of the 20th IEEE International Symposium on Multiple-Valued Logic, 1990,
pp. 418–425.
[13] Q. Cao, O. Ishizuka, Z. Tang, and H. Matsumoto, Algorithm and implementation of a learn-
ing multiple-valued logic network. In Proceedings of the 23rd IEEE International Symposium on
Multiple-Valued Logic, 1993, pp. 202–207.
[14] Z. Tang, O. Ishizuka, Q. Cao, and H. Matsumoto, Algebraic properties of a learning multiple-
valued logic network. In Proceedings of 23rd IEEE International Symposium on Multiple-Valued
Logic, pp. 196–201, 1993.
[15] Z. Obradović and I. Parberry, Computing with discrete multivalued neurons. Journal of Computer
and System Sciences, 45, 471–492, 1992.
[16] Z. Obradović and I. Parberry, Learning with discrete multivalued neurons. Journal of Computer
and System Sciences, 49, 375–390, 1994.
[17] R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis, John Wiley & Sons, New York,
1973.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 423 — #51

24-424 Handbook of Bioinspired Algorithms and Applications

[18] I. Parberry and G. Schnitger, Parallel computation with threshold functions. Journal of Computing
and System Science, 36, 278–302, 1988.
[19] K.Y. Siu, V. Roychowdhury, and T. Kailath, Discrete Neural Computation: A Theoretical Foundation,
Information and System Sciences Series. Thomas Kailath, Series Editor. Prentice-Hall, 1995.
[20] S. Muroga, Threshold Logic and its Applications, Wiley Interscience, New York, 1971.
[21] D. Haring, Multi-threshold threshold elements. IEEE Transactions on Electronic and Computer, 15,
45–65, 1965.
[22] S. Olafsson and Y.A. Abu-Mostafa, The capacity of multilevel threshold functions. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 10, 277–281, 1988.
[23] R. Takiyama, Multiple threshold perceptron. Pattern Recognition, 10, 27–30, 1978.
[24] D. Acketa and J. Žunić, On the number of linear partitions of the (m, n)-grid. Information
Processing Letters, 38, 163–168, 1991.
[25] O. Ishizuka, Multivalued multithreshold networks. In Proceedings of the 6th IEEE International
Symposium on Multiple-Valued Logic, 1976, pp. 44–47.
[26] T. Sasao, On the optimal design of multiple-valued plas. IEEE Computer, 38, 582–592, 1989.
[27] A. Blum and R.L. Rivest, Training a 3-node neural network is NP-Complete. In Proceedings of the
1st Workshop on Computational Learning Theory. Morgan Kaufmann, San Mateo, CA, 1988, p. 9.
[28] J.F. Miller and P. Thomson, Highly efficient exhaustive search algorithm for optimizing canonical
ternary Reed-Muller expansions of Boolean functions. International Journal on Electronics, 76,
37–56, 1994.
[29] C. Yildirim, J.T. Butler, and C. Yang, Multiple-valued PLA minimization by concurrent multiple
and mixed simulated annealing. In Proceedings of the 23rd IEEE International Symposium on
Multiple-Valued Logic, pp. 17–23, 1993.
[30] S. Kirkpatrick, C.D. Gelatt, and M.P. Vecchi, Optimization by simulated annealing. Science, 220,
671–680, 1983.
[31] A. Kaczmarek, V. Antonenko, S. Yanushkevich, and E.N. Zaitseva, Algorithm for network to
realize linear MVL functions using arithmetical logic. In Proceedings of the 12th International
Conference on System Science, pp. 23–30.
[32] Y. Hata, K. Hayase, T. Hozumi, N. Kamiura, and K. Yamato, Multiple-valued logic minimization
by genetic algorithms. In Proceedings of the 27th IEEE International Symposium on Multiple-Valued
Logic, 1997, pp. 97–102.
[33] A. Lloris-Ruiz, J.F. Gomez-Lopera, and R. Roman-Roldan, Entropic minimization of multiple-
valued functions. In Proceedings of the 23rd IEEE International Symposium on Multiple-Valued
Logic, 1993, pp. 24–28.
[34] A. Ngom, C. Reischer, D.A. Simovici, and I. Stojmenović, Learning with permutably homogeneous
multiple-valued multiple-threshold perceptrons. Neural Processing Letters, 12, 2000, Proceedings of
the 28th IEEE International Symposium on Multiple-Valued Logic, May 1998, pp. 161–166,
[35] A. Ngom, I. Stojmenović, and R. Tošić, The computing capacity of three-input multiple-valued
one-threshold perceptrons. Neural Processing Letters, 14, 141–155, 2001.
[36] M.H. Abd-El-Barr, S.G. Zaky, and Z.G. Vranesić, Synthesis of multivalued multithreshold
functions for ccd implementation. IEEE Transactions on Computers, 35, 124–133, 1986.
[37] R. Takiyama, The separating capacity of a multithreshold threshold element. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 7, 112–116, 1985.
[38] A. Ngom, I. Stojmenović, and J. Žunić, On the number of multilinear partitions and the computing
capacity of multiple-valued multiple-threshold perceptrons. IEEE Transactions on Neural Networks,
14, 469–477, 2003, Proceedings of the 29th IEEE International Symposium on Multiple-Valued Logic,
IEEE Computer Society Technical Committee on Multiple-Valued Logic, IEEE Computer Society,
May 1999, pp. 208–213.
[39] H. Edelsbrunner, Pattern Classification and Scene Analysis. Springer-Verlag, Heidelberg, 1987.
[40] D.E. Rumelhart, G.E. Hinton, and R.J. Williams, Learning internal representations by error
propagation. In Parallel Distributed Processing: Explorations in the Microstructure of Cognition,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 424 — #52

Synthesis of Multiple-Valued Circuits by Neural Networks 24-425

(J.L. McClelland, D.E. Rumelhart, and the PDP Research Group, Eds.), Vol. I Foundations.
MIT Press, Cambridge, MA, 1986.
[41] X. Yao, Evolving artificial neural networks. Proceedings of the IEEE, 87, 1423–1447, 1999.
[42] S. Judd, In Proceedings of the IEEE 1st Conference on Neural Networks, Vol. 2, San Diego, CA, 1987,
p. 685.
[43] T.G. Dieterich and G. Bakiri, Solving multiclass learning problems via error-correcting output
codes. Journal of Artificial Intelligence Research, 2, 263–386, 1995.
[44] M. Marchand, M. Golea, and P. Ruján, A convergence theorem for sequential learning in two-layer
perceptrons. Europhysics Letters, 11, 487–492, 1990.
[45] P. Ruján and M. Marchand, A geometric approach to learning in neural networks. Complex Systems,
3, 229–242, 1989.
[46] S.A.J. Keibek, H.M.A. Andree, M.H.F. Savenije, G.T. Barkema, and A. Taal, A fast partitioning
algorithm and a comparison of feedforward neural networks. Europhysics Letters, 18, 555–559,
1992.
[47] M. Marchand and M. Golea, On learning simple neural concepts: From halfspace intersections to
neural decisions lists. Network: Computation in Neural Systems, 4, 67–85, 1993.
[48] R.L. Rivest, Linear decision list. Machine Learning, 2, 229–246, 1987.
[49] S.E. Fahlman and C. Lebière, The cascade-correlation learning architecture. Advances in Neural
Information Processing Systems, 2, 254, 1990.
[50] M. Frean, The upstart algorithm: A method for constructing and training feedforward neural
networks. Neural Computation, 2, 198–209, 1990.
[51] I.K. Sethi, Entropy nets: from decision trees to neural networks. In Proceedings of the IEEE, 78,
1605–1613, 1990.
[52] J.A. Sirat and J.P. Nadal, Neural trees: A new tool for classification. Network, 1, 423–438, 1990.
[53] M. Mezard and J.P. Nadal, Learning in feedforward layered networks: The tiling algorithm. Journal
of Physics A, 22, 2191–2203, 1989.
[54] G.T. Barkema, H.M.A. Andree, and A. Tall, The patch algorithm: Fast design of binary feedforward
neural networks. Network, 5, 393–407, 1993.
[55] F.M. Frattale-Mascioli and G. Martinelli, A constructive algorithm for binary neural networks:
The oil-spot algorithm. IEEE Transactions on Neural Networks, 6, 794–797, 1995.
[56] S. Young and T. Downs, Carve: A constructive algorithm for real-valued examples. IEEE
Transactions on Neural Networks, 9, 1180–1190, 1998.
[57] S.I. Gallant, Three constructive algorithms for network learning. In Proceedings of the
8th Annual Conference on Cognitive Science Society, Amherst, MA, August 15–17, 1986,
pp. 652–660.
[58] M. Golea and M. Marchand, A growth algorithm for neural network decision trees. Europhysics
Letters, 12, 205–210, 1990.
[59] M.M. Muselli, On sequential construction of binary neural networks. IEEE Transactions on Neural
Networks, 6, 678–690, 1995.
[60] J.P. Nadal, Study of a growth algorithm for neural networks. International Journal of Neural Systems,
1, 55–59, 1989.
[61] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
[62] J.H. Holland, Adaptation in Natural and Artificial Systems. Michigan University Press, Ann Arbor,
MI, 1975.
[63] J.D. Schaffer, L.D. Whitley, and L.J. Eshelman, Combinations of genetic algorithms and neural
networks: A survey of the state of the art. In COGANN-92: International Workshop on Combi-
nations of Genetic Algorithms and Neural Networks, (L.D. Whitley and J.D. Schaffer, Eds.), IEEE
Computer Society Press, Washington, 1992.
[64] A. Ngom, I. Stojmenović, and Z. Obradović, Minimization of multiple-valued multiple-threshold
perceptrons by genetic algorithms. In Proceedings of the 28th IEEE International Symposium on

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 425 — #53

24-426 Handbook of Bioinspired Algorithms and Applications

Multiple-Valued Logic, Fukuoka, Japan, May 27–29, 1998, IEEE Computer Society Technical
Committee on Multiple-Valued Logic, IEEE Computer Society, pp. 209–214.
[65] T. Bäck, Evolutionary Algorithms in Theory and Practice. Oxford University Press, Oxford, UK,
1996.
[66] H.-P. Schwefel, Numerical Optimization of Computer Models. Wiley, Chichester, New York, 1981.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C024” — 2005/8/17 — 18:15 — page 426 — #54

 25
On the Computing
Capacity of
Multiple-Valued
Multiple-Threshold
Perceptrons

25.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-427

Multiple-Valued Logic Neural Networks •
The (n, k, s)-Perceptrons • Problem Statement
25.2 Multilinear Separability and Multilinear Partition . . . . 25-430
25.3 Capacity of Multiple-Valued Multiple-Threshold
Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-431
25.4 Linear Partitions and Minimal Pairs . . . . . . . . . . . . . . . . . . . 25-432
25.5 The Computing Capacity of Three-Input
Multiple-Valued One-Threshold Perceptron . . . . . . . . . . 25-433
Two-Dimensional Case • Generalization of the
Two-Dimensional Case • Three-Dimensional Case •
Polynomial Time Algorithm for Computing the Capacity of
Discrete (3, k, 1)-Perceptrons
25.6 On the Number of Multilinear Partitions of
Multiple-Valued Multiple-Threshold Perceptrons . . . . 25-440
Counting Multilinear Partitions of Subsets in General
Position • Counting Multilinear Partitions of the (k, k)-Grid
Alioune Ngom • Complexity of Counting the Number of (2, k 2 , s)-partitions
Ivan Stojmenović 25.7 Conclusion and Open Problems . . . . . . . . . . . . . . . . . . . . . . . . 25-450
Joviša Žunić References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-451

25.1 Introduction
It has been shown in literature that analog neurons of limited precision are essentially discrete multiple-
valued neurons. In this chapter, we study the computational abilities of the Discrete Multiple-Valued
Multiple-Threshold Perceptron. This neuron model extends and generalizes the well-known (two-valued
one-threshold) perceptron and other previously studied models. We introduce the concept of multilinear

25-427

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 427 — #1

25-428 Handbook of Bioinspired Algorithms and Applications

partition of a point set V ⊂ R n and the concept of multilinear separability of a function f : V →

K = {0, . . . , k − 1}. Based on well-known relationships between linear partitions and minimal pairs,
an exact and general formula is derived for the number of linear partitions of a given point set V in
three-dimensional space, depending on the configuration formed by the points of V . The set V can be
a multi-set, that is, it may contain points that coincide. Based on the formula, we obtain an efficient
algorithm for counting the number of k-valued logic functions simulated by a three-input k-valued one-
threshold perceptron. The n-input k-valued s-threshold perceptrons partition the input space V into
s + 1 regions with s parallel hyperplanes. We obtain results on the capacity of a single such perceptron,
respectively, for V ⊂ R n in general position and for V = K 2 . Finally, we describe a fast polynomial-time
algorithm for counting the multilinear partitions of K 2 .

25.1.1 Multiple-Valued Logic Neural Networks

A discrete neuron is a processing unit whose transfer function outputs a discrete value. An example of
such transfer function is the linear threshold function. A discrete n-input multiple-valued neuron has a
discrete transfer function and realizes a function of n variables ranging in the set S ⊆ R with values in
K = {0, . . . , k − 1}(k > 0), that is, computes a function f : S n ⊆ R n → K . For S = K we refer to
the processing unit as a multiple-valued logic neuron since it simulates a multiple-valued logic function
f : K n  → K . Multiple-valued logic neural networks are thus neural networks composed of multiple-
valued logic neurons as processing units. The first model of multiple-valued logic neural networks were
introduced in Reference 1 and since then various other models have been described References 2 to 6.

25.1.2 The (n, k, s)-Perceptrons

Let R be the set of real numbers. Let V ⊆ R n and K = {0, . . . , k − 1}, with k ≥ 2. An n-input k-valued real
function f : V → K is a function with real-valued inputs and k-valued outputs. When V ⊆ K n , we will
refer to f as an n-input k-valued logic function. We denote by Pkn the set of all n-input k-valued real (resp.


logic) functions f : V  → K and by Pk = n≥1 Pkn the set of all k-valued real (logic) functions. Thus Pkn
n
consists of k k k-valued logic functions. The set Pkn is called the set of n-ary operations on K in universal
algebras and the set of n-ary functions of k-valued logic in multiple-valued logic algebras. For instance,
for k = 2, P2 is the set of all two-valued logic functions.
A discrete neuron is a processing unit whose transfer function outputs a discrete value. An example of
such transfer function is the linear threshold function. A discrete n-input multiple-valued neuron has a
discrete transfer function and realizes a function of n variables ranging in a set V ⊆ R n with values in K ,
that is, computes a function f : V  → K . For V ⊆ K n we refer to the processing unit as a multiple-valued
logic neuron since it simulates a multiple-valued logic function f : K n → K . Our model of multiple-valued
logic neuron is the n-input k-valued s-threshold perceptron defined below.
In the theory of multiple-valued logic functions there is an important class of functions called multiple-
valued multiple-threshold functions [7]. Such functions are used in the design of classes of multiple-valued
logic circuits called programmable logic arrays [8]. A k-valued s-threshold function of one variable y ∈ R is
defined as

o0 if y < t1 ,
t , o
gk,s ( y) = oi if ti ≤ y < ti+1 for 1 ≤ i ≤ s − 1, (25.1)

os if ts ≤ y,

where o = (o0 , . . . , os ) ∈ K s+1 is the output vector, t = (t1 , . . . , ts ) ∈ R s is the threshold vector — with
ti ≤ ti+1 (1 ≤ i ≤ s − 1) — and s (1 ≤ s ≤ k n − 1) is the number of threshold values.
Multiple-threshold devices [9] are threshold elements containing multiple levels of excitation
(thresholds). Among their qualities, is that given enough thresholds, a single multiple-threshold element
can realize any given function operating on a finite domain [7].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 428 — #2

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-429

An n-input k-valued s-threshold perceptron [3,10,11], abbreviated as (n, k, s)-perceptron, computes an

t , o ) given by
n (w,
n-input k-valued weighted s-threshold function Fk,s

o0 x < t1 ,
if w
t , o
n
Fk,s t , o )( x ) =
(w, gk,s (w
x ) = oi if ti ≤ w x < ti+1 , (25.2)

os if ts ≤ w x ,

where input vector x = (x1 , . . . , xn ) ∈ V , weight vector w = (w1 , . . . , wn ) ∈ V , threshold vector

t = (t1 , . . . , ts ) ∈ R s and ti ≤ ti+1 (1 ≤ i ≤ s − 1 and 1 ≤ s ≤ k n − 1, the number of threshold values),
and output vector o = (o0 , . . . , os ) ∈ K s+1 .
t , o
The perceptron’s transfer function is a k-valued s-threshold function gk,s : R  → K . A (n, k, s)-perceptron
simulates a k-valued function f : V → K . Depending on V we will either refer to real or logic (n, k, s)-
perceptrons. It is well known that any n-input k-valued logic function can be transformed into a k-valued
s-threshold function [7,12] for some s.
An (n, k, s)-perceptron is monotone if o is monotone, that is o0 ≤ · · · ≤ os or o0 ≥ · · · ≥ os , otherwise
it is nonmonotone. For example, the well-known binary perceptron (which is a [n, 2, 1]-perceptron) studied
in Reference 13 is monotone. A (n, k, s)-perceptron is homogeneous if s = k − 1 and o is the identity
permutation of K (i.e., oi = i for 0 ≤ i ≤ s), it is permutably homogeneous if s ≤ k − 1 and o is
any permutation of s + 1 elements out of K . Functions computed by single such neurons are called
permutably homogeneous functions.
The (n, k, k −1)-perceptrons were introduced by Obradović and Parberry [14] as a model for simulating
continuous perceptrons (i.e., perceptrons containing analog transfer functions) with limited precision on
their inputs. Their computational abilities were extensively studied in Reference 4 whereas Reference 15
examined their learning power. Learning algorithms for some classes of (n, k, s)-perceptrons are inves-
tigated in References 2 and 10. Reference 16 introduced the (n, 2, s)-perceptrons and their computational
power have been intensively studied in References 17 and 18. The neulonets’ units discussed in Reference 1
are a class of (n, 3, 2)-perceptrons with t = (−d, +d), d ∈ R, and o = (1, 0, 2).
Multiple-valued logic neural networks, in our definition, are thus neural networks composed of (n, k, s)-
perceptrons as processing units. It should be noted that the units are not necessarily the same. For example,
one layer of the network may contain (2, 4, 8)-perceptrons while another layer may contain (5, 4, 3)-
perceptrons.
Analog computers are inherently inaccurate due to imperfections in fabrication and fluctuations in
operating temperatures. The classical solution to this problem uses extra hardware to enforce discrete
behavior. However, the brain appears to compute reliably well with inaccurate components without
necessarily resorting to discrete techniques. The continuous neural network is a computational model
based upon certain observed features of the brain. Experimental evidence has shown continuous neural
network to be extremely fault-tolerant; in particular, their performance does not appear to be significantly
impaired when precision is limited. It has been shown by Obradović and Parberry [14] that analog neurons
of limited precision are essentially multiple-valued logic neurons. The first model of multiple-valued logic
neuron (and neural network) was introduced by Chan et al. [1] and since then various other models have
been described (see e.g., References 6, 14, and 19).

25.1.3 Problem Statement

The partitioning of points in n-dimensional space has been well studied. The question of how many
such partitions are possible arises naturally in the theory of pattern classification and machine learning
[13,20–23]. Partitioning by a single hyperplane is fairly limited, and attention has been given to more
complex partitioning methods that arise as simple generalization, such as separation by surfaces with
polynomial equations [24,25], for instance.
In the context of pattern classification and learning, one proposed way of obtaining more powerful
partitioning methods is to use some number s of parallel hyperplanes [17–19,26]. The following question

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 429 — #3

25-430 Handbook of Bioinspired Algorithms and Applications

arises: what is the maximum number of ways in which v points in R n can be partitioned by s parallel
hyperplanes (none of which contains any of the points)?
Answering this question, or obtaining good bounds on the answer, has consequences for the capacity
of generalizations of the threshold functions and threshold elements so central to the theory of artificial
neural networks. (These consequences are discussed in Reference 27.)

25.2 Multilinear Separability and Multilinear Partition

The problem of computing (or simulating) a given function f by a (n, k, s)-perceptron is to determine
s and a vector r = (w, t , o ) ∈ R n+s × K s+1 such that Fk,s n ( r )( x ) = f ( x )(∀ x ∈ V ), that is, f = Fk,sn ( r ).
We will refer to r as an s-representation of Fk,s for f .
n

Let V = { x1 , . . . , x v } ⊆ R n be a set of v vectors (v ≥ 1). A k-valued function f with domain V and

specified by the input-output pairs {( x1 , f ( x1 )), . . . , ( xv , f ( xv ))}, where x i ∈ R n , f ( xi ) ∈ K , is said to be
s-separable if there exist vectors w ∈ R n , t ∈ R s and o ∈ K s+1 such that

o0 x i < t1 ,
if w
f ( xi ) = oj if tj ≤ w x i < tj+1 for 1 ≤ j ≤ s − 1, (25.3)

os if ts ≤ w x i ,

t , o ).
for 1 ≤ i ≤ v. Equivalently, f is s-separable if and only if it has an s-representation defined by (w,
A k-valued function over V is said to be s-nonseparable if it is not s-separable.
In other words, a (n, k, s)-perceptron partitions the space V ⊂ R n into s + 1 distinct classes
[oj ]
H0 , . . . , Hs[os ] , using s parallel hyperplanes, where Hj = { x ∈ V |f ( x ) = oj and tj ≤ w
[o0 ]
x < tj+1 }.
We assume that t0 = −∞ and ts+1 = +∞. Each hyperplane equation denoted by Hj (1 ≤ j ≤ s) is of
the form

x = tj .
Hj : w (25.4)

Multilinear separability (s-separability) extends the concept of linear separability (1-separability of the
common binary 1-threshold perceptron) to the (n, k, s)-perceptron. Linear separability in two-valued
case tells us that a (n, 2, 1)-perceptron can only learn from a space V ⊆ [0, 1]n in which there is a single
hyperplane that separates it into two disjoint halfspaces: H0[0] = { x |f ( x ) = 0} and H1[1] = { x |f ( x ) = 1}.
From the (n, 2, 1)-perceptron convergence theorem [13], concepts that are linearly nonseparable cannot
be learned by a (n, 2, 1)-perceptron. One example of linearly nonseparable two-valued logic function is
the n-input parity function. Likewise, the (n, k, s)-perceptron convergence theorems [10,15] state that a
(n, k, s)-perceptron computes a given function f ∈ Pkn if and only f is s-separable. Figure 25.1 shows an
example of 2-separable 4-valued logic function of P52 .
Let V ⊂ R n with |V | = v ≥ 2. A (n, v, s)-partition is a partition of V by s ≤ v − 1 parallel hyperplanes
(namely (n − 1)-planes), which do not pass through any of the v points. For instance, Figure 25.2 shows
an example of (2, 52 , 2)-partition.
A (n, v, s)-partition determines s + 1 distinct classes S0 , . . . , Ss ⊂ V separated by s parallel (n − 1)-


planes such that si=0 Si = V and si=0 Si = ∅. A (n, v, s)-partition corresponds to an s-separable
k-valued function f ∈ Pkn if and only if all points in the set Si , for 0 ≤ i ≤ s + 1, have the same value
taken out of K . Also, we assume that any two neighboring classes have distinct values. If for a given
(n, v, s)-partition we have Si  = ∅ (0 ≤ i ≤ s + 1) then, clearly, the number of associated functions is
k(k − 1)s . In Section 25.6, we consider only partitions where Si = ∅ for 0 ≤ i ≤ s.
A linear partition of a point set V is a (n, v, 1)-partition, so only a single (n − 1)-plane is required
to separate an n-dimensional space V ⊆ R n into two halfspaces. The enumeration problem for linear
partitions is closely related to the efficiency measurement problem for linear discriminant functions in
pattern recognition [22] and to many other algorithmic problems [28].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 430 — #4

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-431

42 2 2 2 2

32 2 2 0 0

y 20 0 0 0 0

10 0 4 4 4

04 4 4 4 4
0 1 2 3 4
x
FIGURE 25.1 A 2-separable function of P52 .

FIGURE 25.2 A (2, 52 , 2)-partition.

25.3 Capacity of Multiple-Valued Multiple-Threshold

Perceptrons
In this section we explore the computational power of (n, k, s)-perceptrons. More specifically, the question
we will attempt to answer is the following. How many k-valued functions f : V → K can be simulated by
a (n, k, s)-perceptron? This will be referred to as the capacity of a (n, k, s)-perceptron.
There has been intensive interest in threshold logic as the main component of neural network models.
These models provide a direction for pattern recognition systems with distinct natural advantages. The
capacity of these models, as well as their computing power, are directly related to the number of threshold
functions.
The ability of multiple-threshold devices to simulate a larger number of functions compared to single-
threshold devices is vital for the capacity and capabilities of neural network models based on threshold

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 431 — #5

25-432 Handbook of Bioinspired Algorithms and Applications

logic. It is therefore of practical as well as theoretical interest to estimate the number of functions that can
be modeled as multiple-threshold functions for a given number of inputs and threshold levels.
For given s ≥ 2, the capacity of a (n, k, s)-perceptron with domain V (i.e., the number of k-valued
functions over V that can be simulated by the (n, k, s)-perceptron) is approximated by the product of the
number of (n, v, s)-partitions and the number of functions associated with each (n, v, s)-partition of V .
Thus counting the number of (n, v, s)-partitions of V is a first step toward calculating the capacity of
(n, k, s)-perceptrons. We emphasize that the s partitioning (n − 1)-planes of a (n, v, s)-partition do not
pass through any point of V , and therefore we do not obtain the exact number of (n, k, s)-perceptron
computable functions.
Let Ln,v,s be the number of (n, v, s)-partitions of V ⊂ R n and denote by |Fk,s n | the capacity of a (n, k, s)-

perceptron. The capacity of (n, 2, 1)-perceptrons with domain V [29] is well known and is given by
 v
n    2 if n ≥ v − 1,
 v − 1,  
n
|F2,s |=2 =
v−1 v −1 (25.5)
i 
2v − 2 i=n+1 otherwise,
i=0 i

Reference [17] estimated lower and upper bounds for the capacity of (n, 2, s)-perceptrons, using two
essentially different enumeration techniques. The paper demonstrated that the exact number of multiple-
threshold functions depends strongly on the relative topology of the input set. The results corrected a
previously published estimate [18] and indicated that adding threshold levels enhances the capacity more
than adding variables.
In order to answer the question, we first describe well-known relationships between linear partitions
and minimal pairs in Section 25.4. Based on these relationships, we obtain in Section 25.5 an exact and
general formula for the capacity of (3, k, 1)-perceptrons. In Section 25.6 we derive results on the capacity
of (n, k, s)-perceptrons.

25.4 Linear Partitions and Minimal Pairs

An unordered pair ( x , y ) of distinct points of a finite set V ⊆ R n is said to be a minimal pair (with respect
to V ) if there does not exist a third point z of V , which belongs to the open line segment [ x , y ].
A point set V ⊂ R n is said to be in general position if and only if no subset of V of d + 1 points lies on
a (d − 1)-plane (for 1 ≤ d ≤ n) and no two (n − 1)-planes defined from V are parallel (for 2 ≤ d ≤ n).
The number of linear partitions of a finite planar point set, no three points of which are collinear, is well
known [30]. The number of linear partitions of V in general position is a well-known result of pattern
recognition [31,32] and is given by
 n  
v −1
Ln,v,1 = Ln,v−1,1 + Ln−1,v−1,1 = . (25.6)
i
i=0

It should be noted that Equation (25.6) includes the trivial partition.

A relationship between linear partitions and minimal pairs of a finite point set V in the plane was
established in Reference 33. Namely, that the number of linear partitions of V is equal to the number of
minimal pairs in V . Moreover, each minimal pair determines two linear partitions of V (see Figure 25.3)
and conversely, each linear partition of V has exactly two associated minimal pairs (see Figure 25.4) [34].
These facts provide an easier way of counting linear partitions for arbitrary finite planar point sets: it
suffices to count minimal pairs.
In Reference 34 an explicit formula for the number of linear partitions of the (i, j)-grid (a rectangular
part of the infinite grid) is stated, using the correspondence with minimal pairs. The same reference
includes an efficient algorithm for counting linear partitions of the (i, j)-grid in linear time with respect
to the number of its points.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 432 — #6

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-433

P1
D
D

C
C

P2

FIGURE 25.3 Minimal pair (C, D) corresponds to separating lines P1 and P2 .

M1

P
D B

A C

M2

FIGURE 25.4 Separating line P corresponds to minimal pairs (A, B) and (C, D).

Reference 35 derived a formula for the number of linear partitions of a given point set V in two-
dimensional and three-dimensional spaces, depending on the configuration formed by the points of V .
He considered the case where some points of V may coincide.

25.5 The Computing Capacity of Three-Input Multiple-Valued

One-Threshold Perceptron
In this section, an exact and general formula is derived for the number of linear partitions of a given
point set V in three-dimensional space, depending on the configuration formed by the points of V . The
set V can be a multi-set, that is, it may contain points that coincide. Based on the formula, we obtain
an efficient algorithm for counting the number of k-valued logic functions simulated by a three-input
k-valued one-threshold perceptron.
A v-set is a set of cardinality v. The cardinality of a set V is denoted by |V |. Given a point-set V ⊂ R n ,
a linear partition of V is a partition of V into two subsets S0 and S1 , which is induced by an (n − 1)-
dimensional hyperplane H . It is assumed that the intersection of V and H is empty and that the sets S0
and S1 respectively belong to distinct half-spaces with respect to H . In addition, we have S0 ∪ S1 = V and
S0 ∩ S1 = ∅. In the sequel we let |V | = v.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 433 — #7

25-434 Handbook of Bioinspired Algorithms and Applications

The capacity of a (3, k, 1)-perceptron with domain V ⊂ R 3 is then given by

3
|Fk,1 | = k(k − 1)L3,v,1 + k, (25.7)

where the coefficient k(k − 1) is the number of functions associated with each linear partition of V , and
the last term k is the number of functions associated with the trivial linear partition {V , ∅} of V (recall
that (n, v, s)-partitions with empty classes are not included in our definition). Determining L3,v,1 is the
subject of the next sections.
For V ⊂ R 3 in general position, the capacity of (3, k, 1)-perceptrons follows directly from Equations
(25.6) and (25.7). That is,

Corollary 25.1

3 
 
3 v −1
|Fk,1 | = k(k − 1) + k.
i
i=1

In this section, however, V is not necessarily in general position and may even contain points that
coincide, that is, V can be a multi-set. Thus the formula we obtain will be the most general result for
(3, k, 1)-perceptrons.

25.5.1 Two-Dimensional Case

From now on, and for ease of notation, we will denote the number of linear partitions of a point set
V ⊂ R n by LV ,n (it includes the trivial partition {∅, V }). Let V be a planar point set, that is, V ⊂ R 2 , and
MV be its number of minimal pairs. The following well-known statement might be very helpful for the
enumeration of linear partitions of a planar set.

Lemma 25.2 ([2]) LV ,2 = MV + 1.

Formula (25.6) follows directly from Lemma 25.2 since every pair of points in V is a minimal. From Lemma
25.2 and Formula (25.6), the following statement can be easily deduced.

Theorem 25.3 Let p1 , . . . , pd be all lines determined by a planar point-set V (each line contains at least two
points of V ) and let ci,1 denote the number of points of V belonging to the line pi , 1 ≤ i ≤ d. Then

    d  
v −1 v −1 ci,1 − 1
LV ,2 = 1 + + − .
1 2 2
i=1

Proof. If the line pi contains ci,1 points from V , then they determine ( ci,1
2 ) pairs among which exactly
ci,1 −1
ci,1 − 1 are minimal pairs; so the number of pairs that are not minimal is ( ci,1
2 ) − (ci,1 − 1) = ( 2 ).

d ci,1 −1
Then the number of nonminimal pairs of points in V is i=1 ( 2 ). Now, the statement follows from
the fact that the total number of pairs in V is ( v−1 v−1
1 ) + ( 2 ).

25.5.2 Generalization of the Two-Dimensional Case

Before proceeding with the two-dimensional case, let us consider the very special case of one dimension.
If V ⊂ R 1 , then from Formula (25.6), we have
 
v −1
LV ,1 = 1 + = v. (25.8)
1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 434 — #8

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-435

Now, we consider a generalization in R 1 , when some points coincide (multiplicity of points). That is there
j
are points in V with multiple occurrence. By Ai we denote a point Ai ∈ V with multiplicity j, that is, j
points coincide.
c c
Theorem 25.4 Let V = {A11,0 , . . . , Add,0 }, where c1,0 + · · · + cd,0 = v. Then

  d  
v −1 ci,0 − 1
L V ,1 = 1 + − .
1 1
i=1

where L V ,1 is the number of linear partitions of V .

Proof. Obvious.

c VV
cV d ,0
Let V ⊂ R 2 such that |V | = v and V = {A11,0 , . . . , Ad V0 }, where ci,0
V is the multiplicity of the point A
i
0
in V (for 1 ≤ i ≤ d0 ), c1,0 + · · · + cd V ,0 = v and d0 is the number of distinct points in V . Let
V V V V
0

p1 , . . . , pd V (25.9)
1

be all different lines determined by the points of V (each line contains at least two noncoincident points
of V and d1V is the number of all such lines in V ).
V the number of points of V belonging to the line p (with the corresponding multiplicities)
Denote by ci,1 i
cV
V be the number of different lines from (25.9) through the point A i,0 , for 1 ≤ i ≤ d V .
for 1 ≤ i ≤ d1V . Let ri,0 i 0
If we denote by L V ,2 the number of linear partitions of the set V , then the following statement can be
proved.

Theorem 25.5

    d0V 
V −1

v −1 v −1 ci,0
L V ,2 =1+ + −
1 2 1
i=1

d1V   d0V  V 
 V −1
ci,1 ci,0 − 1
V
− + (ri,0 − 1) .
2 2
i=1 i=1

Proof. The proof will be by induction on v. Consider a set of points V such that |V | = v + 1. Let A
be a single point (with multiplicity one) of V , which is a vertex of the convex hull of V . By induction
hypothesis, for the set U = V − {A} we have

LU ,2 = LV −{A},2

    d0U 
U −1

v −1 v −1 ci,0
=1+ + −
1 2 1
i=1

d1U   d0U  U 
 U −1
ci,1 c −1
U
− + (ri,0 − 1) i,0 . (25.10)
2 2
i=1 i=1

Let us determine the number of movable line partitions of U . Project U from the point A into a line a
which separates A from U . Let U  be the projection of U . The number of movable line partitions of U is,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 435 — #9

25-436 Handbook of Bioinspired Algorithms and Applications

according to Theorem 25.4


 d0U   
v −1 U −1
ci,0
L U  ,1 =1+ − . (25.11)
1 1
i=1

From Equations (25.10) and (25.11) it follows that

LV ,2 = LU ,2 + LU  ,1

    d0V 
V −1
 d1V 
V −1
 d0V  V 
v v ci,0 ci,1 V c −1
=1+ + − − + (ri,0 − 1) i,0 . (25.12)
1 2 1 2 2
i=1 i=1 i=1

In the case none of the vertices of the convex hull of V is a single point, we shall see what happen when
the multiplicity of one point A ∈ U (|U | = v), increases by one, producing the set V .


U
U
Denote, for the sake of simplicity, the sums from the right-hand side of (25.10) by U 0 , 1 ,

V
V
V 2
respectively and those from (25.12) by 0 , 1 , 2 . Since

        
v v v −1 v −1
1+ + − 1+ − = v,
1 2 1 2

we need to prove that

 

V 
V 
V 
U 
U 
U
+ − − + − = v. (25.13)
0 1 2 0 1 2

c V +1 cU
Let we have the point Ai i,0 in V instead of Ai i,0 in U . Then


V 
U
− = 1. (25.14)
0 0

V = r U and c V = c U + 1, we have
Taking into account that ri,0 i,0 i,0 i,0


V 
U    U 
U
U
ci,0 c −1
− = (ri,0 − 1) − i,0 ,
2 2
2 2

that is,


V 
U
U U
− = (ri,0 − 1)(ci,0 − 1). (25.15)
2 2

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 436 — #10

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-437

Let {p1 , p2 , . . . , pr U } be the set of lines from Equation (25.9) through the point A. Then
i,0


V 
U i,0  V 
rU
  V −1

c ci,1
− = i,1 −
2 2
2 2 i=1

rU rU rU

i,0 
i,0 
i,0
V V
= (ci,1 − 1) = ci,1 − 1
i=1 i=1 i=1

U U U U
= v − ci,0 + ri,0 ci,0 − ri,0

U U U
= v + ri,0 (ci,0 − 1) − ci,0

U U
= v + (ri,0 − 1)(ci,0 − 1) − 1. (25.16)

Now, Equation (25.13) follows from Equations (25.14) to (25.16).

25.5.3 Three-Dimensional Case

Let V be the set of different points in the three-dimensional space such that |V | = v. Suppose that
p1 , . . . , pd V are all distinct lines determined by the points of V (each line contains at least two different
1
points of S) and R1 , . . . , Rd V are all distinct planes determined by the points of V (each plane contains
2
V , the number of points of
at least three different points of V ). We introduce the following notations: ci,1
V lying on pi ; ci,2 , the number of points of V lying in Ri ; ri,1 , the number of planes from R1 , . . . , Rd V
V V
2
containing the line pi .
If we denote by LV ,3 the number of linear partitions of the set V , then the following statement can be
proved.

Theorem 25.6

     d1V 
V −1
 d2V 
V −1

v −1 v −1 v −1 ci,1 ci,2
LV ,3 =1+ + + − −
1 2 3 2 3
i=1 i=1

d1V
  V 
V c −1
+ (ri,1 − 1) i,1 . (25.17)
3
i=1

Proof. We use induction (with trivial basis) on |V |. Suppose that the statement is valid for |V | = v.
Consider |V | = v + 1.
Take a point A of V . For the sake of simplicity, we may assume (without any loss of generality) that A
is a vertex of the convex hull of V . Denote V − {A} by U . Consider the projection π of U onto a plane
α separating A from U , point A being the center of projection. Those linear partitions of U that can be
established by using planes through the point A are said to be movable with respect to A.
The number of additional linear partition that are obtained after extension of the set U to V (by adding
the point A) is equal to the number of movable (w.r.t. A) linear partitions of U . This last number is equal
to L Y ,2 , that is, to the number of linear partitions of the planar point set Y = π(U ). The corresponding
bijection is established by the projection π . namely, each movable linear partition (w.r.t. A) may be

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 437 — #11

25-438 Handbook of Bioinspired Algorithms and Applications

represented by a plane H through A. The line h = π(H ) corresponds to a linear partition of Y = π(U )
in α. Conversely, given a linear partition of Y with the corresponding line h, the plane through h and A
determines the associated movable linear partition of V . It follows that

LV ,3 = LU ,3 + L Y ,2 . (25.18)

Now Equation (25.17) can be deduced from Equation (25.18) using induction hypothesis and
Theorem 25.5. Namely, by induction hypothesis,
     
v −1 v −1 v −1
LU ,3 = 1 + + +
1 2 3

d1U   d2U   d1U  U 

 U −1
ci,1 U −1
ci,2 c −1
U
− − + (ri,1 − 1) i,1 , (25.19)
2 3 3
i=1 i=1 i=1

while, according to Theorem 25.5,

    d0Y 
Y −1

v −1 v −1 ci,0
L Y ,2 =1+ + −
1 2 1
i=1

d1Y   d0Y  Y 
 Y −1
ci,1 c −1
Y
− + (ri,0 − 1) i,0 . (25.20)
2 2
i=1 i=1

It can be verified that

d1U   d0Y   d1V  

 U −1
ci,1 Y −1
ci,0 V −1
ci,1
+ = , (25.21)
2 1 2
i=1 i=1 i=1

d1U   d1Y   d2V  

 U −1
ci,2 Y −1
ci,1 V −1
ci,2
+ = (25.22)
3 2 3
i=1 i=1 i=1

and

d1

U U −1
 d0Y  Y  d1V  V 
U ci,1 Y ci,0 − 1 V c −1
(ri,1 − 1) + (ri,0 − 1) = (ri,1 − 1) i,1 . (25.23)
3 3 3
i=1 i=1 i=1

Now, the statement follows from Equations (25.18) to (25.23).

25.5.4 Polynomial Time Algorithm for Computing the Capacity of

Discrete (3, k, 1)-Perceptrons
In this section we describe an algorithm, based on Theorem 25.6, for counting the linear partitions of the
three-dimensional square grid K 3 , where K = {0, . . . , k − 1} and k ≥ 2.
To efficiently count the linear partitions, we associate each point (x, y, z) of K 3 with an integer m =
xk + yk + z + 1 ∈ {1, . . . , k 3 }. Also given an integer m ∈ {1, . . . , k 3 } the coordinates of its corresponding
2

point in K 3 are obtained as: x = ((m−1) div k) div k, y = ((m−1) div k) mod k and z = (m−1) mod k.
Let V = K 3 . Next we describe the counting algorithm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 438 — #12

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-439

Step 1: Generate all d2V planes determined by three noncolinear points from V and compute B =

d2V  ci,2
V −1 
i=1 3
as follows.
1. Initialize B and d2V to 0.
3
2. Generate a candidate plane P as a triple m1 < m2 < m3 out of k 3 . There are ( k3 ) candidate
planes among which only d2V planes are valid.
3. Accept the candidate plane P if m1 and m2 are minimal points (i.e., they are the two smallest
values) on it and m3 is minimal point among those noncolinear with m1 and m2 . That is, P
/ {m1 , m2 , m3 } we have m > m2 , and, if
is valid if and only if for each point m in P and m ∈
m is noncolinear with m1 and m2 then also m > m3 .
4. If P is valid, then
(a) d2V ← d2V + 1.
(b) cdVV ,2 ← number of such points m (described above) +3.
2
 c V −1 
(c) B ← B + d2V ,2 .
3
5. Repeat from Step 2 until no more plane can be generated.
Step 2: Generate all d1V lines determined by two points from V and compute the sums A =

d1V  ci,1
V −1 
d1V V  V 
c −1
i=1 2
and C = i=1 (ri,1 − 1) i,13 as follows.
1. Initialize A, C, and d1V to 0.
 3
2. Generate a candidate line L as a pair m1 < m2 out of k 3 . There are k2 candidate lines
among which only d1V lines are valid.
3. Accept the candidate line L if m1 and m2 are minimal points on it. That is, L is valid if and
/ {m1 , m2 } we have m > m2 .
only if for each point m in L and m ∈
4. If L is valid, then
(a) d1V ← d1V + 1.
(b) cdVV ,1 ← number of such points m (described above) +2.
1
 c V −1 
(c) A ← A + d1V ,1 .
2
(d) rdVV ,1 ← number of planes (with third point from V ) that pass through L.
1
  c V −1 
(e) C ← C + rdVV ,1 − 1 d1V ,1 .
1 3
5. Repeat from Step 2 until no more line can be generated.
Step 3: Apply Theorem 25.6 and report the results. That is,
1. LV ,3 = v + ( v3 ) − A − B + C.
3 | = k(k − 1)(L
2. |Fk,1 V ,3 − 1) + k = k(k − 1)LV ,3 − k + 2k.
2

3. Write d1 , d2 , LV ,3 , and |Fk,1 |.

V V 3

Figure 25.5 and Figure 25.6 show, respectively, the codes for Step 1 and Step 2.

Remark 1. There is no need to memorize lines and planes, not even to memorize points on them, the
algorithms work with one plane (resp. line) at a time. Also, to avoid errors or imprecisions, equations of
lines and planes should be made with integer coefficients.

Remark 2. The equation of a plane containing points D, E, F is obtained as follows. Find the normal
−
→ − →
vector n = ED × EF (cross product), which gives the coefficients a, b, c, then find d from ax0 + by0 +
cz0 + d = 0, (where (x0 , y0 , z0 ) is a point on the plane. Also, three points D, E, F are colinear if and only
) × (E −F
if (D −F ) = 0 (cross product).

In Step 1, O(k 9 ) candidate planes are generated and each generated plane is checked for validity at most
k3 times, which gives us a total of O(k 12 ) validity tests. Thus Step 1 has a time complexity polynomial

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 439 — #13

25-440 Handbook of Bioinspired Algorithms and Applications

FIGURE 25.5 Code for step 1.

TABLE 25.1 Results for k = 2, . . . , 8

k K3
d1 d2K
3
A B C LK 3 ,3 3 |
|Fk,1

2 28 20 0 12 0 52 104
3 253 491 49 1,552 0 1,351 8,103
4 1,492 7,502 300 24,422 350 17,356 208,264
5 5,485 52,013 1,338 201,260 4,252 119,529 2,390,565
6 17,092 297,464 3,712 1,031,292 25,852 647,424 19,422,696
7 41,905 1,119,791 10,227 4,322,716 119,598 2,453,869 103,062,463
8 95,140 3,900,890 21,948 14,236,066 418,546 8,399,764 470,386,736

on k. In Step 2, O(k 6 ) lines are generated and each generation consists of at most k 3 validity checks and
O(k 6 ) time to compute the number of planes that contain the given line. Therefore Step 2 has also a time
complexity of O(k 12 ) (Table 25.1).

25.6 On the Number of Multilinear Partitions of

Multiple-Valued Multiple-Threshold Perceptrons
Based on well-known relationships between linear partitions and minimal pairs, we derive formulae
for the number of multilinear partitions of a point set in general position and of the set K 2 . The
(n, k, s)-perceptrons partition the input space V into s + 1 regions with s parallel hyperplanes. We
obtain results on the capacity of a single (n, k, s)-perceptron, respectively for V ⊂ R n in general position
and for V = K 2 . Finally, we describe a fast polynomial-time algorithm for counting the multilinear
partitions of K 2 .

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 440 — #14

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-441

FIGURE 25.6 Code for Step 2.

25.6.1 Counting Multilinear Partitions of Subsets in General Position

In this section, V ⊂ R n is in general position and Ln,v,s denotes its number of (n, v, s)-partitions. Clearly,
every pair of points in V is a minimal pair. We first obtain the number of (2, v, s)-partitions, that is, for
the two dimensional space, and then we find Ln,v,s .
We rewrite the number of linear partitions of V in general position [31,32] with a little modification as

n 
 
v −1
Ln,v,1 = Ln,v−1,1 + Ln−1,v−1,1 = . (25.24)
i
i=1

The difference with the original formula (25.6) is that the index i starts with 1 instead of 0. The reason we
start from i = 1 is that we do not include partitions containing empty classes.
To count the (2, v, s)-partitions, we associate each (2, v, s)-partition with a slope σ ∈ R (or equivalently,
a minimal pair) as follows:

• For each of the s partitioning lines, choose the corresponding minimal pair which has the smaller
slope.
• The associated slope of the given (2, v, s)-partition is the maximum among these smaller slopes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 441 — #15

25-442 Handbook of Bioinspired Algorithms and Applications

 v−2 
Lemma 25.7 The number of (2, v, s)-partitions associated with a given slope σ is s−1 .

Proof: For a given minimal pair, rotate its slope σ to increase it a bit in order to obtain the direction
of separation P that corresponds to σ . Sort the points of V along this direction, that is, according to
 as one point. Then we can choose s − 1
their distance to the line P. Consider the selected minimal pair
additional separating lines (parallel to P) for v − 2 points in v−2
s−1 ways.

Theorem 25.8
  
v −2 v
L2,v,s ≤ .
s−1 2

 
Proof. Since V is in general position, then there are v2 slopes. The inequality is explained by the fact
that two distinct choices of minimal pairs (or slopes) may have sets of associated (2, v, s)-partitions that
intersect each other. Therefore a given partition may be counted many times, depending on the config-
uration of V . For instance, consider one of the partitions in Figure 25.7. Clearly, the same partition can
be obtained either by selecting the upper minimal pair or by selecting the lower minimal pair (indeed in
this example, they both give the same set of associated partitions, even though the points are in general
position). We have equality only when s = 1.

Anthony [36] gave an upper bound on Ln,v,s , refining an upper bound of Olafsson and Abu-Mostafa
[17] which is itself a correction of a claimed upper bound of Takiyama [18]. Their result is as follows:

Theorem 25.9 (Anthony [36]) The maximum possible number of ways in which v points in R n can be
partitioned by s parallel hyperplanes is bounded as follows:

s  
2  n−1
 v −1 
Ln,v,s ≤2 S(v − 1, n − 1 − 2j),
i
i=0 j=0

r
where S(a, b) is the Stirling number of the first kind, the coefficient of x b in j=1 (1 + jx)

Paper [36] also gave another upper bound on Ln,v,s , better than that of Theorem 25.9 as follows:

Theorem 25.10 (Anthony [36]) The maximum possible number of ways in which v points in R n can be
partitioned by s parallel hyperplanes is bounded as follows:

 
n+s−1
vs − 1

Ln,v,s ≤ .
i
i=0

Corollary 25.11 The number of n-input k-valued s-separable functions f : V → K is bounded as follows:

 
n+s−1
vs − 1

n
|Fk,s | ≤ k(k − 1)s .
i
i=0

Proof. Given a (n, v, s)-partition S0 , . . . , Ss , each class can take one of k values from K such that any
two neighboring classes have different values. So there are k(k − 1)s ways to assign values to a (n, v, s)-
partition. Each assignment of values to a (n, v, s)-partition defines a unique k-valued s-separable function.
The inequality is explained by the fact that some functions can be obtained by at least two different
(n, v, s)-partitions (Figure 25.7 shows an example of such function).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 442 — #16

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-443

A p-permutation (or permutation of p elements out) of P = {a0 , . . . , ap−1 } is an arrangement of p > 0

elements into p positions. For example, a0 a3 a2 a4 a1 and a3 a0 a1 a4 a2 are two different five-permutations.
The order of the elements is important. There are p! distinct p-permutations. An (e, p)-permutation (or
permutation of e elements out) of P is an arrangement of e distinct elements of P, with e ≤ p. For instance,
a1 a2 a4 and a3 a0 a1 are two distinct (3, 5)-permutations. The total number of (e, p)-permutations is
p!/(p − e)! When e = p we obtain p-permutations. The permutations we consider here are permutations
without repetitions (i.e., without repeated elements).
An (n, k, s)-perceptron is homogeneous if o , its output vector, is the identity permutation on K , that is,
oi = i for 0 ≤ i ≤ s, otherwise it is heterogeneous. An s-separable function f : V → K is permutably
homogeneous if and only if s ≤ k − 1 and, for any partitioning of V , no two separate classes have equal
value [10]. That is, its output vector is a (s + 1, k)-permutation. These functions are determined by
(n, v, s)-partitions in which each of the s + 1 classes S0 , . . . , Ss maps to a distinct value in K . Thus, for
permutably homogeneous (n, k, s)-perceptrons we necessarily have s ≤ k − 1.
Theorem 25.12 A permutably homogeneous s-separable function f : V → K has a unique (n, v, s)-partition.
Proof. Suppose f has two distinct (n, v, s)-partitions P1 and P2 . Consider two points x and y that are in
the same class with respect to P1 but in different classes with respect to P2 . We have that f ( x ) = f ( y ),
because x and y share the same class in P1 . This means that two distinct classes in P2 have the same value.
This contradicts the definition of permutably homogeneous function.

Corollary 25.13 The number of permutably homogeneous n-input k-valued s-separable functions
f : V → K is

k!  vs − 1
n+s−1
n
|Gk,s |≤ .
(k − s − 1)! i
i=0

Proof. For a (n, v, s)-partition where s ≤ k − 1, the number of ways to map the s + 1 classes S0 , . . . , Ss to
distinct values in K equals the number of (s+1, k)-permutations, that is, k!/(k−s−1)! Each such (s+1, k)-
permutation uniquely determines a permutably homogeneous s-separable function. From Theorem 25.12,
each permutably homogeneous s-separable function uniquely determines a (n, v, s)-partition.

25.6.2 Counting Multilinear Partitions of the (k, k)-Grid

We investigated the computing capacity of the three-input multiple-valued one-threshold perceptron
in Reference 11, and obtained a very general formula for the number of linear partitions of a three-
dimensional point-set (where the points can be in any possible configuration and the set can be a multi-set).
The formula is the most general result for (3, k, 1)-perceptrons. In particular, we studied the special case
of two-dimensional point-sets and derived a general formula for the number of linear partitions of a
planar set. Such planar set can be a multi-set in general or nongeneral position. In Reference 34 an explicit
formula for the number of linear partitions of the (i, j)-grid (a rectangular part of the infinite grid) is
stated, using the correspondence with minimal pairs. The same reference includes an efficient algorithm
for counting linear partitions of the (i, j)-grid in linear time with respect to the number of its points. For
a grid configuration, our two-dimension result is a generalization and an extension of Reference 34 to
a grid multi-set (i.e., a grid set with multiplicity of points). In this section, we generalize the counting
method of Reference 34 to enumerate multilinear partitions of the (k, k)-grid K 2 .
An unordered pair ( x , y ) of distinct points of a finite set V ⊆ R n is said to be a minimal pair (with
respect to V ) if there does not exist a third point z of V which belongs to the open line segment [ x , y ].
Let ( x , y ) be a pair of points from a finite planar grid. Let a = max(|x2 − x1 |, |y2 − y1 |) and b =
min(|x2 − x1 |, |y2 − y1 |). One well-known condition for a pair ( x , y ) to be minimal is that a and b are
relatively prime, that is, their greatest common divisor is 1. a ⊥ b will denote that the integers a and b are
mutually simple.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 443 — #17

25-444 Handbook of Bioinspired Algorithms and Applications

1 0 1 0

0 1 0 1

FIGURE 25.7 Two (2, 4, 2)-partitions for the same function.

→
m y

b/a

→
x

FIGURE 25.8 Rotating slope b/a toward slope m.

Thus the number of minimal pairs corresponds to the number of pairs ( x , y ) of the (i, j)-grid such that
a ⊥ b. Let natural numbers i and j be given so that i ≤ j. The generalized Farey (i, j)-sequence Fi,j [34]
is the strictly increasing sequence of all the fractions of the form b/a, where the integers a and b satisfy:
a ⊥ b, 0 < b < a ≤ j, b ≤ i. Thus the sequence F4,7 is as follows:

1 1 1 1 2 1 2 3 1 4 3 2 3 4
.
7 6 5 4 7 3 5 7 2 7 5 3 4 5

The Farey i-sequence Fi for any positive integer i is the set of irreducible rational numbers b/a, with
0 ≤ b ≤ a ≤ i and a ⊥ b, arranged in increasing order [37]. So the sequence F4 is:

0 1 1 1 2 3 1
.
1 4 3 2 3 4 1

The length of the sequence Fi,j (resp. Fi ) will be denoted by |Fi,j | (resp. |Fi |). Also, Fi,j
d (resp. F d ) stands
i
for the d-th fraction in Fi,j (resp. Fi ), 1 ≤ d ≤ |Fi,j | (resp. |Fi |).
To count the (2, k 2 , s)-partitions, we associate to each (2, k 2 , s)-partition a fraction (b/a) ∈ Fk−1 as
follows:

• For each of the s partitioning lines, choose one of the two minimal pairs that has the smaller slope.
• The associated slope of the given (2, k 2 , s)-partition is the maximum among these smaller slopes.

Then to enumerate or generate the (2, k 2 , s)-partitions associated with a given b/a we will need Lemmas
25.14 and 25.15 below. As in Figure 25.8, rotate a line segment [ x , y ] whose slope b/a is irreducible to
increase it for a small amount, so that we obtain a straight line P with slope m ∈ / Fk−1 . P is the direction
for separation and corresponds to the line segment [ x , y ] with slope b/a.

Lemma 25.14 Slope b/a of the line segment [ x , y ] is greater than or equal to the associated slope of any
(2, k 2 , s)-partition in direction parallel to P (where P is the direction of separation that corresponds to the line
segment [ x , y ]).

Proof. Consider Figure 25.9, where t ∈ Fk−1 is the associated slope of a (2, k 2 , s)-partition parallel to P
and m ∈/ Fk−1 is the slope of P. From our construction of P we have that m > (b/a) ∈ Fk−1 but near b/a.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 444 — #18

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-445

P
m

Convex hull
one side

Convex hull
other side

FIGURE 25.9 t ≤ b/a < m.

d+1
Clearly, from Figure 25.9, t < m. Let d be the rank of b/a in Fk−1 . Since Fk−1
d = (b/a) < m < Fk−1 ,t <
m and, t ∈ Fk−1 , therefore we have t ≤ b/a.

Lemma 25.15 Slope b/a of the line segment [ x , y ] is equal to the associated slope of a (2, k 2 , s)-partition in
direction parallel to P if and only if at least one of the s separating lines intersects a minimal pair with slope
b/a (where P is the direction of separation that corresponds to the line segment [ x , y ]).

Proof. Let t ∈ Fk−1 be the associated slope of a (2, k 2 , s)-partition parallel to P then from Lemma 25.14
we have t ≤ (b/a). ⇒) Clearly, if none of the separating lines intersects any minimal pair of slope b/a
then b/a will not be chosen as the smaller slope for any of the separating lines. Therefore b/a cannot be
the associated slope of any (2, k 2 , s)-partition parallel to P, that is (b/a)  = t . ⇐) Since t is the associated
slope of a (2, k 2 , s)-partition parallel to P, then t is greater than or equal to the slopes of the two minimal
pairs associated with each separating line. In particular, we have t ≥ (b/a) for a separating line that
intersects a minimal pair of slope b/a. Since t ≤ (b/a) and t ≥ (b/a) therefore t = (b/a).

Reference 34 obtained an exact formula for the number of linear partitions of the (i, j)-grid. Substituting
for the (k, k)-grid we obtain the following corollary.

Corollary 25.16

L2,k 2 ,1 = 2k(k − 1) + 2(k − 1)2 + 4 (k − a)(k − b).
a⊥b,0<b<a<k

Proof. If b = 0, then the number of minimal pairs is equal to 2k(k − 1), which is the number of minimal
vertical and horizontal segments of the (k, k)-grid (these are k(k − 1) each). This argument explains the
first term. One can see from Figure 25.10 that there are k − 1 segments in each row and column and that
there are k rows and columns. If b > 0 and a ⊥ b then a = b implies a = b = 1. In that case, the
(k, k)-grid contains obviously 2(k − 1)2 minimal segments with slope 11 , which explains the second term.
In Figure 25.11, it suffices to count the number of squares of size 1 × 1. There are (k − 1)2 such squares and
each square contributes 2 diagonals. Finally, if a > b > 0 and a ⊥ b, then there are exactly 2(k − a)(k − b)
horizontal and vertical rectangles of size a × b in the (k, k)-grid. The third term comes from the fact that
there are two minimal segments per each rectangle (the diagonal segments). In Figure 25.12, the horizontal
rectangle in the lower left corner can be translated by one unit length k − b − 1 times in the horizontal
direction to reach the lower right corner, and k − a − 1 times in the vertical direction to reach the upper
left corner (where k = 5, a = 1, b = 2). This gives (k − a)(k − b) horizontal rectangles, each contributing
two minimal pairs. Similar argument for the vertical rectangle.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 445 — #19

25-446 Handbook of Bioinspired Algorithms and Applications

FIGURE 25.10 How many horizontal and vertical segments of length 1?

FIGURE 25.11 How many squares of size 1 × 1?

FIGURE 25.12 How many rectangles of size a × b?

Lemma 25.17
     
k2 − 1 k −1 2k − 2
L2,k 2 ,s = 4 −2 −2
s s s
    
k −1
2 ak + bk − ab − 1
+4 − .
s s
a⊥b,0<b<a<k

Proof. Let a slope b/a be such that a ⊥ b (we consider slopes for directions between 0◦ and 45◦ . For each
such slope b/a, Lemmas 25.14 and 25.15 give a simple algorithm to construct (i.e., generate) and count
the (2, k 2 , s)-partitions associated with b/a. See Figure 25.13 for an illustration of the proof.

• Rotate slope b/a to increase it a bit (this gives the direction of an associated (2, k 2 , s)-partition P).
• Sort the points along this direction.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 446 — #20

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-447

 2 
• Choose s points x 1 , . . . , x s out of k 2 − 1 points in k s−1 ways (selected points are beginning of
classes S1 , . . . , Ss and the point x 0 = (k − 1, 0) can always be selected for S0 ).
• Eliminate all selections of points where no minimal pair of slope b/a is intersected by a separating
line of direction P.
To intersect one of the s separating lines, the lower end of a minimal pair with slope b/a must be
selected. From Corollary 25.16 we have:
• If b = 0 then there are k(k − 1) lower ends (the number of minimal horizontal segments of the
 2   k−1 
(k, k)-grid). None of them can be selected in k −1−k(k−1)
s = s ways. So the number of ways
 2   
to select a minimal pair with slope 01 to intersect a separating line is 2 k s−1 − 2 k−1 s .
• If a = b = 1 then there are (k − 1)2 lower ends (the number of minimal segments with slope 45◦ ).
 2 2  
None of them can be selected in k −1−(k−1)s = 2k−2 s ways. So the number of ways to select a
 2   
minimal pair with slope 1 to intersect a separating line is 2 k s−1 − 2 2k−2
1
s .
• If a > b > 0 then there are (k − a)(k − b) lower ends (the number of horizontal rectangles of size
 2   ak+bk−ab−1 
a × b of the (k, k)-grid). None of them can be selected in k −1−(k−a)(k−b)
s = s ways.
So the number of ways to select a minimal pair with slope 0 < (b/a) < 1 to intersect a separating
line is
  2   
k −1 ak + bk − ab − 1
4 − .
s s
a⊥b,0<b<a<k

Taking the total sum of all three cases yields the formula. This completes the proof.
Figure 25.13 illustrates the above proof with a = b = 1, s = 2, k = 3. The vector w gives the direction
of separation, which is the dotted line P. The sorted list of points along that direction is 6, 7, 3, 8, 4, 0, 5, 1, 2.
To construct (generate) a (2, 9, 2)-partition, we must select three points out of 9 points in the list (these
will be beginning of classes S0 , S1 , and S2 , respectively). Point 6 can always be selected as x 0 for class
S0 and thus we need only to select two points (x 1 and x 2 ) for the remaining classes. A separating line
is then placed between x 0 and x 1 , and another between x 1 and x 2 (points x 0 , x 1 , x 2 are assumed sorted
along the direction of P). Moreover, if a selected point is a lower end of a minimal pair, we will then
place a separating line that intersects the minimal pair. For example, the (2, 9, 2)-partition in the figure is
generated by selection x0 = 6, x1 = 7, x2 = 0. Point 0 is the lower end of minimal pair (0, 4) so we place
a separating line between 0 and 4. Among the sorted list of points above, only four points are lower ends
of minimal pairs with slope 11 , that is, points 0, 1, 3, 4. To ensure that at least one minimal pair of slope 11
is intersected by a separating line then at least one lower end point must be selected in the construction
of a (2, 9, 2)-partition. To enumerate the number of (2, 9, 2)-partitions parallel to P, it suffices then to
count the number of selections of points (x 1 , x 2 ) that contain at least one lower end point. In Figure 25.13,
selections (6, 7, 8), (6, 8, 5), (6, 8, 2), (6, 7, 5), (6, 7, 2), and (6, 5, 2) for instance, will be eliminated since
none of them contain lower ends. Also, given selection (6, 7, 8) we can surely have a (2, 9, 2)-partition such
as in Figure 25.14, even though (6, 7, 8) is an invalid selection. However, such partition will be generated
by selection (6, 3, 0) and slope 12 , where 0 is the lower end of minimal pair (0, 7) (see dotted separating
lines in Figure 25.14).
Theorem 25.18
   2     
k2 − 1 k −1 k −1 2k − 2
L2,k 2 ,s =4 |Fk−1,k−1 | + 4 −2 −2
s s s s
  
ak + bk − ab − 1
−4 .
s
a⊥b,0<b<a<k

Proof. Follows from Lemma 25.17 and the fact that the last sum is over |Fk−1,k−1 |.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 447 — #21

25-448 Handbook of Bioinspired Algorithms and Applications

2 5 8

1 4 7
w

0 3 6

L2
L1

FIGURE 25.13 How many (2, 9, 2)-partitions for slope 11 ?

2 5 8

1 4 7

0 3 6

L2
L1

FIGURE 25.14 (2, 9, 2)-partition from slope 12 .

Lemma 25.19 ak + bk − ab − 1 < 2k 2 .

Proof. ak + bk − ab − 1 < ak + bk < 2k 2 , since b < a < k.

Corollaries 25.20 and 25.21 below give, respectively, a lower bound and an upper bound on the number
of (2, k 2 , s)-partitions of the (k, k)-grid.

Corollary 25.20

       2
k2 − 1 k −1 2k − 2 2k
L2,k 2 ,s > 4 (|Fk−1,k−1 | + 1) − 2 −2 −4 |Fk−1,k−1 |.
s s s s

Proof. Follows from Lemma 25.19 and Theorem 25.18 because the last sum becomes smaller when
substituting 2k 2 for ak + bk − ab − 1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 448 — #22

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-449

Corollary 25.21
 2     
k −1 k −1 2k − 2
L2,k 2 ,s < 4 (|Fk−1,k−1 | + 1) − 2 −2 .
s s s

Proof. Follows from Theorem 25.18 because the last sum is added.

The asymptotic formula for the length of the generalized Farey (i, j)-sequence (for i ≤ j) is given in
Reference 38 as |Fi,j | = (3i 2 j 2 /π 2 ) + O(i 2 j log j) + O(ij 2 log log j). Hence, substituting for the (k, k)-grid
we obtain |Fk−1,k−1 | = (3k 4 /π 2 ) + O(k 3 log k) + O(k 3 log log k).

Theorem 25.22
  2   2  4  2     
k −1 2k 3k k −1 k −1 2k − 2
L2,k 2 ,s ≈ 4 −2 + 4 − 2 − 2 .
s s π2 s s s

Proof. We take the mean of the lower and upper bounds and replace |Fk−1,k−1 | by its formula given
above.

Corollary 25.23 The number of 2-input k-valued s-separable logic functions is |Fk,s
2 | < k(k − 1)s L
2,k 2 ,s .

Proof. See Corollary 25.11.

The example given in Figure 25.7 explains the inequality in Corollary 25.23. For instance, the XOR
function has exactly two (2, 4, 2)-partitions. For s = 1 we obtain an asymptotic formula.

Corollary 25.24 The number of permutably homogeneous 2-input k-valued s-separable logic functions is
2 | ≈ k!/(k − s − 1)!L
|Gk,s 2,k 2 ,s .

Proof. See Corollary 25.13.

25.6.3 Complexity of Counting the Number of (2, k 2 , s)-partitions

Given a point set V in the plane, there are (v 2 −v)/2 pairs of points. Generally speaking, each pair requires
v − 2 tests to check whether it is minimal. Thus the complexity of the general case for the extraction of
all minimal pairs is O(v 3 ). However, if the considered v = ij points are all the points of the (i, j)-grid,
then this algorithm turns out to be linear with respect to v, namely O(v), on the basis of Corollary 25.16
and the fact that the successor of a member of Fi,j can be calculated in a constant time 34. The simple
linear time algorithm for generating Fi,j is described in Reference 34.
In Figure 25.15 we show the fast algorithm for counting (2, k 2 , s)-partitions of the (k, k)-grid. It is
a modified version of the 34 algorithm for enumerating the linear partitions of the (i, j)-grid, so, the
reader is referred to Reference 34 for a complete proof of correctness of the algorithm. The modifications
include replacing the (i, j)-grid by the (k, k)-grid and computing the binomial coefficients present in
Lemma 25.17.
Let j/i and d/c be two consecutive elements of Fk−1,k−1 . It has been shown in Reference 34 that
the immediate successor b/a of d/c is always determined by the relations b = rd − j and a = rc − i,
where r = [(k − 1 + i)/c]. This implies that the computation of a consecutive member of Fk−1,k−1 can be
completed in constant time. Another consequence of this fact is that initialization is easy: we can always
1
start with the first two Farey numbers Fk−1,k−1 = 1/k − 1 and Fk−1,k−1
2 = (1/k − 2) and then generate
the whole Fk−1,k−1 sequence in a loop using the above relations.
The last sum of the formula given in Lemma 25.17 necessarily reduces to the sum over the Farey
(k −1, k −1)-sequence. Suppose that k 2 and s require each O(log k 2 ) and O(log s) bits for memory storage,
then the complexity of computing the binomial coefficients is O(s). Therefore, for each generation of a
number (b/a) ∈ Fk−1,k−1 , it takes O(s) to compute the number of (2, k 2 , s)-partitions associated with

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 449 — #23

25-450 Handbook of Bioinspired Algorithms and Applications

FIGURE 25.15 Fast algorithm for counting (2, k 2 , s)-partitions.

b/a. Since there are k 2 elements in the (k, k)-grid, then it takes O(sk 2 ) ≤ O(k 4 ) time to calculate L2,k 2 ,s
(recall that s ≤ k 2 − 1). Clearly, the time complexity is polynomial on k.

25.7 Conclusion and Open Problems

The capacity of a neuron is defined as the total number of functions it can simulate. We introduced, in
this chapter, the concept of multilinear partition of a point set V ⊆ R n and the concept of multilinear
separability of a function f : V  → K = {0, . . . , k − 1}. We obtained a very general formula for the number
of linear partitions of a three-dimensional point-set. The points can be in any possible configuration and
the set can be a multi-set. Using a polynomial time algorithm, the formula is applied to compute the
capacity of discrete (n, k, 1)-perceptrons. Based on well-known relationships between linear partitions
and minimal pairs, we derived formulae for the number of multilinear partitions of the (k, k)-grids.
The counting techniques we used for the (k, k)-grid apply as well for the (i, j)-grid, where i ≤ j. From the
number of multilinear partitions of V , we obtained results on the capacity of a single (n, k, s)-perceptron,
respectively for V ⊂ R n in general position and for V = K 2 . Finally, we described a fast polynomial-time
algorithm for counting the multilinear partitions of K 2 .
The following are some interesting open problems related to research in this area.
Another complexity measure of neural networks is the Vapnik-Chervonenkis dimension. It is defined
as the maximum size of a training set T ⊆ V such that a given network realizes all functions defined
on T . If the (n, k, s)-perceptrons have a finite VC-dimension then they may be PAC-learnable. That is,
there may exist an efficient (n, k, s)-perceptron learning algorithm that uses only a polynomial number of
examples and that, with high probability (which can be made as high as desired), the algorithm outputs a
good approximation of a given function within a desired degree of accuracy (which can be made as high
as desired). The VC-dimension of (n, k, s)-perceptrons gives a valid lower bound on the VC-dimension
of linear decision lists and other related learning architectures. The computing capacity of linear decision
lists or any architecture obtained by a partitioning algorithm is not known and is still an open problem.
The technique we have applied for deriving capacity results on (n, k, s)-perceptrons may be extended
or improved to give results for partitioning architectures. The main question is: In how many ways
can we partition a finite set V ⊆ R n using s hyperplanes? The hyperplanes are not necessarily parallel.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 450 — #24

On the Computing Capacity of Multiple-Valued Multiple-Threshold Perceptrons 25-451

The answer to this question gives (bounds on) the capacity of neural networks constructed by partitioning
algorithms. Such neural network architectures include neural trees and neural decision lists. This question
also motivates for the investigation of the VC-dimension and PAC-learnability of such structures. The
combinatorial arguments used to derive results for the number of linear (or multilinear) partitions and
the capacity of (n, k, s)-perceptrons may possibly be extended for the general case of n-dimensional set.
For instance, enumerate partitions for higher dimension grids.

References
[1] S.C. Chan, L.S. Hsu, and H.H. Teh, On neural logic networks. Neural Networks, 1 (Suppl. I) 428,
1988.
[2] V. Milutinović, A. Ngom, and I. Stojmenović, Strip — a strip-based neural network growth
algorithm for learning multiple-valued functions. IEEE Transactions on Neural Networks, 12,
212–227, 2001.
[3] A. Ngom, Synthesis of Multiple-Valued Logic Functions by Neural Networks, Ph.D. thesis,
Computer Science Department, University of Ottawa, Ottawa, Ontario, Canada, 1998.
[4] Z. Obradović, Computing with nonmonotone multivalued neurons. Multiple-Valued Logic — An
International Journal, 1, 271–284, 1996.
[5] Z. Tang, Q. Cao, and O. Ishizuka, A learning multiple-valued logic networks: Algebra, algorithm,
and application. IEEE Transactions on Computers, 47, 247–251, 1998.
[6] G. Wang and H. Shi, Tmlnn: Triple-valued or multiple-valued logic neural network. IEEE
Transactions on Neural Networks, 9, 1099–1117, 1998.
[7] O. Ishizuka, Multivalued multithreshold networks. In Proceedings of the 6th IEEE International
Symposium on Multiple-Valued Logic, 1976, pp. 44–47.
[8] T. Sasao, On the optimal design of multiple-valued plas. IEEE Computer, 38, 582–592, 1989.
[9] D. Haring, Multi-threshold threshold elements. IEEE Transactions on Electronic and Computer, 15,
45–65, 1965.
[10] A. Ngom, C. Reischer, D.A. Simovici, and I. Stojmenović, Learning with permutably homogeneous
multiple-valued multiple-threshold perceptrons. Neural Processing Letters, 12, 2000, Proceedings of
the 28th IEEE International Symposium on Multiple-Valued Logic, May 1998, pp. 161–166.
[11] A. Ngom, I. Stojmenović, and R. Tošić, The computing capacity of three-input multiple-valued
one-threshold perceptrons. Neural Processing Letters, 14, 141–155, 2001.
[12] M.H. Abd-El-Barr, S.G. Zaky, and Z.G. Vranesić, Synthesis of multivalued multithreshold
functions for ccd implementation. IEEE Transactions on Computers, 35, 124–133, 1986.
[13] M. Minsky and S. Papert, Perceptrons: An Introduction to Computational Geometry, MIT Press,
Cambridge, MA, 1969, Expanded edition 1988.
[14] Z. Obradović and I. Parberry, Computing with discrete multivalued neurons. Journal of Computer
and System Sciences, 45, 471–492, 1992.
[15] Z. Obradović and I. Parberry, Learning with discrete multivalued neurons. Journal of Computer
and System Sciences, 49, 375–390, 1994.
[16] R. Takiyama, Multiple threshold perceptron. Pattern Recognition, 10, 27–30, 1978.
[17] S. Olafsson and Y.A. Abu-Mostafa, The capacity of multilevel threshold functions. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 10, 277–281, 1988.
[18] R. Takiyama, The separating capacity of a multithreshold threshold element. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 7, 112–116, 1985.
[19] A. Ngom, I. Stojmenović, and J. Žunić, On the number of multilinear partitions and the computing
capacity of multiple-valued multiple-threshold perceptrons. IEEE Transactions on Neural Networks,
14, 469–477, 2003, Proceedings of the 29th IEEE International Symposium on Multiple-Valued Logic,
IEEE Computer Society Technical Committee on Multiple-Valued Logic, May 1999, IEEE Computer
Society, pp. 208–213.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 451 — #25

25-452 Handbook of Bioinspired Algorithms and Applications

[20] M. Anthony and P.L. Bartlett, Neural Network Learning: Theoretical Foundations, Cambridge
University Press, Cambridge, 1999.
[21] N. Cristiani and J. Shawe-Taylor, An Introduction to Support Vector Machines, Cambridge University
Press, Cambridge, 2000.
[22] R.O. Dud and P.E. Hart, Pattern Classification and Scene Analysis, John Wiley Sons, New York,
1973.
[23] V.N. Vapnik, Statistical Learning Theory, John Wiley Sons, New York, 1998.
[24] M. Anthony, Classification by polynomial surfaces. Discrete Applied Mathematics, 61, 91–103,
1995.
[25] T.M. Cover, The number of linearly inducible orderings of points in d-space. SIAM Journal of
Applied Mathematics, 15, 434–439, 1967.
[26] V. Bohossian and J. Bruck, Multiple threshold neural logic. In Advances in Neural Information
Processing (M. Jordan, M. Kearns, and S. Colla, Eds.), Vol. 10, NIPS’1997, MIT Press, Cambridge,
MA, 1998.
[27] M. Anthony, Analysis of data with threshold decision lists. Center for Discrete and
Applied Mathematics Research Report, CDAM-LSE-2002-12, London School of Economic,
December 2002.
[28] H. Edelsbrunner, Pattern Classification and Scene Analysis, Springer-Verlag, Heidelberg, 1987.
[29] K.Y. Siu, V. Roychowdhury, and T. Kailath, Discrete Neural Computation: A Theoretical Foundation,
Information and System Sciences Series (Thomas Kailath, Series Ed.), Prentice Hall, Upper Saddle
River, New Jersey, 1995.
[30] J.T. Tou and R.C. Gonzalez, Pattern Recognition Principles, Addison-Wesley, Reading, MA, 1974.
[31] T.M. Cover, Geometrical and statistical properties of systems of linear inequalities with applic-
ations in pattern recognition. IEEE Transactions on Electronic and Computer, 14, 326–334,
1965.
[32] J. Nilsson, Learning Machines: Foundations of Trainable Pattern Classifying Systems, McGraw-Hill,
New York, 1968.
[33] J. Koplowitz, M. Lindenbaum, and A. Bruckstein, The number of digital straight lines on an n × n
grid. IEEE Transactions on Information Theory, 36, 192–197, 1990.
[34] D. Acketa and J. Žunić, On the number of linear partitions of the (m, n)-grid. Information
Processing Letters, 38, 163–168, 1991.
[35] R. Tošić, On the number of linear partitions. Review of Research, Mathematical Series, 22, 141–149,
1992.
[36] M. Anthony, Partitioning points by parallel hyperplanes. Discrete Mathematics, to appear.
[37] G.H. Hardy and E.M. Wright, An Introduction to the Theory of Numbers, 5th ed., Clarendon Press,
Oxford, England,1979.
[38] J. Žunić, On the asymptotic number of linear grid square partitions. Bild und Ton, 1991.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C025” — 2005/8/17 — 18:16 — page 452 — #26

 26
Advanced
Evolutionary
Algorithms for
Training Neural
Networks

26.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-453

26.2 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-454
26.3 The Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-456
Backpropagation Algorithm • Levenberg–Marquardt
Algorithm • Evolutionary Algorithms • Hybrid Algorithms
26.4 Representation and Fitness Function . . . . . . . . . . . . . . . . . . . 26-461
Enrique Alba 26.5 Computational Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-462
J. Francisco Chicano Benchmark and Parameters • Analysis of the Results
Francisco Luna 26.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-465
Gabriel Luque Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-466
Antonio J. Nebro References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-466

26.1 Introduction
The interest of making research in Artificial Neural Networks (ANNs) resides in the appealing properties
that ANNs exhibit: adaptability, learning capability, and ability to generalize. Nowadays, ANNs are receiv-
ing a great attention from the international research community: a large number of studies concerning
training, structure design, and real world applications, ranging from classification to robot control or
vision [1].
The neural network training task is a capital process in supervised learning, in which a pattern set
made up of pairs of inputs plus expected outputs is known beforehand. This set of patterns is used to
compute the set of weights that makes the ANN to learn it. To achieve this goal, the algorithm must
modify the weights of the neural network in order to get the desired output for a given input, usually in
an iterative manner, until a minimum error between the actual and the expected output is attained.
One of the most popular training algorithms in the domain of neural networks is the Backpropagation
(BP) technique (generalized delta rule) [2], a gradient-descendent method. Other techniques such as
Evolutionary Algorithms (EAs) have been also applied to the training problem in recent years [3,4],

26-453

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 453 — #1

26-454 Handbook of Bioinspired Algorithms and Applications

trying to avoid the local minima in such a complex problem. Although training is a main issue in ANN’s
design, many other works exist addressing the evolution of the layered structure of the ANN or even the
elementary behavior of the neurons composing the ANN. For example, in Reference 5 a definition of
neurons, layers, and the associated training problem is faced by using parallel genetic algorithms (GAs);
also, in Reference 6 the architecture of the networks and the weights are evolved by using the EPNet
evolutionary system. An exhaustive revision of this topic is really difficult to perform nowadays, however,
the work of Yao [7] represents an excellent starting point to get acquired of the research in training ANNs.
The motivation of the present chapter is manyfold. First, we want to perform a standard presentation
of results that promotes and facilitates future comparisons. This sounds like common sense, but it is
not frequent that authors follow standard rules for comparisons such as the structured Prechelt’s set
of recommendations [8], a “de facto” standard for many ANN researchers. A second contribution is to
include in our study, not only well-known EAs and the BP algorithm, but also the Levenberg–Marquardt
(LM) approach [9], and two additional hybrids. The potential advantages coming from an LM utilization
merit a detailed study. We have selected a benchmark from the field of Medicine, composed of three
classification problems: diagnosis of breast cancer, diagnosis of diabetes in Pima Indians, and diagnosis
of heart disease.
The remainder of the chapter is organized as follows. Section 26.2 introduces the ANN computation
model. In Section 26.3, we give a brief description of the algorithms under analysis. In Section 26.4, we
discuss the mechanisms used for representing solutions and evaluating their quality, a methodological
step needed in the application of EAs. The details of the experiments and the analysis of results are shown
in Section 26.5. Finally, we summarize our conclusions and future work in Section 26.6.

26.2 Artificial Neural Networks

Artificial Neural Networks are computational models naturally performing a parallel processing of infor-
mation [10]. Essentially, an ANN can be defined as a pool of simple processing units (neurons) which
communicate among themselves by means of sending analog signals. These signals travel through weighted
connections between neurons. Each of these neurons accumulates the inputs it receives, producing an
output according to an internal activation function. This output can serve as an input for other neurons,
or it can be a part of the network output. In Figure 26.1 we can see a neuron in detail.
There is a set of important issues involved in the ANN design process. As a first step, the architecture of
the network has to be decided. Initially, two major options are usually considered: feedforward networks
and recurrent networks (additional considerations regarding the order of the ANN exist, but they are out of
our scope). The feedforward model comprises networks in which the connections are strictly feedforward,
that is, no neuron receives input from a neuron to which the former sends its output, even indirectly.

Inputs Weights
W1
A1
W2
A2
W3
A3
Sum-of-product
Output
x
Σ f (x) y
WN
AN Summation Activation
function function
1
θ
Bias

FIGURE 26.1 An artificial neuron.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 454 — #2

Advanced Evolutionary Algorithms for Training Neural Networks 26-455

Output

Output layer

Neuron

Hidden layer
Connection
weights

Input layer

Input pattern

FIGURE 26.2 A MLP with three layers.

The recurrent model defines networks in which feedback connections are allowed, thus inducing complex
dynamical properties in the ANN. In this chapter we concentrate on the first and simpler model, the
feedforward networks. To be precise, we consider the so-called multilayer perceptron (MLP) [11], in which
units are structured into ordered layers, and connections are allowed only between adjacent layers in an
input-to-output sense (see Figure 26.2).
For any MLP, several parameters such as the number of layers and the number of units per layer must
be defined. Then, the last step in the design is to adjust the weights of the network, so that it produces
the desired output when its corresponding input is presented. This process is known as training the ANN
or learning the network weights. Network weights comprise both the previously mentioned connection
weights as well as bias terms for each unit. The latter can be viewed as the weight of a constant saturated
input that the corresponding unit always receives. As initially stated, we will focus on the learning situation
known as supervised training, in which a set of current-input/desired-output patterns is available. Thus,
the ANN has to be trained to produce the desired output according to these examples. The input and
output of the network are both real vectors in our case.
In order to perform a supervised training we need a way of evaluating the ANN output error with
respect to the expected output. A popular measure is the Squared Error Percentage (SEP). This term
measures the proximity of the actual output to the desired output. We can compute this error term just
for one single pattern or for a set of patterns. In this last case, the SEP is the average value of the separate
SEP values. The expression for this global SEP is:

omax − omin

p
P S
p
SEP = 100 · (ti − oi )2 , (26.1)
P ·S
p=1 i=1

p p
where ti and oi are, respectively, the ith components of the expected vector and the current output vector
for the pattern p; omin and omax are the minimum and maximum values of the output neurons, S is the
number of output neurons, and P is the number of patterns. This SEP value is closely related to the Mean
Squared Error (MSE), whose expression is:
P S p p
p=1 i=1 (ti − o i )2
MSE = . (26.2)
P ·S

In classification problems, we could use still an additional measure: the Classification Error Percentage
(CEP). CEP is the percentage of incorrectly classified patterns, and it is a usual complement to any of the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 455 — #3

26-456 Handbook of Bioinspired Algorithms and Applications

other two (SEP or MSE) crude error values, since CEP reports in a high-level manner the quality of the
trained ANN.

26.3 The Algorithms

We use several algorithms to train the networks: the Backpropagation algorithm, the Levenberg–
Marquardt algorithm, a canonical Genetic Algorithm, the CHC method, the Hy3 algorithm, an Evolution
Strategy (ES), a hybrid between Genetic Algorithm and Backpropagation, and a hybrid between Genetic
Algorithm and Levenberg–Marquardt. We briefly describe them in the following paragraphs.

26.3.1 Backpropagation Algorithm

The BP algorithm [2] is a classical domain-dependent technique of supervised training. It works by
measuring the output error, calculating the gradient of this error, and adjusting the ANN weights (and
biases) in the descending gradient direction. Hence, BP is a gradient-descent local search procedure
(foreseeable stagnation in local optima in complex landscapes).
First, we define the squared error of the ANN for a set of patterns:

P

S
p p
E= (ti − oi )2 . (26.3)
p=1 i=1

The actual value of the previous expression depends on the weights of the network. The basic BP
algorithm calculates the gradient of E and updates the weights by moving them in the gradient-descendent
direction. This can be summarized with the expression:

∂E
wij (t + 1) = wij (t ) − η , (26.4)
∂wij

where the parameter η > 0 is the learning rate that controls the learning speed. A more general BP
algorithm adds to the previous expression a momentum term in order to increase the stability of the
search process. Then, the final expression for the BP algorithm is:

∂E
wij (t + 1) = wij (t ) + αwij (t ) − η , (26.5)
∂wij

where wij (t ) is the change in the weight wij at step t , and α is the momentum constant, whose value
must match 0 ≤ α < 1. With this term, the algorithm accelerates the minimization of the error when the
error function does not change (in smooth zones of the function). The pseudo-code of the BP algorithm
is shown in Figure 26.3.

26.3.2 Levenberg–Marquardt Algorithm

The LM algorithm [9] is an approximation to the Newton method used also for training ANNs. The
Newton method approximates the error of the network with a second order expression, which contrasts
to the first order information used in the BP algorithm. LM is very popular in the ANN domain (even
being adopted as the initially more promising approach for an unseen MLP training task). Curiously, it is
not that popular in the metaheuristic field. LM updates the ANN weights as follows:
 −1


P
w = − µI + J p (w)T J p (w) ∇E(w), (26.6)
p=1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 456 — #4

Advanced Evolutionary Algorithms for Training Neural Networks 26-457

FIGURE 26.3 Pseudo-code of the BP algorithm.

FIGURE 26.4 Pseudo-code of the LM algorithm.

where J p (w) is the Jacobian matrix of the vector ep (w) evaluated in w, and I is the identity matrix. The
vector ep (w) is the error of the network for pattern p, that is, ep (w) = tp − op (w). The parameter µ is
increased or decreased at each step. If the error is reduced then µ is divided by a factor β and it is multiplied
by β in other case. LM performs the steps included in Figure 26.4. It calculates the network output, the
error vectors, and the Jacobian matrix for each pattern. Then, it computes w using Equation (26.6) and
recalculates the error with w + w as network weights. If the error has decreased, µ is divided by β, the
new weights are maintained, and the process starts again; otherwise, µ is multiplied by β, w is calculated
with a new value, and it probes again.

26.3.3 Evolutionary Algorithms

Evolutionary Algorithms [12] are heuristic search techniques loosely based on the principles of natural
evolution, namely adaptation and survival of the fittest. These techniques have been shown to be very
effective in solving hard optimization tasks with similar properties to ANN training, that is, problems
in which gradient-descent techniques get trapped into local minima, or are fooled by the complexity
and non-differentiability of the search space. The underlying idea in EAs is making individuals represent

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 457 — #5

26-458 Handbook of Bioinspired Algorithms and Applications

FIGURE 26.5 Pseudo-code of a GA.

the weights and biases of the ANN, using the network error function to be minimized. Some general
considerations must be taken into account when using an evolutionary approach for ANN training. The
search features of EAs contrast with those of the BP and LM in that they are not trajectory driven, but
population driven. An EA is expected to avoid local optima frequently by promoting exploration of the
search space, in opposition to the exploitative trend usually allocated to local search algorithms such
as BP or LM. In this chapter, we use four EAs: a canonical GA, the CHC method, the Hy3 algorithm,
and an ES.

26.3.3.1 Genetic Algorithms

A GA proceeds in an iterative manner by generating new populations of strings from the old ones
(Figure 26.5). Every string is the encoded (binary, real, etc.) version of a tentative solution. The canonical
algorithm applies stochastic operators such as selection, crossover, and mutation on an initially random
population in order to compute a new population. In generational GAs all the populaton is replaced with
new individuals. In steady-state GAs only one new individual is created and it replaces the worst one in the
population if it is better. In our study, the experimental section explains the used parameters and details
deeper.

26.3.3.2 CHC
The CHC acronym stands for “Cross generational elitist selection, Heterogeneous recombination, and
Cataclysmic mutation” [13]. A CHC is a non-canonical GA that combines a conservative replace strategy
(elitist replace) with a highly disruptive recombination (HUX). In Figure 26.6 we present the pseudo-code
for the CHC method. The main differences between a canonical GA and a CHC are described in the
following paragraphs.
First, in CHC, the bias in favor of the best structures occurs in the replacement stage (or survival-
selection) rather than in reproduction-selection. More precisely, during the selection for reproduction,
each member of the current population is copied to the parent set, that is, the candidates for reproduction
are identical to the current population except that the order of the structures has been shuffled. During the
replacement, the offspring and the old population are merged and ranked according to the fitness value,
the new population is created by selecting the best µ members out of the merged population (where µ is
the population size). This mechanism always preserves the best individuals found so far.
Second, a new bias is introduced against mating individuals who are similar (incest prevention). The
initial threshold (minimum Hamming distance between the parents) for allowing mating is often set to 1/4
of the chromosome length. If no offspring is inserted into the new population during the reproduction,
this threshold is reduced by 1.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 458 — #6

Advanced Evolutionary Algorithms for Training Neural Networks 26-459

FIGURE 26.6 Pseudo-code of the CHC algorithm.

Third, the recombination operator in CHC is a variant of uniform crossover (HUX), a highly disruptive
form of crossover. This operator crosses over exactly half of the bits that differ between the two parent
strings. HUX guarantees that the children are always the maximum Hamming distance from their parents.
Finally, a restart process reintroduces diversity whenever convergence is detected (i.e., the difference
threshold of the incest prevention bias has dropped to zero). The population is restarted by using the best
individual found so far as a template for creating the new population.

26.3.3.3 Hy3 Algorithm

In this section we discuss a recent algorithm proposal called Hy3 [14], a kind of distributed technique with
heterogeneous components to enhance the search. Distributed GAs (dGAs) are a kind of structured GAs
where the population is partitioned into several subpopulations, each one processed by a GA, indepen-
dently of the others. In addition, sparse migrations of individuals produce an exchange of genetic material
between the subpopulations that enhances diversity and usually improves the accuracy and efficiency
of the algorithm. Making different decisions on the subalgorithms of a dGA through the application of
different search strategies, we obtain the so-called heterogeneous dGAs.
The Hy3 algorithm is a heterogeneous dGA that runs eight populations in a cubic topology. This model
is suitable for the optimization of continuous functions, because it includes in the basic improvement loop
of the algorithm the utilization of crossover operators for real genes (variables), engineered with fuzzy
logic technology to deal with the traditional “fuzzy” GA concepts of exploration and exploitation. Hy3 is
included into the class of heterogeneous algorithms since it applies a different crossover operator in each
of its component subpopulations. There are two important faces in the topology (Figure 26.7) that have
to be considered:

• The front side is devoted to exploration. It is made up of four subpopulations E1 , . . . , E4 , in which

several exploratory crossovers are applied.
• The rear side promotes exploitation. It is composed of subpopulations e1 , . . . , e4 , that apply
exploitative crossover operators.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 459 — #7

26-460 Handbook of Bioinspired Algorithms and Applications

Exploitation plane
e4 e1

E4 E1

e3 e2

E3 E2
Exploration Plane

FIGURE 26.7 Structure of Hy3.

FIGURE 26.8 Pseudo-code of an ES.

The resulting structure is a parallel-suited multi-resolution method using several crossover opera-
tors that allow to achieve simultaneously a diversified search (reliability) and an effective local tuning
(accuracy).

26.3.3.4 Evolution Strategy

In ES [15] each individual is composed of a vector of real numbers representing the problem variables (x),
a vector of standard deviations (σ ), and, optionally, a vector of angles (ω). These two last vectors are
used as parameters for the Gaussian mutation and undergo evolution together with the problem variables,
thus allowing the algorithm to self-adapt the way the search is performed.
Figure 26.8 outlines the evolution strategy. First, a random initial population of µ individuals is
generated. Then, several individuals are selected from the population for recombination and mutation
to obtain λ offsprings. Finally, the new population is formed selecting µ individuals from the offsprings
and the current population. If the new population is only formed with the offsprings we have a (µ, λ)-ES,
where it must hold µ ≤ λ. If the current population is taken into account in addition to the offsprings,
we have a (µ + λ)-ES.
The recombination and mutation affect the three potential vectors contained in the individual, that is,
the problem variables and the self-adapted parameters. The mutation of the individual follows the next

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 460 — #8

Advanced Evolutionary Algorithms for Training Neural Networks 26-461

equations:

σi = σi exp(τ N (0, 1) + ηNi (0, 1)), (26.7)

ωi = ωi + ϕNi (0, 1), (26.8)

x  = x + N(0, C(σ  , ω )), (26.9)

where C(σ  , ω ) is the covariance matrix associated with σ  and ω , N (0, 1) is the standard univariate
normal distribution, and N(0, C) is the multivariate normal distribution with mean 0 and covariance
matrix C. The subindex i in the standard normal distribution indicates that a new random number is
generated for each component of the vector. The notation N (0, 1) is employed to indicate that the same
random number is used for all the components. The parameters τ , η, and ϕ are set to (2n)−1/2 , (4n)−1/4 ,
and 5π/180, respectively as mentioned in Reference 16.
For the recombination many alternatives may be used. In Reference 17 Fogel summarizes some of
them. The three vectors of the individual are recombined in an independent way, that is, for each vector,
a different recombination scheme can be chosen.

26.3.4 Hybrid Algorithms

Hybridization is almost “a must” in complex applications if one expects to get efficient and accurate
algorithms. In our study, “hybridization” refers to the inclusion of problem-dependent knowledge in a
general search template [18,19]. We can distinguish two kinds of hybridization: strong and weak hybrid-
ization. In the first one, the knowledge is included using specific operators or representations. In the latter
several algorithms are combined somehow. In this last case, an algorithm can be used to improve the
results of another one separately or it can be used as an operator of the other.
The hybrid algorithms that we use in this chapter are combinations of two algorithms (weak
hybridization), where one of them acts as an operator in the other. We combine a GA with the BP algorithm
(GABP), and a GA with LM (GALM). In both cases the problem-specific algorithm (BP and LM) is used
as a mutation operation of the general search template (GA). Therefore, we can say that GAxx is a GA
(see Figure 26.5) in which the mutation has been replaced by the “xx” algorithm that is applied with
probability pt .

26.4 Representation and Fitness Function

One of the first topics that has to be addressed to solve a problem with an algorithm is the representation
of the solutions. This is a very important step in tailoring a general search sheet as an EA for the problem
at hands. In our case, a solution is composed of the values of all the weights and biases in the ANN.
These values can be included in a real vector where each weight (and bias) is allocated always the same
position in it. This k-dimensional real vector could be codified in different ways. In the present analysis we
consider a linear encoding of the vector, that is, first, each variable is conveniently encoded in an algorithm-
dependent way; subsequently, the codified vector is constructed by concatenating the encoding of each
variable into a linear string. This linear encoding of variables raises a question: the best distribution of the
variables within the string. This distribution is important in connection with the particular recombination
operator used in the EAs. In effect, if this operator breaks the strings into large blocks using them as units
for exchange, this distribution might be relevant. On the contrary, using a recombination operator that
breaks the strings into very small chunks makes the distribution irrelevant. A good piece of advice is

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 461 — #9

26-462 Handbook of Bioinspired Algorithms and Applications

grouping together the input weights and bias for each unit. This way, the probability of transmitting them
as a block is increased. Obviously, recombination is not used in many EAs, so this consideration does not
apply to all situations.
In BP, LM, ES, and Hy3 each variable is encoded using a machine-dependent codification for real
numbers. In the rest, the encoding of solutions is approached via binary strings. More precisely, m bits
are used to represent each single variable; subsequently, the k m-bit segments are concatenated into an
l-bit binary string, where l = k · m. This encoding of the network variables raises a number of issues. Two
of them are the choice of m and the encoding mechanism for individual variables, that is, pure binary,
Gray-coded numbers, magnitude-sign, etc. In this work we use 16-bit pure binary encoding. The integer
value of each variable is mapped linearly into an interval and the result is the weight value.
Now, we discuss the alternatives for the fitness function in order to evaluate the quality of the solutions
in EAs. The objective of the network is to classify all the patterns correctly, that is, to obtain 0% of CEP.
We can use the CEP as a function to minimize. However, among two networks with the same CEP for the
pattern set, the output of one of them can be nearer to the desired output than the other. For this reason,
the SEP can better guide the search. In our presentation, the fitness function to be minimized is the SEP
for the training pattern set. Conversely, a maximization approach should consider the inverse of the SEP
as fitness.

26.5 Computational Experiments

After discussing the algorithms, the representation, and the fitness function, we present in this section
the experiments performed and their results. The benchmark for training and the parameters of the
algorithms are presented in the next subsection. The analysis of the results is shown in Section 26.5.2.

26.5.1 Benchmark and Parameters

We tackle three classification problems, which consist in determining the class that a certain input vector
belongs to. Each pattern from the training set contains an input vector and its desired output vector.
These vectors are composed of real numbers. For the classification task we must interpret the network
output as a class. The interpretation can be performed in different ways [8]. We adopt here the so called
winner-takes-all in which an output neuron is associated to each class. When an input vector is presented
to the network, the network response is the class associated with the output neuron with the larger value.
The instances solved here belong to the Proben11 benchmark [8]: Cancer, Diabetes, and Heart. We now
briefly detail them:

• Cancer: Diagnosis of breast cancer. Classify a tumor as either benign or malignant based on cell
descriptions gathered by microscopic examination. There are 699 examples that were obtained by
Dr. William H. Wolberg at the University of Wisconsin Hospitals, Madison [20–23].
• Diabetes: Diagnose diabetes of Pima Indians. Based on personal data and the results of medical
examinations, decide whether a Pima Indian individual is diabetes positive or not. There are 768
examples from the National Institute of Diabetes and Digestive and Kidney Diseases by Vincent
Sigillito [24].
• Heart : Predict heart disease. Decide whether at least one of four major vessels is reduced in diameter
by >50%. This decision is made based on personal data and results of medical examinations.
There are 920 examples from four different sources: Hungarian Institute of Cardiology in Budapest
(Andras Janosi, M.D.), University Hospital of Zurich in Switzerland (William Steinbrunn, M.D.),
University Hospital of Basel in Switzerland (Mathhias Pfisterer, M.D.), V.A. Medical Center of
Long Beach and Cleveland Clinic Foundation (Robert Detrano, M.D., Ph.D.) [25,26].

1 Available from ftp://ftp.ira.uka.de/pub/neuron/proben1.tar.gz.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 462 — #10

Advanced Evolutionary Algorithms for Training Neural Networks 26-463

The structure of the MLP used for all the instances accounts for three layers (input-hidden-output)
having six neurons in the hidden layer. The number of neurons in the input and output layers depends
on the concrete instance. The activating function of the neurons is the sigmoid function. Table 26.1
summarizes the network architecture for each instance.
To evaluate an ANN, we split the pattern set into two subsets: the training one and the test one. The
ANN is trained with the different algorithms by using the training pattern set, and then it is evaluated on
the unseen test pattern set. The training set for each instance is approximately made of the first 75% of
the examples, while the last 25% constitutes the test set. The exact number of patterns for each instance is
presented in Table 26.1 to ease future comparisons.
After presenting the problems, we now turn to describe the parameters for the algorithms. The para-
meters for BP, LM, and the hybrid algorithms are shown in Table 26.2. The hybrid algorithms use the same
parameters as their elementary components. However, the mutation operator of the GA is not applied but
it is replaced by BP or LM, respectively. The BP and LM are applied with an associated probability pt , just
like any other classical genetic operator. When applied, BP/LM only performs one single epoch.
The parameters of the rest of the algorithms are shown in Tables 26.3 and 26.4. The ES uses neither
vector of angles nor recombination, thus making irrelevant the distribution of variables inside the vector.
The ES and Hy3 algorithms employ real representation for the variables of the network, but the GA, the
CHC, and the hybrid algorithms use binary vectors. These vectors are 16-bit length and represent a real

TABLE 26.1 MLP Architecture and Patterns

Distribution for All Instances

Patterns
Instance Architecture Training Test

Cancer 9 — 6 — 2 525 174

Diabetes 8 — 6 — 2 576 192
Heart 35 — 6 — 2 690 230

TABLE 26.2 Parameters for BP, LM, and the Hybrid

Algorithms

BC DI HE

Epochs 1000 1000 500

BP η 0.01 0.01 0.001
α 0.0 0.0 0.0
Epochs 1000 1000 500
LM µ 0.001 0.001 0.001
β 10 10 10
pt 1.0 1.0 0.5
GAxx
Epochs of BP/LM 1 1 1

TABLE 26.3 Parameters for GA, CHC, Hy3, and ES

GA CHC Hy3 ES

Pop. size 64 64 160 (8 × 20) 1

Selection Roulette (2 inds.) CHC selection LR + SUS 10 inds.
Recombination SPX (pc = 1.0) HUX (pc = 1.0) FCBX (pc = 0.6) None
Mutation Bit-Flip (pm = 1/length) NU (pm = 0.125) GM (pm = 1.0)
Replacement Elitist
Stop criterion 1,064 evals. 10,048 evals. 800,160 evals. 10,001 evals.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 463 — #11

26-464 Handbook of Bioinspired Algorithms and Applications

TABLE 26.4 Abbreviations Used in Table 26.3

Abbreviation Meaning

LR Linear Ranking
SUS Stochastic Universal Sampling
FCBX Fuzzy Connective-Based Crossover
SPX Single Point Crossover
HUX Highly Disruptive Uniform Crossover
GM Gaussian Mutation
NU Non-Uniform Mutation

TABLE 26.5 Results of ANN Training

CEP(%) BP LM GA CHC Hy3 ES GABP GALM

x̄ 0.98 3.03 16.59 8.56 49.00 0.90 0.75 0.02

Cancer
σn 0.26 1.34 6.47 0.02 35.40 0.38 3.92 0.10
x̄ 21.81 25.94 36.46 37.44 53.00 21.30 36.46 28.32
Diabetes
σn 0.37 3.50 0.00 0.01 13.50 1.35 0.00 1.21
x̄ 27.13 34.81 39.87 38.06 52.00 31.45 53.83 22.55
Heart
σn 1.54 3.77 11.51 0.03 25.30 4.70 21.01 0.73

60
BP LM GA CHC Hy3 ES GABP GALM

50

40
CEP(%)

30

20

10

0
Cancer Diabetes Heart
Instances

FIGURE 26.9 Comparison among the algorithms.

value in the interval [−1, +1]. The weights (variables) associated to input links for a neuron are placed
contiguously in the chromosome. The GA uses a steady-state strategy.

26.5.2 Analysis of the Results

In this section we present the results obtained after the application on the three instances of the eight
algorithms. We report the mean and the standard deviation of the CEP for the test pattern set after
performing 30 independent runs. Table 26.5 and Figure 26.9 show the results.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 464 — #12

Advanced Evolutionary Algorithms for Training Neural Networks 26-465

There are many interesting works related to neural network training that also solve the instances tackled
here. But unfortunately, some of the results are not comparable with ours, because they use a different
definition of the training and test sets; this is why we consider a capital issue to adhere to a standard
way of evaluations like the one proposed by Prechelt [8]. In fact, we did find some works for meaningful
comparisons.
For the Cancer instance the best results are obtained by BP, ES, GABP, and GALM. They are followed in
a less accurate ranking by LM and, finally, by the GAs. This is not a surprising fact, since the GAs perform
a rather explorative search in this kind of problems. For this instance, we found that the best mean CEP
in Reference 27 is 1.1%, which represents a lower accuracy compared to our 0.02% obtained with the
GALM hybrid, so far the best solution to our knowledge. In Reference 28, a CEP close to 2% for this
instance is achieved, and the reader should notice that our GALM is one hundred times more accurate.
The mentioned work uses 524 patterns for the training set and the rest for the test set, that is, almost
exactly our configuration with only one pattern changed (a minor detail), and therefore the results can be
compared. The same happens for the work of Yao and Liu [6], where their EPNet algorithm gets neural
networks reaching 1.4% of CEP.
In Diabetes, BP and ES show again the highest accuracy, followed by LM and GALM. This time, GA,
CHC, and GABP obtain similar CEP, always below that of Hy3. For this instance, a CEP of 30.11% is
reached in Reference 29 (outperformed by many of our algorithms like BP, LM, ES, and GALM) with the
same network architecture as here. In Reference 6 the authors report a 22.4% of CEP for this problem,
clearly outperformed by our BP and ES techniques.
As to the Heart problem, we found approximately the same behavior as for the two previous instances.
However, Hy3 reaches lower CEP than GABP, and GALM gets the lowest CEP. In Reference 29 the
authors reported a 45.71% of CEP for the Heart instance using the same architecture. All our algorithms
outperform this CEP measure (except Hy3 and GABP).
We observed that BP and ES performed slightly more accurate than LM for all the instances. This is also
an unexpected result, since LM is quite accurate in many applications. We conclude that the three selected
instances represent problems that do not really need such a complex in-depth local search. In the near
future we plan to consider larger and more complex instances to further test LM. With respect to the
hybrid algorithms, the results do confirm this hypothesis: GALM is more accurate than GABP.
In summary, we have found some of the more accurate results for the three instances, but there is still
need to get ahead in other instances and improve the accuracy on the addressed ones. A learned lesson
is to always keep in mind the importance of reporting results in a standard way for meaningful future
comparisons.

26.6 Conclusions
In this chapter we have tackled the neural network training problem with eight algorithms: two well-
known problem-specific algorithms such as BP and LM, four general metaheuristics such as GA, CHC,
Hy3 Algorithm, ES, and two hybrid algorithms combining the GA with the problem-specific techniques.
To compare the algorithms we solved three classification problems from the domain of Medicine: the
diagnosis of breast cancer, the diagnosis of diabetes in the Pima Indians, and the diagnosis of heart
disease.
Our results show that the problem-specific algorithms (BP and LM) and the ES get lower classification
error than the other more general search procedures. It is surprising that the ES, also a general search
procedure, gets a lower CEP than the LM, a sophisticated problem-specific algorithm.
With respect to hybrids, the algorithm GALM outperforms in two of the three instances the classification
error of the problem-specific algorithms. This makes GALM a promising algorithm for neural network
training. On the other hand, many of the classification errors obtained in this work are below those found
in the literature, what represents a contribution as a reference work for these medical problems. As a future
work we plan to add new algorithms to the analysis, and to apply them to more instances, especially in the
broader domain of Bioinformatics.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 465 — #13

26-466 Handbook of Bioinspired Algorithms and Applications

Acknowledgments
This work has been partially funded by the Ministry of Science and Technology (MCYT) and Regional
Development European Fund (FEDER) under contract TIC2002-04498-C05-02 (the TRACER project,
http://tracer.lcc.uma.es).

References
[1] J.T. Alander. Indexed Bibliography of Genetic Algorithms and Neural Networks. Technical
report 94-1-1NN, University of Vaasa, Department of Information Technology and Production
Economics, 1994.
[2] D. Rumelhart, G. Hinton, and R. Williams. Learning representations by backpropagation errors.
Nature, 323: 533–536, 1986.
[3] E. Cantú-Paz. Pruning neural networks with distributions estimation algorithms.
In Erick Cantú-Paz et al., Eds., Proceedings of GECCO 2003, Vol. 2733 of Lecture Notes in Computer
Science. Springer-Verlag, 2003, pp. 790–800.
[4] C. Cotta, E. Alba, R. Sagarna, and P. Larrañaga. Adjusting weights in articficial neural networks
using evolutionary algorithms. In P. Larrañaga and J.A. Lozano, Eds., Estimation of Distribu-
tion Algorithms. A New Tool for Evolutionary Computation. Kluwer Academic Publishers, 2001,
pp. 357–373.
[5] E. Alba, J.F. Aldana, and J.M. Troya. Full automatic ANN design: A genetic approach. In J. Mira,
J. Cabestany, and A. Prieto, Eds., New Trends in Neural Computation. Springer-Verlag, 1993,
pp. 399–404.
[6] X. Yao and Y. Liu. A new evolutionary system evolving artificial neural networks. IEEE Transactions
on Neural Networks, 8: 694–713, 1997.
[7] X. Yao. Evolving artificial neural networks. Proceedings of the IEEE, 87: 1423–1447, 1999.
[8] L. Prechelt. Proben1 — A Set of Neural Network Benchmark Problems and Benchmarking Rules.
Technical report 21, Fakultät für Informatik Universität Karlsruhe, 76128 Karlsruhe, Germany,
September, 1994.
[9] M.T. Hagan and M.B. Menhaj. Training feedforward networks with the Marquardt algorithm.
IEEE Transactions on Neural Networks, 5: 989–993, 1994.
[10] J.L. McClelland and D.E. Rumelhart. Parallel Distributed Processing: Explorations in the Microstruc-
ture of Cognition. MIT Press, Cambridge, MA, 1986.
[11] F. Rosenblatt. Principles of Neurodynamics. Spartan Books, New York, 1962.
[12] T. Bäck. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary
Programming, Genetic Algorithms. Oxford University Press, New York, 1996.
[13] L.J. Eshelman. The CHC adaptive search algorithm: How to have safe search when engaging in
nontraditional genetic recombination. In Foundations of Genetic Algorithms. Morgan Kaufmann,
San Mateo, CA 1991, pp. 265–283.
[14] E. Alba, F. Luna, A.J. Nebro, and J.M. Troya. Parallel heterogeneous genetic algorithms for
continous optimization. Parallel Computing, 2004.
[15] I. Rechenberg. Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen
Evolution. Fromman-Holzboog Verlag, Stuttgart, 1973.
[16] G. Rudolph. Evolutionary Computation 1. Basic Algorithms and Operators, Vol. 1. IOP Publishing
Ltd, 2000, chap. 9, pp. 81–88.
[17] D.B. Fogel. Evolutionary Computation 1. Basic Algorithms and Operators, Vol. 1. IOP Publishing
Ltd, 2000, chap. 33.2, pp. 270–274.
[18] C. Cotta and J.M. Troya. On decision-making in strong hybrid evolutionary algorithms. In Tasks
and Methods in Applied Artificial Intelligence, Vol. 1415 of Lecture Notes in Artificial Intelligence,
1998, pp. 418–427.
[19] L. Davis, Ed. Handbook of Genetic Algorithms. Van Nostrand Reinhold, New York, 1991.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 466 — #14

Advanced Evolutionary Algorithms for Training Neural Networks 26-467

[20] K.P. Bennett and O.L. Mangasarian. Robust linear programming discrimination of two linearly
inseparable sets. Optimization Methods and Software, 1: 23–34, 1992.
[21] O.L. Mangasarian, R. Setiono, and W.H. Wolberg. Pattern recognition via linear programming:
Theory and application to medical diagnosis. In Thomas F. Coleman and Yuying Li, Eds., Large-
Scale Numerical Optimization. SIAM Publications, Philadelphia, PA, 1990, pp. 22–31.
[22] W.H. Wolberg. Cancer diagnosis via linear programming. SIAM News, 23: 1–18, 1990.
[23] W.H. Wolberg and O.L. Mangasarian. Multisurface method of pattern separation for medical
diagnosis applied to breast cytology. In Proceedings of the National Academy of Sciences, Vol. 87,
U.S.A., December 1990, pp. 9193–9196.
[24] J.W. Smith, J.E. Everhart, W.C. Dickson, W.C. Knowler, and R.S. Johannes. Using the ADAP learn-
ing algorithm to forecast the onset of diabetes mellitus. In Proceedings of the Twelfth Symposium
on Computer Application in Medical Care. IEEE Computer Society Press, 1988, pp. 261–265.
[25] R. Detrano, A. Janosi, W. Steinbrunn, M. Pfisterer, J. Schmid, S. Sandhu, K. Guppy, S. Lee, and
V. Froelicher. International application of a new probability algorithm for the diagnosis of coronary
artery disease. American Journal of Cardiology, 64: 304–310, 1989.
[26] J.H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial
Intelligence, 40: 11–61, 1989.
[27] T. Ragg, S. Gutjahr, and H. Sa. Automatic determination of optimal network topologies based on
information theory and evolution. In Proceedings of the 23rd EUROMICRO Conference, Budapest,
Hungary, September 1997.
[28] W.H. Land and L.E. Albertelli. Breat cancer screening using evolved neural networks. In IEEE
International Conference on Systems, Man, and Cybernetics, 1998, Vol. 2, IEEE Computer Society
Press, October 1988, pp. 1619–1624.
[29] W. Erhard, T. Fink, M.M. Gutzmann, C. Rahn, A. Doering, and M. Galicki. The improvement
and comparison of different algorithms for optimizing neural networks on the MasPar MP-2.
In M. Heiss, Ed., Neural Computation — NC’98. ICSC Academic Press, 1998, pp. 617–623.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C026” — 2005/8/8 — 17:12 — page 467 — #15

 27
Bio-Inspired
Data Mining

27.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-469

27.2 A Brief Overview of Bio-Inspired Data Mining . . . . . . . 27-471
Evolutionary Approaches to Data Mining • Ant Colonies for
Data Mining • Artificial Immune Systems for Data Mining •
Particle Swarm Algorithms for Data Mining
27.3 Covering Algorithm and Validation Procedure . . . . . . . . 27-474
Validation Procedure • Covering Algorithm • Postprocessing
Routines — Rule Pruning and Rule Set Cleaning
27.4 Basic Particle Swarm Data Miners . . . . . . . . . . . . . . . . . . . . . . 27-478
The PSO Algorithm • Rule Representation •
Rule Evaluation • Data Mining with Particle Swarms
27.5 An Adaptive PSDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-481
The Predator–Prey Interaction • The Adaptation Scheme •
Using SAPPO for Data Mining
Tiago Sousa 27.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-484
Arlindo Silva Results
Ana Neves 27.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-487
Ernesto Costa References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-487

27.1 Introduction
The dissemination of information systems in organizations and technological progress in terms of
computational power, communications technology, and storage capability has the side effect of producing
repositories with huge amounts of data. Databases — as these repositories are called — have grown in
number and size and it quickly became apparent that all the stored information constituted a valuable
resource, especially if reliable methods could be found to recover useful information — knowledge —
from the raw data in storage. A completely new interdisciplinary field has grown around this general goal
and is now known as Knowledge Discovery in Databases (KDD). As a broadly accepted definition, we
can say that KDD is a complex process that aims to extract implicit, previously unknown, and potentially
useful information from data, in a nontrivial way.
The central step in the KDD process is usually called data mining and consists of the actual search for
interesting regularities or patterns in the data. This step is preceded by a processing stage, where data is

27-469

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 469 — #1

27-470 Handbook of Bioinspired Algorithms and Applications

prepared for the application of a data mining algorithm. After the data mining algorithm is executed,
a postprocessing stage occurs, where the algorithm’s results can be refined and simplified.
From the data mining view point, the essential task is to build computer programs capable of search-
ing through the data for nuggets of knowledge. These nuggets are usually represented as data patterns
that are expected to have some beneficial characteristics, namely: validity on new data with a high degree
of certainty, novelty, potential usefulness measured by some utility function, and “comprehensibility to
humans,” allowing them a better understanding of the underlying data.
Data used in this task is usually divided into instances described by a set of attributes. The most common
problem in data mining is trying to predict the value of a user-defined attribute based on the values of
some other attributes. This is called a classification problem. When we want to predict the values of several
attributes (instead of just one goal attribute), the problem is now an association one. In some problems
we do not really have an attribute that defines the class of a given instance but still would like to know if
it were possible to group the data in different groups or classes. This grouping (and class discovery) must
then be done by the algorithm and this becomes a clustering problem. Useful overviews of the field can
be found in References 1 and 2.
There are several research areas that provide the methods in data mining, the most important being stat-
istics and Artificial Intelligence (AI). Other areas include pattern recognition, databases, data visualization,
etc. Within the AI area, the main contributions came from the machine learning community and represent
the three main AI paradigms: symbolic, connectionist, and evolutionary. From these, the last two bring us
back to the title — and content — of this chapter, since both have a strong biological inspiration. A useful
introduction to machine learning approaches to data mining is presented in Reference 3.
Connectionist approaches in data mining are usually synonyms of neural networks. This is one of the
most widely known (if not widely understood) technique for data mining. It provides a very flexible way
of fitting a model to observed data. This model can then be used to classify new data or make some kind
of prediction on it. The major drawback of neural networks is that the resulting models cannot easily
fulfill the “comprehensibly to humans” requirement, since their most friendly representation is a network
of weighted connections between nodes where some nonlinear operation occurs.
In this chapter, we focus mainly on bio-inspired approaches for data mining that are not neural network
related. We start by giving a slight overview of these approaches, especially evolutionary ones. We also
describe some novel approaches that are based in new computational models, inspired by biological
mechanisms, but not completely accommodated by the connectionist or evolutionary paradigms.
Following this overview, we introduce a new approach to the data mining task of finding classification
rules based on the Particle Swarm Optimizer (PSO) algorithm.
The particle swarm algorithm has been originally presented in Reference 4 as a population-based
function optimizer in the n-dimensional space of real numbers. Bird flock flight simulations initially
inspired the algorithm, and biological inspiration is still present in the current denomination. The swarm
or flock metaphor applies to the PSO in the way particles fly in a somewhat coordinated way through an
n-dimensional space (the space of parameters of the function being optimized) in search of some desired
place (function optimum).
While being considered as a form of evolutionary algorithm, there is no use of genetic operators
such as mutation or recombination, as is the case in other evolutionary paradigms such as evolutionary
programming, evolutionary strategies, or genetic algorithms. Explicit selection is also not present. Instead,
in each iteration, the position of every particle in the search space is updated accordingly to the particle’s
velocity. The velocity of a particle in a given iteration is a function of the velocity in the previous iteration,
its best previous position in the search space, and the best previous position in the search place of all of the
particle’s neighbors. The behavior of a particle in the swarm is the result of balancing the desire of flying
toward the best point in the search place according to its own experience or conforming to the swarm
knowledge of where the current best point is. For an extensive discussion of the cultural model behind
the PSO, as well as of the PSO itself and several variants see Reference 5.
Our method uses a specialized particle swarm algorithm to build rules for classification tasks. We call the
resulting algorithm a Particle Swarm Data Miner (PSDM). We present several variants of the PSDM, more

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 470 — #2

Bio-Inspired Data Mining 27-471

explicitly, a discrete version, two real-valued versions using different stop criteria, and an adaptive version.
This last version is based on the Simple Adaptive Predator–Prey Optimizer (SAPPO) algorithm [6,7],
initially developed to improve performance on real-valued optimization problems. SAPPO introduces
a predator–prey mechanism in the basic swarm algorithm in order to maintain diversity during the
search. It also includes a swarm of symbiotic particles, designed to allow the adaptation of the algorithm
parameters to the problem being tackled.
These algorithms are compared between themselves and against two industry standard classification
algorithms (J48 and PRISM) on a set of benchmark problems. The results are then used to compare and
discuss the different characteristics of each variant.
Section 27.2 presents the overview on bio-inspired data mining methodologies. In Section 27.3 we
describe the higher level classification algorithm, which can transparently make use of any of the lower
level PSO-based algorithms to search for classification rules. The first three variants of the PSO algorithm
are discussed in Section 27.4. The SAPPO data miner is presented in Section 27.5. The experimental
setup used to test the algorithms is outlined in Section 25.6, where we also report and discuss the results
obtained. Finally, we draw some conclusions in Section 27.7.

27.2 A Brief Overview of Bio-Inspired Data Mining

As mentioned previously, we are not going to consider the neural-network-based approaches in this
overview. Neural networks were the first bio-inspired computational mechanism to be used in data
mining, probably because of being so directly applicable to classification tasks. Its applications are now
numerous and well known.
Evolutionary approaches, on the other hand, are relatively new to data mining. We describe the most
common ones in the following paragraph. More recently, other biological mechanisms, such as the
intelligent behavior of ants and the working of the human immune system have inspired new algorithms
and the first attempts of applying these to data mining tasks have been made. We report on these in the
following paragraphs. In the rest of this chapter, we introduce and discuss the use of particle swarm-based
algorithms for data mining. These algorithms also have a biological base, as their origin is closely related
with the study of coordinated movement in “swarms” of living creatures, for example, flock of birds
and schools of fish.
The main motivation for using these algorithms in data mining comes from the way in which patterns
are searched by more traditional approaches (e.g., tree induction and rule discovery algorithms). These
are usually based on some greedy search criterion that is used to fix a part of the pattern, with the search
then continuing for the rest. It is well known that greedy search techniques are easily tricked into local
optima of the search space. Data mining tasks involving numerous attributes that are strongly related
provide search spaces where these sub-optima abound. Biological inspired search procedures take a more
global approach to search, usually incorporating stochastic mechanisms that can enable the avoidance of
such optima.

27.2.1 Evolutionary Approaches to Data Mining

Evolutionary algorithms are based on the principles behind the theory of evolution, mainly the idea of
survival of the fittest. These algorithms act over a group (population) of data structures representing
tentative solutions for some problem we want to solve. The quality of these solutions is evaluated (using
a fitness function) so that better solutions have a higher probability of being totally or partially copied
onto a new group of solutions — the next generation. This mechanism is called selection.
However, selection alone is not enough, since you will only have the individuals in the initial population
to work with. We also need variation, so that better individuals can appear, and have their characteristics
spread through the population by selection. In evolutionary computation, variation is mainly achieved

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 471 — #3

27-472 Handbook of Bioinspired Algorithms and Applications

by randomly changing parts of a solution (mutation) or recombining existing solutions into new ones
(recombination).
These basic principles have given origin to many flavors of evolutionary algorithms of which the
most widely known (and used) are genetic algorithms [8], evolutionary programming [9], evolutionary
strategies [9], and genetic programming (GP) [10]. In spite of their differences, all these algorithms are
built around what is usually called the generic evolutionary algorithm loop:

P(0) = Generate_initial_population()

Evaluate_population(P(0))

While not Some_stop_criterion() do

P’(t) = Select_individuals_to_reproduce(P(t))

P’’(t) = Create_offspring_using_recombination(P’(t))

P(t+1) = Merge_old_population_with_offspring(P’’(t))

Evaluate_population(P(t+1))

Of the algorithms given, genetic algorithms and GP are the most widely used for data mining tasks.
In Sections 27.2.1.1 and 27.2.1.2 we present, in a generic way, how these algorithms can be used for
data mining, more specifically their application to rule discovery. An extended survey on the use of
evolutionary algorithms not only for data mining but also for other tasks in the KDD progress can be
found in Reference 11.

27.2.1.1 Genetic Algorithms

The classical genetic algorithm acts on populations of fixed length binary strings and uses a one-point
crossover recombination operator, as well as some form of bit mutation. Most data mining applications of
genetic algorithms focus on using the algorithm to discover prediction (if–then) rules. These applications
can be divided accordingly to the way in which the rules (the solution) are encoded. Two main approaches
are possible: one individual can represent a single rule — this is sometimes called the Michigan approach,
for example, References 12 and 13 or it can represent a whole set of rules — the Pittsburgh approach, for
example, References 14 and 15.
The choice of an approach can depend on both the task being tackled and the way the evolutionary
algorithm is integrated in the data mining system. As a general guideline, if a rule can be effectively
evaluated in an isolated way — such as in association problems — it can be simpler to encode just one
rule in each individual. If rules are highly dependent (as is often the case with rule sets for classification
tasks), encoding the whole rule set might be more advisable, even if at a cost in terms of representation
complexity.
In terms of representation, a rule is usually represented as a bit string. In a typical if-then rule, some
attributes are tested in the conditional part of the rule and the value of one or more attributes are predicted
in the consequent part. Assuming categorical attributes, each one is represented in the rule by a number of
bits equal to the number of its possible values. A test on attribute a of the form (a = a1 or a = a4)
can then be represented by the substring 1001 (for four possible values). An individual is built by concat-
enating the substrings for each attribute tested by a rule. If the representation includes several rules then
this process is repeated for each one.
Standard operators can be used to create offspring, but specialized ones are often implemented,
for example, to specialize or generalize rules.
An essential aspect of every application of a genetic algorithm is the design of the evaluation function.
While it is out of the scope of this overview a detailed discussion of evaluation functions, it should be

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 472 — #4

Bio-Inspired Data Mining 27-473

pointed out that not only the accuracy of a rule (or rule set) is important in its evaluation, but also other
aspects can be taken into account, for example, its simplicity.

27.2.1.2 Genetic Programming

GP is an evolutionary algorithm designed to evolve computer programs [10]. These are usually represented
as parse trees and recombination and mutation involve the exchanging and replacement of sub-trees.
Functions, variables, and constants are user chosen, so GP can easily be used to evolve rules for data
mining tasks. With some effort these rules can accommodate not only attribute/values comparisons
(e.g., a = a1), but also complex relations between attributes (e.g., a < c). Primary attributes can also
be combined into new, more expressive, attributes (e.g., numerical attributes can be combined using
arithmetic functions).
Typical recombination (sub-tree exchanging between individuals) and mutation (sub-tree replacement)
operators can be used if care is taken to guarantee that correct parameters are given to every function.
Especially when mining databases with both numerical and nominal attributes, when functions capable of
dealing with both types of attributes must be used, the operators need to guarantee that nominal values are
not passed to numerical functions and vice versa (to avoid expressions such as ‘‘high’’+‘‘low’’).
Due to its flexibility, as well as its ability to use any user-chosen functions to test and combine attributes,
many applications have emerged in data mining using GP, mainly for classification tasks, see, for example,
References 16 to 19.
As with genetic algorithm-based data mining applications, the design of the fitness function can
incorporate not only an accuracy measure, but also other factors, such as simplicity. With GP this is
usually related with the size of the program tree.

27.2.2 Ant Colonies for Data Mining

Another biological mechanism that has served as inspiration for a data mining application is the one
imitated by the Ant Colony Optimization algorithm [20]. This algorithm is inspired in the way real ants
optimize their tracks. When faced with an obstacle in a food track, an ant initially chooses its way around
it in a random fashion. However, while moving, every ant deposits on the ground a chemical trail of
pheromone. Ants that choose the shortest path around the obstacle will take less time to reach the food
than the ones that follow the longest trajectory, so, with all ants moving at a similar rate and depositing
similar quantities of pheromone there will be a tendency to accumulate more pheromone in the shortest
path. This tendency will be exacerbated by the fact that all ants have a preference, when in a crossroad,
to choose the way with larger deposits of pheromone. This will obviously lead to a rapid increase in the
number of ants that choose the shortest path, with a corresponding increase in pheromone depositing.
The result is that the final path to the food source, followed by almost all the ants, will emerge as the
shortest path between the nest and the food source.
Replacing obstacles by decisions and paths by costs (or gains), this mechanism can be implemented
as the Ant Colony Optimization algorithm [21], which has been successful in several combinatorial
optimization tasks, as well as in shortest path problems (e.g., in the telecommunications domain). Its use
in data mining, for finding if-then rules for classification tasks, was proposed by Parpinelli et al. [22], with
the decisions corresponding to the choice of an attribute/value test. A heuristic was used for estimating
the cost of a given test. A path represents a rule and pheromone is deposited in the path in proportion
to the rule fitness.

27.2.3 Artificial Immune Systems for Data Mining

Artificial immune systems are a new computational paradigm inspired in theoretical immunology
and observed immune functions, principles, and models, applied to solve problems [23]. There is not
just one artificial immune system inspired algorithm. Instead, we can talk about a framework for the
development of artificial immune system algorithms. A specific algorithm includes a representation to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 473 — #5

27-474 Handbook of Bioinspired Algorithms and Applications

create abstract models of immune organs, cells, and molecules; a set of affinity functions, which quantify
the interactions of these elements; and a set of algorithms to rule the dynamic behavior of the artificial
immune system.
Several mechanisms of the immune system have been used as inspirations for specific algorithms
[23,24]. Clonal selection, a mechanism by which immune cells that have a higher affinity to a given
pathogen are allowed to reproduce more and with lower mutation rates than cells with lower affinity,
has been used in optimization. The “affinity measure” is given by the function being optimized. There is
no special reason why this mechanism cannot be used to search for binary represented if–then rules, with
affinity being measured by rule fitness.
Another theoretical model of the immune system, immune networks, have been used to develop
learning networks, where nodes correspond to immune cells and connection strength represent affinity
between the pathogen recognized by each cell. This model has already been used in data analysis, namely
for data clustering [25,26].

27.2.4 Particle Swarm Algorithms for Data Mining

In the following sections we present — to the best of our knowledge — the first attempt at the use of
particle swarm algorithms for data mining. Being a recent addition to the biological inspired computation
area [4,5], particle swarm algorithms have been mostly used for real-valued optimization (e.g., References
27 and 28) but applications to other areas have been appearing quickly, as can be observed in this extensive
bibliography of the field [29].
In our work several variants of the PSO algorithm were used in the inner loop of a classification
covering algorithm in the search for classification rules. In Section 27.3 we describe the high level covering
algorithm and validation procedure. The following sections describe the use of PSO-based algorithms
for rule search and present experimental results when compared with industry standard classification
techniques.

27.3 Covering Algorithm and Validation Procedure

Our research project has focused in applying, testing, and tuning PSO-based algorithms to classification
rule discovery problems.
In a classification problem having a dataset in which each instance is characterized by a set of attributes
and belongs to a specific class, the objective is to extract a set of rules that can accurately predict the class
of a new instance from its attribute values.
Classification rules are no more than conditional clauses, involving two parts: the antecedent and
the consequent. The former is a conjunction of logical tests, and the later gives the class that applies to
instances covered by this rule. These rules usually take the following format:

IF attribute_a=value_1

AND attribute_b=value_2

AND attribute_n=value_I

THEN class_x

A rule set is constructed with rules performing different attributes tests, predicting different classes and
a default rule, to capture and classify instances not classified by the previous rules.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 474 — #6

Bio-Inspired Data Mining 27-475

A sequential rule set would conform to this structure:

Rule #1

Rule #2

Rule #n

Default Rule

Containing no attribute tests and predicting the same class as the one predominant in the remaining
instances, the default rule takes the form:

IF true

THEN class_x.

As an illustrating example, one of the datasets commonly used as a benchmark, describes animals accord-
ingly to whether they have hair, feathers, or tail, among other attributes, and classifies them as mammals,
birds, fish, etc. The resulting set of rules could be:

=>RULE #1

IF milk = true

THEN type IS mammal

=>RULE #2

IF feathers = true

THEN type IS bird

=>RULE #3

IF fins = true

THEN type IS fish

=>RULE #4

IF airborne = false

AND tail = false

THEN type IS invertebrate

=>RULE #5

IF legs=6

THEN type IS insect

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 475 — #7

27-476 Handbook of Bioinspired Algorithms and Applications

=>RULE #6

IF TRUE

THEN type IS reptile

A new instance, the characterization of one animal that we might need to classify, would be iteratively
tested against each rule, until its attributes fully matched the attribute tests of one rule. This new instance
would be classified with the rule’s predicted class.
This type of database that extracted high level knowledge is greatly valued in areas such as loan granting,
fraud detection, marketing, etc., where huge databases already exist and predictions are most welcome.
Once found, a set of if–then rules is easily comprehended and utilized by humans, which need no special
training.
The complexity of this problem — finding the rule set and the set of attribute tests for each rule —
lies in the exponentially high number of attribute-value pairs, which generate a wide multidimensional
search space.
In this kind of search problem, evolutionary algorithms such as GA and PSO have already proven to be
reliable and efficient due to their parallel, population-based search strategies.
Following a typical architecture of the Michigan approach in classification rule discovery, the over-
all structure of our work was designed contemplating three imbricated procedures, each one fulfilling
a specific task.

Validation Procedure

| Covering Procedure

| |

| | Classification Rule Discovery Procedure

| | |

| | A Quality/iteration number criterion is met

| |

| A minimum number of uncovered instances is met

10 times

The classification rule discovery procedure aims to find and return the rule that better classifies the
predominant class in a given instance set. One rule is better than another, if it can match a higher
number of instances with the specified class and segregate instances with a different class. It is at this
level that the PSO-based algorithms are used. The covering procedure, on receiving an instance set (the
training set), invokes the classification rule discovery procedure to reduce this set by removing instances
correctly classified by the rule returned by the classification rule discovery procedure. This process is
repeated until only a predefined number of instances are left to classify in the training set. A sequential
rule set is therefore created — hence the Michigan approach. The aim of the validation procedure is, not

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 476 — #8

Bio-Inspired Data Mining 27-477

only to determine the accuracy of a rule set returned by the covering procedure, but also to gauge the
liability of the complete classifying procedure — classification rule discovery and covering procedures
altogether. This is achieved by iteratively dividing the initial dataset into different test and training sets
and computing average indicators, such as accuracy, rule number per set, and attribute tests number
per rule.
Having this architecture, clearly segregating tasks to each of the three imbricated procedures, and using
alternatively different search strategies in its core procedure — classification rule discovery — we end up
having an unbiased platform for testing and benchmarking each of these search strategies.

27.3.1 Validation Procedure

The purpose of the validation procedure is to statistically evaluate the accuracy of the rule set obtained by
the covering procedure. This is done using a method known as tenfold cross validation.
The tenfold cross validation consists of dividing the dataset into ten equal partitions and iteratively
using one of this sets as a test set and the remaining nine as training sets. In the end, ten different rule sets
are obtained and average indicators, such as accuracy, rule number per set, and attribute tests number per
rule are computed.
Rule set accuracy is evaluated and presented as the percentage of instances in the test set correctly
classified. An instance is considered correctly classified, when the first rule in the rule set whose antecedent
matches this instance and the consequent (predicted class) matches this instance’s class.

27.3.2 Covering Algorithm

The covering procedure is basically a divide-and-conquer technique. Being given an instance training set,
it calls on the rule discovery procedure in order to obtain the highest quality rule for the predominant
class in the training set.
Once found, this rule goes through a pruning process where unnecessary attribute tests are removed.
This simple process iteratively removes attribute tests if the pruned rule has a better quality value than the
original rule.
Correctly classified instances are then removed from the training set and the rule discovery procedure
is run once more. A sequential rule set is built iteratively, and the covering procedure runs until only
a predefined number of instances are left to classify. This threshold criteria value is user defined as
a percentage and it is typically set to 10%.

27.3.3 Postprocessing Routines — Rule Pruning and Rule Set Cleaning

A final aspect of the higher level algorithm worth discussing is the postprocessing routines. Recall that
high level knowledge extracted from databases must conform to three main requisites: accuracy, compre-
hensibility, and interest for the user [1]. In classification rule discovery problems, the number of attribute
tests per rule and the number of rules per set is a major contributor for the comprehensibility of the
obtained results — fewer attribute tests and rules ease comprehensibility.
After a rule is returned from the classification rule discovery algorithm it goes through a pruning process
in order to remove unnecessary attribute tests. This is done by iteratively removing each attribute test
whenever the newly obtained rule has the same or higher quality value than the original rule.
Just after the covering algorithm returns a rule set, another postprocessing routine is used: rule set
cleaning, where rules that will never be applied are removed from the rule set.
As rules in the rule set are applied sequentially, in this routine, rules are removed from the rule set if:

• There is a previous rule in the rule set that has a subset of the rule’s attribute tests.
• If it predicts the same class as the default rule and is located just before it.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 477 — #9

27-478 Handbook of Bioinspired Algorithms and Applications

Therefore, in the example given, rule number 2 and 3 will be removed and the rule set will be reduced to
the first and last rules:

Rule #1

If attribute_a = x_a

Then class=c_1

Rule #2

If attribute_a = x_a and attribute_b = x_b

Then class=c_2

Rule #3

If attribute_c = x_c

Then class=c_3

Rule #4 - Default Rule

If TRUE

Then class=c_3.

27.4 Basic Particle Swarm Data Miners

As already mentioned, this architecture, which clearly segregates tasks to each of the three imbricated
procedures, provides an unbiased platform for testing and benchmarking different search strategies in its
core procedure — classification rule discovery.
Several variants of PSO algorithm were therefore implemented and tested. Sections 27.4.1 to 27.4.4 give
an overview of these algorithms.

27.4.1 The PSO Algorithm

The PSO is inspired in the intelligent behavior of beings as part of an experience sharing community as
opposed to an isolated individual reactive response to the environment. The Adaptive Culture Model [5],
which is PSO’s framing theory, states that the process of cultural adaptation is rooted into three principles:
evaluate, compare, and imitate.
In PSO algorithms a particle decides where to move next, considering its own experience, which is the
memory of its best past position, and the experience of its most successful neighbor.
There may be different concepts and values for neighborhood; it can be seen as spatial neighborhood
where it is determined by the Euclidean distance between the positions of two particles, or as a sociometric
neighborhood (e.g., the index position in the storing array). The latter is the most commonly used.
In PSO, a population of point particles “fly” in an n-dimensional real number search space, where each
dimension corresponds to a parameter in a function being optimized. A vector X represents the position
of the particle in each dimension of the search space, and another vector V represents the velocity of the
particle, that is, its change in position. The particle “flies” in the search space by adding the velocity vector
to its position vector in order to change its position.
The particle’s trajectory is determined by V and depends on two “urges” for each particle i: flying toward
its best previous position and flying toward its neighbors’ best previous position. Different neighborhood

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 478 — #10

Bio-Inspired Data Mining 27-479

definitions have been tried [30]; here, we assume that every particle is a neighbor to every other particle
in the swarm. The general equations for updating the position and velocity for some particle i are the
following:

Vi (t ) = χ (wVi (t − 1) + ϕ1i (Pi − Xi (t − 1)) + ϕ2i (Pg − Xi (t − 1))),

(27.1)
Xi (t ) = Xi (t − 1) + Vi (t ).

In this formula, χ is the constriction coefficient described in Reference 27, ϕ1 and ϕ2 are random numbers
distributed between 0 and an upper limit and different for each dimension in each individual, Pi is the
best position particle i has found in the search space and g is the index of the best individual in the neigh-
borhood. The velocity is usually limited in absolute value to a predefined maximum, Vmax . The parameter
w is a linear decreasing weight. The swarm is usually run for a limit number of iterations or until an error
criterion is met.
From Equation (27.1) we can derive the two most usual ways in which convergence and, as a result,
the balance between exploration and exploitation is controlled. Reference 28 uses χ = 1 and weight w
decreasing linearly from wmax to wmin during the execution of the algorithm. In Reference 27 convergence
is guaranteed by choosing appropriated values for χ and ϕ = ϕ1 + ϕ2 . w is fixed and equal to 1 in this
approach.
Although some actions differ from one variant of PSO to another, its basic pseudo-code is as follows:

Initiate_Swarm()

Loop

For p=1 to number of particles

Evaluate(p)

Update_past_experience(p)

Update_neighbourhood_best(p,k)

For d=1 to number of Dimensions

Move(p,d)

Until Criterion

The output of this algorithm is the best point in the hyperspace the swarm visited.
There are several variants of PSO, typically differing in the representation: discrete or continuous
PSO [5]; in the mechanism used to avoid spatial explosion of the swarm and guaranteeing convergence:
linear decreasing weight [28] or constricted PSO [27]; or in the mechanism used to avoid premature
convergence to local optima: predator–prey interactions [6] or collision avoiding swarms [31].

27.4.2 Rule Representation

In our particle swarm data miners, each particle encodes a rule, mapping the attribute tests in dimension
coordinates. In a preprocessing routine, a data image is created normalizing all attribute values to the
range [0.0, t ] with 0.0 < t < 1.0, t being a user predefined value. t stands for the indifference threshold
where a higher value will trigger the omission of the corresponding attribute test.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 479 — #11

27-480 Handbook of Bioinspired Algorithms and Applications

Nominal attributes are normalized assigning to each different attribute value an enumerated index #idx
and applying Equation (27.2)

idxv × t
vnorm = . (27.2)
#idx

idxv is the index of the attribute value v and #idx the total number of different attribute values.
Both integer and real types are normalized using Equation (27.3)

(v − vmin ) × t
vnorm = . (27.3)
vmax − vmin

vmin and vmax are the lower and higher attribute values found for this attribute.
Rules/Particles are encoded as a floating-point array and each attribute is represented by either one or
two elements on the array, according to its type: nominal attributes are assigned with one element on the
array and attribute-matching tests are defined as follows:


true if vr × #idx = vi × #idx ,
m(vr , vi ) = (27.4)
false otherwise.

t being the indifference threshold value, vr the attribute value stored in the rule for testing and vi the
instance value stored in the normalized image of the dataset.
Integer and real attributes are assigned with an extra element in the array in order to implement a value
range instead of a single value,


true if vr1 ≥ t or (vr1 − vr2 ) ≤ vi or (vr1 + vr2 ) ≥ vi ,
m(vr1 , vr2 , vi ) = (27.5)
false otherwise.

vr1 can be seen as the center and vr2 as a neighborhood radius, inside which matching will occur.

27.4.3 Rule Evaluation

Rules must be evaluated during the training process in order to establish points of reference for the training
algorithm: best particle positioning. The rule evaluation function must not only consider instances
correctly classified, but also the ones left to classify as well as the wrongly classified ones.
The formula used to evaluate a rule is expressed in the given Equation [22]:


 TP TN
× if 0.0 ≤ xi ≤ 1.0, ∀i ∈ d,
Q(X ) = TP + FN TN × FP (27.6)

−1.0 otherwise,

where
• TN (True Positives) = number of instances covered by the rule that are correctly classified, that is,
its class matches the training target class.
• FP (False Positives) = number of instances covered by the rule that are wrongly classified, that is,
its class differs from the training target class.
• TN (True Negatives) = number of instances not covered by the rule, whose class differs from the
training target class.
• FN (False Negatives) = number of instances not covered by the rule, whose class matches the
training target class.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 480 — #12

Bio-Inspired Data Mining 27-481

This formula penalizes a particle, which has moved out of legal values, assigning it with a negative value
(−1.0), forcing it to return to the search space.

27.4.4 Data Mining with Particle Swarms

The basic particle swarm data miner simply uses the updated equations presented for the PSO algorithm
in Section 27.4.1 to search the rule space defined by the earlier representation. The covering algorithm
described in Section 27.4.3 iteratively calls the particle swarm data miner with receeding datasets to build
the final rule set.
In our experiments, we implemented and tested three different loop end criteria for the basic PSO
data miner:
• Iteration number: the loop stops when a predefined number of iterations is achieved (usually 2000).
• Convergence platform: the loop stops when the swarm’s best particle’s quality value is constant for
a predefined number of iterations (usually 1000).
• Convergence radius: the loop stops when all the swarm’s particles are within a predefined normalized
Euclidean distance from the swarm’s best particle (usually 0.01), as a safeguard against simultaneous
multiple optima exploitation, which causes the splitting of the swarm and therefore making this
condition unachievable, the loop stops when the number of iterations reaches a predefined value
(usually 2000).
The termination criterion is an important factor in the algorithm, since the PSDM is computationally
very intensive. An early termination, achieved by radius or platform convergence can help decrease the
number of evaluations but it can also result in the algorithm stopping before the best rule is found. Results
from early versions of our algorithm can be found in our earlier work [32,33].

27.5 An Adaptive PSDM

The traditional PSO models, such as the ones used for the algorithm described in Section 27.4, while
successful in many applications, especially in optimization, suffered from some limitations. One of these
was that the basic PSO had no mechanism to introduce diversity after convergence to an optimum had
occurred. These meant that if the algorithm converges to a local optimum it cannot escape it. The
PSO algorithms presented in References 27 and 28 used carefully chosen parameters to promote a large
exploration phase followed by convergence, to increase the probability of finding the global optimum.
Another limitation was the algorithm’s sensibility to its parameter’s values, which can be observed in the
results obtained by Clerc and Kennedy [27] for several benchmark problems and parameters sets.
In the previous work [6,7], we introduced a new mechanism in the basic PSO algorithms to help
overcome these deficiencies. A predator–prey mechanism was introduced to maintain diversity even after
convergence, so that some exploration of the search space was possible in the later stages of the algorithm’s
execution. A swarm of symbiotic particles was also used to allow the adaptation of the parameters to the
problem being solved. The resulting algorithm was called SAPPO. The initial inspiration for the particle
swarm optimizer, as described in Reference 4, was the coordinated movement of groups of animals in
nature, for example, schools of fish or flocks of birds. The new mechanisms introduced in SAPPO also
get their inspiration from nature. The predator-prey interaction is based on the disturbance caused by
predators to the groups of animals being hunted. The adaptation scheme used mimes, the symbiotic
associations between species so frequently found in nature. Both mechanisms are explained in detail in
Sections 27.5.1 and 27.5.2.

27.5.1 The Predator–Prey Interaction

In standard PSO, both the constriction factor version [27] and the linear weight version [28], promote
convergence of the swarm to a local optimum by damping the particles’ velocity by a deliberate choice of

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 481 — #13

27-482 Handbook of Bioinspired Algorithms and Applications

the algorithm’s parameters. An undesirable side effect of these strategies is that when the local optimum is
not the global optimum being sought, particles cannot gain velocity to jump to another optimum in the
search space. This phenomenon is similar to premature convergence in other evolutionary algorithms.
Our motivation for introducing the predator–prey interaction was mainly to introduce a mechanism
for creating diversity in the swarm at any moment during the run of the algorithm, by adding velocity to
some particles, not depending on the level of convergence already achieved. This would allow the “escape”
of particles even when convergence of the swarm around a local sub-optimum had already occurred.
There are other mechanisms used for the same effect in the literature (see Reference 31), but two main
reasons led us to prefer the predator–prey scheme. The first reason is its computational simplicity when
compared with other approaches. As it will be seen when the mechanism is explained, it only introduces
a new particle and little computational effort in the basic algorithm. The simple adjective in SAPPO
comes from this fact. A second, and less technical, motive was to maintain the swarm intelligence philos-
ophy behind the algorithm. It seemed more appropriate to introduce a mechanism that could also be
implemented as a distributed behavior in the swarm.
The predator–prey model is based on the disturbance caused by a predator to the group of animals
being hunted. Animals are driven from their favorite locations, for example, pastures and water sources,
by fear of nearby predators. Eventually, this process will result in the finding of even better locations where
the arriving animals will also be chased by nearby predators.
It is this predator/prey interaction that we try to reproduce in SAPPO. Here, a new particle is introduced
into the swarm to mime the predators’ behavior. This particle, called the predator particle, is attracted by
the best (fittest) particle in the swarm, according to the following equations:

Vp (t ) = ϕ3 (Xg (t − 1) − Xp (t − 1)),
(27.7)
Xp (t ) = Xp (t − 1) + Vp (t ).

In Equation (27.7), ϕ3 is another random number distributed between 0 and an upper limit, usually 1,
and Xg is the present position of the best particle in the swarm.
The predator particle can influence any particle in the swarm by changing its velocity in one or
more dimensions. This influence is controlled by a “fear” probability f , which is the probability of
a particle changing its velocity in one of the available dimensions due to the presence of the predator. For
some particle i, if there is no change in the velocity in a dimension j, the update rules in that dimension
still are:

vij (t ) = wvij (t − 1) + ϕ1ij (pij − xij (t − 1)) + ϕ2ij (pg j − xij (t − 1)),
(27.8)
xij (t ) = xij (t − 1) + vij (t ).

The only differences from the other approaches are that w is fixed and χ is not explicitly used. However,
if the predator “scares” the prey (particle), that is, if there is a change in velocity in dimension j, the rule
becomes:

vij (t ) = wvij (t − 1) + ϕ1ij (pij − xij (t − 1)) + ϕ2ij (pg j − xij (t − 1)) + D(d),
(27.9)
xij (t ) = xij (t − 1) + vij (t ).

This process is repeated for all dimensions, that is, there can be simultaneous changes in velocity in several
dimensions.
The fourth term in the first equation of (27.9) quantifies the repulsive influence of the predator. This
term is a function of the difference between the positions of the predator and the particle. The Euclidean
distance between predator and prey is d. D(x) is an exponential decreasing distance function:

D(x) = ae−x/b . (27.10)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 482 — #14

Bio-Inspired Data Mining 27-483

D(x) makes the influence of the predator grow exponentially with proximity. The objective of its use
is to introduce more perturbation in the swarm when the particles are near the predator, which usually
happens when convergence to a local optimum occurs. The a and b parameters define the form of the
D function: a represents the maximum amplitude of the predator effect over a prey and b allows to control
the distance at which the effect is still significant.
During the initial stages of the algorithm, its behavior is similar to a traditional PSO, since the particles
are scattered and the predator’s influence is negligible. As convergence occurs and particles start to move
to a local optimum, their distance to the best particle diminishes and the predator effect exponentially
increases, accelerating some particles in new directions. When the current local optimum is not the same
as the global optimum being searched, this will hopefully allow one particle to jump to another near
local optima, becoming the new best particle and leading the swarm to a new exploration phase. Several
repetitions of this process could lead the swarm through a chain of local optima until the global optimum
is found.

27.5.2 The Adaptation Scheme

The analysis of previous work with the PSO has made it clear that a good parameter choice for the PSO,
is not only important for the algorithm performance, over a set of problems, but also that the performance
of the algorithm, for a specific problem, varies significantly for different set of parameters.
Choosing parameters to generally guarantee convergence to a local optimum can be done using the
constriction factor method. Nevertheless, if what we are looking for is not simply convergence to a local
optimum, but converge to the global optimum or at least to a “good” local optimum, finding a set of good
parameters basically remains a trial and error process.
As in any other optimization or search algorithm, the ideal solution would be to have a built-in procedure
to adapt the algorithm itself to the problem that it is trying to solve. Other evolutionary algorithms,
for example, evolutionary strategies, tried to do this by encoding some of the algorithm’s parameters
(e.g., mutation probability, variance, etc.) together with the individuals, thus trying to coevolve solutions
and algorithm’s parameters.
In spite of the only relative success of past instances of this idea, we believe that PSO-based algorithms,
in general, and specially predator–prey based ones, can greatly gain by the inclusion of coevolution-
based adaptive schemes. For instance, the a parameter in the D(x) function has its best influence on the
algorithm when it is near the average distance between sub-optima in the search space. The algorithm
could substantially gain if this parameter could be learned during execution. The same could be said for
other parameters, with different roles in the algorithm.
Another important reason is the simple (one almost could say natural) way in which such a mechanism
can be included in the PSO framework, as we hope to show in the text that follows.

27.5.2.1 Symbiotic Particles

We introduced a new species of particles in SAPPO as the base for the adaptation mechanism. To the
original swarm — constituted by solution particles — we now add a swarm of parameter particles,
encoding the algorithm’s parameters. These particles interact between themselves as a normal swarm
using Equation (27.1) with a constriction factor χ of 0.729 and ϕ = 4.1, as recommended in Reference 27,
with only the two adjacent particles being considered as neighbors. Each particle encodes seven real-valued
parameters corresponding to seven adaptable parameters in SAPPO: a, b, and f for the predator–prey
effect, N for the number of neighbors of each particle and c1 , c2 , c3 for the velocity and position equations,
which become:

Vi (t ) = c1 Vi (t − 1) + c2 (Pi − Xi (t − 1)) + c3 (Pg − Xi (t − 1)),

(27.11)
Xi (t ) = Xi (t − 1) + Vi (t ).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 483 — #15

27-484 Handbook of Bioinspired Algorithms and Applications

The aim of the adaptation scheme is to find, simultaneously, a solution for some problem and a set of
parameters that increase the performance of the algorithm in the search for that solution. To achieve
this, we must link in some way the two swarms in the algorithm. We modeled this link on the symbiotic
relations so common in nature, where two species live in a close relation from which both get some
advantage.
In SAPPO each solution particle lives in symbiosis with a parameter particle, which encodes the
parameters used when the algorithm’s update equations are applied to that solution particle. The symbiotic
relation is implemented through the definition of the parameter particle fitness function.
A parameter particle has a “slower” life cycle than its companion does. While a solution particle is
evaluated and has its velocity and position updated every iteration, to a parameter function this only
happens every i iterations (usually 10). The parameter particle is then evaluated by comparing the actual
fitness of its solution particle companion with the same fitness i iterations ago. If there was an improve-
ment, its value is stored as the parameter particle fitness, unless it is smaller than the improvement value
already stored. When a new improvement value replaces an older one, the P vector is also updated with
the current parameter particle’s position in the search space. Velocity and position updates will then be
made using Equation (27.11). As a final element, the fitness of a parameter particle will decay slowly over
every iteration of the algorithm (usually being multiplied by a decay factor α = 0.98). This decaying
ensures that a parameter particle has to keep producing improvements in the associated solution particle
to maintain a high fitness.
Our approach to adaptation in SAPPO tries to maintain, as we already did with the predator–prey
principle, the biological inspiration of the original PSO algorithm. We also made an effort to develop
a mechanism that could be implemented following principles and ideas underlying the paradigm of
swarm intelligence. Both the mechanisms introduced are based in the interaction of simple particles,
with simple update rules and no centralized control. Both the solution and the behavior of the algorithm
emerge from the interactions between the individuals in this ecology of particles, and if there is intelligence
in the system it is clearly not preprogrammed but emerges as a property of the system of swarms.

27.5.3 Using SAPPO for Data Mining

We believe that the search space of rules in classification tasks will have irregularities and local optima
capable of attracting the basic PSO (and especially traditional, greedier, approaches) to sub-optimal
solutions. In Section 27.6 we empirically compare a SAPPO-based data miner to both the other PSO
based approach already described and classical data mining algorithms, to investigate if the increase in
search space complexity (new dimensions added for the algorithm’s parameters) is compensated by an
increase in performance.
For these experiments, the SAPPO algorithm was transparently integrated in our PSDM, which, as was
explained in the previous sections can accommodate any algorithm for rule search whiteout changes to
the higher levels of the PSDM.

27.6 Experimental Results

To evaluate the PSDM introduced in the previous sections we decided to test their performance in terms
of both prediction accuracy and comprehensibility in a set of benchmark datasets. These datasets, listed
in Table 27.1, were chosen to present different kinds of problems to the algorithms.
In Table 27.1 each dataset is described in terms of number of instances, attributes, classes, possible
attribute/value pairs, and, finally, approximate size of the search space for an individual rule. While these
values cannot be used directly to evaluate the difficulty of each task, they provide, at least, a coarse grained
assessment of that difficulty, specially attending to the size of the search space and the number of instances.
Obviously, this can be misleading, since, for example, a large number of examples neatly grouped by its
class in the search space will provide a simpler classification task than a smaller number of instances
uniformly scattered in the same space.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 484 — #16

Bio-Inspired Data Mining 27-485

TABLE 27.1 Dataset Characterization

Dataset Instances Attributes Classes Attribute/Values S.S. size

Zoo 101 17 7 43 1.38E+6

Breast cancer 286 10 2 54 1.01E+6
Wisconsin breast cancer 699 10 2 92 2.00E+9
Promoters 106 58 2 230 4.15E+34
Splice 1000 61 3 243 3.99E+36

Zoo is a simple dataset that acts as a basic test. It is useful to access the algorithms’ ability to deal with
multiple classes. Breast Cancer is a typical dataset from the field’s literature with only two possible
classes. Wisconsin Breast Cancer presents a similar problem but with a larger search space and
number of instances. Promoters, with instances being DNA sequences, presents a qualitatively different
problem, with much more attributes to be taken into account then the previous datasets. This also results
in a much larger search space. Splice is again a DNA classification problem, but with the availability of
a significantly larger number of examples.
The PSO-based data miners were empirically compared with two standard classification algorithms,
namely J48 and PRISM. J48 is a Java implementation of a decision-tree building algorithm (C4.5) and
PRISM is a rule discovering algorithm. Both are widely used for classification tasks. J48 was chosen for
this comparison since it is probably the best known “industry standard” data mining algorithm. PRISM
uses a similar covering strategy to the one used in the higher level of our algorithms, so it is especially
useful for accessing the advantages of using swarm particle algorithms for the rule discovering level. Both
algorithms can produce rule sets as their final classification model, which provides significant advantages
in comparison fairness. Not only do the algorithms produce the same type of knowledge, but also rule
sets can be easily compared between themselves in terms of comprehensibility and simplicity.
For each dataset we ran ten tenfold cross-validations for each of the algorithms being compared. This
means that 100 classification models where built for each algorithm and dataset. For each model, 90% of
the available instances were effectively used in its construction, while the remaining 10% were used for
testing. The results are presented in Section 27.6.1.

27.6.1 Results
The tables presented show the results obtained for the described datasets. For each experiment with a given
algorithm/dataset combination, we present the average accuracy and variance over the ten tenfold cross-
validations. Averages and standard deviations are also presented for the number of rules and attribute test
per rule over the same cross-validations. For J48 the number of rules represents the number of tree leafs
and the attribute/rule ratio is equal to the number of nodes in the tree divided by the number of leafs.
These numbers allow a fair comparison between the complexity of rule set and tree-based models.
Experiments were done with J48, PRISM, three variants of the basic PSDM with different stop criterions
and, finally a SAPPO-based data miner. The algorithms with platform and radius stop criterions were
investigated in order to find out if temporal complexity — always a major constraint in evolutionary
computation — could be reduced without compromising accuracy. The other PSDM algorithms stopped
(for each rule) when a 2000 iteration limit was achieved.
On an individual analysis of the results, for the zoo dataset, we can observe that, with exception of
PRISM, all algorithms performed at similar levels. From these the simplest models were built by the
PSO-based data miners, with just around seven rules against an average 10.98 by J48. In the PSO-based
approaches, each rule tested just one attribute against an average 1.67 by the J48 algorithm. This makes the
models found by the PSDM algorithms significantly simpler for this problem than the ones found by J48.
For the Wisconsin breast cancer dataset, the accuracies are again similar, with a slight advan-
tage for PRISM. Regarding simplicity, we find that the tendency observed for the previous dataset is
maintained. While J48 (the nonPSDM algorithm with simpler models) uses an average 37.18 rules, all the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 485 — #17

27-486 Handbook of Bioinspired Algorithms and Applications

PSDM obtained models with a number of rules slightly below or around 10. The number of attributes
per rule is similar.
The results for the breast cancer dataset show that several PSO data miners are around 2% more
accurate than J48, the best standard data miner. The models found are again significantly simpler for the
PSDM algorithms.
The experiences with the promoters dataset show that the best PSDM algorithm presented an
accuracy of 3% above the most accurate nonPSO approach — J48. The other PSDM algorithms all
obtained similar accuracy. In terms of simplicity, the PSO-based approaches used, again, a significantly
inferior number of rules and attribute tests.

Experimental Results

Algorithm Particles Acc StdDev #Rules StdDev #Att/Rule StdDev

Promoters
J48 — 73.33 11.95 14.60 3.07 1.31 0.28
PRISM — 72.02 11.43 16.00 0.00 1.94 0.00
PSDM platform 100 76.30 8.60 6.60 0.49 0.87 0.05
PSDM radius 100 75.80 9.00 6.60 0.51 0.88 0.07
PSDM 2000 it. 100 73.60 8.82 6.64 0.50 0.88 0.06
SAPPO 100 74.80 8.70 6.61 0.51 0.92 0.08
Splice
J48 — 91.14 3.31 105.21 6.11 1.24 0.07
PRISM — 81.17 3.71 108.00 0.00 3.03 0.00
PSDM platform 100 76.26 5.71 6.67 1.70 0.88 0.08
PSDM radius 100 78.24 5.46 7.04 1.76 0.92 0.11
PSDM 2000 it. 100 77.56 5.59 6.67 1.62 0.90 0.09
SAPPO 100 83.06 4.62 6.92 1.91 1.17 0.11
Zoo
J48 — 92.63 7.36 10.98 2.65 1.67 0.31
PRISM — 62.43 5.04 40.00 0.00 1.00 0.00
PSDM platform 100 93.20 7.23 7.02 0.14 1.04 0.09
PSDM radius 100 93.90 7.09 7.00 0.00 1.05 0.10
PSDM 2000 it. 100 93.40 7.81 7.01 0.10 1.04 0.09
SAPPO 100 93.30 8.41 7.00 0.00 1.00 0.00
Wbc
J48 — 93.24 2.85 37.18 13.90 1.11 0.42
PRISM — 94.30 2.63 86.00 0.00 1.57 0.00
PSDM platform 100 92.22 5.49 9.98 0.75 0.91 0.03
PSDM radius 100 92.30 5.40 9.92 0.85 0.91 0.03
PSDM 2000 it. 100 92.26 5.46 10.00 0.74 0.91 0.03
SAPPO 100 92.44 4.80 9.87 0.97 0.96 0.04
Bc
J48 — 74.91 6.43 11.21 8.84 1.30 0.91
PRISM — 69.65 7.58 126.00 0.00 2.5 0.00
PSDM platform 100 77.26 8.31 6.51 0.89 1.35 0.18
PSDM radius 100 76.96 8.64 6.52 0.96 1.39 0.18
PSDM 2000 it. 100 77.03 8.55 6.46 0.93 1.38 0.17
SAPPO 100 77.40 8.56 7.00 1.01 1.48 0.15

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 486 — #18

Bio-Inspired Data Mining 27-487

For the final dataset, splice, the accuracy of J48 is almost 8% above the best PSO data miner, which
was the SAPPO-based approach. This difference is reflected in terms of simplicity, with the SAPPO models
requiring a mere 6.92 average rules, against an average 105.21 from J48. The next best PSDM approach
obtained a 5% worse accuracy than the SAPPO-based PSDM.
On a more global analysis, there are also some conclusions to be made. The most relevant one is
clearly the competitiveness of the PSO-based data miners — especially SAPPO — with the classical, more
established, approaches in terms of accuracy. Inclusively, in two of the datasets (three if we include zoo
where the results were very similar, but with all PSDM approach getting an accuracy around 1% above the
accuracy of J48), the best results were obtained by PSO-based approaches. Only in the splice dataset
was the best PSO miner clearly outperformed by a classical approach.
When discussing comprehensibility of the results, and assuming simpler models are more comprehen-
sible, results are extremely favorable to the PSO-based approaches. The PSO miners always present the
solutions with the least number of rules for each of the tested datasets and the number of attribute tests
per rule is also always, at least, slightly lower than the values obtained by the classic approaches. Probably,
it is this bias toward simplicity that is accountable for the worse results with the splice dataset, this
being an issue worthy of further investigation.
Between the PSO-based approaches, in terms of simplicity, all the algorithms performed in a similar way.
Regarding accuracy, SAPPO was probably the most robust PSDM, with results near the best PSDM for all
datasets, and significantly superior for the last, more complex, problem.

27.7 Conclusions
In this chapter, after briefly discussing other biologically inspired methods for data mining, more specifi-
cally for classification problems, we presented a first approach to classification based on particle swarm
algorithms. We tested several variants of our algorithm, including an adaptive one, which uses symbiotic
particles to adapt its parameters to the problem, and a predator–prey mechanism to maintain diversity in
the swarm.
The results obtained until now, for a limited number of datasets, lead us to believe that the PSO-based
approaches are clearly competitive with the classical algorithms, used for comparison, in terms of accuracy,
and clearly superior in terms of simplicity and, consequently, comprehensibility.
Results with the last dataset seem to indicate that an increase in number of instances can result in a loss
of accuracy of the PSDM approaches, probably caused by the identified bias toward simplicity, which can,
in these cases become prejudicial by not allowing more complex models to be found by the algorithm.
We plan to address this issue in future work.
Especially relevant for us is the fact that, for all but one of the datasets, the PSDM approaches outper-
formed the PRISM algorithm, which uses a similar covering strategy, but a greedy approach to the search
for new rules. The strengths of evolutionary search algorithms are clearly patent in these results.
If a choice was to be made between the PSDM algorithms, if temporal complexity was an issue,
the PSDM-radius and PSDM-platform variants are clearly competitive in terms of accuracy — with a
possible slight advantage for the platform version — and obviously take less time to find solutions than
the remaining approaches. If accuracy was the essential choice criterion, these first results seem to indicate
that the SAPPO approach can be more robust over a large set of application domains.

References
[1] U.M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy. Advances in Knowledge Discovery
and Data Mining. AAAI/MIT Press, Cambridge, MA, 1995.
[2] S. Weiss and N. Indurkhya. Predictive Data Mining: A Practical Guide. Morgan Kaufmann,
San Francisco, CA, 1998.
[3] I.H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Techniques with Java
Implementations. Morgan Kaufmann, San Francisco, CA, 1999.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 487 — #19

27-488 Handbook of Bioinspired Algorithms and Applications

[4] R.C. Eberhart and Y. Shi. Particle swarm optimization. In Proceedings of the International
Conference on Neural Networks and the Brain. Beijing, China, 1998, pp. PL5–PL13.
[5] J. Kennedy, R.C. Eberhart, and Y. Shi. Swarm Intelligence. Morgan Kaufmann, San Francisco, CA,
2001.
[6] A. Silva, A. Neves, and Ernesto Costa. An empirical comparison of particle swarm and predator prey
optimisation. In Proceedings of Artificial Intelligence and Cognitive Science: 13th Irish International
Conference, AICS 2002. Vol. 2464/2002 of Lecture Notes in Computer Science, September 12–13,
2002, Springer-Verlag, Limerick, Ireland.
[7] A. Silva, A. Neves, and E. Costa. SAPPO: A simple, adaptive, predator prey optimiser. In Proceedings
of the EPIA’03 — 11th Portuguese Conference on Artificial Intelligence, Workshop on Artificial Life
and Evolutionary Algorithms (ALEA), December 4–7, 2003, Beja, Portugal.
[8] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
[9] T. Back. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary
Programming, Genetic Algorithms. Oxford University Press, Oxford, 1996.
[10] J.R. Koza. Genetic Programming: On the Programming of Computers by Means of Natural Selection.
MIT Press, Cambridge, MA, 1992.
[11] A.A. Freitas. A survey of evolutionary algorithms for data mining and knowledge discovery.
Advances in Evolutionary Computation (A. Ghosh and S. Tsutsui, Eds.), Springer-Verlag, 2002.
[12] A. Giordana and F. Neri. Search-intensive concept induction. Evolutionary Computation, 3: 375–
419, 1995.
[13] D.P. Greene and S.F. Smith. Competition-based induction of decision models from examples.
Machine Learning, 13: 229–257, 1993.
[14] K.A. DeJong, W.M. Spears, and F.D. Gordon. Using genetic algorithms for concept learning.
Machine Learning, 13: 161–188, 1993.
[15] C.Z. Janikow. A knowledge-intensive genetic algorithm for supervised learning. Machine Learning,
13: 189–228, 1993.
[16] M.L. Wong and K.S. Leung. Data Mining Using Grammar Based Genetic Programming and
Applications. Kluwer, Dordrecht, 2000.
[17] J. Eggermont, A.E. Eiben, and J.I. van Hemert. A comparison of genetic programming variants for
data classification. Proceedings of the Intelligent Data Analysis (IDA-99), 1999.
[18] C.C. Bojarczuk, H.S. Lopes, and A.A. Freitas. Discovering comprehensible classification rules by
using genetic programming: A case study in a medical domain. GECCO, 953–958, 1999.
[19] M.D. Ryan and V.J. Rayward-Smith. The evolution of decision trees. In Proceedings of the Third
Annual Conference on Genetic Programming, Morgan Kaufmann, San Francisco, CA, 1998.
[20] E. Bonabeau, M. Dorigo, and G. Théraulaz. Swarm Intelligence: From Natural to Artificial Systems.
Oxford University Press, Oxford, 1999.
[21] M. Dorigo, V. Maniezzo, and A. Colorni. The ant system: optimization by a colony of cooperating
agents. IEEE Transactions on Systems, Man, and Cybernetics — Part B, 26: 29–41, 1996.
[22] R.S. Parpinelli, H.S. Lopes, and A.A. Freitas. An ant colony algorithm for classification rule
discovery. In Data Mining: A Heuristic Approach (H. Abbass, R. Sarker, and C. Newton, Eds.),
Idea Group Publishing, London, 2002, pp. 191–208.
[23] L.N. de Castro and J. Timmis. Artificial Immune Systems: A New Computational Intelligence
Approach. Springer-Verlag, 1998.
[24] D. Dasgupta, Z. Ji, and F. González. Artificial immune system (AIS) research in the last five years.
In Proceedings of the International Conference on Evolutionary Computation Conference (CEC),
Canbara, Australia, December 8–12, 2003.
[25] L.N. de Castro and Fernando José Von Zuben. aiNet: An artificial immune network for data
analysis. In Data Mining: A Heuristic Approach (Hussein A. Abbass, Ruhul A. Sarker, and Charles
S. Newton, Eds.), Idea Group Publishing, USA, March, 2001.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 488 — #20

Bio-Inspired Data Mining 27-489

[26] O. Nasraoui, F.A. González, C. Cardona, C. Rojas, and D. Dasgupta. A scalable artificial immune
system model for dynamic unsupervised learning. GECCO, 219–230, 2003.
[27] M. Clerc and J. Kennedy. The particle swarm-explosion, stability, and convergence in a
multidimensional complex space. IEEE Transactions on Evolutionary Computation, 6: 58–73, 2002.
[28] Y. Shi and R.C. Eberhart. Empirical study of particle swarm optimization. In Proceedings of the
IEEE Congress on Evolutionary Computation (CEC 1999), Piscataway, NJ, 1999, pp. 1945–1950.
[29] http://web.ics.purdue.edu/∼hux/bibliography.shtml
[30] J. Kennedy. Small worlds and mega-minds: Effects of neighborhood topology on particle
swarm performance. In Proceedings of IEEE Congress on Evolutionary Computation (CEC 1999),
Piscataway, NJ, 1999, pp. 1931–1938.
[31] T.M. Blackwell and P.J. Bentley. Don’t push me! collision-avoiding swarms. In Proceedings of the
IEEE Congress on Evolutionary Computation (CEC 2002), Honolulu, Hawaii, USA, 2002.
[32] T. Sousa, A. Neves, and A. Silva. Swarm optimisation as a new tool for data mining. In Proceedings
of the Parallel and Distributed Processing Symposium, 2003, pp. 144–149.
[33] T. Sousa, A. Silva, and A. Neves. A particle swarm data miner. In Proceedings of 11th Protuguese
Conference on Artificial Intelligence (EPIA 2003), Vol. 2902 of Lecture Notes in Computer Science,
Progress in Artificial Intelligence, Beja, Portugal, 2003, pp. 43–53.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C027” — 2005/8/17 — 18:17 — page 489 — #21

 28
A Hybrid Evolutionary
Algorithm for
Knowledge Discovery
in Microarray
Experiments

28.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-491

28.2 Microarray Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-492
28.3 Rule Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-493
Association Rules • A Constructive Method for Rule Mining •
Bioinspired Methods and Rule Mining • Interest of Rule
Mining for Microarray Analysis
28.4 A Hybrid Multiobjective Metaheuristic . . . . . . . . . . . . . . . . 28-495
A Multicriteria Model for Rule Mining • The Genetic
Algorithm • Multicriteria Mechanisms • Hybridization •
Adaptive Rate
L. Jourdan 28.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-501
M. Khabzaoui Data • Evaluation Protocol and Parameters • Results
C. Dhaenens 28.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-505
El-Ghazali Talbi References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-505

28.1 Introduction
Experiments with DNA microarray technology generate a large amount of data as they are able to measure
the expression levels of thousands of genes in a single experiment [1]. To explore these data it is necessary
to develop knowledge discovery techniques that can extract biological significance and use the data
to assign functions to genes. This is the goal of data mining (also known as Knowledge Discovery in
Databases [KDDs]) that has been defined as “The nontrivial extraction of implicit, previously unknown,
and potentially useful information from data” [2].
Several kinds of representations to express knowledge that can be extracted from microarray data can be
found in the literature [6]. Most current approaches in this area group genes, using clustering algorithms
such as hierarchical clustering [3] or Kohonen maps, that is, self-organizing maps [4]. Association rule

28-491

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 491 — #1

28-492 Handbook of Bioinspired Algorithms and Applications

mining is an advanced data mining technique that is useful in deriving meaningful rules from a given
dataset. Rule mining, as well as other datamining approaches, may be seen as a NP-hard combinatorial
problem that can be solved with combinatorial optimization methods such as evolutionary algorithms.
In this chapter, we propose a multicriteria hybrid metaheuristic for microarray data analysis. We first
introduce the biological context of the work. Then we set the rule mining context. After this positioning,
we present a multicriteria hybrid algorithm and describe every feature of the algorithm (representation,
operators, etc.). Finally, results on real datasets are presented and analyzed.

28.2 Microarray Analysis

Microarray technology (also called DNA microarrays, DNA arrays, DNA chips, gene chip, etc.) is now
widely used in many areas of biomedical research. It provides access to expression levels of thousands of
genes at once in order to identify coexpressed genes, relationships between genes, patterns of gene activity,
changes in gene activity under some medical treatments, etc. A microarray is typically a glass (or some
other material) slide, on to which probes are attached at fixed locations (spots). There may be tens of
thousands of spots on a single array. The technology involves hybridizing DNA of a reference sample and
a test sample, previously labeled with different fluorescent dyes, with probes that match a single gene.
Hence, measuring the fluorescence allows us to calculate the relative abundance of DNA of each gene and
evaluate their expression level. Arrays are scanned and images are produced and analyzed to obtain an
intensity value for each probe.
There are two variants of the DNA microarray technology: Synteni/Stanford chips and Affymetrix
chips. They differ in

• How DNA sequences are put down (spotting/photolithography)

• Length of DNA sequences (complete sequences or a series of fragments)

In the Synteni/Stanford chips, probes of cDNA (500/5000 bases long) are immobilized to a solid surface
such as glass using robot spotting and exposed to a set of targets either separately or in a mixture.
Affymetrix chips are powerful technology for resequencing DNA and polymorphisms detection. They use
photolabile agents and photolithography techniques.
Although fundamental differences exist between those two technologies, their strength lies in a massively
parallel analysis of thousands of genes and the generation of a lot of data.
Analyzing DNA microarray data requires a preprocessing phase [1,5,6]. Figure 28.1 recalls the different
steps required to produce new biological assumptions from microarray experiments. Major steps are
described as follows:

Relative gene expression: The differential gene expression is calculated by dividing the intensity of the
gene in the sample under study by its intensity level in the control. This intensity ratio has a highly
asymmetric distribution. To avoid this problem, a log 2 -transformation is usually used to make a
normal-like distribution:

Expression of a gene in the sample

Gene expression = log2 . (28.1)
Expression of the same gene in the control

Normalization: There are many sources of variations of measures in microarray experiments (variations
in cells or individuals, mRNA extraction, isolation, hybridization condition, optical measurement,
scanner noise, etc.). The purpose of normalization is to adjust (or correct) a signal in order to make
the comparison with other signals more meaningful. Many techniques for normalization aim to
make the data more normally distributed (log transformation per chip and per gene). This is an
important issue for data analysis.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 492 — #2

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-493

New hypothesis

Model Validation

DB
Analysis
Normalization
DB DB

Microarray Selection
DB Discretization
Gene filtering

Image analysis
Normalization

FIGURE 28.1 Process of hypothesis extraction from microarray.

Gene filtering and discretization: We are interested in finding genes that show significant differences
between two groups of patients. Hence, the filtering process may remove genes that do not differ-
entiate the sample under study from the test sample (their relative expression is not significant).
Most of the time, gene expression data is discretized into under/over expressed genes thanks to
cutoffs.

Hence after these different phases, data may be considered as large tables indicating, with discretized
values, the relative gene expression for thousands of genes under different experiments.

28.3 Rule Mining

In this chapter, we propose to use a rule mining approach to explore the data provided by microarray
experiments. We first introduce what rule mining is, and a classical rule mining approach. Then, we present
a state-of-the-art bioinspired methods for rule mining. Finally, we justify the use of rule mining, in analysis
of microarray experiments.

28.3.1 Association Rules

The search of association rules is an important task in knowledge discovery from databases [7–9].
Originally proposed for Market Basket data to study consumer-purchasing patterns, it has potential
applications in many areas. Given a database describing instances (e.g., transactions) according to differ-
ent attributes (e.g., items purchased), this problem involves discovering rules in the form: IF C THEN P
or in a more detailed expression, IF Cond1 AND Cond2 AND . . . AND Condm THEN P. The condi-
tion C is a conjunction of terms. For binary representation, a term indicates whether the corresponding
attribute has been chosen or not. In nominal data, where an attribute may take different values, a term
may be of the form attribute operator value. The prediction P is also represented by a conjunction
of terms.
The association rule search problem is a complex combinatorial problem as every attribute may be
candidate to participate, with one of its possible values, to the rule. Hence, the number of possible
combinations is exponential to the number of attributes. Moreover, this problem has been shown to
be NP-hard by Angiulli et al. [10]. The problem has been reduced into a MAX clique problem that is a
well-known NP-hard problem [11]. Therefore, algorithms dedicated to this problem have to be carefully
designed.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 493 — #3

28-494 Handbook of Bioinspired Algorithms and Applications

28.3.2 A Constructive Method for Rule Mining

A classical approach for finding association rules is to look for frequent itemsets (set of items — or
terms — that often occur simultaneously) as done by the Apriori [12] algorithm. This algorithm relies
on a fundamental property of frequent itemsets, called the Apriori property: “Every subset of a frequent
itemset is also a frequent itemset.” This monotony allows to incrementally construct interesting sets of
items in a very efficient way. Then items belonging to a same set are combined in order to obtain rules.
There exists numerous variants of Apriori [13–15] that allow to deal with problems of interesting sizes,
but each of them requires the monotony of the criterion used.

28.3.3 Bioinspired Methods and Rule Mining

As the number of rules that are candidate for extraction is exponential [16], the use of an evolutionary
algorithm such as a genetic algorithm is well adapted to explore the large search space of candidate rules.
Those algorithms have already shown their ability to solve large combinatorial optimization problems and
have already been applied with success to the rule discovery problem [17].
The paradigm of evolutionary algorithms consists of stochastic search algorithms that are based on an
abstraction of the process of Darwinian evolution [18]. An evolutionary algorithm maintains a population
of “individuals,” each of them representing a potential solution to a given problem. Each individual is
evaluated by a fitness function, which measures the quality of the solution (its adaptation to the problem).
Individuals evolve toward better and better individuals via a selection procedure, based on the fitness of
individuals, thanks to crossover (recombination) and mutation operators, and the replacement procedure.
In previous studies on association rule discovery with genetic algorithms, we can find two major
representations: the Michigan [19] representation and the Pittsburg [20] one. Each representation has
its drawbacks and advantages. The Pittsburg representation, where an individual is a set of rules, is
more expensive in memory but has been used in several genetic algorithms: GABIL [21], GIL [22], and
HDPDCS [23]. In the Michigan representation each individual of the population of the genetic algorithm
represents a candidate rule. The main drawback of the Michigan approach is that it could, without the use
of specific mechanisms, converge to a single rule and not to a set of rules. For example, in Reference 24
Weiss uses the sharing fitness to avoid the convergence to a single rule in order to predict rare events.
Classical genetic algorithms for rule mining using the Michigan representation are COGIN [25] and
REGAL [26].

28.3.4 Interest of Rule Mining for Microarray Analysis

Microarray data format is very similar to the Market Basket data format. Microarray data may be
represented in two different ways [27]:
Gene table: The genes constitute the rows of the data table whereas experiments to which genes were
exposed are the columns. The values represent the abundance of transcript for each spot on the
microarray. Clustering and classification may be applied to the gene table to divide the dataset into
clusters/classes by grouping the rows (genes).
Treatment table: The reverse way to present data is to flip the gene table. Thus, genes are now the
columns whereas treatments are the rows. The values are the gene expression level. The objective
is then to find associations between columns (genes). Therefore, association rule mining is a great
challenge. This data mining model is general and allows to find, for example, associations between
subsets of genes. Moreover, thanks to the condition and prediction notions, relations obtained are
precise and give more information than only grouping genes together.
There has been some recent works on using association rule mining to analyze gene expression data.
Chen et al. [28] apply association rules to mine the transcription factors essential to certain gene expres-
sions using the Apriori algorithm. They manage, with a small value of support to extract a small number
of rules they can analyze. Creighton and Hanash [29] propose to also use the Apriori algorithm to

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 494 — #4

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-495

reveal biological relevant associations between different genes or between environmental effects and gene
expression. They apply it with success to a yeast database. Kotala et al. [27] introduce a new approach
to mine association rules from microarray gene expression data using Peano count tree. Icev et al. [30]
focus on the combinatorial analysis of motifs involved in transcriptional control and introduce a notion
of association rules with distance information.
In our study we will consider data in the treatment table form (genes are the columns, treatments — or
comparison of individuals of different status — are the rows. Our goal is to look for rules combining genes
where a term can be in the form gene = value. The value belongs to the discretized gene expression
level. An example of a rule could be: IF (gene12 = over_expressed) AND (gene504 = under_expressed)
THEN (gene8734 = over_expressed).

28.4 A Hybrid Multiobjective Metaheuristic

The algorithm proposed has been developed thanks to the open source framework PARADISEO (PARAllel
and DIStributed Evolving Objects) [31] http://www.lifl.fr/∼cahon/logiciels.html that
allows, once you have designed your algorithm, to easily develop multicriteria evolutionary algorithms and
to use parallelism and distributed mechanisms. The originality of the proposed approach is to deal with
rule mining through a multiobjective point of view. Hence, we first recall some principles of multiobjective
combinatorial optimization problem, then we present the modeling of the rule mining problem as a
multiobjective problem. Finally, we present the algorithm and its hybridization with an exact enumeration.

28.4.1 A Multicriteria Model for Rule Mining

28.4.1.1 Multicriteria Definitions
We first describe and define Multi-Objective combinatorial Problems (MOPs) in a general case. Assume
that a solution to such a problem can be described by a decision vector (x1 , x2 , . . . , xn ) in the decision
space X . A fitness function f : X → Y evaluates the quality of each solution by assigning it an objective
vector (y1 , y2 , . . . , yp ) in the objective space Y (Figure 28.2) [32]. So, multiobjective optimization consists
in finding the solutions, in the decision space, optimizing (minimizing or maximizing) p objectives.
For the following definitions, we consider the minimization of p objectives. For maximization problems,
definitions are similar.
In the case of a single objective optimization, the comparison between two solutions x 1 and x 2 is
immediate. If y 1 < y 2 then x 1 is better than x 2 . For multiobjective optimization, comparing two solutions
x 1 and x 2 is more complex. Here there exists only a partial order relation, known as the Pareto dominance
concept illustrated in Figure 28.3.

Decision space Objective space

x2 y2

x1 y3 y1

(x1,x2, . . . ,xn) f (y1,y2, . . . ,yp)

FIGURE 28.2 Example of a MOP.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 495 — #5

28-496 Handbook of Bioinspired Algorithms and Applications

f 2 min

4
3

0 f 1 min

FIGURE 28.3 Example of dominance. In this example, points 1, 3, and 5 are nondominated. Point 2 is dominated
by point 3, and point 4 by points 3 and 5.

Definition 28.1. A solution x i dominates a solution x j if and only if:

∀k ∈ [1..p], fk (x i ) ≤ fk (x j ),
(28.2)
∃k ∈ [1..p]/fk (x i ) < fk (x j ).

Definition 28.1. A solution is Pareto optimal if it is not dominated by any other solution of the feasible set.

The set of optimal solutions in the decision space X is denoted as the Pareto set, and its image in the
objective space is the Pareto front. In MOP, we are looking for all the Pareto optimal solutions.
In the Pareto front two types of solutions may be distinguished: the supported solutions (that are on
the convex hull of the set of solutions and that may be found by a linear combination of criteria), and
nonsupported solutions [33]. These solutions are important, because for some problems only few Pareto
solutions are supported (the extremes) and to get a good compromise between the two criteria, it is
necessary to choose one of the nonsupported solution.

28.4.1.2 Multicriteria Association Rules

In order to solve association rule discovery problem as a combinatorial optimization problem, the optim-
ization criterion has to be defined. A lot of measures exist for estimating the quality of association rules.
For an overview, readers can refer to Freitas [34], Tan et al. [35], or Khabzaoui et al. [36]. In Reference 36,
we made a statistical study of different criteria found in the literature. This study lead us to determine five
groups where each group represents correlated criteria. We choose to select one criterion of each group
and obtain five complementary criteria that allow to evaluate rules in a complete way: Support, Jmeasure,
Interest, Surprise, and Confidence. These criteria are described below.
Rules are evaluated for a set of N instances, where |C| (respectively |P|) represents the number of
instances satisfying the C (respectively P) part of the rule and |C&P| the number of instances satisfying
simultaneously the C and the P parts of the rule.

Support (S): It is the classical measure of association rules. It enables to measure rule frequency in
the database. It is the percentage of transactions containing, both the C part and the P part,
in the database. It is used to find frequent itemsets in Apriori

|C&P|
S= . (28.3)
N

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 496 — #6

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-497

Confidence (Cf ): The Confidence measures the validity of a rule. It is the conditional probability of P
given C. It is used in Apriori to find interesting rules in frequent itemsets

|C&P|
Cf = . (28.4)
|C|

J-measure (Jm): Smyth and Goodman [37] have proposed the Jmeasure, which estimates the degree of
interest of a rule and combines support and confidence. It is used in optimization methods [38,39]
 
|P| |C&P| N × |C&P|
Jm = × log . (28.5)
N |P| |C| × |P|

Interest (I ): The Interest measures the dependency while privileging rare patterns in the region of weak
support

N × |C&P|
I= . (28.6)
|C| × |P|

Surprise (R): It is used to measure the affirmation. It enables the search for surprising rules

|C&P| − |C& P̄|

R= . (28.7)
|P̄|

Hence, the use of these five criteria allow an evaluation of a rule in a multicriteria manner. As they
are complementary, this model is interesting to select interesting rules and to reduce the numbers of
possible rules.

28.4.2 The Genetic Algorithm

We present the general scheme of the algorithm and the genetic operators used. The algorithm presented
here is a multicriteria version of ASGARD presented in References 17 and 40.
28.4.2.1 General Scheme
Figure 28.4 presents the scheme of the genetic algorithm with its multicriteria aspects. The algorithm
starts with a set of randomly generated solutions. Then, solutions are selected according to their fitness
quality to create new solutions (offsprings). Best solutions encountered over generations are archived into
a secondary population called the “Pareto archive.”
28.4.2.2 Genetic Operators
Genetic operators allow diversification and intensification of the search:
Crossover: The proposed crossover operator has two versions:
• Exchange crossover: If two individuals X and Y have one or several common attribute(s) in
the C parts, one common attribute is randomly selected. The value of the selected attribute
in X is exchanged with its counterpart in Y (see Figure 28.5).
• Insert crossover: Conversely, if X and Y have no common attribute, one term is randomly
selected in the C part of X and inserted in Y with a probability inversely proportional
to the length of Y . The similar operation is performed to insert one term of Y in X
(see Figure 28.6).
Mutation: There are four mutation operators designed. The mutation operator called “Value mutation”
replaces the value of an attribute by a randomly chosen one (see Figure 28.7). The second one
called “Attribute mutation” replaces an attribute by another; the value of this attribute is randomly

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 497 — #7

28-498 Handbook of Bioinspired Algorithms and Applications

1–Pm
Pm1
Value
Pm
mutation
Offsprings Mutation
Attribute
mutation 1–Pm1

Elitist Roulette wheel selection Crossover

insertion with Pareto ranking

Update Pareto archive

FIGURE 28.4 General scheme of the genetic algorithm.

Parent 1 (X ) Att 1 = Val2 AND Att 3 = Val4 THEN Att 7 = Val3

Parent 2 (Y ) Att 1 == Val1 THEN Att 5 = Val2

Offspring 1 Att 1 = Val1 AND Att 3 = Val4 THEN Att 7 = Val3

Offspring 2 Att 1 = Val2 THEN Att 5 = Val2

FIGURE 28.5 Exchange crossover.

chosen (see Figure 28.8). The third one is a reduction mutation that randomly removes one term
of the rule. The last one is an augmentation mutation that randomly adds a term to the rule.

28.4.3 Multicriteria Mechanisms

In order to deal with multicriteria optimization problems, different mechanisms have to be used.
For example, the notion of dominance has to be defined (to be able to compare solutions) and population
management has to be carefully studied.
Selection operator: The classical roulette selection based on the ranking notion has been used. The
probability of selection of a solution is proportional to its rank. We use two ranking methods:
1. Pareto ranking : The rank of a solution corresponds to the number of solutions, in the current
population, by which it is dominated (see Figure 28.9 for a minimization problem) [4].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 498 — #8

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-499

Parent 1 (X ) Att 1 = Val2 AND Att 3 = Val4 THEN Att 7 = Val3

Parent 2 (Y ) Att 4 = Val3 THEN Att 5 = Val2

Offspring 1 Att 1 = Val1 AND Att 3 = Val4 AND Att 4 = Val3 THEN Att 7 = Val3

Offspring 2 Att 1 = Val2 AND Att 4 = Val3 THEN Att 5 = Val2

FIGURE 28.6 Insert crossover.

Individual Att 1 = Val2 AND Att 3 = Val4 THEN Att 7 = Val3

New individual Att 1 = Val10 AND Att 5 = Val2 THEN Att 7 = Val3

FIGURE 28.7 Value mutation.

Individual Att 1 = Val2 AND Att 3 = Val4 THEN Att 7 = Val3

New individual Att 4 = Val10 AND Att 5 = Val2 THEN Att 7 = Val3

FIGURE 28.8 Attribute mutation.

2. Nondominated sorting GA (NSGA): This method assigns ranks to solutions by first finding
the set of nondominated solutions in the current population. Those solutions are removed
from the population and assigned rank 1. As these solutions are removed, a new so-called
front of nondominated solutions is now present in the remainder of the original population.
This second front is extracted and assigned rank 2. This procedure is repeated till there is no
solution present in the population (see Figure 28.10 for a minimization problem) [42].
Experiments have shown that with the proposed algorithm, both ranking methods give interesting
results with a small superiority for the Pareto ranking. Therefore we will use this Pareto ranking
for the selection process.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 499 — #9

28-500 Handbook of Bioinspired Algorithms and Applications

0
* 5
¤
+ 2

Criterion1
0
* 0 1
* +
0
*
0
*
Criterion2

FIGURE 28.9 Pareto ranking.

1
* ¤
3
2
1 +
Criterion1

3
* ¤
1 2
* +
1
* 1 + 2
*
Criterion2

FIGURE 28.10 NSGA ranking.

Replacement operator: We use the elitist nondominated sorting replacement. The worst ranked solutions
are replaced by the dominating solutions (if there exists any) generated by mutation and crossover
operators (offsprings). The size of the population remains unchanged.
Archive: Nondominated association rules are archived into a secondary population called the “Pareto
Archive” in order to keep track of them. It consists of archiving all the Pareto association rules
encountered over generations. When a new Pareto solution is added to the archive, an update has
to be done (some solutions may become dominated).
Elitism: The Pareto solutions (best solutions) are not only stored permanently, they also take part in
the selection and may participate to the reproduction.

28.4.4 Hybridization
In order to increase the robustness of the approach, we hybridize it with an exact enumeration procedure.
As the search space is large, this enumeration is realized for a small subspace defined thanks to solutions of
the population. Hence, we design an operator which makes an exhaustive search of all the possible rules
generated with attributes selected in two rules.
This operator may be seen as a quadratic crossover operator. It takes as input two individuals, each
coding a rule. It examines all the possible itemsets that can be derived from the items composing the
rules, while taking into account the different possible values of the attributes. All the possible rules that
can be constructed from the generated itemsets are evaluated and introduced if necessary in a local Pareto
archive. Finally, only the global Pareto solutions of the local Pareto archive are introduced in the global
Pareto archive with the replacement operator. Two offsprings are candidates to take part in the population
(Figure 28.11).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 500 — #10

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-501

Individual 1 Individual 2

Enumeration

Offspring 1 Local Global

Offspring 2 Pareto archive Pareto archive

FIGURE 28.11 The enumeration operator.

28.4.5 Adaptive Rate

Probabilities of mutation are hard to set when several mutation operators compete and are often set
experimentally. To overcome this problem, we implement an adaptive strategy for calculating the rate of
application of each mutation operator. Many authors have worked on setting automatically probabilities of
applying operator [43–45]. Hong et al. [46] proposed to compute the new rate of mutation by calculating
the progress of the jth application of mutation Mi , for an individual ind mutated into an individual mut
as follows:

progressj (Mi ) = Max(fitness(ind), fitness(mut)) − fitness(ind). (28.8)

Then for each mutation operator Mi , assume Nb_mut (Mi ) applications of the mutation are done
during a given generation ( j = 1, . . . , Nbmut (Mi )). Then we can compute the profit of a mutation Mk :

j progressj (Mk )/Nb_mut(Mk )
Profit(Mk ) =    . (28.9)
i j progressj (Mi )/Nb_mut(Mi )

We set a minimum rate δ and a global mutation rate pmutation for N mutation operators. The new mutation
ratio for each Mi is calculated using the following formula [46]:

p(Mi ) = Profit(Mi ) × (pmutation − N × δ) + δ. (28.10)

The sum of all the mutation rates is equal to the global rate of mutation pmutation . The initial rate of
application of each mutation operator is set to pmutation /N .

28.5 Experiments
28.5.1 Data
In order to evaluate the algorithm, we evaluate it on two microarray databases:
• A confidential microarray data containing 22,376 human genes for 45 Affymetrix chips (DB1).
• A public database, the “MIPS yeast genome database” containing 2467 genes for 79 chips (YeastDB).
Genes expressions have been discretized and may take five values: Increase (I), Marginal Increase (MI),
when the gene is over-expressed, Decrease (D) and Marginal Decrease (MD) when it is under-expressed
and No Change (NC), when the difference of expression is not significant.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 501 — #11

28-502 Handbook of Bioinspired Algorithms and Applications

For an initial study, a set of 514 genes (numbered from 1 to 514) that show an interesting differential
expression over the set of experiments (filtered on the number of No Changes), have been selected for
DB1. For the yeastDB all the genes (2467) have been considered.

28.5.1.1 Missing Value

Missing values lead to problems in the data analysis, so that they influence the computation of statistical
tests and quality criteria of association rules problems. There may have numerous missing values in
microarray data due to the empty spots or because the background intensity is higher than the spot
intensity. Two methods are commonly used for treatment of missing values: they may be replaced by
estimated values (Median for example), or corresponding instances are deleted. In an association rule
problem we propose that all genes that have missing values are kept without modification but when these
missing values are on the attributes of a rule, these genes are excluded from the computation of the quality
of this rule. So |N | may be different for each rule.

28.5.2 Evaluation Protocol and Parameters

28.5.2.1 Configurations and Parameters
Several propositions have been done in this chapter in order to improve the standard genetic algorithm:
elitism, adaptive strategy for mutation rates, hybridization. In order to evaluate the contribution of each
of these mechanisms, several configurations of the genetic algorithm have been compared:
• Conf A: Pareto ranking
• Conf B: Pareto ranking + Adaptive strategy
• Conf C: Pareto ranking + Elitism
• Conf D: Pareto ranking + Adaptive strategy + Elitism
• Conf E: Conf D + Hybridization
• Conf E  : Conf D + Hybridization 1/10
In the Conf E configuration, the crossover operator has been replaced by the enumeration operator
presented for the hybridization. Using this operator for every crossover is very time consuming, hence
we compared with Conf E  algorithm where the hybridization operator replaces the classical crossover by
only one iteration over ten.
For all of these configurations, the same parameters have been used:
• Population size: around NbAttributes /2, which gives 250 for DB1, 1200 for YeastDB
• Selection in Pareto archive: 0.5
• Global mutation rate: 0.5
• Crossover rate: 0.8
• Number of generations: 500

28.5.2.2 Evaluation Measure Used

In multicriteria optimization, solutions quality can be assessed in different ways. Some approaches com-
pare the obtained front with the optimal Pareto front [47]. Other approaches evaluate a front with a
reference point [48]. Some performance measures do not use any reference point or front to evaluate an
algorithm [49,50], especially when the optimal Pareto front is not known at all.
Here, we use the contribution metric [51,52] to evaluate the proportion of Pareto solutions given by
each front.
The contribution of a set of solutions PO1 relatively to a set of solutions PO2 is the ratio of nondominated
solutions produced by PO1 in PO ∗ , where PO ∗ is the set of Pareto solutions of PO1 ∪ PO2 .
• Let PO be the set of solutions in PO1 ∩ PO2 .
• Let W1 (respectively W2 ) be the set of solutions in PO1 (respectively PO2 ) that dominate some
solutions of PO2 (respectively PO1 ).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 502 — #12

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-503

• Let L1 (respectively L2 ) be the set of solutions in PO1 (respectively PO2 ) that are dominated by
some solutions of PO2 (respectively PO1 ).
• Let N1 (respectively N2 ) be the other solutions of PO1 (respectively PO2 ): Ni = POi \
(PO ∪ Wi ∪ Li ).

(PO/2) + W1  + N1 

Cont (PO1 /PO2 ) = . (28.11)
PO ∗ 

Let us remark that PO ∗  = PO + W1  + N1  + W2  + N2  and Cont (PO1 /PO2 ) +
Cont (PO2 /PO1 ) = 1 (with Cont (PO1 /PO2 ) ∈ [0, 1]). Hence a contribution greater than 0.5 indicates
that the Pareto front has been improved.
For example, we evaluate the contribution of the two sets of solution PO1 and PO2 in Figure 28.12:
solutions of PO1 (respectively PO2 ) are represented by circles (respectively crosses). We obtain
Cont (PO1 , PO2 ) = 0.7 and Cont (PO2 , PO1 ) = 0.3.

28.5.3 Results
Genetic algorithms are stochastic methods. Hence to evaluate the proposed approach, we have executed
10 runs for each configuration. Results are given with different indicators: mean of the value, minimum,
maximum, and its standard deviation. This allows a more efficient comparison.

28.5.3.1 Evaluation of the Mechanisms

In order to evaluate the different mechanisms, comparisons of the final Pareto fronts obtained with
the different configurations described before (Conf A, B, C, D, E, E  ) are reported in Table 28.1. This
table indicates the average of contribution as well as the minimum contribution, the maximum, and the
standard deviation.

C=4
Y2 W1 = 4
W2 = 0
N1 = 1
N2 = 1

Y1

FIGURE 28.12 Example of contribution.

TABLE 28.1 Quality Comparison of the Different

Configurations (Contribution Metric)

Contribution Average Minimum Maximum σ

Conf B/Conf A 0.55 0.39 0.78 0.09
Conf C/Conf A 0.62 0.43 0.90 0.09
Conf D/Conf B 0.72 0.52 0.86 0.07
Conf D/Conf C 0.61 0.43 0.76 0.08
Conf E/Conf D 0.74 0.59 0.87 0.06
Conf E  /Conf D 0.60 0.45 0.81 0.09
Conf E/Conf E  0.58 0.41 0.74 0.09

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 503 — #13

28-504 Handbook of Bioinspired Algorithms and Applications

28.5.3.1.1 Elitism
The gain of elitism may be evaluated by comparing similar configurations with and without elitism (here,
C/A and D/B). Table 28.1 indicates that on an average, the Pareto fronts obtained using elitism are of
better quality than those obtained without elitism (Cont (C/A) = 0.62 and Cont (D/B) = 0.72). Moreover
this contribution may reach 0.9 for best improvements. This shows the interest of this mechanism.

28.5.3.1.2 Adaptive Strategy

In a similar way, the gain of using the adaptive strategy may be evaluated by comparing similar configur-
ations with and without it (here, B/A and D/C). Table 28.1 shows that using the adaptive strategy also
improves, but in a smaller way, the results obtained (Cont (B/A) = 0.55 and Cont (D/C) = 0.61). Let us
remark that the improvement is greater for more complete configurations (Pareto ranking + elitism with
and without adaptive strategy).

28.5.3.1.3 Hybridization
Two questions may arise concerning the use of the hybridization. Is it interesting to use such an operator
and is it worth using it at each iteration. Hence we can compare configurations E/D and E/E  . Table 28.1
shows that using the hybridization allows to improve the more complete configuration (Pareto ranking +
elitism + adaptive strategy). In this case, the minimum contribution encountered is equal to 0.59, which
means that the hybridization always improves the Pareto front. When applying the operator only one
generation over ten (Conf E  ), the results are still interesting (Cont (E  /D) = 0.60), but Pareto front
obtained are in general dominated by those produced by Conf E (Cont (E/E  ) = 0.58), where the operator
is applied at each generation. However, the drawback of using Conf E is the large amount of time required
and the compromise between the quality of the solution and the computing time allowed has to be
considered.

28.5.3.2 Example of Rules Obtained

Table 28.2 describes some rules of the Pareto front obtained thanks to Conf E for the YeastDB, whereas
Table 28.3 presents values of these rules for the five criteria considered (Support, Interest, Surprise,
Confidence, and Jmeasure). Interest of such rules is difficult to estimate from a biological point of view;
however one idea would be to compare with relations already known between genes belonging to a
same rule.
Table 28.2 shows that there exists genes that appear in several rules, and we may suppose that those
genes are relevant for the problem under study. Those genes should be the first the biologists have to study.
Moreover, Table 28.3 shows the interest of using the proposed multicriteria model as good solutions
for one criteria which are not always interesting for another one.

TABLE 28.2 Description of Some Pareto Solutions Obtained with Conf E (YeastDB)

Rules Description

R1 IF ((VPS35 = D) and (PRB1 = I)) then (VPH1 = D)

R2 IF ((DAK2 = I) and (SVS1 = D) and (KTR7 = D) and (RPS27A = D)
and (PRO2 = D)) then (ARE2 = I)
R3 IF ((POP3 = D) and (TAF67 = I) and (VPS35 = D) and (SRP14 = D))
then (MF(ALPHA)2 = I)
R4 IF ((RPS6A = D) and (RPO26 = D) and (SVS1 = D) and (KTR7 = D))
then (DAK2 = I)
R5 IF ((RNR4 = I) and (IPT1 = I) and (GPM1 = I) and (YTA12 = I)
and (PUP3 = I) and (SAP185 = D) and (ESC2 = D)) then (GAR1 = I)
R6 IF ((PUS1 = D)) then (POP3 = D)

Rules indicate the level of expression of the genes: I, Increase; D, Decrease.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 504 — #14

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-505

TABLE 28.3 Quality of Pareto Solutions Presented in Table 28.2

Rules S Cf I R Jm

R1 0.582 0.978 1.288 2.368 0.147

R2 0.177 0.875 3.638 0.200 0.228
R3 0.278 0.916 2.263 0.425 0.227
R4 0.253 0.909 2.762 0.339 0.257
R5 0.063 1.000 7.900 0.072 0.130
R6 0.835 0.929 1.005 10.16 0.004

S, Support; I , Interest; R, Surprise; Cf , Confidence; Jm, Jmeasure.

28.6 Conclusion
In this chapter we have presented a multiobjective genetic algorithm for rule mining problems. There-
fore, a multicriteria model has been proposed for association rules mining. We have proposed a genetic
algorithm which helps look for the Pareto solutions regarding the five selected criteria. We have presented
its application to analyze microarray experiment data. Through the experiments, the advanced mech-
anisms proposed have been validated, and in particular the hybridization with an exact enumerative
procedure.
In order to improve the use of such an algorithm and to speed up executions that may take several
hours, we now work on a parallel implementation of the method. The parallelism should allow for results
in a more reasonable time, but should also allow to an execution of more intense searches with the
hybridization operator. Hence, we will be able to give biologists different hypothesis for their evaluation.

References
[1] D.P. Berrar, W. Dubitzky, and M. Granzow, Eds. A Practical Approach to Microarray Data Analysis.
Kluwer Academic Publishers, New York, 2003.
[2] U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth. From data mining to knowledge discovery: An
overview. Advances in Knowledge Discovery. MIT Press, Cambridge, MA, 1996, pp. 1–34.
[3] P.T. Spellman, G. Sherlock, M.Q. Zhang, V.R. Iyer, K. Anders, M.B. Eisen, P.O. Brown, D. Botstein,
and B. Futcher. Comprehensive identification of cell cycle regulated genes of yeast Saccharomyces
cervisiae by microarray hybridization. Molecular Biology of the Cell, 9: 3273–3297, 1998.
[4] P. Tamayo, D. Slonim, J. Mesirov, Q. Zhu, S. Kitareewan, E. Dmitrovsky, E. Lander, and T. Golub.
Interpreting patterns of gene expression with self-organizing maps: Methods and application
to hematopoietic differentiation. Proceedings of National Academy of Sciences, 96: 2907–2912,
1999.
[5] B. Phimister. The chipping forecast. Nature Genetics, 21(Suppl.): 1–60, 1990.
[6] Collective works. The human genome project. Nature, 409: 813–959, 2001.
[7] R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of items in large
databases. In P. Buneman and S. Jajodia, Eds., Proeceedings of the 1993 ACM SIGMOD International
Conference on Management of Data. ACM Press, Washington, DC, May 1993, pp. 207–216.
[8] S. Morishita and A. Nakaya. Parallel branch-and-bound graph search for correlated association
rules. In Large-Scale Parallel Data Mining, 1999, pp. 127–144.
[9] T. Scheffer. Finding association rules that trade support optimally against confidence. In Principles
of Data Mining and Knowledge Discovery, 2001, pp. 424–435.
[10] F. Angiulli, G. Ianni, and L. Palopoli. On the complexity of mining association rules. In Atti del
Nono Convegno su Sistemi Evoluti per Basi di Dati (SEBD), Venise, Italie, 2001, p. 8.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 505 — #15

28-506 Handbook of Bioinspired Algorithms and Applications

[11] M.R. Garey and D.S. Johnson. Computers and Intractability. The Guide to the Theory of
NP-Completeness. W.H. Freeman and Company, San Francisco, CA, 1979.
[12] R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In J.B. Bocca, M. Jarke,
and C. Zaniolo, Eds., Proceedings of the 20th International Conference on Very Large Data Bases,
VLDB, Morgan Kaufmann, San Francisco, CA, 12–15 1994, pp. 487–499.
[13] S. Brin, R. Motwani, and C. Silverstein. Beyond market baskets: Generalizing association rules to
correlations. In ACM SIGMAD, 1997, pp. 265–276.
[14] A. Savasere, E. Omiecinski, and S.B. Navathe. An efficient algorithm for mining association rules
in large databases. In The VLDB, Zurich, Switzerland, 1995, pp. 432–444.
[15] H. Toivonen. Sampling large databases for association rules. In T.M. Vijayaraman, A.P. Buchmann,
C. Mohan, and Nandlal L. Sarda, Eds., Proceedings of the 1996 International Conference on Very
Large Data Bases. Morgan Kaufmann, San Francisco, CA, 1996, pp. 134–145.
[16] M.J. Zaki. Parallel and distributed association mining: A survey. IEEE Concurrency, 7:14–25, 1999.
[17] L. Jourdan, C. Dhaenens, and E.-G. Talbi. Rules extraction in linkage disequilibrium mapping
with an adaptive genetic algorithm. In European Conference on Computational Biology (ECCB)
2003, Paris, France, 2003, pp. 29–32.
[18] C. Darwin. On the Origin of Species. John Murray, London, 1859.
[19] J. Holland. Adaptation in Natural and Artificial Systems. University of Michigan Press, Ann Arbor,
MI, 1975.
[20] S. Smith. Flexible learning of problem solving heuristics through adaptive search. In Proceedings
of the Eighth International Joint Conference on Artificial Intelligence, Karlsruhe, Germany, 1983,
pp. 422–425.
[21] K.A. De Jong, W.M. Spears, and D.F. Gordon. Using genetic algorithms for concept learning.
Machine Learning, 13: 161–188, 1993.
[22] C.Z. Janikow. A knowledge-intensive genetic algorithm for supervised learning. Machine Learning,
13: 189–228, 1993. J.J. Grefenstette, Ed., Kluwer Academic Publishers, Massachusetts.
[23] M. Pei, E.D. Goodman, and W.F. Punch III. Pattern discovery from data using genetic algorithms.
In Proceedings of the First Pacic-Asia Conference on Knowledge Discovery and Data Mining, February
1997. Available via www URL: http://garage.cps.msu.edu/papers/papers-index.html.
[24] G.M. Weiss. Timeweaver: A genetic algorithm for identifying predictive patterns in sequences
of events. In W. Banzhaf, J. Daida, A.E. Eliben, M.H. Garzon, V. Honavar, M. Jakiela, and
R.E. Smith, Eds., Proceedings of the Genetic and Evolutionary Computation Conference, Vol. 1.
Morgan Kaufmann, Orlando, Florida, USA, San Francisco, CA, 1999, pp. 718–725.
[25] D.P. Greene and S.F. Smith. Competition-based induction of decision models from examples.
Machine Learning, 13: 229–257, 1993. J.J. Grefenstette, Ed., Kluwer Academic Publishers,
Massachusetts.
[26] A. Giordana and F. Neri. Search-intensive concept induction. Evolutionary Computation, 3:
375–419, 1995.
[27] P. Kotala, P. Zhou, S. Mudivarthy, W. Perrizo, and E. Deckard. Gene expression profiling of
dna microarray data using peano count trees (p-trees). In Online Proceedings on the First Virtual
Conference on Genomics and Bioinformatics, 2001.
[28] R. Chen, Q. Jiang, H. Yuan, and L. Gruenswald. Mining association rules in analysis of transcription
factors essential to gene expressions. In Atlantic Symposium on Computational Biology and Genome
Information Systems and Technology, 2001.
[29] C. Creighton and S. Hanash. Mining gene expression databases for association rules.
Bioinformatics, 19: 79–86, 2003.
[30] A. Icev, C. Ruiz, and E.F. Ryder. Distance-enhanced association rules for gene expres-
sion. In ACM SIGKDD Workshop on Data Mining in Bioinformatics (BIOKDD03), 2003,
pp. 34–40.
[31] S. Cahon, N. Melab, E.-G. Talbi, and M. Schoenauer. Paradiseo based design of parallel and
distributed evolutionary algorithms. In Evolutionary Algorithms EA’03, 2003, pp. 195–207.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 506 — #16

Hybrid Evolutionary Algorithm for Knowledge Discovery 28-507

[32] E. Zitzler, M. Laumanns, and S. Bleuler. A tutorial on evolutionary multiobjective optimization.

In Metaheuristics for Multiobjective Optimisation, Lecture Notes in Economics and Mathematical
Systems, Vol. 535. Springer-Verlag, Berlin, 2004, pp. 3–37.
[33] E.L. Ulungu, J. Teghem, and Ph. Fortemps. Heuristics for multi-objective combinatorial
optimization by simulated annealing. In J. Gu, G. Chen, Q. Wei, and S. Wang, Eds., Proceedings of
the Sixth National Conference on Multiple Criteria Decision Making. Sci-Tech, 1995, pp. 228–238.
[34] A.A. Freitas. On rule interestingness measures. Knowledge-Based Systems Jounral, 12: 309–315,
1999.
[35] P.-N. Tan, V. Kumar, and J. Srivastava. Selecting the right interestingness measure for association
patterns. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Dis-
covery and Data Mining, 2002. http://www.cse.unsw.edu.au/∼qzhang/proceedings/research.html.
[36] M. Khabzaoui, C. Dhaenens, A. N’Guessan, and E.-G. Talbi. Etude exploratoire des critères de
qualité des règles d’association en datamining. In XXXVèmes JOURNEES DE STATISTIQUE, 2003,
pp. 583–586.
[37] P. Smyth and R.M. Goodman. Rule induction using information theory. In G. Piatetsky-Shapiro
and J. Frawley, Eds., Knowledge Discovery in Databases, MIT Press, Cambridge, MA, 1991,
pp. 159–176.
[38] D.L.A. Araujo, H.S. Lopes, and A.A. Freitas. A parallel genetic algorithm for rule discovery in large
databases. In K. Ilto, Ed., Proceedings of the 1000 IEEE Systems, Man and Cybernetics Conference,
Vol. III. IEEE, Tokyo, October 1999, pp. 940–945.
[39] K. Wang, S.H.W. Tay, and B. Liu. Interestingness-based interval merger for numeric association
rules. In R. Agrawal, P.E. Stolorz, and G. Piatetsky-Shapiro, Eds., Proceedings of the Fourth Inter-
national Conference on Knowledge Discovery and Data Mining, KDD. AAAI Press, New York, USA,
27–31 1998, pp. 121–128.
[40] L. Jourdan. Métaheuristiques pour l’extraction de connaissances: Application à la génomique.
Ph.D. thesis. Université des Sciences et Technologies de Lille, November 2003.
[41] C.M. Fonseca and P.J. Fleming. An overview of evolutionary algorithms in multiobjective
optimization. Evolutionary Computation, 3: 1:1–16, 1995.
[42] K. Deb. Introduction to selection. Evolutionary Computation 1: Basic Algorithms and Operators,
Institute of Physic, 2000.
[43] L. Davis. Adapting operator probabilities in genetic algorithms. In J.D. Schaffer, Ed., Proceedings
of the Third International Conference on Genetic Algorithms. Morgan Kaufmann, San Mateo, CA,
1989, pp. 61–69.
[44] F. Herrera and M. Lozano. Adaptation of genetic algorithm parameters based on fuzzy logic
controllers. In F. Herrera and J.L. Verdegay, Eds., Genetic Algorithms and Soft Computing. Physica-
Verlag, Heidelberg, 1996, pp. 95–125.
[45] B.A. Julstrom. What have you done for me lately? Adapting operator probabilities in a steady-
state genetic algorithm. In L.J. Eshelman, Ed., Proceedings of the Sixth International Conference on
Genetic Algorithms. Morgan Kaufmann, San Francisco, CA, 1995, pp. 81–87.
[46] T.P. Hong, H.S. Wang, and W.C. Chen. Simultaneously applying multiple mutation operators in
genetic algorithms. Journal of Heuristics, 6: 439–455, 2000.
[47] D.A. Van Veldhuizen and G.B. Lamont. On measuring multiobjective evolutionary algorithm
performance. In 2000 Congress on Evolutionary Computation, Vol. 1. Piscataway, NJ, July 2000,
pp. 204–211.
[48] A. Jaszkiewicz. On the Performance of Multiple Objective Genetic Local Search on the 0/1 Knapsack
Problem. A Comparative Experiment. Technical report RA-002/2000, Institute of Computing
Science, Poznan University of Technology, Poznan, Poland, July 2000.
[49] J.D. Knowles, D.W. Corne, and M.J. Oates. On the assessment of multiobjective approaches
to the adaptive distributed database management problem. In Proceedings of the Sixth Inter-
national Conference on Parallel Problem Solving from Nature (PPSN VI), September 2000,
pp. 869–878,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 507 — #17

28-508 Handbook of Bioinspired Algorithms and Applications

[50] E. Zitzler, L. Thiele, M. Laumanns, M.C. Fonseca, and V.G. da Fonseca. Performance Assess in IEEE
Transactions on Evolutionary Computation, 7(2): 117–132, 2003.
[51] M. Basseur, F. Seynhaeve, and E.-G. Talbi. Design of multi-objective evolutionary algorithms:
Application to the flow-shop scheduling problem. In Congress on Evolutionary Computation
(CEC’02), Honolulu, USA, 2002, pp. 1151–1156.
[52] H. Meunier, E.G. Talbi, and P. Reininger. A multiobjective genetic algorithm for radio
network optimisation. In CEC, Vol. 1, IEEE Service Center, Piscataway, NJ, July 2000, pp. 317–324.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C028” — 2005/8/6 — 14:12 — page 508 — #18

 29
An Evolutionary
Approach to Problems
in Electrical
Engineering Design

29.1 Problems in Engineering Design. . . . . . . . . . . . . . . . . . . . . . . . 29-510

Universal Motor Design • Design of Integrated Circuits
29.2 Evolutionary Optimization in Problems of
Engineering Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-514
The Genetic Algorithm in UM and IC Design
29.3 Procedures and Evaluations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-518
Gregor Papa UM Design Procedure • IC Design Procedure
Jurij Šilc 29.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-528
Barbara Koroušić-Seljak References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-528

[Genetic algorithms] have been shown to rapidly converge to near-optimal solutions in a wide variety
of application domains. Further, they have been shown to be computationally efficient, and to be
well suited for solving problems characterized by local minima.

Karr, Yakushin, and Nicolosi,

Engineering Applications of Artificial Intelligence, 2000.

This chapter presents two engineering design problems, both of which were solved with biology-inspired
algorithms. The evolutionary approach is used in universal electromotor geometry optimization (UM
design) and integrated circuits area/time optimization (IC design).
In the UM design we improve the efficiency of a universal motor; here the goal is to find a new set of
independent geometrical parameters for the rotor and the stator with the aim of reducing the motor’s
power losses, which occur in the iron and the copper. In the IC design we improve some parts of the
high-level synthesis process of integrated circuits by considering the concurrency of operation scheduling
and resource allocation constraints to ensure a globally optimal solution in a reasonable time.

29-509

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 509 — #1

29-510 Handbook of Bioinspired Algorithms and Applications

FIGURE 29.1 A UM used in a vacuum cleaner, showing its rotor and stator parts.

29.1 Problems in Engineering Design

29.1.1 Universal Motor Design
Home appliances, such as vacuum cleaners and mixers, as well as power tools, such as drills and saws, are
generally powered by a UM [1] (see Figure 29.1). This type of motor has many advantages that make the
UM such a popular choice for home appliances and power tools: a large output power in relation to its
small size, high starting and running torque, variable speed that can be regulated in a simple way, and low
manufacturing costs.
Home appliances and power tools need as low an energy consumption, that is, input power, as possible,
while still satisfying the needs of the user by providing sufficient output power. The ratio of the output
power to the input power defines the efficiency of the motor, which can be improved by reducing some
of the main power losses in the motor, that is, those that originate in the iron and the copper. This can be
done by optimizing the geometry of both the rotor and the stator. Because of the high magnetic saturation
of the iron in a UM the problem is a highly nonlinear one.
29.1.1.1 The Geometry of the Rotor and the Stator
The rotor-and-stator unit of a UM is constructed by stacking the rotor/stator iron laminations (see
Figure 29.2). The shape and the profile of the rotor/stator lamination are described by several two-
dimensional geometrical parameters. There are two types of parameter: the invariable and the variable.
Invariable parameters are fixed; they cannot be altered, either for technical reasons or because of the
physical constraints of the motor. Some of the more important invariable parameters of a UM are:
the air gap, the external radius of the stator, the radius of the rotor’s shaft, the radius on the stator’s side
hole, and the width of the rotor’s jag.
Variable parameters are those that do not have predefined optimum values. Some of these variable
parameters are mutually independent and without any constraints, they include: the thickness of the
horizontal of the stator’s yoke, the thickness of the vertical of the stator’s yoke, the width of the stator,
the height of the stator, the radius of the stator’s internal edge, the thickness of the stator’s yoke at the
hole, the length between the bisector and the slot edge, the radius of the stator’s teeth, the external radius
of the rotor, the width of the rotor’s pole, the thickness of the rotor’s yoke, and the height of the rotor’s
teeth. Other variable parameters are dependent, either on some invariable parameters or on mutually
independent ones.
In our case we optimize 12 mutually independent variable parameters, while taking into account the
following constraints:
• The parameters (both independent and dependent) should be changed simultaneously to achieve
proper electromagnetic conditions in the material.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 510 — #2

An Evolutionary Approach to Problems in Electrical Engineering 29-511

FIGURE 29.2 Geometrical parameters of the stator and rotor of a UM.

• Each parameter’s dimension should only be varied within a predefined feasible limit.
• Parameter transformations and their evaluation should be done as quickly as possible.
29.1.1.2 The Efficiency of a UM
The efficiency of a UM is defined as the ratio of the output power to the input power, and it depends on
various power losses, which include:
• Copper losses: the joule losses in the windings of the stator and the rotor.
• Iron losses: including the hysteresis losses and the eddy-current losses, which are primarily in the
armature core and in the saturated parts of the stator core.
• Other losses: such as brush losses, ventilation losses, and friction losses.
The overall copper losses (in all stator and rotor slots) are as follows:


PCu = ( J 2 · A · ρ · l turn )i , (29.1)
i

where i stands for each slot, J is the current density, A is the slot area, ρ is the copper’s specific resistance,
and lturn is the length of the winding turn.
Because of the nonlinear magnetic characteristic, the calculation of the iron losses is less exact. The iron
losses are separated into two components: the hysteresis losses and the eddy-current losses. Consequently,
a motor’s iron losses can be expressed by the following equation [1]:

P Fe = ke · B 2 · f rot
2
· m rot + ke · B 2 · f stat
2
· m stat + k h · B 2 · f stat · m stat , (29.2)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 511 — #3

29-512 Handbook of Bioinspired Algorithms and Applications

where k e is an eddy-current material constant of 50 Hz, kh is a hysteresis material constant of 50 Hz, B is

the maximum magnetic flux density, f is the frequency, and m is the mass.
Besides the iron and the copper losses, three additional types of losses also occur in a UM, that is, brush
losses PBrush , ventilation losses PVent , and friction losses PFrict . All three types of losses mainly depend on
the speed of the motor. When optimizing the geometry of the rotor and the stator, the motor’s speed is
fixed, so PBrush , PVent , and PFrict have no impact on the efficiency of the motor. Therefore, these parameters
are not significantly affected by the geometry of the rotor and the stator.
The output power P2 of the motor is a product of the electromagnetic torque, T , and the angular
velocity, ω,
P2 = T · ω, (29.3)

where ω is set by the motor’s speed, and T is a vector product of the distance from the origin, r, and the
electromagnetic force, F .
When considering all the mentioned losses and the output power, the overall efficiency of a UM can be
defined as follows:

P2
η= . (29.4)
P2 + PCu + PFe + PBrush + PVent + PFrict

29.1.2 Design of Integrated Circuits

Whenever a new IC is designed, the problem of selecting the best register-transfer level (RTL) specification
has to be faced. And as circuits get bigger, so too does the problem: more combinations have to be examined
before the optimum combination is found. Therefore, automatic circuit optimization is required, as this
speeds up the whole design process and eliminates some of the errors. More specifically, the case is
focused on application-specific integrated circuits (ASICs), which need an even more sophisticated design
(in terms of size and speed) because of their specific uses.
High-level synthesis [2] is an automatic design process that transforms the initial behavioral description
into the final specification of the RTL. The process consists of the following: compilation, transformation,
scheduling, allocation, and binding. Of these, the operation scheduling and the resource allocation are the
most important tasks of the high-level synthesis because they are at the core of the design and crucially
influence both the design and the final layout. Due to the interdependence of these two tasks, the solution
of one task depends on an estimation of the solution of the other task, which is not solved yet. The
scheduling of the operation into different control steps therefore affects the allocation of operations to
different units. The interaction of these two tasks presents formidable obstacles to the goal of optimization
[3]. There are, however, some approaches to concurrent solving, but their solutions, to some extent, are
less than optimal.
Optimally scheduled operations are not necessarily optimally allocated to units. To enable optimal
allocation we need to consider some allocation criteria while the scheduling is being done. There-
fore, algorithms with concurrent scheduling and allocation produce the best results. However, these
algorithms are also time-consuming. It is obvious that we have to deal with a trade-off between the quality
of the solution and its design time: an algorithm might be fast, but its solutions will not be as good, and
vice versa.

29.1.2.1 Data-Flow Graphs of ICs

An IC described with a hardware description language, for example, VHDL, can be presented as a data-
flow graph (DFG) [2] (see Figure 29.3). Each node i represents the operation to be executed, and the type
of node determines the type of operation. The edges e represent dependencies between operations. An
edge eij , between nodes i and j, represents the data produced by the node i and used by the node j. All
edges are unidirectional.
In the final design there is a group of resources (functional units, storage units or registers, and
connection units or buses) that define the implementation. Functional units (adders, multipliers, etc.)
perform transformations on data values; connection units transport values from one unit to another; while

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 512 — #4

An Evolutionary Approach to Problems in Electrical Engineering 29-513

storage units preserve those values over time. We have N different functional units Fi , i = 1, 2, . . . , N , nr
registers, and nb buses, while T represents the execution time of the DFG.
The parameters are calculated as follows:

• The number nfi is the highest number of the i-th functional unit needed in a separate
control step.
• The number nr is the highest number of variables needed in a separate control step. We consider
variables that are needed by the functional unit as input data, variables that are returned as output
data, and variables that are not used at the moment but will be used in some of the later control
steps or must be available until the end of the execution of all operations.
• The number nb is the highest number of data transmissions (into or from the functional units) in
a separate moment.
• The execution time, T , is the time needed to execute all the operations of the schedule.
• The weights wfi , wr , wb , and wt are the weights of functional units, registers, buses, and time,
respectively, to be considered in the IC quality-evaluation cost function. The first three weights are
proportional to their silicon area in the IC, while wt reflects our IC speed constraints.

29.1.2.2 Concurrency of Scheduling and Allocation

Algorithms that perform concurrent scheduling and allocation are better in terms of the optimality of
the solutions. These algorithms are, however, very time-consuming. Therefore, we have to deal with a
trade-off between the quality of the final solution and the length of the design-time for it. There are some
approaches to concurrent solving, but their solutions are, to some extent, less than optimal.

• Landwehr et al. [4] suggest solving with integer linear programming. The procedure ensures
optimal solutions, but is computationally expensive and therefore practically useful only for small
circuits.
• Zhu and Gajski [5] suggest a “soft scheduling,” where operations are temporarily scheduled, but
are later adjusted according to some physical characteristics. Operations are finally scheduled after
the allocation and binding tasks, when all the parameters that affect the scheduling optimality are
known. Actually, there is no concurrency, but there is some iterative refinement of the scheduled
operations.
• Mandal and Zimmer [6] suggest concurrent scheduling and allocation using a genetic algorithm,
and with a stress on the reduction of the number of connections. In the optimization process, the
whole circuit is partitioned into blocks consisting of a functional unit, a storage unit and internal
connections. Operations are scheduled according to the units in those blocks, and according to
additional global storage units and interconnections, which are to be minimized. Here, the problem
of various functional units being present in each block arises, to ensure that all the operations of
the block are executed.
• Kim [7] suggests a concurrent procedure where the operations subset is chosen for each control
step. These operations are later bound to functional units and the variables are bound to storage
units in order to achieve the highest possible connection utilization. The order of operations to be
scheduled is based on their readiness for execution and their influence on the number of storage
and connection units, while considering the lowest possible number of control steps.
• Grajcar [8] describes the concurrency of scheduling and allocation on a multiprocessor system.
The procedure is based on a GA and list scheduling, which considers only some values (critical
points) in the chromosome code.

As mentioned earlier, when tasks are performed separately (Figure 29.4[a]) the solution is not neces-
sarily optimal; it is better to use an approach with iterative repetition of the scheduling and allocation
(Figure 29.4[b]). Here, the problem of the next operation/unit to be changed appears, since the order of
changes can influence the final solution. It is a similar situation with the approach that involves partition-
ing of the operations into small groups, within which there is an iterative repetition of the scheduling and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 513 — #5

29-514 Handbook of Bioinspired Algorithms and Applications

Evolutionary

optimization
time/area

FIGURE 29.3 DFG and its IC.

the allocation (Figure 29.4[c]). Since there are fewer operations in the group there is no problem with the
order of changes, but there is a problem with the appropriate partitioning of the operations. Obviously,
the best approach is the one with purely concurrent scheduling and allocation (Figure 29.4[d]), where the
iterative-refinement order does not influence the quality of the solution [9]. Concurrency is achieved by
using algorithms that do not depend on the order of the transformations. Therefore, there is no influence
of the altered start time on the allocated unit, nor is there any influence of the allocated unit on the start
time. When all the transformations are made, then the appropriateness of the changes is checked.

29.2 Evolutionary Optimization in Problems of

Engineering Design
Evolutionary techniques are used in various search methods for a range of different optimization areas.
Their undetermined approach gives them an advantage when it comes to multicriteria problems and
problems with more local optima [10]. For this reason we adopted an artificial approach to the solving
of our design problems using a genetic algorithm (GA) [11,12]. The GA, as a frequent implementation
of evolutionary techniques, is an optimization method based on the mechanism of evolution and natural
genetics. This algorithm has already proved to be very efficient in a wide range of different optimization
procedures where the exact equations are not available or some nonlinearities are present [13]. In spite of its
simplicity, the GA has proved to be an efficient method for solving various optimization and classification
problems, in areas ranging from economics and game-theory to control-system design [14–18].

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 514 — #6

An Evolutionary Approach to Problems in Electrical Engineering 29-515

(a) (b)

Scheduling
Scheduling

Allocation

Allocation

(c) (d)

Partitioning
Scheduling
and allocation
Scheduling Scheduling Scheduling

Allocation Allocation Allocation

Binding

FIGURE 29.4 Scheduling and allocation concurrency.

29.2.1 The Genetic Algorithm in UM and IC Design

When reducing the main power losses in a UM design by optimizing the geometry of the rotor/stator
lamination, we have to deal with a complex search space and its nonlinear behavior. In IC design we apply
the concurrency of interdependent tasks with their opponent constraints. Because traditional search-and-
optimization methods have proved to be inefficient at finding the solution under such conditions, we
decided to apply a GA. This heuristic method requires only a little information to provide a robust, yet
flexible, search in a wide and complex search space.

29.2.1.1 The Basics of the GA

The GA codes parameters of the problem’s search space as finite-length strings over some finite alphabet.
It works with a coding of the parameter set, not the parameters themselves. The algorithm employs an

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 515 — #7

29-516 Handbook of Bioinspired Algorithms and Applications

initial population of strings, which evolve into the next generation under the control of probabilistic
transition rules — known as randomized genetic operators — such as selection, crossover, and mutation.
The objective function evaluates the quality (or fitness) of solutions coded as strings. This information
is then used to perform an effective search for better solutions. There is no need for any other auxiliary
knowledge. The GA tends to take advantage of the fittest solutions by giving them greater weight, and
concentrating the search on the regions of the search space that show likely improvement.

29.2.1.2 Encoding
One of the most important parts of the GA is the encoding. By encoding the proper parameters and using
the proper encoding type we can significantly influence the efficiency of the algorithm.
In the UM design, the mutually independent variable geometrical parameters of the rotor/stator lam-
ination are coded as strings over the alphabet IR of real values. There is no need to normalize the physical
parameters. They do differ in ranges but the crossover operation switches the values of the same parameter,
no matter where the crossover point is, since each parameter is always encoded at the same place in the
chromosome.
In the IC design the chromosome string consists of the numbers that represent the starting time of
each operation and the allocated unit for each operation, where the position in the string depends on
the order of the operations in the input IC description. This means that the chromosome consists of
pairs of time/unit information for each operation. And the genetic operators can influence both parts
of that information, either together or separately. The selected encoding type is chosen because of its
convenience. When strings have to be further transformed, checked, and analyzed, there is no need for
any additional conversion of their values. In addition, the used implementation of genetic operators can
check the changed values (their feasibility) instantly, without any transformation. The correctness of the
transformation can therefore be checked within the function itself.

29.2.1.3 Cost Function

In the UM design each string of the population is decoded into a set of rotor and stator parameters.
Its fitness is estimated by performing a finite-element numerical simulation to calculate the copper and
the iron power losses (using Equations [29.1] and [29.2], respectively). Their sum corresponds to the
solution’s fitness.
In the IC design a multiobjective [19,20] function is used, which means it takes control over more criteria
or objectives. The cost function, represented by Equation (29.5), considers the number of functional units,
the number of registers, the number of buses, and the execution time of all the operations in the DFG.


N


Cost =  (costfi )2 + cost2r + cost2b + cost2t ,
i=1

costfi = wfi Fi ,
(29.5)
costr = wr nr ,

costb = wb nb ,

costt = wt T .

To obtain the cost of a certain DFG, the algorithm has to evaluate the required number of resources. In
contrast to the other multiobjective functions that give more than one final solution, this one already
includes the decision-making part, that is, it chooses one solution from all the solutions on the Pareto
front. The chosen solution has the shortest distance to the origin, where the origin represents the ideal,
costless, solution and the axis represents the considered objectives.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 516 — #8

An Evolutionary Approach to Problems in Electrical Engineering 29-517

29.2.1.4 The Genetic Operators

To evolve the best solution candidate, the GA employs the genetic operators for manipulating the strings
in a population. The GA uses these operators to combine the strings of the population in different
arrangements, seeking a string that optimizes the objective function. In the IC design, operators also
consider data dependencies and the given library of available functional units. This combination of
strings results in a new population.

29.2.1.4.1 Basic Operators

• The selection operator is used for creating a new generation. We apply the elitism strategy, where
a number of least-fit members of the current population are interchanged with an equal number
of the best-ranked strings. This approach ensured better starting positions for the best-ranked
solutions, as all solutions have equal chances for reproduction.
• The crossover operator is used for exchanging information between the selected strings. In our case
it proceeds in two steps. First, the strings are mated randomly to pair off the couples. Second, the
mated string couples cross over, using a given probability pc . Whether the one-point or two-point
crossover method is chosen, characteristic values are swapped between the mated couples to create
two new strings. Moreover, we might use a constant probability pr to select a case in which the
values of the swapped characteristics are calculated as a mean value of the parent characteristic
values. While the first crossover approach ensures that the offspring solutions preserve the “genetic
material” from both parents, the second approach helps to seek other solutions close to the solutions
that appear to be good. In the IC design, after crossover points are determined, either the unit
information is changed between the two chromosomes and the start times are adapted, or the
start times are changed and a suitable unit is allocated. So the dominancy is expressed either in
functional units or the start times of operations.
• The mutation is a process by which strings resulting from selection and crossover are perturbed;
the process serves to create random diversity in the population. We use a mutation approach
where each string is subjected to the mutation operator. The values of the chromosome string are
either slightly changed in the UM design, or start times are changed with a respective faster/slower
unit allocated in the IC design. Mutation is performed on a characteristic-by-characteristic basis,
each characteristic mutating with a probability pm . We also apply the possibility of annealing the
mutation rate, where pm is the variable mutation probability that decreased linearly with each
new population. In other words, we assume that each new population is generally fitter than the
previous one. Such an approach is used to overcome a possible disruptive effect of mutation, and
to speed up the convergence of the GA to the optimum solution.
• The variation operator swaps some values of the similar chromosome positions. After two oper-
ations are selected, and when they are of the same type, for example, additions in the IC design,
their functional units are switched. If needed, their start times are also updated.

29.2.1.4.2 Independent Operators

The advantage of the independent GA approach [9] is that there is no need to preset some working
parameters, for example, the number of generations, the population size, and the probabilities of crossover,
mutation, and variation. These parameters are set automatically during the optimization phase, depending
on the progress and the speed of the optimization.

• Setup. If the chromosome that represents a solution is large, then the population size also has
to be large enough to ensure that many different chromosomes will be involved in a search. The
population size therefore depends on the size of the chromosome or the complexity of the problem.
• Crossover. Considering four candidates — two parents and their two offspring — only the first and
the third, rated according to their fitness, pass to the next generation. This forces, very probably,
at least one of the offspring to be passed to the next generation in addition to the best candidate.
Otherwise the offspring have only a small influence on new generations, since the crossing of two

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 517 — #9

29-518 Handbook of Bioinspired Algorithms and Applications

good parents probably produces offspring that are not so good. They might, however, be good after
a few more transformations.
• Mutation. Chromosomes with low fitness are mostly exposed to mutation. Each position in the
chromosome string is mutated if that position of the chromosome is of the same value in the
majority of “poorly fitted” chromosomes in the population. This is the way to change the bad
characteristics in “poorly fitted” chromosomes and to redirect the search to another direction. In
the case of “well-fitted” chromosomes, values are mutated if they differ from the majority of values
in other good chromosomes at the same position. This ensures faster convergence in the final stages
of the optimization.
• Variation. The interchange of the values of two positions, as described for the basic operators, is
performed if the frequency of the value in that position in the population of one position is high
and the frequency of another bit is low.

29.3 Procedures and Evaluations

29.3.1 UM Design Procedure
29.3.1.1 The Conventional versus the Evolutionary Design Procedure
In a conventional design procedure for a motor:

1. The initial estimation for the geometry of the rotor and the stator is made based on experience.
2. The appropriateness of this geometry is then usually analyzed by means of a numerical simula-
tion of the electromagnetic field. In our case, the analysis is performed with commercial ANSYS
software [21], which applies a finite-element method (FEM) with an automatic finite-element-mesh
generation. The result is a magnetic vector potential on every node of the finite-element mesh.
3. If the results of the numerical simulation show an inconvenient electromagnetic field structure, the
direct design procedure is repeated until the motor geometry is optimized.

The advantage of this approach is that engineers can significantly influence the progress of the design
process by using their experience, and they can react intelligently to any noticeable electromagnetic
response with proper geometry redesign. The drawback of this approach is that an experienced engineer
and a large amount of time are needed.
The described conventional motor design can be upgraded with a genetic algorithm. The concept of
this evolutionary design can be roughly explained as follows:

1. The GA can start its optimization from any configuration of geometrical parameters while it defines
a feasible solution.
2. Each geometrical configuration is analyzed using the ANSYS finite-element program. This step
requires a prior decoding of the strings into a set of geometrical parameters for the rotor and the
stator.
3. After the calculation of the fitness, the reproduction of the individuals and the application of the
genetic operators to a new population are made. The GA repeats this procedure until a predefined
number of iterations have been accomplished.

The advantages of this approach are that: there is no need for an experienced engineer to be present during
the whole process, only at the beginning to decide on the initial design, and there is no need to know the
mechanical and physical details of the problem. The problem can be solved without any knowledge of
the problem.
The drawbacks of this approach are that it can lead to the improper use of genetic operators, and an
initial solution that is too loosely set, can lead to a longer convergence time.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 518 — #10

An Evolutionary Approach to Problems in Electrical Engineering 29-519

29.3.1.2 The Optimization Software

In accordance with the proposed evolutionary design approach described earlier we developed software
that links together: (a) the genetic algorithm that optimizes the geometry of the rotor and the stator of a
UM, and (b) the finite-element program needed to evaluate the optimized geometry.
The DOptiMeL software [22] is a tool for experimenting with different settings of the GA’s parameters,
which may significantly influence the quantity and the quality of the program solutions. The tool consists
of two parts:

• A setup part for setting the GA parameters and the geometry limits of the rotor/stator lamination
of an initial UM.
• An optimization part for optimizing the geometry of the rotor/stator lamination.

The program was developed using the Microsoft® Visual C++® programming tool and runs under the
Microsoft® Windows® operating systems.

29.3.1.3 Parameters
29.3.1.3.1 Geometrical Parameters
As stated earlier, there are 12 mutually independent geometrical parameters that need to be optimized.
These parameters can only be varied within their predefined dimension limits to find an optimum con-
figuration that will increase the motor’s efficiency. Solutions in which the parameters exceed the limits are
rejected as being inoperable.
There are some invariable parameters that have a strong influence when defining the outline of the
lamination:

1. The external radius of the stator: which roughly defines the amount of iron and copper, and
consequently, the price of the motor. This is held constant during the optimization to ensure
cost-comparable solutions.
1. The radius of the rotor’s shaft : which we fixed at 5.5 mm. From our experience we know that when
the rotor-shaft radius is less than 5.5 mm, the rotor’s natural frequency can fall below the maximum
frequency of the motor and the resulting resonance would cause the rotor’s vibrations to exceed
allowable limits.
3. The radius of the stator’s side-hole: which would be 0 mm in the ideal case, is set to a small positive
value because holes for the rivets are required in order to bind the stator.
4. The air gap: The angles of the symmetrical and tangential parts of the air gap are set to fixed values
because the commutation, which is conditioned by the air gap, is not taken into account during
the optimization.

29.3.1.3.2 GA Parameters
For the GA to work well, robust parameter settings have to be found for the population size, the number
of generations, the selection criteria, and the genetic operator probabilities:

• If the population is too small, the GA converges too quickly to a local optimum solution and can
miss the best solution. On the other hand, a large population requires a long time to converge to
a region of the search space with significant improvement. The best results are obtained when the
population size is between 30 and 50.
• By applying the elitism strategy, fitter solutions have a greater chance of reproducing. But when
the ratio of least-fit solutions to be exchanged with best-fit ones is too high, the GA is trapped too
quickly into a local optimum solution. However, this number is subject to the population size, and
appears to be acceptable within 20 to 30% of the population size.
• As a crossover probability that is too low preserves solutions from being interchanged and a longer
time is required for them to converge, the probability is set to at least 60%, so that the algorithm
can act satisfactorily.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 519 — #11

29-520 Handbook of Bioinspired Algorithms and Applications

• A mutation rate that is too high introduces too much diversity and takes a longer time to reach an
optimum solution. A mutation rate that is too low tends to miss some near-optimum solutions.
Using the annealing strategy — a linearly decreasing mutation probability rate with each new
generation — the effects of a too high or too low mutation rate can be overcome. The operator is
useful if the value of the probability is 0.1%, while in the annealing strategy it starts with 1% and
ends at 0.1%.

29.3.1.4 The UM Evaluation

The proposed evolutionary design approach is evaluated by estimating the actual improvement in the
efficiency of an initial UM that is designed using the conventional and evolutionary design approaches.

1. With the ANSYS software we calculated the efficiency of an initial UM. An outline of the rotor/stator
lamination of this motor is shown in Figure 29.5(a). The power losses of this motor were calculated
to be 313 W and the output power was calculated to be 731 W (Table 29.1). In the outline, the
levels of magnetic flux density through the rotor/stator lamination are shown, expressed as T. The
darkest gray color indicates the areas with the highest level of magnetic flux density, which results
in high iron losses. The copper losses are not shown in this area.
2. After several runs of the DOptiMeL software, a set of promising solution candidates was collected.
We applied the following settings for the GA parameters: population size 30, number of generations
100, selection ratio 0.3, crossover probability 0.7, and mutation rate 0.01. For each candidate we
make a finite-element numerical simulation followed by the calculation of the objective function
value (fitness). Because each design is verified with finite-element numerical calculations, the
optimization is a lengthy process. It takes around 3000 runs for the optimization to converge. Most
of the solutions that are given by the DOptiMeL program show a significant reduction in the iron
and the copper losses in comparison with the losses in the initial motor. The best solution results
in a power-loss reduction of 24%, and gives us a motor with iron and copper losses of 239 W
(see Figure 29.5[b]).

(a) (b) (c)

FIGURE 29.5 Stator/rotor lamination outline: (a) initial, (b) optimized, (c) cost-invariant optimized.

TABLE 29.1 UM Evaluation Results

Analytic calculation Prototype measurement

Initial Optimized 1 Optimized 2 Initial Optimized 1 Optimized 2

Input power (W) 1044 970 982 1050 990 1000
Efficiency (%) 70.0 75.8 74.8 69.5 73.7 73.0
Output power (W) 731 731 731 730 730 730
Power losses (W) 313 239 251 320 260 270
Efficiency difference (%)∗ — 5.8 4.8 — 4.2 3.5
Power losses difference (W)∗ — 74 62 — 60 50

∗ According to the initial UM.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 520 — #12

An Evolutionary Approach to Problems in Electrical Engineering 29-521

The main differences between the initial design (Figure 29.5[a]) and the optimized design (Figure 29.5[b])
are: (a) the height of the rotor-and-stator laminations is increased by 13%, (b) the rotor radius is increased
by 5%, (c) the slot (copper) areas in the stator and the rotor are larger, and (d) the iron area in the rotor
is larger.
A comparison of the magnetic flux densities in the initial and the optimized motor shows a clear
reduction of the areas with the highest levels of magnetic flux density in the optimized motor.

29.3.1.4.1 Copper and Iron Losses

In the optimized lamination the copper losses in the rotor and the stator are significantly lower than in the
initial lamination. The reason for this effect is that the slot area of the optimized stator lamination is larger.
The number of ampere-turns in the optimized lamination that are necessary for obtaining the required
torque are lower in comparison with the initial lamination due to the optimized lamination design.
The copper losses increase with the square of the number of ampere-turns and they are inversely
proportional to the slot area [1]. In the rotor, the slot area is increased, and this results in a reduction
of 52% for the copper losses in the rotor. For the same reason, the stator’s copper losses are 52% lower
as well.
In contrast, the iron losses are higher than that in the initial lamination. The iron losses are increased
in both the rotor and the stator, mainly because of the slightly larger iron area in the optimized design.
The magnetic flux density in the stator remains in the same range as in the initial design. In the rotor, the
magnetic flux density is decreased, but the iron area of the rotor is larger.
Overall, the optimized design has a much better torque capability, and this results in a total loss
reduction of 24%.
After considering these results we might be tempted to think that the next step after a reduction of
the copper losses would be a reduction of the iron losses. However, if we design a motor that has lower
iron losses, the copper losses would automatically increase — and at an even higher rate. It should be
remembered that a lamination is optimized when the sum of the copper and the iron losses is a minimum.

29.3.1.4.2 Material Costs

The results presented earlier consider an optimized UM. During the optimization procedure the material
costs of the motor are not taken into account, that is, the external radius of the stator is not considered
to be a variable parameter. The solution resulted in much larger stator dimensions than that in the initial
lamination. As the amount of iron used for manufacturing the motor depends only on the width and the
height of the stator, the 13% increase in the stator’s height (as in our optimized solution) would cause the
same increase in the amount, and consequently the cost, of the iron required.
The amount of copper that is used in the motor depends on the size of the slot areas; assuming the
copper-fill factor is constant during the optimization. In the optimized solution, the slot area is 18% larger
than in the initial one. As a consequence, the copper costs increased by the same percentage.
The figures for the estimated material costs suggest that the optimized design would be much more
expensive than the initial design. Therefore, we repeated the optimization procedure, this time by con-
sidering the material price criterion. We extended the definition of the technical quality of the UM to
be a function of both power losses and materials costs. The new criterion is based on the product of
efficiency and an estimation of the material costs. A unity factor is set to the material costs of the initial
design: an increase in material costs decreased the material costs factor linearly and vice versa. The goal of
the optimization is to maximize this new criterion. The outline of the new costs-optimized rotor/stator
lamination is shown in Figure 29.5(c).
This lamination has a different profile than the first optimized design. The stator covers the same area
as the initial design, so the amount of iron is not increased. The slot areas are slightly larger than in the
initial design, so 6.5% more copper would be needed. This would mean a more expensive motor, but the
increases in cost would be negligible in comparison with the much better performance. Such a motor
would have a power loss of 251 W, which is 20% better than the initial design. Comparing the optimized
design that does not consider the material costs criterion with this new costs-optimized design, the first

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 521 — #13

29-522 Handbook of Bioinspired Algorithms and Applications

one is better: its power loss is estimated to be 12 W lower. By fixing the stator’s outer radius, we mostly
lose the gain of the first design in terms of the decrease of the stator’s copper losses.

29.3.1.4.3 Prototyping
We made prototypes of both the optimized and the costs-optimized motors and measured the real power
losses and the efficiencies of the motors. These values are shown in Table 29.1. The results are only
slightly different from those calculated with the ANSYS finite-element program. The main reason for
this difference can be explained by the non-exact calculation of the iron losses, due to a variation in the
material’s properties.

29.3.2 IC Design Procedure

The promising results of different evaluations [14–16] led us to the Evolutionary Concurrent Scheduling
and Allocation (ECSA) design approach [9]. This approach considers scheduling and allocation con-
straints, allows a short design time, and can find globally optimal solutions. The input description of the
circuit is transformed into two basic (initial) schedules, obtained with the ASAP and ALAP algorithms.
The functional units used in the first case are those that are the fastest for each operation, and in the
second case are those that are the slowest for each operation. These two schedules present some kind of
boundary solutions, since all the other solutions are executed in between the time limits defined by these
two schedules. In other words, no other solution can be faster or slower, irrespective of the combinations
of used units.
Each solution has to be properly encoded, that is, each operation’s start time and functional unit have
to exist in the chromosome. The initial population is built upon the two initial solutions, which are
multiplied to form the population with the so-called boundary solutions. The optimal solution has to
be somewhere in between the boundaries, therefore genetic operators (crossover, mutation, variation)
transform those encoded solutions. With the transformations their start times and allocated functional
units are changed.
The appropriateness of the proposed approach is tested by a computer implementation of the ECSA
algorithm, which is used with test-bench ICs.

29.3.2.1 Test-Bench Circuits

The ICs used for the evaluation are chosen on the basis of their appearance in the literature and similar
studies. They differ in terms of size and the number of operation types.

29.3.2.1.1 Differential Equation

The relatively small circuit of differential equation [23] has only 11 operations, but 4 different operation
types (6 multiplications, 2 additions, 2 subtractions, 1 comparison). This circuit is useful when testing
libraries with different implementations of the same operation types.

29.3.2.1.2 Elliptic Filter

This filter [24] consists of 34 operations, but only 2 operation types: 26 additions and 8 multiplications.
The circuit is suitable for comparison, due to its size and operation dependencies, since they form two
independent, similar critical paths, both influencing the circuit delay.

29.3.2.1.3 Bandpass Filter

One of the implementations of the bandpass filter [25] is the circuit used for our evaluation. It consists of
29 operations: 11 multiplications, 10 additions, and 8 subtractions. Due to data dependencies, almost all
the operations influence the circuit delay.

29.3.2.1.4 Least-Mean-Square Filter

This filter for signal adaptation (noise reduction) is based on the least-mean-square method [26]. It
consists of 47 operations: 24 multiplications and 23 additions. This test-bench circuit is useful because of
its size and unique data dependencies.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 522 — #14

An Evolutionary Approach to Problems in Electrical Engineering 29-523

TABLE 29.2 Technical Characteristics of the Functional Units

Functional unit Delay No. of

{operations} [No. of steps] gates

FE1 {+, −} 6, 6 370

FE2 {+, −} 1, 1 665
FE3 {<} 6 353
FE4 {+, −, <} 1, 1, 1 696
FE5 {×} 4 3040
FE6 {×} 2 7296
FE7 {÷} 21 3040
FE8 {÷} 9 7296
FE9 {×, ÷} 4, 21 3344
FE10 {×, ÷} 2, 9 8025
FE11 {+, −, <, ×, ÷} 1, 1, 1, 4, 21 3692
FE12 {+, −, <, ×, ÷} 1, 1, 1, 2, 9 8373

29.3.2.2 Functional Units

For an easier and more realistic comparison of different algorithms when testing the size and delay of
implemented circuits we made a library of different functional units, which differ in terms of their sizes
and delays. Table 29.2 shows the sizes and delays of various implementations of the arithmetic logic
operations. Here, different types of logic are used to make the units with different delays and sizes. The
values are based on an analysis of data on circuits and their complexities [27–29]. The number of gate
transitions defines the delay, whereas the overall number of gates needed to implement the unit defines
its size. These values are just for orientation, since the real numbers depend on the chosen technology
[27]. The delays presented in Table 29.2 are relative, for example, normalized to the fastest functional unit
among all the operations. Most of the units are multifunctional, that is, they can perform different types
of operation.

29.3.2.3 Parameters
By considering 18,750 different schedules of each circuit with the ECSA algorithm and 3,125 different
combinations of the parameters, we statistically compared (using the procedure described in Reference 30)
the results according to their cost function (Equation [29.5]). To ensure that most solutions are time-
constrained, that is, executed in the shortest possible time, the weight wT is set to an extremely high
value.
Figure 29.6 presents the results of different parameter-set evaluations for different test circuits.
Figure 29.6(a) shows the influence of the number of generations on the percentage of good/bad solu-
tions within the set of solutions with a specified number of generations. Here, and in other figures and
subfigures, the column marked as high represents high-quality (good) solutions, while the column marked
as low represents the low-quality (bad) solutions. So, from among all the solutions we count the number
of solutions with high fitness and the number of solutions with low fitness for each number of genera-
tions. Figure 29.6(b) shows the influence of the population size on the percentage of good/bad solutions
within the set of solutions with a specified population size. Figure 29.6(c) shows the influence of the
crossover probability on the percentage of good/bad solutions within the set of solutions with a specified
crossover probability. Figure 29.6(d) shows the influence of the mutation probability on the percentage
of good/bad solutions within the set of solutions with a specified mutation probability. Figure 29.6(e)
shows the influence of the variation probability on the percentage of good/bad solutions within the set of
solutions with a specified variation probability.
As shown in Figure 29.6 and Table 29.3, the high-quality solutions are mostly obtained with the
following values of the parameters: probability of crossover equal to 0.7, probability of mutation equal to
0.04, and probability of variation equal to 0.03. In addition, taking into account the sizes of the circuits,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 523 — #15

29-524 Handbook of Bioinspired Algorithms and Applications

(a) 50
Solution quality
High
Low
40
% of solutions

30

20

10

0
20
30
40
50
60
60
70
80
90
100
60
70
80
90
100
100
110
120
130
140
Number of
generations
Differential Fifth-order elliptic Bandpass filter Least-mean-
equation filter square filter

(b) 50
Solution quality
High
Low
40
% of solutions

30

20

10

0
30
40
50
60
70
80
90
100
110
120
80
90
100
110
120
130
140
150
160
170

Polulation
size
Differential Fifth-order elliptic Bandpass filter Least-mean-
equation filter square filter

FIGURE 29.6 Test-bench parameter evaluation: (a) number of generations, (b) population size, (c) probability of
crossover, (d) probability of mutation, (e) probability of variation.

the number of generations and the population size should be set to 3- and 3.5-times the size of the circuit,
respectively.
The values of the parameters in this combination are referred to as the optimal values. These optimal
values are determined on the basis of the percentage of solutions with certain parameters from among the
good solutions. A parameter value that is to be considered as optimal should have at least a 25% share of
the high-quality solutions, as well as having a <10% share of the low-quality solutions.
The ECSA algorithm is used with the values of the parameters as presented in Table 29.3. Other
parameters needed to run the force-directed scheduling (FDS) and ECSA algorithms and the cost function
depend on the sizes of the functional units.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 524 — #16

An Evolutionary Approach to Problems in Electrical Engineering 29-525

(c) 50
Solution quality
High
Low
% of solutions 40

30

20

10

0.6

0.6
0.7
0.8
0.9
1.0
0.6
0.7
0.8
0.9
1.0
0.6
0.7
0.8
0.9
1.0

0.7
0.8
0.9
1.0
Probability of
crossover
Differential Fifth-order elliptic Bandpass filter Least-mean-
equation filter square filter

(d) 50
Solution quality
High
Low
40
% of solutions

30

20

10

0
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05

Population
size
Differential Fifth-order elliptic Bandpass filter Least-mean-
equation filter square filter

FIGURE 29.6 Continued.

29.3.2.4 The IC Evaluation

The ECSA algorithm is evaluated by a comparison with nearly optimal [31] FDS [32]. FDS tries to schedule
optimally the DFG considering a uniform distribution of the operations of the same type over the available
control steps.
Table 29.4 shows the results of the following evaluations: FDS with fast units, FDS with slow units, and
ECSA with basic and independent genetic operators. There are two types of DFGs for each circuit. The
first, or plain, is an ordinary data-flow graph with nodes that represent operations, as described in similar
studies; and the second, or improved [9], considers the input variables (start registers) via some additional
nodes to ensure a more accurate estimation of the registers and the buses needed to implement the circuit.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 525 — #17

29-526 Handbook of Bioinspired Algorithms and Applications

(e) 50
Solution quality
High
Low
% of solutions 40

30

20

10

0
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05
0.01
0.02
0.03
0.04
0.05
Probability of
crossover
Differential Fifth-order elliptic Bandpass filter Least-mean-
equation filter square filter

FIGURE 29.6 Continued.

TABLE 29.3 Optimal Values of the Parameters for Different Test-Bench Circuits

Differential Fifth-order Bandpass Least-mean-square Average optimal

equation elliptic filter filter filter values
Number of generations 40 100 90 130 3 × DFG size
Population size 55 120 110 160 3.5 × DFG size
Probability of crossover 0.8 0.6 0.7 0.6 0.7
Probability of mutation 0.04 0.05 0.04 0.05 0.04
Probability of variation 0.04 0.02 0.02 0.05 0.03

29.3.2.4.1 Differential Equation

Because of the small circuit size there is no improvement in the solutions obtained with the ECSA algorithm
(either basic or independent) when considering an ordinary DFG — all the solutions are of a larger size.
But when we consider the start registers (input variables) there are some ECSA solutions with a slightly
larger size and a smaller number of buses.

29.3.2.4.2 Fifth-Order Elliptic Filter

The evolutionary method with a basic approach finds a smaller circuit with a smaller number of buses and
a slightly longer execution time for the ordinary DFG, whereas the independent approach could not find
any improved solution. When dealing with the improved DFG, both approaches (basic and independent)
find considerably smaller circuits with a slight increase in the execution time, while the independent
approach also finds the solution with a substantial decrease in the required number of registers and buses.

29.3.2.4.3 Bandpass Filter

Both ECSA methods find, when dealing with the ordinary DFG, the solutions with a smaller number of
registers and buses; the basic approach also finds the smaller circuit, but with a slightly longer execution
time. When dealing with the improved DFG, both approaches find the solutions with the same circuit
size and execution time as the comparable FDS solution, but the required number of registers and buses
is considerably smaller for the ECSA solutions.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 526 — #18

An Evolutionary Approach to Problems in Electrical Engineering 29-527

TABLE 29.4 The Evaluation Results of the ECSA Algorithm with Different Test-Bench ICs

Algorithm Functional units Size Registers Buses Delay Runtime [sec]

Differential equation
FDS-fast 1 × FE2 + 1 × FE4 + 3 × FE6 23,249 17 6 6 0.01
FDS-slow 2 × FE1 + 1 × FE3 + 2 × FE5 7,173 18 6 20 0.01
ECSA-basic 2 × FE2 + 1 × FE4 + 3 × FE6 23,914 18 8 6 0.11
ECSA-independent 2 × FE2 + 1 × FE4 + 3 × FE6 23,914 17 6 6 0.09

Differential equation with start registers

FDS-fast 1 × FE2 + 1 × FE4 + 3 × FE6 23,249 10 9 6 0.01
FDS-slow 2 × FE1 + 1 × FE3 + 2 × FE5 7,173 11 9 20 0.01
ECSA-basic 2 × FE2 + 1 × FE4 + 4 × FE6 31,210 10 7 6 0.15
ECSA-independent 2 × FE2 + 1 × FE4 + 3 × FE6 23,914 10 7 6 0.35
Fifth-order elliptic filter
FDS-fast 3 × FE2 + 3 × FE6 23,883 26 8 17 0.02
FDS-slow 5 × FE1 + 3 × FE5 10,970 30 6 78 0.03
ECSA-basic 2 × FE2 + 1 × FE5 + 2 × FE6 18,962 29 4 21 3.80
ECSA-independent 4 × FE2 + 4 × FE6 31,844 30 8 17 1.60
Fifth-order elliptic filter with start registers
FDS-fast 3 × FE2 + 3 × FE6 23,883 21 16 17 0.02
FDS-slow 5 × FE1 + 3 × FE5 10,970 25 16 78 0.04
ECSA-basic 2 × FE2 + 1 × FE5 + 2 × FE6 18,962 24 16 19 4.80
ECSA-independent 2 × FE2 + 2 × FE6 15,922 18 9 21 3.40
Bandpass filter
FDS-fast 3 × FE2 + 4 × FE6 31,179 34 10 10 0.01
FDS-slow 4 × FE1 + 3 × FE5 10,600 35 8 44 0.04
ECSA-basic 3 × FE2 + 3 × FE6 23,883 33 8 11 1.70
ECSA-independent 3 × FE2 + 1 × FE5 + 4 × FE6 34,219 33 8 10 1.30
Bandpass filter with start registers
FDS-fast 3 × FE2 + 4 × FE6 31,179 25 23 10 0.02
FDS-slow 4 × FE1 + 3 × FE5 10,600 26 23 44 0.04
ECSA-basic 3 × FE2 + 4 × FE6 31,179 23 19 10 2.40
ECSA-independent 3 × FE2 + 4 × FE6 31,179 23 19 10 2.90
Least-mean-square filter
FDS-fast 3 × FE2 + 6 × FE6 45,771 68 12 13 0.40
FDS-slow 3 × FE1 + 4 × FE5 13,270 72 8 70 2.79
ECSA-basic 3 × FE2 + 6 × FE5 + 3 × FE6 42,123 67 10 14 6.30
ECSA-independent 9 × FE2 + 6 × FE5 + 9 × FE6 89,889 69 30 15 8.05
Least-mean-square filter with start registers
FDS-fast 3 × FE2 + 6 × FE6 45,771 33 29 13 0.48
FDS-slow 3 × FE1 + 4 × FE5 13,270 37 27 70 3.52
ECSA-basic 4 × FE2 + 2 × FE5 + 5 × FE6 45,220 32 25 13 9.20
ECSA-independent 5 × FE2 + 3 × FE5 + 7 × FE6 63,517 33 25 13 12.30

29.3.2.4.4 Least-Mean-Square Filter

At the expense of a small increase in the delay, the basic ECSA is able to decrease the size and lower the
number of register and buses of the ordinary DFG; but the independent ECSA is not able to improve
any parameter. When dealing with the improved DFG, the basic ECSA is able to keep the initial delay, to
decrease the circuit size and to lower the number of required registers and buses. The independent ECSA
is only able to decrease the number of buses while increasing the circuit size.
There are slightly longer runtimes when the ECSA algorithm is used. But considering the speed (a few
seconds) and the evaluation presented in Reference 33, where the runtimes for larger circuits increase
enormously (exponentially) when the FDS algorithm is used, we can conclude that small and large

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 527 — #19

29-528 Handbook of Bioinspired Algorithms and Applications

circuits can be designed and optimized with the use of the proposed evolution-based algorithm, which
exhibits a linear increase in the design time with an increase of circuit size.

29.4 Conclusions
In the UM geometry optimization we used an evolutionary approach to improve the efficiency of a UM,
the motor that is typically used in home appliances and power tools. The goal of our optimization was
to find the new set of independent geometrical parameters of the rotor and the stator with the aim of
reducing the motor’s power losses, which occur in the iron and the copper. The approach proves to be
a simple and efficient search-and-optimization method for solving this day-to-day design problem in
industry. It outperforms, by a significant improvement of the motor’s efficiency, a conventional design
procedure that was used previously. By using the GA we are able to reduce the iron and the copper losses
of an initial UM by at least 20%, and increasing the GA running time or setting its parameters more
appropriately could improve on this result.
In the IC area/time optimization we used an evolutionary approach to some parts of IC design. The work
was focused on ASICs that need an even more sophisticated design due to their specific use. Optimally
scheduled operations are not necessarily optimally allocated to functional units. To ensure optimum
allocation we need to consider some allocation criteria while the scheduling is being done. The evolutionary
approach considers scheduling and allocation constraints and ensures a globally optimal solution in a
reasonable time. To evaluate our method we built an algorithm and implemented it with a computer. It
is used with a group of test-bench ICs. These circuits are chosen because the same types were used in
similar studies. They differ in terms of their size and the number of operation types. The results of the
evaluation of a computer-implemented algorithm show that the evolutionary methods are able to find a
solution that is more appropriate in terms of all the considered and important objectives than the classical
deterministic methods.

References
[1] Sen, P.C. Principles of Electric Machines and Power Electronics. John Wiley & Sons, New York, 1996.
[2] Gajski, D. et al. High-Level Synthesis: Introduction to Chip and System Design. Kluwer Academic
Publishers, Boston, MA, 1992.
[3] Armstrong, J.R. and Gray, F.G. VHDL Design: Representation and Synthesis. Prentice Hall PTR,
Upper Saddle River, NJ, 2000.
[4] Landwehr, B., Marwedel, P., and Dömer, R. Optimum simultaneous scheduling, allocation and
resource binding based on integer programming. In Proceedings of EuroDAC, Grenoble, France,
1994, p. 90.
[5] Zhu, J. and Gajski, D. Soft scheduling in high level synthesis. In Proceedings of the 36th ACM/IEEE
Design Automation Conference, New Orleans, LA, 1999, p. 219.
[6] Mandal, C. and Zimmer, R.M. A genetic algorithm for the synthesis of structured data paths.
In Proceedings of the IEEE VLSI Design, Calcutta, India, 2000, p. 206.
[7] Kim, T. Scheduling and Allocation Problems in High-Level Synthesis, Ph.D. thesis, University of
Illinois at Urbana-Champaign, 1993.
[8] Grajcar, M. Genetic list scheduling algorithm for scheduling and allocation on a loosely coupled
heterogeneous multiprocessor system. In Proceedings of the 36th ACM/IEEE Design Automation
Conference, New Orleans, LA, 1999, p. 280.
[9] Papa, G. Concurrent operation scheduling and unit allocation with an evolutionary technique in
the process of integrated-circuit design, Ph.D. thesis, Faculty of Electrical Engineering, University
of Ljubljana, 2002.
[10] Karr, C.L., Yakushin, I., and Nicolosi, K. Solving inverse initial-value, boundary-value problems
via genetic algorithm. Engineering Applications of Artificial Intelligence, 13, 625, 2000.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 528 — #20

An Evolutionary Approach to Problems in Electrical Engineering 29-529

[11] Bäck, T. Evolutionary Algorithms in Theory and Practice. Oxford University Press, New York, 1996.
[12] Goldberg, D.E. Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley, Reading, MA, 1989.
[13] Drechsler, R. Evolutionary Algorithms for VLSI CAD. Kluwer Academic Publishers, Boston, MA,
1998.
[14] Papa, G. et al. Universal motor efficiency improvement using evolutionary optimization. IEEE
Transactions on Industrial Electronics, 50, 602, 2003.
[15] Filipič, B. and Štrancar, J. Tuning EPR spectral parameters with a genetic algorithm. Applied Soft
Computing, 1, 83, 2001.
[16] Papa, G. and Šilc, J. Automatic large-scale integrated circuit synthesis using allocation-based
scheduling algorithm. Microprocessors and Microsystems, 26, 139, 2002.
[17] Koroušić-Seljak, B. Heuristic methods for a combinatorial optimization problem — real-time
task scheduling problem. In Smart Engineering System Design: Neural Networks, Fuzzy Logic,
Evolutionary Programming, Data Mining, and Complex Systems, Dagli, C.H. et al., Eds., ASME
Press, New York, 1999, p. 1041.
[18] Dasgupta, D. and Michalewicz, Z. Evolutionary Algorithms in Engineering Applications. Springer-
Verlag, Berlin, 1997.
[19] Coello Coello, C.A. A comprehensive survey of evolutionary-based multiobjective optimization
techniques. Knowledge and Information Systems, 1, 269, 1999.
[20] Van Veldhuizen, D. and Lamont, G.B. Multiobjective evolutionary algorithms: analyzing the state-
of-the-art. Evolutionary Computation, 8, 125, 2000.
[21] ANSYS User’s Manual, ANSYS version 5.6, 2000.
[22] Papa, G. DOptiMeL — User’s Manual. Technical report IJS DP 8440, Jožef Stefan Institute,
Ljubljana, 2001.
[23] Paulin, P.G., Knight, J.P., and Girczyc, E.F. HAL: a multiparadigm approach to automatic data
path synthesis. In Proceedings of the 23rd ACM/IEEE Design Automation Conference, Las Vegas, NV,
1986, p. 263.
[24] Kung, T., Whitehouse, H.J., and Kailath, T. VLSI and Modern Signal Processing. Prentice Hall,
New York, 1985.
[25] Grewal, G.W. and Wilson, T.C. An enhanced genetic algorithm for solving the high-level syn-
thesis problems of scheduling, allocation, and binding. International Journal of Computational
Intelligence and Application, 1, 91, 2001.
[26] Benesty, J. and Duhamel, P. A fast exact least square adaptive algorithm. IEEE Transactions on
Signal Processing, 40, 2904, 1992.
[27] Parhami, B. Computer Arithmetic: Algorithms and Hardware Designs. Oxford University Press, New
York, 2000.
[28] Šilc, J., Robič, B., and Ungerer, T. Processor Architecture: From Dataflow to Superscalar and Beyond.
Springer-Verlag, Berlin, 1999.
[29] Thornton, M.A., Gaiche, J.D., and Lemieux, J.V. Trade-off analysis of integer multiplier circuits
implemented in FPGAs. In Proceedings of the IEEE Pacific Rim Conference on Communications,
Computers and Signal Processing, Victoria, Canada, 1999, p. 301.
[30] Papa, G. and Šilc, J. Evolutionary synthesis algorithm — genetic operators tuning. In Advances in
Intelligent Systems, Fuzzy Systems, Evolutionary Computation, Grmela, A. and Mastorakis, N., Eds.,
WSEAS Press, Athens, 2002, p. 256.
[31] Paulin, P.G. and Knight, J.P. Scheduling and binding algorithms for high-level synthesis. In
Proceedings of the 26th ACM/IEEE Design Automation Conference, Las Vegas, NV, 1989, p. 1.
[32] Paulin, P.G. and Knight, J.P. Force-directed scheduling in automatic data path synthesis. In
Proceedings of the 24th ACM/IEEE Design Automation Conference, Miami, FL, 1987, p. 195.
[33] Papa, G., Šilc, J., and Wyrzykowski, R. Scheduling algorithms based on genetic approach. In
Proceedings of the 4th Conference on Neural Networks and Their Application, Zakopane, Poland,
1999, p. 469.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C029” — 2005/8/6 — 14:13 — page 529 — #21

 30
Solving the
Partitioning Problem
in Distributed Virtual
Environment Systems
Using Evolutive
Algorithms

30.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30-531

30.2 The Partitioning Problem in DVE Systems . . . . . . . . . . . .30-533
Architectures for DVE Systems • Linear Optimization
Technique (LOT) • Other Approaches to the Partitioning
Problem
30.3 Metaheuristic Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30-536
Obtaining the Initial Population • Genetic Algorithms (GA) •
Improving the Performance of the Proposed GA Approach •
Ant Colony System • Simulated Annealing (SA) • Greedy
Randomized Adaptive Search (GRASP)
Pedro Morillo 30.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30-549
Marcos Fernández 30.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30-551
Juan Manuel Orduña References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30-551

30.1 Introduction
The widespread use of both fast Internet connections and also high-performance graphic cards have made
possible the current growth of Distributed Virtual Environment (DVE) systems. These systems allow
multiple users, working on different computers that are interconnected through different networks (and
even through Internet) to interact in a shared virtual world. This is achieved by rendering images of the
environment as the user would perceive them if he was located at that point of the virtual environment.
Each user is represented in the shared virtual environment by an entity called avatar, whose state is
controlled by the user. Since DVE systems support visual interactions between multiple avatars, every
change in each avatar must be notified to the neighboring avatars in the shared virtual environment.

30-531

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 531 — #1

30-532 Handbook of Bioinspired Algorithms and Applications

DVE systems are currently used in different applications [1], such as collaborative design [2], civil and
military distributed training [3], e-learning [4], or multiplayer games [1,5–7].
Designing an efficient DVE system is a complex task, since these system show an inherent hetero-
geneousness. Such heterogeneousness appears in several elements:
Hardware: Each client computer controlling an avatar may have installed different hardware: a very
different range of resources like processor speed, memory size, and graphic card technology can be
specified for different client computer.
Connection: Different connections can be found in a single system. From shared medium topologies
like Ethernet or Fast-Ethernet to other network connections like ISDN, fiber-optic or ATM can be
simultaneously found in some DVEs.
Communication rate of avatars: Depending on the application, different communication rates of avatars
can be found. For example, the communication rate of avatars in a collaborative three-dimensional
(3D) environment may greatly differ from the communication rate of avatars in a 3D virtual
military battle.
Additionally, other factors help to increase the complexity of designing an efficient DVE system. Each
of them have now become an open research field:
Data model: This concept describes some conceivable ways of distributing persistent or semipersistent
data in a DVE [8]. Data can be managed in a replicated, shared, or distributed methodology.
Communication model: Network bandwidth determines the size and performance of a DVE. The system
behavior is related to the way all the scene clients are connected. Broadcast, peer-to-peer, or unicast
schemes define different network latency values for exchanging information between avatars.
View consistency: This problem has already been defined in other computer science fields such as
database management [9]. In DVE systems, this problem involves ensuring that all avatars sharing
a virtual space with common objects have the same local vision of them.
Message traffic reduction: Keeping a low amount of messages allows DVE systems to efficiently scale
with the number of avatars in the system. Traditionally, techniques like dead-reckoning described
in Reference 1 offered some level of independence to the avatars. With network support, broadcast
or multicast solutions [10,11] decrease the number of messages used to keep a consistent state of
the system.
Most of the issues described above are related to the partitioning problem or p-problem. This problem
consists of efficiently distributing the workload (avatars) among different servers in the system [12]. The
partitioning problem may seriously affect the overall performance of the DVE system, since it deter-
mines not only the workload that each server must support, but also the inter-server communication
requirements (and therefore the network traffic).
Some methods for solving the partitioning problems have been already proposed [13–15]. These
methods provide efficient solutions even for large DVE systems. However, there are still some features in
the proposed methods that can still be improved. For example, different heuristic search methods can be
used for finding the best assignment of clients to servers, instead of using ad hoc heuristics. In this chapter,
we present a comparison study of several evolutive heuristics for solving the partitioning problem in DVE
systems. We have implemented five different heuristics, ranging over most of the current taxonomy of
heuristics: Genetic Algorithms (GAs) [16], two different implementations of Simulated Annealing [17],
Ant Colony Systems (ACSs) [18], and Greedy Randomized Adaptive Search (GRASP) [19]. Performance
evaluation results show that the execution cost of the partitioning algorithm (in terms of execution times)
can be dramatically reduced, while providing similar or even better solutions than the ones provided by
the ad hoc heuristic proposed in [14].
The rest of the chapter is organized as follows: Section 30.2 describes the partitioning problem and
the existing proposals for solving it. Section 30.3 describes the proposed implementations of the heur-
istics considered for this study. Next, Section 30.4 presents the performance evaluation of the proposed
heuristics. Finally, Section 30.5 presents some concluding remarks.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 532 — #2

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-533

30.2 The Partitioning Problem in DVE Systems

30.2.1 Architectures for DVE Systems
Several architectures have been traditionally used for simulating a large set of avatars sharing the same
virtual world. Internet multiplayer games as Quake [5] and Kali [7] or educational systems as VES [4]
are examples of client–server systems (Figure 30.1[b]). In these applications, each client computer has a
single connection to the only existing server in the system. This server maintains the global state of the
simulation, but it becomes a single point of failure in the system. Instead of sending messages to a central
server, in peer-to-peer architectures (Figure 30.1[a]) avatars exchange messages directly. Several systems
have been developed with this architecture, such as NPSNET [10]. Although these systems obtain low
latencies, they do not properly scale. When the number of avatars greatly increases, clients are not able to
handle the amount of messages from other avatars and simultaneously offering an interactive 3D virtual
world to the user. In order to improve scalability, peer–server and server–network architectures group sets
of avatars. Following a peer–server scheme (Figure 30.1[d]), systems like ATLAS [11] reduces the volume
of information using multicast messaging. However, this architecture will be useful only when multicast
protocols are fully available in the Internet [13].
Architectures based on networked servers are becoming a de facto standard for DVE systems
[1,12,15]. In these architectures, the control of the simulation relies on several interconnected servers.
Figure 30.1(c) depicts how multi-platform client computers are attached to only one of the servers of the
simulation.
In this architecture, when a client modifies an avatar, it also sends an updating message to its server, that
in turn must propagate this message to other servers and clients. Servers must render different 3D models,
perform positional updates of avatars and transfer control information among different clients. Thus,
each new avatar represents an increase in both the computational requirements of the application and
also in the amount of network traffic. When the number of connected clients increases, the number of
updating messages must be limited in order to avoid a message outburst. In this sense, concepts like
Areas of Interest (AOI) [1], locales [20], or auras [21] have been proposed for limiting the number of
neighboring avatars that a given avatar must communicate with. All these concepts define a neighborhood
area for avatars, in such a way that a given avatar must notify his movements (by sending an updating
message) only to those avatars located in that neighborhood. Depending on their origin and destination
avatars, messages in a DVE system can be intra- or inter-server messages (see Figure 30.2). Inter-server
messages are those messages involving two or more servers. On the contrary, intra-server messages are
those messages exchanged between avatars whose client computers are attached to the same server. In order
to design a scalable DVE systems, the number of intra-server messages must be maximized. Effectively,
when clients send intra-server messages they only concern a single server. Therefore, they are minimizing

(a) (c)
P P C C
S S

P P C C

(b) (d) P
C C
P P

S S

P P
C C
P

FIGURE 30.1 Architectures: (a) peer-to-peer, (b) server–network, (c) client–server, and (d) peer-to-server.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 533 — #3

30-534 Handbook of Bioinspired Algorithms and Applications

Inter-server communication

Server1 Inner-server
communication
Server2

LAN–WAN
Server3

LAN–WAN

AOI of avatars

FIGURE 30.2 A multi-server architecture for a basic DVE.

the computing, storage, and communication requirements for maintaining a consistent state of the avatars
in a DVE system.
The partitioning problem consists of efficiently distributing the workload (assigning avatars) among
the different servers in the system. Lui and Chan have shown the key role of finding a good assignment
of avatars to servers in order to ensure both a good frame rate and a minimum network traffic in DVE
systems [12,14]. They propose a quality function, denoted as Cp , for evaluating each partition (assignment
of avatars to servers). This quality function takes into account two parameters. One of them consists of the
computing workload generated by clients in the DVE system, and is denoted as CpW . In order to minimize
this parameter, the computing workload should be proportionally shared among all the servers in the DVE
system, according to the computing resources of each server. The other parameter of the quality function
consists of the overall number of inter-server messages, and it is denoted as CpL . In order to minimize this
parameter, avatars sharing the same AOI should be assigned to the same server. Thus, quality function Cp
is defined as

Cp = W1 CpW + W2 CpL , (30.1)

where W1 + W2 = 1. W1 and W2 are two coefficients that weighs the relative importance of the com-
putational and communication workload, respectively. These coefficients should be tuned according to
the specific features of each DVE system. Using this quality function (and assuming W1 = W2 = 0.5)
Lui and Chan propose a partitioning algorithm that reassigns clients to servers [14]. The partitioning
algorithm should be periodically executed for adapting the partition to the current state of the DVE
system as it evolves (avatars can join or leave the DVE system at any moment, and they can also move
everywhere within the simulated virtual world). Lui and Chan also have proposed a testing platform for
the performance evaluation of DVE systems, as well as a parallelization of the partitioning algorithm [14].
The partitioning method proposed by Lui and Chan, known as LOT or Linear Optimization Technique,
currently provides the best results for DVE systems. However, it uses an ad hoc heuristic. We propose
a comparative study of several heuristics, ranging over most of the current taxonomy of heuristics,
in order to determine which one provides the best performance when applied to the partitioning problem
in DVE systems. In this study, we propose the same approach as Lui–Chan: using the same quality
function, we will obtain an initial partition (assignment) of avatars to servers, and then we will test the
implementation of each heuristic to provide a near optimal assignment.

30.2.2 Linear Optimization Technique (LOT)

Linear Optimization Technique (LOT) was initially published by Lui et al. [12] and revisited in Reference 14
from their ideas published in Reference 22 about graph theory.
This ad hoc approach models the 3D virtual scene as a graph. Each avatar is modeled by a node and
two avatars are linked by an edge if their AOI collide. Using Cp as the evaluation function, LOT provides
an efficient partition of this graph following three steps. The first step is named the Recursive Bisection
Procedure (RBP), and it consists of using a divide-and-conquer procedure that provides an initial partition.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 534 — #4

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-535

(a) (b) (c)

Server 2
Server 2
3 2 3 2 3 2
5 5 Server 2 5
3 3 3
5 5 5
4 10 4 10 4 10
7 7 7 Server 0
Server 0 Server 0
3 9 3 9 4 3 9
4 3 3 4 3
1 2 1 2 1 2
9 2 9 2 9 2
7 5 7 7
5 1 5
2 1 2 2 1

Server 1 Server 1 Server 1

FIGURE 30.3 Partition of a DVE performed by LOT method: (a) RBP, (b) CRP, and (c) LP.

Then, the layering partitioning procedure (LP) and the Communication Refinement Partitioning (CRP) are
applied on that initial partition of the graph. Each of these procedures performs workload balancing and
minimizes the number of inter-server messages, respectively.
Figure 30.3 shows the partitions that this method would provide when applied to a small DVE system.
In this example, a DVE composed by 10 avatars is simulated with three identical servers. Nodes and edges
obtained from the associated graph has been labeled. The label of a node (avatar) represents a estimation
of the workload generated by this avatar to the server where it is going to be assigned. Ranging from 1 to 10,
the label of each edge represents the nearness of these two avatars. Figure 30.3(a) represents the result
obtained by a RBP phase. Although the number of avatars assigned to each server seems to be balanced, the
workload that those avatars generate must be also uniform in order to achieve actual workload balancing.
By adding the labels of all the nodes assigned to the same server, we obtain that the workload assigned
to each server is 16, 7, and 12 units of workload, respectively. Then, CRP balances (Figure 30.3[b]) the
existing workload in sets of 12, 11, and 12 units. At that point, the number of inter-server messages
generated by these sets of avatars is reduced by LP, as shown in Figure 30.3(c). Since strategies of CRP and
LP techniques could be opposed, this couple of steps is repeated three times.

30.2.3 Other Approaches to the Partitioning Problem

In addition to LOT technique, two other relevant approaches have been also proposed for solving the
partitioning problem. One of them divides the whole virtual scene into hexagonal cells, and each cell
is mapped to a multicast group [13]. With this division, all avatars located in the same cell share the
same multicast address. Since the size of AOI is bigger than the area of the cells, avatars can list different
multicast addresses but they can only send information to the cell where they are located.
Figure 30.4 shows an example of this technique. In this figure, the represented avatar is assigned to
group A and also lists multicast information from groups B, C, E, F, and G. In order to ensure efficient
management, this algorithm generates a group leader for each cell. The oldest avatar is selected as the
group leader, using a timestamp mechanism. This leader controls how avatars join and leave the multicast
group, and he also implements a flow control mechanism for the messages.
Despite its apparent scalability, this approach presents several problems. First, it does not obtain good
performance when avatars are located following a nonuniform distribution. Second, system performance
is very low when avatars can quickly cross the virtual scene. This is due to the associated cost of leaving
and joining different multicast groups. Finally, multicast is not fully implemented within the Internet.
Another different approach was proposed by Tam [15]. In this method, dynamic concepts associated
with avatars like AOI, aura, or locale are rejected. This partitioning method divides the scene in square cells.
The side length of the cells S is related to the radius of the AOI of avatars, following the next relation:

S
≤ AOI ≤ S. (30.2)
2

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 535 — #5

30-536 Handbook of Bioinspired Algorithms and Applications

DVE world
B
G C 4 3 2

F D
E 3 2 3

2 1 5

Cell division of the virtual scene Graph representation

FIGURE 30.4 Other approaches: cell division and graph representation.

A graph representation of the virtual scene is obtained from the volume of avatars contained in each
cell, as also shown in Figure 30.4. Next, this graph is divided in partitions and each partition is assigned
to a server of the DVE. In order to accomplish this division an exhaustive and greedy algorithms are
compared. Using a quality different from Cp , these algorithms take also into account the number of inter-
server messages and workload balancing among the servers. Although this approach provides a fast way of
solving the partitioning problem, the performance of the static partitioning is quite low when avatars show
a clustered distribution. In this case, the servers controlling the crowded areas are overloaded, increasing
the overall cost of the quality function.

30.3 Metaheuristic Procedures

Evolutive computing has become a important paradigm in the development of high-quality solution for
NP-Complete problems. This paradigm, which models computational problems as natural processes,
is used to provide solutions for complex problems that in most cases could not be solved by using
deterministic computing techniques. Problems such as multi-objective optimization have been success-
fully adapted using these strategies [23]. In this section, we present five implementations of different
heuristics for solving the partitioning problem in DVE systems. Following the approach presented by Lui
and Chan (and using the same quality function Cp ), the idea is dynamically applying a heuristic search
method that provides a good assignment of clients to servers as the state of the DVE system changes.
All the presented solutions have been adapted in the same way. That is, a initial partition obtains good
assignments only for several avatars, while the evolutive algorithm itself performs successive refinements
of the initial partition that lead to a near optimal partition of the DVE system. Also, the evolutive
method is used periodically for updating the obtained partition to the current state of the DVE system
(avatars change their locations, new avatars can join the system, and some avatars can leave the system at
any time).
In this section, we describe the implementation of each heuristic search method and the tuning of
its parameters for solving the partitioning problem. Concretely, this tuning has been performed on the
DVE evaluation test described by Lui and Chan in [14]. This test defines two kinds of DVE systems:
a SMALL virtual world, composed of 13 avatars and 3 servers and also a LARGE world, composed of
2500 avatars and 8 servers. Effectively, the purpose of solving the partitioning problem is to provide
scalable DVE systems. Therefore, the partitioning method must efficiently work in LARGE virtual worlds.
On other hand, since the performance of the method may heavily depend on the location of avatars,
this evaluation test also considers three different distributions of avatars: uniform, skewed, and clustered
distribution.

30.3.1 Obtaining the Initial Population

All of the implemented heuristics start from an initial partition (assignment) of the n avatars in the
DVE system. We tested several clustering algorithms for obtaining this initial partition. Although they

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 536 — #6

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-537

are not shown here due to space limitations, we obtained the best results for a Density-Based Algorithm
(DBA) [24].
This algorithm divides the virtual 3D scene in square sections. Each section is labeled with the number
of avatars that it contains (na), and all the sections are sorted (using Quick-sort algorithm) by their na
value. The first S sections in the sorted list are then selected and assigned to a given server, where S is the
number of servers in the DVE system. That is, all the avatars in a selected region are assigned to a single
server. The next step consists of computing the mass-center (mc) of the avatars assigned to each server.
Using a round-robin scheme, the algorithm then chooses the closest free avatar to the mc of each server,
assigning that avatar to that server, until all avatars are assigned.
The proposed implementation of the DBA method consists of the following steps (expressed as pseudo-
code statements):
program Initial_Partition (avatar, Int n, Int S)

const
n_sections = 25x25

type Cell
sum,idx: Int

var
assigned,represent:Int[]
pivot,elect,ncentr:Int
min_dis,dist_tmp :Real
na :Cell[n_sections]

begin
DivideSceneInSquareSections(n_sections)
for i:=0 to n_sections do
na[i].sum:=CountAvatarsInSection(i)
na[i].idx:=i
end_for
QuickSort (na)
for i=0 to S do
representant [i]= ObtainMC (na[i].idx)
end_for;
pivot := 0, elect := -1, ncentr:=0
for i:=0 to n do
elect:=-1,min_dis := 100000
for j:=0 to n do
if (assigned[j] = NOT_ASSIGNED)
dist_tmp:= EuclideanDistance(avatar[j],represent([ncentr])
if (dist_tmp < min_dis)
pivot := j
min_dis := dist_tmp
endif
endif
end_for
avatar[pivot].assignment := assigned[pivot] := ncentr
ncentr : (ncentr + 1) mod S
end_for
end

Since the assignment of avatars follows a round-robin scheme, this algorithm provides a good bal-
ancing of the computing workload (the number of avatars assigned to each server does not differ in

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 537 — #7

30-538 Handbook of Bioinspired Algorithms and Applications

more than one). On other hand, avatars that are grouped in a small region and close to the mass-center of
a server will be assigned to that server by the DBA. Additionally, since these avatars are located so closely,
they will probably will share the same AOI. Therefore, the DBA also provides an initial partition with low
inter-server communication requirements for those avatars.
However, the assignment of avatars equidistant (or located far away) from the mass-center of the
servers is critical for obtaining a partition with minimum inter-server communication requirements
(minimum values of the quality function Cp ), particularly for SMALL virtual worlds with only a few
servers. DBA inherently provides good assignments for clustered avatars, but it does not properly focus
on the assignment of these critical avatars. Each of the following evolutive methods should be used at this
point to search a near optimal assignment that properly reassigns these avatars.

30.3.2 Genetic Algorithms (GA)

This heuristic consists of a search method based on the concept of evolution by natural selection [16].
GA starts with an initial population (the initial partition) and then it evolves a certain number of gen-
erations (iterations), providing an evolved population (final solution). The proposed implementation for
solving the partitioning problem, proposed initially in Reference 25, starts with a population composed of
a set of elements called genomes or chromosomes. The number of chromosomes is the number of partial
solutions that each iteration must maintain. Each chromosome is defined by a descriptor vector containing
a given assignment of avatars to servers. The length of this vector is equal to the number of Border Avatars
(BA) within the DVE system. BA is an avatar that, although it lies within the AOI of another avatar, it is
not assigned to the same server. These kinds of avatars offers a higher probability of giving a successful
permutation, since they minimize inter-server communication.
Starting from the initial population, this approach applies an auto-fertilization mechanism for gener-
ating new chromosomes [26]. This mechanism is based on a single-point random crossover, where each
generation is found by exchanging some elements of the restricted population in such a way that in each of
the N chromosomes, two border avatars assigned to different servers are randomly chosen and exchanged.
Thus, an iteration performed on a population of N chromosomes will produce a new population of 2N .
From these 2N chromosomes, the N elements with the lower value of Cp will be chosen.
GA is also capable of escaping from local minima due to the mutation process. Once the child-vector
has been obtained, mutation involves changing at random the server assigned to one of the elements of the
population. Some other mutations such as the modification of several elements or the crossover between
the characteristics of pairs of chromosomes have also been tested. But after several tests it has been found
that the mutation of just one “bit” is the one offering the best results. It is important to mention that the
mutation is a random process controlled by a parameter. This parameter needs to be carefully chosen for
each specific experiment in order to achieve solutions with low Cp System Cost.
Next code describes the behavior of the proposed approach based on GA. It has been expressed as
pseudo-code statements:
program GA (Int chromo, Int iterations, Real mut_rate)

var
B,temp_cost,Cp_GA :Real
av_i,av_j,av_k :Integer

begin
Initial_Partition (DBA)
B := ObtainBorderAvatars()
Cp_GA := Compute_Cp()
For i:=0 to iterations do
For j:=0 to chromo do
SelectAndCopyChromosome(j)
Choose2DifRamdomAvatars(B,av_i,av_j)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 538 — #8

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-539

ExchangeServerAssignment(av_i,av_j)
temp_cost := Compute_Cp()
if (HaveItoMutate(mut_rate))
Chose1RandomAvatar(av_i)
ForceServerAssignment(av_i)
endif
AddChromosomeToPopulation(this)
end_for
CWPSortPopulationByCp(2*chromo)
Cp_GA := SelectBestIndividuals(chromo)
end_for
end

Figure 30.5 shows a example of generation of new avatars in a SMALL DVE system where six BA have
been obtained from the full set of avatars. These avatars define a chromosome, and they can be assigned
to any of the three servers in the DVE system. This figure represents a possible crossover and mutation on
this chromosome.
Although this basic GA approach performs reasonably well, it is based on the generation of an ini-
tial population obtained by deriving a unique solution provided by a clustering algorithm described in
Section 30.3.1. Despite the fact that this feature offers an improved initial population of feasible solutions,
it does not focus on maximizing the structural diversity of chromosomes. As described in References 27
and 28, this low level of structural diversity can lead the algorithm to reach a local minimum or even
a poorer approximation of this value. Additionally, the crossover mechanism used by this algorithm is
based on an auto-fertilization technique, where chromosomes are derived following a single-point cross-
over [26]. This crossover mechanism is excessively generalist, and it is possible to offer new crossover
strategies more oriented to problem specifications.

30.3.3 Improving the Performance of the Proposed GA Approach

In order to improve the performance of our current approach for solving the partitioning problem in DVE
systems, we propose a refined version of the genetic algorithm. This version focuses on maximizing the
structural diversity of initial population and improving the crossover operator. The new algorithm replaces
the current generation of the initial population, described in Section 30.3.1 with a new method based
on random projections. Additionally, the crossover is performed by randomly choosing a mechanism
from a list composed of five different crossover operators. All these five operators are very oriented to the
specifications of partitioning problem in DVE systems.

A2 A10 A12 A16 A34 A36 S0 S1 S2

Boarder avatars (clients) Servers

S0 S1 S0 S2 S2 S0 Chromosome

S2 S1 S0 S2 S0 S0 Crossover (swapping)

S0 S1 S0 S2 S2 S0 Mutation (Gene)

S1

FIGURE 30.5 Generation of new individuals in GA approach.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 539 — #9

30-540 Handbook of Bioinspired Algorithms and Applications

30.3.3.1 Generation of a Heuristic Initial Population

Most metaheuristics are based on the fast generation of an initial population of elements [27,29]. This
initial population usually represents a set of poor solutions to the problem, and it is evolved through
a crossover operator until a stopping criterion (for example, a given number of iterations) is reached.
If the initial population has been correctly defined, then the metaheuristic algorithm easily obtains
a good approximation to the global optimum. Moreover, as described in Reference 27, if the initial
population is not randomly selected then the algorithm should maintain a certain level of structural
diversity among all the chromosomes, in order to properly represent the whole set of feasible solutions
and to avoid the premature convergence of the search.
Taking into account these considerations, we propose a new mechanism, called Projections algorithm
(PA), in order to generate the new population of initial solutions. This fast algorithm provides a set of
independent and well-diversified initial solutions to the problem.
PA consist of a given number of iterations nc that defines the number of chromosomes in the population
(population size). Each one of the nc chromosomes represents a complete partitioning solution to the
problem where all the N avatars (A0 , . . . , AN −1 ) in the DVE are assigned to the M servers (S0 , . . . , SM −1 )
in the system. For the sake of clearness, we will consider in the following description that avatars move
across a 2D virtual world. The extrapolation to 3D worlds is reasonably trivial. Each one of the nc iterations
consists of four steps illustrated in Figure 30.6.
First, each avatar in the system is assigned to server S0 (Figure 30.6[a]). Next, a random value θ between
0 and π/2 is generated. Since all avatars are located on a Cartesian plane, PA draws a straight line which
passes through the zero coordinates (0,0) with a slope of θ radians (Figure 30.6[b]). This line and its
perpendicular define a new coordinate axis that is rotated θ radians with respect to the original position.
Using a simple affine transformation (Figure 30.6[c]), the old coordinates (Xi , Yi ) of each avatar can now
be expressed with respect to the rotated axis by the new coordinates (Xi , Yi ). At this point, PA generates
two different binary search trees with Xi and Yi search keys of each avatar. Once both trees are created,
the N /M different avatars in both trees with the highest and the lowest keys are put into four different
sets [30]. In order to assign a set of avatars to a server, the third step of PA evaluates separately the different
sets of avatars using the Cp function. Since both sets have the same cardinality N /M , PA algorithm only
computes the CpL term. The term CpW is not computed, since it evaluates the standard deviation of the
assigned avatars with respect to the perfect balancing. The last step is to select the set with the lowest
CpL value, and all avatars contained in this set are assigned to server S1 (Figure 30.6[d]). At this point,
N (M − 1)/M avatars are still assigned to server S0 , and N /M avatars are assigned to server S1 . Next
iteration allows the server S0 to lose another N /M avatars, which are assigned to S2 . PA algorithm finishes
when the last group of avatars (with a size less or equal to N /M avatars) is assigned to server SM −1 . At this
point, a number of avatars very close to N /M are assigned to each server in the DVE system.
It is important to mention that the use of four binary search trees allows to improve diversity when
selecting the sets of avatars to be assigned to the target server for each iteration. The random generation of
rotation angles guarantees inherently independent solutions in the generation of individuals of the initial
population. Moreover, this population allows the genetic approach to evolve solutions with good values
of quality function Cp . These good Cp values are achieved by balancing the number of avatars among the
servers and assigning to the same server the avatars that are closely located in the virtual scene.

(a) (b) (c) (d)

y 120
y9 117
x9
u
x
86
94

FIGURE 30.6 Generation of the initial population based on Projection Algorithm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 540 — #10

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-541

30.3.3.2 Providing Randomness to Chromosome Crossover

In order to evolve chromosomes, the first approach for solving the partitioning problem in DVE systems,
based on a genetic algorithm uses an auto-fertilization technique [25]. In this technique, each chromosome
generates a new chromosome following a single-point crossover with a probability equal to one [26].
When the crossover operator is applied to all individuals of the population (the child population has
been completely created) a elitist selection guarantees the survival of the best individuals. This crossover
operator is excessively generalist and has been used often for solving different combinatorial problems.
The proposed technique is based on the random selection of crossover operators from a list. This list,
which is accessed for each derivation, consists of five operators very oriented to problem specifications.
The operator list consists of the following elements:

Operator 1: Random exchange of the current assignment for two border avatars. A given avatar Ai is a
border avatar if it is assigned to a certain server Sr in the initial partition and any of the avatars in
its AOI is assigned to a server different from Sr [31].
Operator 2: Once a border avatar Ai has been randomly selected, it is randomly assigned to one of the
servers Sf hosting the border avatars of Ai .
Operator 3: Besides the step described in the previous operator, if there exists an avatar Aj such that Aj
is assigned to Sf and it is a neighbor avatar of Ai , then Aj is assigned to Sr .
Operator 4: Since each avatar generates a certain level of workload in the server where it is assigned
to [12], then it is possible to sort the servers of a DVE system according to the level of workload
they support. If Sm and Sn are the servers with the highest and the lowest level of workload in the
system, respectively, then a random avatar Ak assigned to Sm is assigned to Sn .
Operator 5: Besides the step described in the previous operator, a random avatar Al , initially assigned
to Sn , is now assigned to Sm .

The main parameters to be tuned in GA method are the population size P, the number of iterations N
and the mutation rate M . Figure 30.7 depicts the values of que quality function Cp (denoted as system
cost) as the number of generations and individuals grows. In both cases Cp decreases significantly while
the algorithm does not reach a threshold value. From this threshold value (close to 300 for the number
of generations and 15 for the population size) the quality of the obtained solutions remains constant,
showing the impossibility of finding better solutions in this search domain. Therefore it has no sense in
spending more time in searching new partitioning solutions.
Figure 30.8 shows the tuning of the mutation rate in a given DVE system. The behavior of the algorithm
is different for this parameter. In this case, the algorithm is able to provide a high-quality solution when
mutation rate is close to 1%.
If GA approach uses values much lower than 1% for this parameter, the system can be trapped in a local
minimum. In the opposite case, when GA selects rates higher than this threshold value, then the search
method spends too much CPU time testing useless solutions.

5000 5000
Uniform Uniform
System cost Cp

4000 Skewed
System cost Cp

Skewed 4000
Clustered Clustered
3000 3000

2000 2000

1000 1000

0 0
60 120 180 240 300 360 420 480 540 600 660 3 6 9 12 15 18 21 24 27 30 33
Number of generations Number of individuals

FIGURE 30.7 Values of the quality function Cp for different number of generations and population sizes.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 541 — #11

30-542 Handbook of Bioinspired Algorithms and Applications

Uniform
5000 Skewed
1250 Clustered

System cost Cp
1000
750
500
250
0
1 2 4 6 8 10 12 14 16 18 20
Mutation rate (%)

FIGURE 30.8 Values of the quality function Cp for different mutation coefficients.

As a conclusion of this tuning phase, we can state that for this particular application GA method
provides good results for a population of 15 individuals, a mutation probability of 1%, and 100 iterations
as stop criterion.

30.3.4 Ant Colony System (ACS)

Ant Colony Systems (ACSs) search method is based on the behavior shown by ant colonies when searching
possible paths to their food [18]. They use a hormone called pheromone to communicate among them.
The path that a given ant follows when searching food depends on the amount of pheromone each possible
path contains. Additionally, when a given ant chooses a path to the food, she adds pheromone on that path,
thus increasing the probability for the ants behind her to choose the same path. This system makes the food
search to be initially random. Nevertheless, since the ants that choose the shortest path will add pheromone
more often, the probability of choosing the shortest path increases with time (positive feedback). On the
other hand, pheromone evaporates at a given rate. Therefore, the associated pheromone for each path
decreases if that path is not used during certain period of time (negative feedback). Evaporation rate
determines the ability of the system for escaping from local minima.
ACS search method has been implemented in different ways as it has been used for solving several
discrete optimization problems [32]. Concretely, we propose a new implementation of the ACS search
method, to be used for solving the partitioning problem in DVE systems. This implementation was
originally proposed in Reference 31. As with the rest of evolutive approaches, this implementation follows
two steps. First, the DBA initial partition (described in Section 30.3.1), obtains good assignments only for
several avatars. At that point, the ACS algorithm performs successive refinements of the initial partition
that lead to a near optimal partition.
The first step in the ACS method is to select the subset of border avatars from the set of all the avatars
in the system. A given avatar is selected as a border avatar if it is assigned to a certain server S in the initial
partition and any of the avatars in its AOI is assigned to a server different from S. For each of the border
avatars, a list of candidate servers is constructed, and a level of pheromone is associated to each element of
the list. This list contains all the different servers that the avatars in the same AOI are assigned to (including
the server that the avatar is currently assigned). Initially, all the elements in the list of candidate servers
have associated the same pheromone level.
ACS method consists of a population of ants. Each ant consists of performing a search through the
solution space, providing a given assignment of the B border avatars to servers. The number of ants N is
a key parameter of the ACS method that should be tuned for a good performance of the algorithm. Each
iteration of the ACS method consists of computing N different ants (assignments of the B border avatars).
When each ant is completed, if the resulting assignment of the B border avatars produces a lower value of
the quality function Cp , then this assignment is considered as a partial solution, and a certain amount of
pheromone is added to the servers that the border avatars are assigned to in this assignment (just the same
way that each ant adds pheromone to the search path that she follows). Otherwise, the ant (assignment)
is discarded. When each iteration finishes (the N ants have been computed), the pheromone level is then

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 542 — #12

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-543

equally decreased in all the candidate servers of all of the border avatars, according to the evaporation
rate (the pheromone evaporates at a given rate). The ACS method ends when all the iterations have been
performed.
In the process described earlier, each ant must assign each border avatar to one of the candidate servers
for that avatar. Thus, a selection value is computed for each of the candidate servers. The selection value
Sv is defined as

Sv = α × pheromone + β × Cp , (30.3)

where pheromone is the current pheromone level associated to that server, Cp is the resulting value of
the quality function when the border avatar is assigned to that server instead of the current server, and
α and β are weighing coefficients that must be also tuned. The server with the highest selection value will
be chosen by that ant for that border avatar.
On other hand, when a partial solution is found then the pheromone level must be increased in those
servers where the border avatars are assigned to in that solution. The pheromone level is increased using
the following formula:

1
pheromone = pheromone + Q × . (30.4)
Cp

Following this description, the proposed implementation of the ACS search method consists of the
following steps:
program ACS (Int Ants, Int iterations, Real evap_rate)

const
alpha = 1.0
beta = 7.0
Q = 1000

var
temp_sol :Real[Number_of_Avatars]
L :Integer[]
B,Cp_ACS,temp_cost :Real

begin
Initial_Partition (DBA)
B := ObtainBorderAvatars()
Cp_ACS := Compute_Cp()
For i:=0 to iterations do
For j:=0 to N do
For k:=0 to B do
L:=ChooseServer(alpha,beta,Q)
end_for
temp_sol := Compose_Solution(B)
temp_cost:= Obtain_Cp(temp_sol)
if (temp_cost < Cp_ACS)
Cp_ACS := temp_cost
IncreasePheromone (B,Q)
endif
end_for
DecreasePheromone(evap_rate)
end_for
end

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 543 — #13

30-544 Handbook of Bioinspired Algorithms and Applications

5000 5000

4000 4000
System cost Cp

System cost Cp
3000 3000

2000 2000
Uniform Uniform
1000 Skewed 1000 Skewed
Clustered Clustered
0 0
5 25 50 75 100 125 150 175 200 225 250 2 5 10 15 20 25 30 35 40 45 50
Number of ants Number of ants

FIGURE 30.9 Values of the quality function Cp for different number of iterations.

5000

4000
System cost Cp

3000

2000

Uniform
1000
Skewed
Clustered
0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5
Coefficient of evaporation (%)

FIGURE 30.10 Values of the quality function Cp for different evaporation rates.

Like in GA approach, there are some parameters in ACS search method that must be properly tuned.
In particular, the values for the number of ants N , the pheromone evaporation rate and the number of
iterations that ACS method must perform should be tuned.
Figure 30.9 shows the values of the quality function Cp (denoted as system cost ) reached by the ACS
method when different number of ants and iterations are considered. It shows that Cp decreases when
the number of iterations increases, until value of 25 iterations is reached. The same behavior manifests
when the number of ants is grown. In this case is 100 the top value. From that point, system cost Cp
slightly increases or remain constant, depending on the considered distribution of avatars. The reason
for this behavior is that the existing pheromone level keeps the search method from finding better search
paths even when more iterations are performed. Thus, the number of iterations and ants selected for ACS
method has been 25 and 100, respectively.
Finally, Figure 30.10 shows the values of Cp reached by the ACS method when different pheromone
evaporation rates are considered. This figure shows on the x-axis the percentage decreasing in pheromone
level that all candidate servers suffer after each iteration. It shows that for all the considered distributions
Cp decreases when the evaporation rate increases, until a value of 1% is reached. The reason for this
behavior is that for evaporation rates lower than 1% the pheromone level keeps the search method from
escaping from local minima, thus decreasing performance. From that point, system cost Cp also increases,
since pheromone evaporation is too high and the search method cannot properly explore good search
paths. Thus a coefficient of 1% has been selected as the optimal value of evaporation rate.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 544 — #14

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-545

Additionally, we have performed empirical studies in order to obtain the best values for α, β, and
Q coefficients. Although the results are not shown here for the sake of shortness, we have obtained the
best behavior of the ACS method for α = 1.0, β = 7.0, and Q = 1000. These are the values that the
pseudo-code shown earlier uses for ACS algorithm.

30.3.5 Simulated Annealing (SA)

SA was initially proposed for solving complex optimization problems [17]. Although some authors do
not consider SA as an evolutive technique, it can be modeled as a simplified version of GA approach. SA
has shown to be effective in the solution of a large set of applications such as staff scheduling problems,
timetabling problems and locomotive allocation problem in railway networks [33].
This heuristic search method is based on a thermodynamic theory establishing that the minimum
energy state in a system can be found if the temperature decreasing is performed slowly enough. SA
is a heuristic search method that always accepts better solutions than the current solutions, and also
accepts worse solutions according to a probability system based on the system temperature. SA starts
with a high temperature (a high probability of accepting a worsening movement), and in each iteration
system the temperature is decreased. Thus, SA can leave local minima by accepting worsening movements
at intermediate stages. The search method ends when system temperature is so low that worsening
movements are practically impossible. Since the method cannot leave local minima, it cannot find better
solutions neither (the algorithm ends when certain amount of iterations N are performed without finding
better solutions).
The proposed partitioning method based on SA was previously presented in Reference 25. As with
the rest of evolutive procedures, this approach starts from an initial population. This initial population
has been obtained with DBA algorithm. In SA method each iteration consists of randomly selecting two
different border avatars assigned to different servers. Then, the servers where this two border avatars
are assigned to are exchanged. If the resulting value of the quality function Cp is higher than the previous
one plus a threshold T , that change is rejected. Otherwise, it is accepted (the search method must decrease
the value of the quality function Cp associated with each assignment). The threshold value T used in
each iteration i of the search depends on the rate of temperature decreasing R, and it is defined in this
implementation as


R×i
T =R− , (30.5)
N

where N determines the finishing condition of the search. When N iterations are performed without
finding a partition that decreases the value of quality function Cp , then the search finishes.
The next code shows the described implementation based on SA:
program SA (Int iterations, Real dec_t_rate)

var
B,temp_cost,Cp_SA :Real
delta_sup :Real
av_i,av_j, :Integer

begin
Initial_Partition (DBA)
B := ObtainBorderAvatars()
Cp_SA := Compute_Cp()
For i:=0 to iterations do
Choose2DifRandomAvatars(B,av_i,av_j)
ExchangeServerAssignment(av_i,av_j)
temp_cost := Compute_Cp()

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 545 — #15

30-546 Handbook of Bioinspired Algorithms and Applications

delta_sup := dec_t_rate - (dec_t_rate*i/interations)

if (temp_cost - delta_sup < Cp_SA)
Cp_SA := temp_cost
else
ExchangeServerAssignment(av_i,av_j)
end_for
end

In order to improve the performance of the SA method, we measured the impact of exchanging
groups of avatars, instead of exchanging one avatar in each iteration. Table 30.1 compares the results
(in terms of both the value for the quality function Cp and also in terms of execution times) of exchanging
groups of two, three, and four avatars. We tested each option under three different distributions of avatars
in the virtual world (uniform, skewed, and clustered). These distributions are detailed in Section 30.4.
On the other hand, these results have been obtained for a LARGE world composed by 2500 avatars and
8 servers. This table shows that the best option for SA method is to exchange the lowest number of avatars
as possible in each permutation. Therefore, we exchanged a single avatar in all the simulations performed
in our study.
On other hand, the two key issues for properly tuning this heuristic search method are the number of
iterations N and the temperature decreasing rate R [33]. Figure 30.11 shows the performance (in terms
of Cp values) obtained with SA algorithm for a LARGE world when the number of iterations increases.
From Figure 30.11 we can conclude that performing more iterations results in providing better values
of Cp . However, the slope of this plot decreases from a certain number of iterations. We have con-
sidered the value of 3000 iterations at that point, and we have tested GA method with this number of
iterations.
System temperature shows a different behavior in terms of Cp . Figure 30.12 shows the values of Cp
obtained when the temperature decreasing rate is modified in a LARGE world. It clearly shows that the

TABLE 30.1 Different Types of Exchanges Performed on SA Approach

Two avatars Three avatars Four avatars

Cp Time (sec) Cp Time (sec) Cp Time (sec)

Uniform 1707.62 6.35 1808.38 6.51 1817.90 7.02

Skewed 2628.46 13.79 2826.44 14.32 2992.30 15.82
Clustered 4697.61 29.62 5046.27 30.25 5283.705 33.98

6000

5000
System cost Cp

4000

3000

2000
Uniform
1000 Skewed
Clustered
0
0

00

00

00

00

00

00

00

00

00
40

80

12

16

20

24

28

32

38

40

44

Number of iterations

FIGURE 30.11 Values of the quality function Cp for different number of iterations.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 546 — #16

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-547

6000

5000

System cost CP
4000

3000

2000
Uniform
1000 Skewed
Clustered
0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5
Temperature rate (%)

FIGURE 30.12 Values of the quality function Cp for different temperature decreasing rate.

quality of the obtained solutions do not follow a lineal progression. Effectively, since the temperature
decreasing rate allows SA approach to escape from local minima, a threshold value appears when this
parameter is modified. As this rate comes closer to 1.15, algorithm abandons local minima much more
fast, and therefore the quality of the obtained solution increases. Beyond this value, the risk of accepting
inefficient exchange of avatars is much too high and thus the algorithm is unable to find the right path.

30.3.6 Greedy Randomized Adaptive Search (GRASP)

Greedy Randomized Adaptive Search (GRASP) is a constructive technique designed as a multi-start
heuristic for combinatorial problems [34]. Although GRASP is not an evolutive technique, it has been
elected as a metaheuristic method to be compared with LOT and the set of evolutive approaches. Several
problems such as circuit partitioning, set covering, or graph planarization have been successfully addressed
by GRASP algorithms [35]. GRASP implementations are very robust and it is unusual to find examples
where this method performs badly.
GRASP implementation starts with a initial partition of the DVE. Unlike the evolutive techniques
described above, the initial partition used by GRASP does not provide any assignment for the border
avatars. In this case, GRASP exploits an important property of the DBA method presented in Section 30.3.1.
In DBA algorithm the assignment of avatars to servers is calculated iteratively. The avatar with the
lowest distance to a given server of the DVE (following a round-robin scheme) is assigned in each
iteration. In order to measure these distances, servers are represented by mass-centers (see DBA algorithm).
Therefore, as avatars are assigned they are clustered around DVE servers.
Because DBA obtains a perfect balancing of avatars (CpW ), system cost (Cp ) depends only (see
Equation 30.1) on the evaluation of inter-server communications (CpL ). Thus, the best partitioning solu-
tions are obtained when neighboring avatars are assigned to the same server. This property makes avatars
that are located far away or equidistant of different mass-centers to be critical. These avatars are denoted as
critical avatars. In order to properly assign critical avatars it is necessary to spend more resources. In order
to avoid such spending, a different version of DBA algorithm is used. This version, denoted as DBA-R,
does not assign critical avatars. These avatars are the inputs to GRASP algorithm.
Figure 30.13 shows an example of a DBA execution. In this example 66 avatars are simulated with three
servers in a DVE system. Each server received a balanced group of 18 elements each located in a round
pattern. The remaining 12 avatars have been intentionally nonassigned.
The GRASP-based method for solving the partitioning problem in DVE systems consists of several
iterations [36]. Each iteration assigns a critical avatar and consists of two steps: construction and local
search. The construction phase builds a feasible solution choosing one border avatar by iteration, and the

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 547 — #17

30-548 Handbook of Bioinspired Algorithms and Applications

Nonassigned

Server C

Server B

Server A

FIGURE 30.13 Result from a initial solution obtained by DBA-R algorithm.

local search also provides a server assignment of that border avatar in the same iteration, following the
next procedure: first, the resulting cost Cp of adding each nonassigned critical avatar to the current (initial)
partition is computed. Since each border avatar can be assigned to different servers, the cost for assigning
each border avatar to each server is computed, forming the List of Candidates (LCs) (each element in this
list has the form (nonassigned border avatar, server, resulting cost ). This list is sorted (using Quick-sort
algorithm) by the resulting cost Cp in descendent order, and then is reduced to its top quartile. One
element of this reduced list of candidates (RLCs) is then randomly chosen (construction phase). Next,
an extensive search is performed in the AOI of that selected avatar. That is, all the possible assignments
of the avatars in the AOI of the selected avatars are computed, and the assignment with the lowest Cp
is kept.
Next code describes how the GRASP approach works when n avatars are assigned to S servers in a DVE
system:

program GRASP (Int threshold)

type New_sol
idx_av,idx_ser: Int
new_cost : Real
avatar,server : Int

var
tmp_cost,Cp_GRASP :Real
Non_assig :Integer
list :New_sol[]

begin
Initial_Partition (DBA-R,threshold)
non_assig := n - threshold
For i:=0 to non_assig do
for j:=0 to n do
for k:=0 to S do
tmp_cost=TestSolution(j,k)
AddToList(list,tmp_cost,j,k)
end_for
end_for
QuickSort(list)
ReduceToFirsQuartile(list)
ChooseRandomElement(list,avatar,server)

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 548 — #18

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-549

200 3500
180
3000
160
140 2500
Computation (s)

System cost CP
120 2000
100
80 1500

60 1000
40
500
20
0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Number of iterations Number of iterations

FIGURE 30.14 Variation of the performance measures for different threshold values.

Cp_GRASP = tmp_cost = 10000

for j:=0 To AvatarsInAOI(avatar) do
for k:=0 to S do
tmp_cost=TestSolution(j,k)
AddToList(list,tmp_cost,j,k)
if (tmp_cost < Cp_GRASP)
Cp_GRASP := tmp_cost
savej := j
savek := k
end_for
end_for
end_for
AssignSolutionServer(savej,savek)
end_for
end

The quality of the solutions provided by GRASP search method depends on the quality of the elements
in the RLC, and the range of solutions depends on the length of the RLC. Thus, the main parameter to be
tuned in this case is the number of nonassigned N or critical avatars that the initial partition must leave.
Figure 30.14 shows the results in this tuning phase in order to compose an intermediate solution. In this
example a LARGE world composed of 2500 avatars are assigned to eight servers. The avatars are located
following a uniform distribution. This figure represents the variations of two performance measures as
the number of critical avatars (iterations) is increased. The quality of the obtained solutions Cp and the
execution time for GRASP algorithm have been elected as performance measures.
Figure 30.14 shows that as the number of critical avatars increases the quality of the provided solutions
also increases (Cp values decreases), but the execution time for GRASP algorithm (labeled as com-
putations) also increases. We chose in this case a compromise solution of 250 iterations. It is worth
mentioning that a larger number of iterations result in higher execution times and do not reach significant
solutions.

30.4 Performance Evaluation

In this section, we present the performance evaluation of the heuristics described in the previous section
when they are used for solving the partitioning problem in DVE systems. Following the evaluation
methodology shown in Reference 14, we have empirically tested these heuristics in two examples of a
DVE system: a SMALL world, composed of 13 avatars and 3 servers, and a LARGE world, composed of
2500 avatars and 8 servers. We have considered two parameters: the value of the quality function Cp for
the partition provided by the search method and also the computational cost, in terms of execution time,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 549 — #19

30-550 Handbook of Bioinspired Algorithms and Applications

(a) (b) (c)

FIGURE 30.15 Distributions of avatars: (a) uniform, (b) skewed, and (c) clustered.

TABLE 30.2 Results for a SMALL DVE System

Uniform distribution Skewed distribution Clustered distribution

Time (sec) Cp Time (sec) Cp Time (sec) Cp

Exhaustive 3.4110 6.54 3.8430 7.04 4.7830 7.91

LOT 0.0009 6.56 0.0010 8.41 0.0011 8.89
SA 0.0040 6.82 0.0050 7.46 0.0050 7.91
ACS 0.0007 6.59 0.0030 7.61 0.0024 8.76
GA-basic 0.0020 6.54 0.0030 7.04 0.0050 7.91
GA-improved 0.0030 6.54 0.0030 7.04 0.0060 7.91
GRASP 0.0002 7.42 0.0002 8.63 0.0003 11.88

required by the search method in order to provide that partition. For comparison purposes, we have also
implemented the LOT [14]. This method currently provides the best results for the partitioning problem
in DVE systems. In the case of SMALL worlds we have also performed an exhaustive search through the
solution space, obtaining the best partition possible. The hardware platform used for the evaluation has
been a 1.7 GHz Pentium IV with 256 Mbytes of RAM.
Since the performance of the heuristic search methods may heavily depend on the location of avatars
in the virtual world, we have considered three different distributions of avatars: uniform, skewed, and
clustered distribution. Figure 30.15 shows an example of how avatars would be located in a 2D world
when following each one of these distributions.
Table 30.2 shows the Cp values corresponding to the final partitions provided by each heuristic search
method for a SMALL virtual world, as well as the execution times required for each method in order
to obtain that final partition. For this proposed approach based on genetic algorithms, GA-B represents
the basic approximation described in Section 30.3.2 and GA-I incorporates both the PA method and the
proposed crossover operator detailed in Section 30.3.3. It can be seen that all of the heuristics provide better
(lower) Cp values than the LOT search method for a uniform distribution of avatars. For the skewed and
clustered distributions, most of the heuristics also provides better Cp values than the LOT search method,
and some of them (GA and SA methods) even provide the minimum value. However, the execution times
required by most of the heuristics are longer than the ones required by the LOT method. Only GRASP
method provides worse Cp values than the LOT method, but it requires much shorter execution times.
Although these results do not clearly show which heuristic provides the best performance, they validate
any of the proposed heuristics as an alternative to the LOT search method.
However, in order to design a scalable DVE system, the partitioning method must provide good
performance when the number of avatars in the system increases. That is, it must provide a good per-
formance specially for LARGE virtual worlds. Table 30.3 shows the required execution times and the Cp
values obtained by each heuristic search method for a LARGE virtual world.
When uniform distributions of avatars are considered, GA-I not only obtains the best partitioning
solutions (in terms of Cp ) but also achieved them by spending the minimum execution time. The rest

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 550 — #20

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-551

TABLE 30.3 Results for a LARGE DVE System

Uniform distribution Skewed distribution Clustered distribution

Time (sec) Cp Time (sec) Cp Time (sec) Cp

LOT 30.939 1637.04 32.176 3460.52 43.314 5903.80

SA 6.350 1707.62 13.789 2628.46 29.620 4697.61
ACS 5.484 1674.08 14.050 2286.16 23.213 3736.69
GA-B 6.598 1832.21 14.593 2825.64 29.198 4905.93
GA-I 6.410 321.31 14.590 450.80 28.740 791.94
GRASP 6.622 1879.76 13.535 2883.84 26.704 5306.24

of heuristics provides similar values of Cp than the one provided by LOT method, while requiring much
shorter execution times. When nonuniform distributions of avatars are considered, then all the heuristics
provide much better Cp values than the LOT method and they also require much shorter execution
times than the LOT method. In particular, GA-I method provides the best Cp values for nonuniform
distributions, requiring also the shortest execution time in the case of a skewed and clustered distribution
of avatars.
These results show that the performance of the partitioning algorithm can be significantly improved
by simply using any of the proposed heuristics instead of the LOT method, thus increasing the scalability
of DVE systems. In particular, GA-I method provides the best performance as a partitioning algorithm
for LARGE worlds.

30.5 Conclusions
In this chapter, we have proposed a comparison study of modern heuristics for solving the partitioning
problem in DVE systems. This problem is the key issue that allows to design scalable and efficient DVE
systems. We have evaluated the implementation of different metaheuristics, ranging over most of the
current taxonomy of modern heuristics. We have tested the proposed heuristics when applied for both
SMALL and LARGE DVE systems, with different distributions of the existing avatars in the system. We have
compared these results with the ones provided by the LOT, the partitioning method that currently provides
the best solutions for DVE systems. For SMALL virtual worlds, we can conclude that in general terms
any of the implemented heuristics provides a partition with similar values of the quality function Cp , but
the execution times required by the implemented heuristics are longer than the time required by the LOT
search method. Although SA and GA methods provide the minimum value of the quality function, only
GRASP method provides execution times shorter than the ones required by the LOT method for all the
tested distributions of avatars. These results validate any of the proposed heuristics as an alternative to
the LOT search method when considering SMALL DVE systems. However, for LARGE virtual worlds any
of the proposed heuristics provides better Cp values and requires shorter execution times than the LOT
method for nonuniform distributions of avatars. In particular, GA-I method provides the best results.
Since a scalable DVE system must be able to manage large amounts of avatars, we can conclude that these
results validates GA-I search method as the best heuristic method for solving the partitioning problem in
DVE systems.

References
[1] S. Singhal and M. Zyda. Networked Virtual Environments. ACM Press, 1999.
[2] J.M. Salles, R. Galli, A.C. Almeida, C.A.C. Belo, and J.M. Rebordão. mworld: A multiuser 3d virtual
environment. IEEE Computer Graphics, 17(2): 55–65, 1997.
[3] D.C. Miller and J.A. Thorpe. Simnet: The advent of simulator networking. Proceedings of the IEEE,
83: 1114–1123, 1995.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 551 — #21

30-552 Handbook of Bioinspired Algorithms and Applications

[4] C. Bouras, D. Fotakis, and A. Philopoulos. A distributed virtual learning centre in cyber-
space. In Proceedings of the Fourth International Conference on Virtual Systems and Multimedia
(VSMM’98), November 1998.
[5] M. Abrash. Quake’s game engine. Dr. Dobb’s Journal, 51–63, Spring 1997.
[6] M. Lewis and J. Jacboson. Game engines in scientific research. Communications of the ACM,
45: 17–31, 2002.
[7] Kali networked game support software. http://www.kali.net.
[8] M. Macedonia. A taxonomy for networked virtual environments. IEEE Multimedia, 4(1): 48–56,
1997.
[9] P.A. Berstein, V. Hadzilacos, and N. Goodman. Concurrency, Control and Recovery in Database
Systems. Addison-Wesley, Reading, MA, 1997.
[10] J. Falby, M. Zyda, D. Pratt, and R. Mackey. Npsnet: Hierarchical data structures for real-time
three-dimensional visual simulation. Computers & Graphics, 17(1): 65–69, 1993.
[11] D. Lee, M. Lim, and S. Han. Atlas — A scalable network framework for distributed virtual envi-
ronments. In Proceedings of ACM Collaborative Virtual Environments (CVE 2002), September 2002,
pp. 47–54.
[12] J.C.S. Lui, M.F. Chan, and K.Y. Oldfield. Dynamic Partitioning for a Distributed Virtual
Environment. Department of Computer Science, The Chinese University of Hong Kong, 1998.
[13] P. Barham and T. Paul. Exploiting reality with multicast groups. IEEE Computer Graphics and
Applications, 15: 38–45, 1995.
[14] J.C.S. Lui and M.F. Chan. An efficient partitioning algorithm for distributed virtual environment
systems. IEEE Transaction on Parallel and Distributed Systems, 13: 193–211, 2002.
[15] P.T. Tam. Communication Cost Optimization and Analysis in Distributed Virtual Environ-
ment. Technical report RM1026-TR98-0412, Department of Computer Science and Engineering.
The Chinese University of Hong Kong, 1998.
[16] R.L. Haupt and S.E. Haupt. Practical Genetic Algorithms. John Wiley & Sons, New York,
1997.
[17] S. Kirkpatrick, C.D. Gelatt, and M.P. Vecchi. Optimization by simulated annealing. Science, 220:
671–679, 1983.
[18] M. Dorigo, V. Maniezzo, and A. Coloni. The ant system: Optimization by a colony of operation
agent. IEEE Transactions on Systems, Man and Cybernetics, 96: 1–13, 1996.
[19] H. Delmaire, J.A. Díaz, E.M. Fernández, and M. Ortega. Comparing New Heuristics for the Pure
Integer Capacitated Plant Location Problem. Technical report DR97/10, Department of Statistics
and Operations Research, Universitat Politecnica de Catalunya (Spain), 1997.
[20] D.B. Anderson, J.W. Barrus, and J.H. Howard. Building multi-user interactive multimedia
environments at MERL. IEEE Multimedia, 2(4): 77–82, 1995.
[21] F.C. Greenhalgh. Analysing movement and world transitions in virtual reality tele-conferencing. In
Proceedings of Fifth European Conference on Computer Supported Cooperative Work (ECSCW’97),
1997, pp. 313–328.
[22] J.C.S. Lui and W.K. Lam. General methodology in analysing the performance of parallel/distributed
simulation under general computational graphs. In Proceedings of Third International Conference
on the numerical Solution of Markov Chain, September 1996.
[23] C. Coello, G. Lamont, and D. Van Veldhuizen. Evolutionary Algorithms for Solving Multi-Objective
Problems. Kluwer Academic Publishers, 2002.
[24] R. Duda, P. Hart, and D. Stork. Pattern Classification. Wiley Interscience, 2000.
[25] P. Morillo, M. Fernández, and N. Pelechano. A grid representation for distributed virtual envi-
ronments, acrossgrid’2003. In Proceedings of the First European Across Grids Conference, February
2003.
[26] K.E. Kinnear. Alternatives in automatic function definition: A comparison of performance. In
K.E. Kinnear, Ed., Advances in Genetic Programming. MIT Press, Cambridge, MA, 1994,
pp. 119–141.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 552 — #22

Solving the Partitioning Problem in Distributed Virtual Environment Systems 30-553

[27] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs, 2nd ed. Springer-
Verlag, Heidelberg, 1992.
[28] J.H. Holland and D.E. Goldberg. Genetic algorithms and machine learning: Introduction to the
special issue on genetic algorithms. Machine Learning, 3, 1988.
[29] P. Morillo, M. Fernández, and J.M. Orduña. A comparison study of modern heuristics for solving
the partitioning problem in distributed virtual environment systems. In International Conference
in Computational Science and Its Applications (ICCSA’2003) Vol. 2669 of Lecture Notes in Computer
Science, Springer-Verlag, Heidelberg, 2003, pp. 458–467.
[30] R. Sedgewick. Algorithms in C, 3rd ed. Addison-Wesley, Reading, MA, 1998.
[31] P. Morillo, M. Fernández, and J.M. Orduña. An ACS-based partitioning method for distributed
virtual environment systems. In Proceedings of 2003 IEEE International Parallel and Distributed
Processing Symposium, NIDISC-IPDPS’2003, April 2003, p. 148.
[32] M. Dorigo, G. Di Caro, and M. Sampels. Ant Algorithms: Third International Workshop, Ants 2002.
Springer-Verlag, Heidelberg, 2002.
[33] C. Koulamas, R. Jaen, and S.R. Antony. A survey of simulated annealing applications to operations
research problems. International Journal of Management Science, 22: 41–56, 1994.
[34] T.A. Feo and M.G.C. Resende. Greedy randomized adaptive search procedures. Journal of Global
Optimization, 6: 109–133, 1995.
[35] M. Resende and C. Ribeiro. Handbook of Metaheuristics. Kluwer Academic Publishers.
[36] P. Morillo and M. Fernández. A grasp-based algorithm for solving DVE partitioning problem. In
Proceedings of 2003 IEEE International Parallel and Distributed Processing Symposium, IPDPS’2003,
April 2003, p. 60.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C030” — 2005/8/6 — 14:13 — page 553 — #23

 31
Population Learning
Algorithm and Its
Applications

31.1 Population-Based Algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . .31-555

31.2 Population Learning Algorithm . . . . . . . . . . . . . . . . . . . . . . . .31-558
31.3 Example Applications of the PLA to Solving
Combinatorial Optimization Problems . . . . . . . . . . . . . . . .31-562
Permutation Scheduling • Other Combinatorial Optimization
Problems
31.4 Example Applications of the PLA to Machine
Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .31-576
Application of the Population Learning Algorithm to Training
Feed-Forward ANN • An Approach to Instance Reduction in
Supervised Learning
Piotr Jedrzejowicz References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .31-586

31.1 Population-Based Algorithms

The techniques used to solve difficult combinatorial optimization problems have evolved from constructive
algorithms to local search techniques, and finally to population-based algorithms. Population-based
methods have become very popular. They provide good solutions since any constructive method can
be used to generate the initial population, and any local search technique can be used to improve each
solution in the population. But population-based methods have the additional advantage of being able to
combine good solutions in order to get possibly better ones. The basic idea behind this way of doing is
that good solutions often share parts with optimal solutions (Hertz and Kobler, 1998).
Population based methods are optimization techniques inspired by natural evolution processes. They
handle a population of individuals that evolves with the help of information exchange procedures. Each
individual may also evolve independently. Periods of cooperation alternate with periods of self-adaptation.
Among best-known population-based method are evolutionary algorithms. These can be perceived as
a generalization of genetic algorithms, which operate on fixed-length binary strings, which need not be
the case for evolutionary algorithms (Michalewicz, 1994). Evolutionary algorithms are stochastic search
methods that mimic the metaphor of natural biological evolution. Evolutionary algorithms operate on
a population of potential solutions applying the principle of survival of the fittest to produce better and
better approximations to a solution.

31-555

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 555 — #1

31-556 Handbook of Bioinspired Algorithms and Applications

Evolutionary algorithms model natural processes, such as selection, recombination, mutation,

migration, locality, and neighborhood. Evolutionary algorithms work on populations of individuals
instead of single solutions. At the beginning of the computation, a number of individuals (the pop-
ulation) are randomly initialized. The objective function is then evaluated for these individuals. The
first/initial generation is produced. If the optimization criteria are not met the creation of a new gen-
eration starts. Individuals are selected according to their fitness for the production of offspring. Parents
are recombined to produce offspring. All offspring will be mutated with a certain probability. The fitness
of the offspring is then computed. The offspring are inserted into the population replacing the parents,
producing a new generation. This cycle is performed until the optimization criteria are reached.
Such a single population evolutionary algorithm is powerful and performs well on a broad class of
problems. However, better results can be obtained by introducing many populations, called islands. Every
island evolves for a number of generations isolated (like the single population evolutionary algorithm)
before one or more individuals are exchanged between the islands. The island-based evolutionary algorithm
models the evolution of a species in a way more similar to nature than the single population evolutionary
algorithm (Gordon and Whitley, 1993).
Development and proliferation of evolutionary algorithms have been initiated by and advanced in
seminal works of Holland (1975), Goldberg (1989), Davis (1991), and Michalewicz (1994). Despite the
fact that evolutionary algorithms, in general, lack a strong theoretical background, the application results
were more than encouraging. This success has led to emergence of numerous techniques, algorithms, and
their respective clones that are now considered as belonging to the population-based class, and which
differ in some respects from “classic” evolutionary algorithms.
Among well known population-based methods are scatter search techniques (Glover, 1977) maintaining
a population of reference points and generating offspring by weighted linear combinations. Different type
of population-based techniques are adaptive memory algorithms (Golden et al., 1997). Here a central
memory is used to keep track of the best components of the solutions visited during the search. These
components are combined to create new solutions, which are subject to repair to assure feasibility. Finally,
the components of the new solutions are considered as candidates that may be selected for replacing old
components in the central memory.
Another population based approach — ant colony optimization (ACO) studies artificial systems that take
inspiration from the behavior of real ant colonies and are used to solve discrete optimization problems.
Recently, the ant colony optimization metaheuristic was defined in Dorigo and Di Caro (1999). Ant colony
algorithms are inspired by biological observations on the behavior of ant colonies. Simple agents called
ants explore a region of the search space and share information gathered throughout their individual
searches. In a basic ant system (Colorni et al., 1992), each ant is a constructive procedure generating
solution to the considered problem. At each step of the construction, each ant has to decide how to make
one more step towards the completion of a partial solution. Such choice is based on two factors. The trace
factor guides the ants to choices that gave good results in earlier constructions. The intensity of this factor
(an analogue of the pheromone trail) informs about the quality of the solutions generated by making the
corresponding choice. The desirability factor guides the ants to choices that induce the best value of a
partial solution. Once all ants have completed their construction, the trace factors are updated. A choice
that led to good solutions increases the corresponding trace factor. The solutions constructed by ants are
further submitted to an improvement algorithm.
Particle swarm optimization (PSO) is a population based stochastic optimization technique developed
by Eberhart and Kennedy (1995), inspired by social behavior of bird flocking or fish schooling. PSO shares
many similarities with evolutionary computation techniques such as genetic algorithms (GAs). The system
is initialized with a population of random solutions and searches for optima by updating generations.
However, unlike GA, PSO has no evolution operators such as crossover and mutation.
An idea that culture might be symbolically encoded and transmitted within and between popula-
tions as another inheritance mechanism have triggered the development of cultural algorithms (CAs).
Reynolds (1994) developed a computational model in which cultural evolution is seen as an inherit-
ance process that operates at two levels: the microevolutionary and the macroevolutionary. Genetic

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 556 — #2

Population Learning Algorithm and Its Applications 31-557

algorithms are used to model the microevolutionary process, and so-called Version Spaces are used
to model the macroevolutionary process of a cultural algorithm. A cultural algorithm models the
evolution of the culture component of an evolutionary computational system over time. The cul-
ture component provides an explicit mechanism for acquisition, storage, and integration of individual
and group’s problem solving experience, and behavior. In a cultural algorithm, there are two main
spaces: the normal population of the evolutionary algorithm and the belief space, which is the place
where the shared acquired knowledge is stored during the evolution of the population (Chung and
Reynolds, 1998).
Moscato and Norman (1989) have introduced the term memetic algorithm (MA) to describe evolution-
ary algorithms in which local search plays a significant part. This term is motivated by Richard Dawkins’s
notion of a meme as a unit of information that reproduces itself as people exchange ideas (Dawkins,
1976). Moscato and Norman liken this thinking to local refinement, and therefore promote the term
“memetic algorithm” to describe genetic algorithms that use local search heavily. While genetic algorithms
have been inspired by a biological evolution, (MAs) would try to mimic cultural evolution. MA is a
marriage between a population-based global search and the heuristic local search made by each of the
individuals.
Given a representation of an optimization problem, a certain number of individuals are created. The
state of these individuals can be randomly chosen or set according to a certain initialization procedure.
A heuristic can be chosen to initialize the population. After that, each individual makes local search. After
that, when the individual has reached a certain development, it interacts with the other members of the
population. The interaction can be a competitive or a cooperative one. The cooperative behavior can be
understood as the mechanisms of crossover in GA or other types of breeding that result in the creation
of a new individual. More generally, cooperation is understood as an interchange of information. The
local search and cooperation (mating, interchange of information) or competition (selection of better
individuals) is repeated until a stopping criterion is satisfied. Usually, it should involve a measure of
diversity within the population.
An important addition to the family of the population-based method is GRASP. A greedy random-
ized adaptive search procedure is a metaheuristic for combinatorial optimization. It is a multi-start
or iterative process in which each iteration consists of two phases, a construction one, in which a
feasible solution is produced and a local search one, in which a local optimum in the neighborhood
of the constructed solution is sought (Feo and Rosende, 1995). In the construction phase a feas-
ible solution is iteratively constructed, one element at a time. At each construction step, the choice
of the next element to be added is determined by ordering all candidate elements in a candidate list
C with respect to a greedy function measuring the benefit of selecting each element. The heuristic
is adaptive because the benefits associated with every element are updated at each iteration of the
construction phase to reflect the changes brought on by the selection of the previous element. The
probabilistic component of a GRASP is characterized by randomly choosing one of the best candidates
in the list, but not necessarily the top one. The list of best candidates is called the restricted candidate
list (RCL).
The solutions generated by a GRASP construction are not guaranteed to be locally optimal. Hence,
it is almost always beneficial to apply a local search in attempt to improve each constructed solution.
It terminates when no better solution is found in the neighborhood. While such local optimization
procedures can require exponential time from an arbitrary starting point, empirically their efficiency
significantly improves as the initial solution improves. The result is that often many GRASP solutions
are generated in the same amount of time required for the local optimization procedure to converge
from a single random start. Furthermore, the best of these GRASP solutions is generally signific-
antly better than the single solution obtained from a random starting point. GRASP can be easily
implemented in parallel. Each processor can be initialized with its own copy of the procedure, the
instance data, and an independent random number sequence. The GRASP iterations are then per-
formed in parallel with only a single global variable required to store the best solutions found over
all processors.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 557 — #3

31-558 Handbook of Bioinspired Algorithms and Applications

The above review, by no means exhaustive, allows to draw the following general conclusions:
• Population-based algorithms and techniques have become a standard tool to deal effectively with
computationally complex problems.
• There are several ingredients common to all population-based approaches. Among them, the most
characteristic include diversification by means of generating a population of individuals that are
solutions or parts of solutions and introducing some random noises at various stages of searching
for the solution. All population-based algorithms are equipped with some tools allowing to exploit
information gathered during computation with a view to drop less promising direction of search.
Finally, all population-based approaches mimic some natural biological or social processes.
• A common framework of properties characterizing population-based methods still does allow
for design flexibility and development of new population-based algorithms, each having its own
distinctive features and strengths, which may prove effective in solving target problem types.

31.2 Population Learning Algorithm

Population learning algorithm (PLA), was proposed in Jedrzejowicz (1999) as another population-based
method, which can be applied to solve combinatorial optimization problems. PLA has been inspired by
analogies to a social phenomenon rather than to evolutionary processes. Whereas evolutionary algorithms
emulate basic features of natural evolution including natural selection, hereditary variations, the survival
of the fittest, and production of far more offspring than are necessary to replace current generation, PLAs
take advantage of features that are common to social education systems:
• A generation of individuals enters the system.
• Individuals learn through organized tuition, interaction, self-study, and self-improvement.
• Learning process is inherently parallel with different schools, curricula, teachers, etc.
• Learning process is divided into stages.
• More advanced and more demanding stages are entered by a diminishing number of individuals
from the initial population (generation).
• At higher stages, more advanced education techniques are used.
• The final stage can be reached by only a fraction of the initial population.
In PLA, an individual represents a coded solution of the considered problem. Initially, a number of
individuals, known as the initial population is randomly generated. Once the initial population has
been generated, individuals enter the first learning stage. It involves applying some, possibly basic and
elementary, improvement schemes. These can be based, for example, on some simple local search proced-
ures. The improved individuals are then evaluated and better ones pass to a subsequent stage. A strategy of
selecting better or more promising individuals must be defined and duly applied. In the following stages
the whole cycle is repeated. Individuals are subject to improvement and learning, either individually
or through information exchange, and the selected ones are again promoted to a higher stage with the
remaining ones dropped-out from the process. At the final stage, the remaining individuals are reviewed
and the best represents a solution to the problem at hand. Basic idea of the population learning algorithm
is shown in Figure 31.1.
Figure 31.1. covers a simple, nonparallel version of the algorithm. For a parallel PLA the following
features need to be defined:
• A set of rules controlling how individuals are grouped into concurrent populations at various
stages.
• The respective rules for running learning and improvement algorithms in parallel at various stages.
• Rules for information exchange and coordination between concurrent processes.
Three example types of the PLA structure are shown in Figure 31.2. Scheme (a) is a simple sequential,
nonparallel, implementation of the algorithm. Schemes (b) and (c) depict parallel implementations.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 558 — #4

Population Learning Algorithm and Its Applications 31-559

START
LEARNi (P )

Set the number of learning

stages N. Set the initial
population size m SELECTi (P )

Define learning/improvement
procedures LEARNi (P ), i := i + 1
i = 1...N, operating on a
population of individuals P
NO
Define selection procedures i>N
SELECTi (P) i = 1...N,
operating on a population of YES
individuals P
Consider the best individual
from P as a solution
i := 1

END
Generate the initial population
Set P := initial population

FIGURE 31.1 General idea of the population learning algorithm.

(a) Stage 1 Stage k Final stage

Initial Intermediate Final

population Solution
population population

(b)
Intermediate Final
Initial population 1
population 1 population 1
Solution

Intermediate Final
Initial population n population n population n

(c)
Intermediate
Initial population 1
population 1
Final Solution
population n
Intermediate
Initial population n
population n

FIGURE 31.2 Three example types of the PLA structure.

Scheme (b) is a simple parallel implementation, assuming that the learning and improvement procedures
used at various stages do not require any information exchange between the concurrent popula-
tions. Finally, Scheme (c) is a parallel implementation with information exchange. Such scheme (or
its variants) should be used in case some information is exchanged between the concurrent popula-
tions during the learning and improvement process. Choice of the appropriate PLA scheme depends
on both — the computational resources available and characteristics of learning and improvement
procedures used.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 559 — #5

31-560 Handbook of Bioinspired Algorithms and Applications

TABLE 31.1 Population Learning Algorithm Design Elements

Design element Comment

Definition of an individual An individual is a coded feasible solution or part of a solution. The
designer should aim at achieving ease of manipulation and storage, ease
of transformation into a solution and ease of fitness evaluation
Procedure for generating the initial The designer should aim at achieving unbiased individuals, if possible
population of individuals representing all regions of the feasible solution space. A good construc-
tion heuristics can be used to produce a seed from which the initial
population can be generated
Size of the initial population Must be set at the PLA fine-tuning stage. Should be chosen as a comprom-
ise between the requirement of the sufficient and adequate represent-
ation assuring good quality of results and the available computational
resources
Number of learning and Depends on availability of learning and improvement algorithms. Should
improvement stages be chosen with a view of finding a satisfactory compromise between
quality of solutions and computation time
Number and size of parallel groups Should be chosen considering the available computational resources,
of individuals at each stage complexity of the problem at hand, and complexity of the learning and
improvement algorithms used. Another compromise between quality
of solutions and computation time is involved
Learning and improvement Basic and simple procedures at earlier stages should be replaced by more
algorithms for each stage/each complex and sophisticated at later ones. The designer should try to
group assure adequate diversity of learning and improvement algorithms used
Fitness function of an individual Should be simple (easily computable) and directly related to the quality
of the solution represented by an individual
Selection strategy Must be set at the PLA fine-tuning stage. Rules for rejection and pro-
motion of individuals may differ at various stages. The designer should
aim at achieving good efficiency of the algorithm not losing, too early,
representation sufficiency and adequacy. Another compromise between
quality of solutions and computation time is involved

Designing population learning algorithm intended for solving a particular problem type allows the
designer a lot of freedom (as, in fact, happens in case of majority of other population-based algorithms).
Moreover, an effective PLA would certainly require a lot of fine-tuning and experimenting. This could
be considered as a disadvantage, at least as long as the process of setting different parameters of the
population learning algorithm is rather based on heuristics instead of some theoretical or statistical rules,
which, unfortunately, are not yet appropriately developed. Main PLA design elements are summarized in
Table 31.1.
Although PLA shares many features with other population-based approaches, it clearly has its own
distinctive characteristics. Brief comparison of several example algorithms belonging to the discussed
class is shown in Table 31.2.
In the following sections several example implementations of the PLA applied to solving different com-
putationally difficult problems are discussed. The PLA is seen here as a general framework for constructing
hybrid solutions to difficult computational problems. From such a perspective the PLA role is to struc-
ture, organize, sequence, and eventually help to apply in a parallel environment variety of techniques.
Strength of the PLA stems from combining in an “intelligent” manner the power of population-based
algorithms using some random mechanism for diversity assurance, with efficiency of various local search
algorithms. The later may include, for example, reactive search, tabu search, simulated annealing as well
as the described earlier population based approaches.
Computational intelligence embedded into the population learning algorithm scheme is based on using
the following heuristic rules:

• To solve difficult computational problems apply a cocktail of methods and techniques including
random and local search techniques, greedy and construction algorithms, etc., building upon their
strengths and masking weaknesses.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 560 — #6

Population Learning Algorithm and Its Applications 31-561

TABLE 31.2 Comparison of Example Population-based Algorithms

Iterative search Information

Algorithm Processing object mode Intensification Diversification exchange

PLA Population of Two level A cocktail of Random initial Different

potential iterative search local search and population, schemes used
solutions (stages with a random search plus random by different
decreasing procedures local search at learning and
population size some stages improvement
and iterative algorithms
local search
procedures)
EA Population of Generation None Random initial Crossover
potential replacement population, operator
solutions plus mutation
operator
ACO Population of Search at each Improvement Random initial Trace factor
construction construction algorithm population mechanism
processes step
PSO Population of Generation None Random initial Through pbest,
potential replacement population lbest, and gbest
solutions values
CA Population of Generation None Random initial Knowledge
potential replacement population, sharing
solutions plus mutation through version
spaces
MA Population of Two level Local search Random initial Cooperative
potential iterative search procedure population, interactions
solutions (generation plus mutation mechanism
replacement operator
through
competitive
interactions
plus local
search)
GRASP Population of Two level Local search Multiple start None
construction iterative search procedure construction
processes (iterative local processes
search
procedure plus
multiple
construction
processes)

• To escape getting trapped into a local optimum generate or construct an initial population of
solutions called individuals, which in the following stages will be improved, thus increasing chances
for reaching a global optimum.
• Another means of avoiding getting trapped into local optima is to apply at various stages of search
for a global optimum some random diversification algorithms.
• To increase effectiveness of searching for a global optimum divide the process into stages retaining
after each stage only a part of the population consisting of “better” or “more promising” individuals.
• Another means of increasing effectiveness is to use at early stages of search, improvement algorithms
with lower computational complexity as compared to those used at final stages.

To conclude this section it should be noted that the PLA is an addition to the family of the population-
based techniques, which can be seen as a hybridization framework allowing for an effective integration of
different deterministic and random local search techniques. Figure 31.3 summarizes main features of the
population learning algorithm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 561 — #7

31-562 Handbook of Bioinspired Algorithms and Applications

Decreasing population size

First stage of Final stage of

Intermediate learning and
learning and stages of learning
improvement improvement
and improvement Solution
Final population
Intermediate populations
Initial population

Increasing complexity of learning algorithms

FIGURE 31.3 Main features of the PLA.

31.3 Example Applications of the PLA to Solving Combinatorial

Optimization Problems
This section contains a review of several example PLA implementations applied to solving combinatorial
optimization problems. In all cases the problem description is followed by a presentation of some main
features of the respective implementation and a discussion of the computational experiments carried out.

31.3.1 Permutation Scheduling

Permutation scheduling problems are defined as scheduling problems where the solution is a permutation
of tasks or jobs to be processed on one or more machines (processors). In the review, four permutation
scheduling problems are considered — common due date, permutation flow shop, a single machine
weighted tardiness, and multiprocessor task scheduling in multistage hybrid flowshops.
In a common due date problem there is a set of n non-preemptable tasks available at time zero. Tasks
are to be processed on a single machine. Processing times of the tasks pi , i = 1, . . . , n are known. There
is also a common due date d. Earliness and tardiness of tasks are defined as Ei = max{d − Ci , 0} and
Ti = max{Ci − d, 0}, respectively (i = 1, . . . , n) where Ci is the completion time of job i. Penalties per
time unit of the task i being early or tardy are ai and bi , respectively. The objective is to minimize the sum
of earliness and tardiness penalties:

i ai Ei + i bi Ti → min, i = 1, . . . , n.

It should be also noted that the due date is called unrestrictive if the optimal sequence of tasks can
be constructed without considering the value of the due date. Otherwise the common due date is called
restrictive. Obviously a common due date for which d ≥ i pi holds is unrestrictive.
In a permutation flow shop there is a set of n jobs. Each of n jobs has to be processed on m machines
1, . . . , m in this order. The processing time of job i on machine j is pij , where pij is fixed and nonnegative.
At any time, each job can be processed on at most one machine, and each machine can process at most one
job. The jobs are available at time 0 and the processing of a job may not be interrupted. In permutation
flow shop problem (PFSP) the job order is the same on every machine. The objective is to find a job
sequence minimizing schedule makespan (i.e., completion time of the last job).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 562 — #8

Population Learning Algorithm and Its Applications 31-563

TABLE 31.3 Neighborhood Structures used in the PLA Implementations

Notation Move Neighbourhood pace
N1 (x) Exchange of two consecutive tasks in x All possible exchanges
N2 (x) Exchange of two nonconsecutive tasks in x All possible exchanges
N3 (x) Finding the order of four consecutive tasks in x by All quadruplets of consecutive tasks
enumeration
N4 (x) A single step rotation of three random tasks in x All possible triples
N5 (x) Moving a random task in x to another location and adjusting All possible moves
other tasks

In a single machine weighted tardiness problem there is a set of n jobs. Each of n jobs (numbered
1, . . . , n) is to be processed without interruption on a single machine that can handle no more than one
job at a time. Job j(j = 1, . . . , n) becomes available for processing at time zero, requires an uninterrupted
positive processing time pj on the machine, has a positive weight wj , and has a due date dj by which it
should ideally be finished. For a given processing order of the jobs, the earliest completion time Cj and
the tardiness Tj = max{Cj − dj , 0} of job j( j = 1, . . . , n) can readily be computed. The problem is to find
a processing order of the jobs with minimum total weighted tardiness j wj Tj .
Finally, a multiprocessor task scheduling in multistage hybrid flow shops is considered. The discussed
problem involves scheduling of n jobs composed of tasks in a hybrid flow shop with m stages. All jobs
have the same processing order through the machines, that is, a job is composed of an ordered list of
multiprocessor tasks where the i-th task of each job is processed at the i-th flow shop stage (the number
of tasks within a job corresponds exactly to the number of flow shop stages). Processing order of tasks
flowing through stages is the same for all jobs. At each stage i, i = 1, . . . , m, there are available mi identical
parallel processors. For the processing at stage i, task i being a part of the job j, j = 1, . . . , n, requires
sizei,j processors simultaneously. That is, sizei,j processors assigned to task i at stage i start processing the
task simultaneously and continue doing so for a period of time equal to the processing time requirement
of this task, denoted pi,j . Each subset of available processors can process only the task assigned to them at
a time. The processors do not break down. All jobs are ready at the beginning of the scheduling period.
Preemption of tasks is not allowed. The objective is to minimize makespan, that is the completion of the
lastly scheduled task in the last stage.
To solve permutation scheduling problems three versions of the population learning algorithm, denoted
respectively as PLA1, PLA2, and PLA3 have been designed and implemented. PLA1 has been used to solve
the restricted instances of the common due date scheduling problem, as well as instances of flow shop
and total tardiness problems. PLA2 has been used to solve the unrestricted instances of the common
due date problem. PLA3 has been designed to solve instances of permutation flow shop, total weighted
tardiness problem, as well as instances of multiprocessor task scheduling in multistage hybrid flow shops.
The proposed algorithms make use of different learning and improvement procedures, which, in turn,
are based on five neighborhood structures shown in Table 31.3 In what follows x denotes an individual
encoded as a set of natural numbers (a permutation of tasks) and g (x) its fitness function.
All learning and improvement (l & p) procedures operate on the population of individuals P, and
perform local search algorithms based on the above defined neighborhood structures. General structure
of a l & p procedure is shown in the following pseudo-code:

Procedure LEARN(i,P):
begin
for each individual x in P do
Local_ search(i,x);
end for
end

Local search algorithms used are shown in Table 31.4.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 563 — #9

31-564 Handbook of Bioinspired Algorithms and Applications

TABLE 31.4 Local Search Algorithms used in the PLA Implementations

i Idea of the local search algorithm

1 Perform all moves from the neighborhood structure N1 (x); accept moves improving g (x); stop when no
further improvements of g (x) are possible
2 Mutate x producing x {mutation procedure is selected randomly from the two available ones — the two
point random exchange or the rotation of all tasks between two random points}; perform all moves from the
neighborhood structure N1 (x ); accept moves improving g (x ); stop when no further improvements of g (x )
are possible
3 Repeat k times {k is a parameter set at the fine-tuning phase; in the reported experiment k = 3∗ initial
population size}; generate offspring y and y by a single point crossover of x and a random individual x ;
perform all moves from the neighborhood structure N1 (y) and N1 (y ); accept moves improving g (y) and
g (y ); stop when no further improvements of g (y) and g (y ) are possible; adjust P by replacing x and x with
the two best individuals from {x, x , y, y }
4 Perform all moves from the neighborhood structure N2 (x); accept moves improving g (x); stop when no
further improvements of g (x) are possible
5 Perfor Perform all moves from the neighborhood structure N3 (x); accept moves improving g (x); stop when
no further improvements of g (x) are possible
6 Perform SIMULATED ANNEALING (x) based on N4 (x) neighborhood structure
7 Perform TABU SEARCH (x) based on N5 (x) neighborhood structure

Simulated annealing is a metaheuristic introduced in Kirkpatrick et al. (1983). Implementation of

simulated annealing used within PLA as the sixth local search algorithm is based on random moves from
the N4 (x) neighborhood structure.
Tabu search is yet another metaheuristic (see Glover, 1990). General idea of the present implementation
using a short term memory (STM) and a long term memory (LTM), is shown in the following pseudo code
(STM and LTM contain parameters of the moves and are managed according to the reverse elimination
method):

Procedure TABU SEARCH (x):

begin
xbest := x;
Gbest := g(x);
STM, LTM := ∅;
for j = 1 to it do {where it is the number of iterations}
Do := - ∞;
for i = 1 to lo do {where lo is the number of iterations}
Perform current_move, which is a random move from N5 (x) and not
in STM, producing y from x;
D := g(x) -- g(y);
Update STM;
if D > Do then Do = D; x’ := y; best_move := current_move;
end for
if g(x’) < Gbest then Gbest := g(x’); xbest := x’; x := x’;
else if best_move ∈/ LTM then x := x’;
end if
Update LTM;
Update STM;
end for
x := xbest;
end

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 564 — #10

Population Learning Algorithm and Its Applications 31-565

In the PLA implementations a simple procedure SELECT (P, s) is used, where s is a percentage of best
individuals from a current population promoted to higher stages. For any population P, let LEARNING(P)
stand for the following:

Procedure LEARNING(P):
begin
for i = 1 to 5 do
LEARN(i,P);
SELECT(P, s);
end for
end

The last learning procedure LEARN_REC(x) is recursive and operates on overlapping partitions of the
individual x. For example, for x = (5, 2, 4, 6, 1, 3, 7) the individuals s = (5, 2, 4), t = (4, 6, 1), u = (1, 3, 7)
make an overlapping partition of x, and the merge of (s, t , u) is x.

Procedure LEARN_REC(x):
begin
let (y1 ,. . ., ys ) be an overlapping partition of x
for i = 1 to s do
generate randomly a ‘‘small’’ population Pi of individuals for yi ;
LEARNING(Pi );
yi ’:=the best individual from Pi ;
end for
output the merge of (y1 ’,. . .,ys ’);
end

Now, the structure of the implemented population learning algorithms can be shown as:

Procedure PLA1:
begin
generate randomly initial_population;
P := initial_population;
LEARNING (P);
for each individual x in P do
LEARN_REC (x);
end for
output the best individual from P;
end

For the unrestrictive common due date scheduling not only the sequence of tasks but also the
starting time of the first task need to be found. To achieve this PLA2 is based on the following
assumptions:

• In the course of computation it is assumed that the deadline equals 0 and each task can be scheduled
prior to and after time 0.
• Each individual x = (x0 , x1 , . . . , xn ) in the population consists of a sequence of n tasks, 1 ≤ xj ≤ n
for j = 1, . . . , n and x0 = i ≤ n, which is the number of the last task completed before deadline
(x1 , . . . , xi are completed before time 0 and xi+1 , . . . , xn after time 0).

0
• starting time of the schedule is (− xj=1 pxj ). In what follows x0 is called the boundary task.
• In the final step of the computation the real starting time of the first task is computed from:

0
t0 = d − xj=1 pxj .

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 565 — #11

31-566 Handbook of Bioinspired Algorithms and Applications

PLA2 uses slightly modified learning and improvement procedures as compared with PLA1. For instance,
LEARN (3, P) produces an offspring using the following rule: If two individuals x and x produce y, then
the boundary task of y equals that of x if the crossover point is greater than the boundary task of x, or the
value: (the crossover point + boundary task of x–1) otherwise. Same applies to y .
Further modifications include two new learning and improvement procedures using the following local
search procedures:

Procedure LEARN (8, P):

begin
for each individual x in P do
Local_ search (8); {exchanges two tasks i, j, not necessarily

consecutive, where i (j) is less (greater) than the boundary task}

end for
end

Procedure LEARN (9, P):

begin
for each individual x in P do
Local_ search (9); {if the boundary task is less than the number
of tasks then it is incremented by one}
end for
end

Now PLA2 is defined as follows:

Procedure PLA2:
begin
generate randomly initial_population;
P := initial_population;
for i = 1,2,3,4,5,8 do
LEARN (i, P);
SELECT (P, s);
end for
LEARN (9, P);
output the best individual from P;
end

Finally, PLA3 has the following structure:

Procedure PLA3:
begin
generate randomly initial_population;
P := initial_population;
for i = 1, 7 do
LEARN (i, P);
SELECT (P, s);
end for
LEARN (6, P);
output the best individual from P;
end

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 566 — #12

Population Learning Algorithm and Its Applications 31-567

TABLE 31.5 Benchmark Data Sets

Problem Data set description Source

Common due date 240 problem instances with 20, 50, 100, 200, 500, and 1000 www.wiwi.uni-bielefeld.de/∼
tasks (40 instances for each problem size). Due dates for kistner/Bounds.html
instances in each problem size group are calculated as: d =
hi pi with h = 0.2, 0.4, 0.6, and 0.8, respectively. Problems
with h = 0.6 and 0.8 are considered as unrestrictive. Upper
bounds are provided
Flow shop 120 problem instances; 10 instances for each combination OR-LIBRARY,
of number of tasks and number of machines (20-5, 20-10, http://people.brunel.ac.uk/∼
20-20, 50-5, 50-10, 50-20, 100-5, 100-10, 100-20, 200- mastjb/jeb/info.html (see also
10, 200-20 and 500-20 tasks-machines). Upper bounds are [1993])
provided
Weighted tardiness 375 problem instances; 125 instances for each problem size OR-LIBRARY,
(40, 50, and 100 tasks). Optimal solutions are provided for http://people.brunel.ac.uk/∼
instances with 40 and 50 tasks. Upper bounds are provided mastjb/jeb/info.html
for instances with 100 tasks
Multiprocessor task 160 problem instances. The dataset includes 80 instances OR-LIBRARY,
scheduling in a with 50 jobs each and 80 instances with 100 jobs each. http://people.brunel.ac.uk/∼
multistage hybrid Each group of 80 problem instances is further partitioned mastjb/jeb/info.html
flow shops into four subgroups of 20 instances each with 2, 5, 8,
and 10 stages, respectively. Processor availability at various
stages and processor requirements per task varied between
1 and 10

In case of the multiprocessor task scheduling in a multistage hybrid flow shop problem, the PLA3 uses
the following fitness function:

Procedure FITNESS:
begin
for i = 1 to m do
for j = 1 to n do
allocate task π (j) at stage i to the required number of
processors, scheduling it as early as feasible;
end for
end for
g(π ) := finishing time of task number π (n) at stage m;
end

To evaluate the proposed algorithms several computational experiments have been carried out using
a PC with Pentium III, 850 Mhz processor (experiment results were originally reported in Jedrzejowicz
& Jedrzejowicz, (2003a, 2003b). The results have been compared with upper bounds or optimal solutions
(if known) of several sets of benchmark problems. Table 31.5 contains short descriptions of benchmark
data sets.
Table 31.6 shows mean relative errors (MREs) and mean computation times (MCT) for the experiment
with common due date scheduling instances. These have been calculated by comparing the best result
out of the three experimental runs with the upper bounds (i.e., best up-to-now known results for bench-
mark problems). MRE and MCT values in Table 31.6 have been calculated taking the average from ten
instances available for each problem size and due date class. Negative value of the MRE shows, in fact,
the relative improvement in comparison to previously established upper bounds. In the experiment the
initial population size has been set to 2000 and the selection coefficient value s has been set to 0.5.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 567 — #13

31-568 Handbook of Bioinspired Algorithms and Applications

TABLE 31.6 MRE and MCT for the Common due Date
Scheduling Instances

MRE (PLA1) MRE (PLA2)

n h = 0.2 h = 0.4 h = 0.6 h = 0.8 Overall MCT
(%) (%) (%) (%) (%) (%)

20 0.00 0.01 −0.63 −0.41 −0.26 6.3s.

50 −0.05 −0.01 −0.33 −0.24 −0.16 28.9s.
100 −0.01 −0.01 −0.17 −0.17 −0.09 91.5s.
200 −0.51 −0.45 −0.11 −0.11 −0.29% 205.2s.
500 −5.84 −5.91 −2.34 −2.34 −4.11 21.2m.
1000 −7.42 −4.39 −0.03 −0.04 −2.97 43.3m.
Overall −2.31 −1.80 −0.60 −0.55 −1.31 —

TABLE 31.7 MRE, SDE and MCT for the Flow Shop Scheduling

Measure Processors n = 20 n = 50 n = 100 n = 200 n = 500

5 0.0000% 0.0000% −0.0097% — —

MRE 10 −0.0073% 0.0566% 0.1456% 0.1471% —
20 0.0084% 0.4885% 0.2944% 0.3405% 0.3818%
5 0 0 0.000442 — —
SDE 10 0 0.001897 0.001234 0.003874 —
20 0.000279 0.000957 0.002313 0.001238 0.001631
5 2.7m. 9.75m. 28.4m. — —
MCT 10 8.5m. 23.0m. 29.1m. 81.0m. —
20 15.2m. 33.4m. 38.8m. 140.0m. 295m.

Experiment results prove that PLA1 and PLA2, as applied to the common due date scheduling, perform
well. Mean relative error for the whole considered population of problem instances is negative and equals
–1.31% , which means that the total amount of penalties to be paid as a result of scheduling all 240
problem instances has decreased by 1.31% . Out of 240 problem instances involved in the experiment in
77.5% of cases, that is, in 186 instances, it has been possible to find a better upper bound than previously
known. Only in case of 8 problem instances out of 240, results obtained by applying PLA were worse than
respective benchmark values. The discussed results have been obtained with a reasonable computational
effort.
Table 31.7 shows mean relative errors, standard deviation of errors (SDEs), and mean computation
times for the experiment with flow shop scheduling instances. These have been calculated by comparing
results of a single experimental run of the PLA3 with the upper bounds, that is best up-to-now known
result for benchmark problems. MRE and MCT values in Table 31.7 have been calculated taking an average
from 10 benchmark instances available for each problem size and number of processors class. The initial
population size has been set to 100 and the selection coefficient value s has been set to 0.5.
Experiment results prove that PLA3 as applied to permutation flow shop scheduling perform quite
well. Mean relative error for the whole considered population of benchmark problem instances equals
only 0.1776% and it has been obtained in a single run. Out of 120 problem instances involved in the
experiment in 10% of cases, that is, in 12 instances, it has been possible to find a better upper bound than
previously known. Considering the research effort invested during the last 20 years into finding algorithms
solving flow shop problems the quality of PLA3 is more than satisfactory. The discussed results, however,
required a substantial computational effort.
Table 31.8 shows mean relative errors and standard deviation of errors for the total tardiness scheduling.
These have been calculated by comparing results of a single experimental run of PLA3 with the optimal
solutions in case of 40 and 50 tasks and upper bounds in case of 100 tasks instances (strongly suspected to
be optimal solutions). MRE values in Table 31.8 have been calculated for different initial population sizes

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 568 — #14

Population Learning Algorithm and Its Applications 31-569

TABLE 31.8 MRE and SDE for the Total Tardiness Scheduling Experiment

Initial n = 40 n = 50 n = 100
population size MRE SDE MRE(%) SDE MRE(%) SDE

25 0.0000 0 0.0107 0.000526 0.0870 0.005079

50 0.0000 0 0.00134 0.000144 0.0632 0.004727
100 0.0000 0 0.00116 0.000130 0.0227 0.001916
200 0.0000 0 0.00116 0.000130 0.0164 0.001453
400 0.0000 0 0.00000 0 0.00074 0.000045
800 0.0000 0 0.00000 0 0.00030 0.000027

TABLE 31.9 MRE, SDE, and MTC for the Multiprocessor Task Scheduling in
a m-h Flow Shop Experiment

Measure Number of jobs Number of stages

2 5 8 10
MRE 50 −0.9814% −0.4036% −0.3367% 0.3712%
100 −0.6127% −0.3180% −0.2535% −0.4540%
SDE 50 0.0174% 0.2971% 0.4104% 0.5992%
100 0.0116% 0.2253% 0.3042% 0.2122%
MCT(m.) 50 1.20 3.35 5.76 10.40
100 3.55 8.80 14.15 27.42

taking the average from 125 instances available for each problem size. The initial population sizes have
been set to 25, 50, 100, 200, 400, and 800 individuals with the selection factor set to 0.5.
Experiment results prove that PLA3 applied to total tardiness scheduling perform very well. However,
the computational effort involved is rather high (from an average 72s. per instance of 100 tasks and the
initial population size set to 25, up to about 3000s. with the initial population size set to 800).
Finally, MRE, SDE, and MCT calculated after a single run of the PLA3 applied to solving benchmark
instances of the multiprocessor task scheduling in a multistage hybrid (m-h) flow shop problem are shown
in Table 31.9. Negative values of MRE show a percentage of improvement achieved over known upper
bounds within the respective clusters of benchmark dataset.
Application of the population learning algorithm has resulted in improving the total of upper bounds
of the considered cases by 0.34% . Out of 160 instances solved it has been possible to improve currently
known upper bounds in 73 instances, which is more than 45% of all considered instances.

31.3.2 Other Combinatorial Optimization Problems

PLA has been used to solve a number of instances of the following combinatorial optimization problems
belonging to the NP-hard class:
• Scheduling multiple variant programs (Jedrzejowicz and Wierzbowska, 2000).
• Scheduling multiple processor tasks with correlated failures, variants, ready times, deadlines, and
variant reliabilities on identical multiple processors to maximize schedule reliability (Jedrzejowicz
and Ratajczak, 2001; Czarnowski et al., 2003).
• Generalized segregated storage problem (GSSP) (Barbucha and Jedrzejowicz, 2000).
• Resource constrained project-scheduling problem with makespan as an objective function
(Jedrzejowicz and Ratajczak, 2003).
• Optimization of system reliability structure under multiple criteria (Jedrzejowicz et al., 2002).
In all of the above listed implementations, PLA proved to be an effective framework for designing a
hybrid algorithms producing competitive solutions. In this section, because of the length constraints, only
two example implementations out of the above listed ones are briefly described.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 569 — #15

31-570 Handbook of Bioinspired Algorithms and Applications

A GSSP is a combinatorial optimization problem that involves allocation of certain number of goods to
the available compartments subject to segregation (physical separation) constraints. Let m be a number
of divisible consignments consisting of the homogenous goods, n be a number of available (internal)
compartments. The size of each consignment to be stored — ai (i = 1, . . . , m) and the capacity of each
compartment, bj (j = 1, . . . , n) are known. Let cij be the cost of storing a unit of goods from consignment
i in the compartment j(i = 1, . . . , m, j = 1, . . . , n). For a given set of goods a segregation matrix, S
is introduced. Each element sij ∈ Z + of the matrix S(i, j = 1, . . . , m) defines the required segregation
distance between goods. Element sij is equal to 0, if and only if, good i can be stored together with j
(one without any restrictions) and element sij is greater than 0, if some segregation type between cargos
i, j(i, j = 1, . . . , m) is required (for example sij = 2 means that goods i and j must be stored in two
separated compartments). On the other hand, for a given set of compartments an additional compartment
segregation matrix, CS is defined. Element csij ∈ Z + of CS matrix (i, j = 1, . . . , n) specifies segregation
type guaranteed when storing goods in two compartments i and j(i, j = 1, . . . , n). It is also required
that all consignments will be stored. To meet this constraint, it is assumed that an external storage space,
denoted as (n + 1)st compartment, is also available. It can accommodate any consignment at a higher unit
cost ci,n+1 (i = 1, . . . , m). In this compartment segregation requirements do not have to be satisfied. Let
xij denote the amount of cargo i stored in the compartment j(i = 1, . . . , m, j = 1, . . . , n + 1). Generalized
Segregated Storage Problem can be then formulated as follows:


m 
n+1
Z = min cij xij
i=1 j=1

subject to:


n+1
xij = ai i = 1, . . . , m,
j=1


m
xij ≤ bj j = 1, . . . , n
i=1

m 
m
xik xjl hijkl = 0 k, l = 1, . . . , n,
i=1 j=1

xij ≥ 0 i = 1, . . . , m, j = 1, . . . , n + 1,

where

0 if sij ≤ cskl ,
hijkl =
1 if sij ≥ cskl ,

Elements hijkl form a binary matrix H . Each element of H is defined for two pairs: segregation requirement
for goods and segregation type for compartments, (i, k) and (j, l), where i, j = 1, . . . , m, and k, l =
1, . . . , n.
Let I = {1, 2, . . . , m} be a set of consignments, each consisting of a certain number of units and
H = {1, 2, . . . , n} be a set of internal compartments. Let g be a function g : {1, . . . , mn} → I × H defined
as follows:
g (v) = ((v − 1) /n + 1, (v − 1) mod n + 1).

An ordered pair (p,q) is called allocation (distribution) of some units of goods p to compartment q(p ∈
I , q ∈ H ).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 570 — #16

Population Learning Algorithm and Its Applications 31-571

PLA for GSSP is based on the following assumptions:

• Individuals are represented by permutations of numbers from the {1, . . . , mn} set where each num-
ber can be transformed to an allocation of goods to a compartment. For example, the chromosome
chk = [2 5 3 8 1 7 4 6] with m = 2, n = 4 is transformed into an allocation (1, 2) (2, 1) (1, 3) (2, 4)
(1, 1) (2, 3) (1, 4) (2, 2).
• Each solution represented by an individual is directly evaluated in terms of its fitness by the
decoding procedure (DECODE). The fitness of the individual, f , reflects the cost of allocating all
consignments to available compartments:


m 
n+1
f (chk ) = cij xyk
i=1 j=1

where xijk DECODE(ch)k .

• Initial population J consists of randomly generated permutations from the {1, . . . , mn} set.
• There are three learning stages.
• Selection criterion requires that all individuals with fitness below an average be rejected.
The respective learning-improvement procedures are shown in the following pseudo code:

Procedure L1 :
begin
for each individual in J do
Create string Zj containing allocations (p,q) for which xkpq >0
and Nj containing allocations for which xkpq =0. Create new
individual by substituting the most expensive allocation i.e.
allocation (p,q) for which the expression xkpq (cp,n+1 –cpq ) has
the greatest value with element randomly chosen from Nj string.
Accept exchange if the new individual is an improvement,
otherwise recover an old individual;
end for
end
Procedure L2 :
begin
for each individual in J do
Choose randomly the position tj (tj ∈ {1, . . . ,mn}). Create a new
individual from the old one cyclically shifting its elements
to the left by tj positions. Accept if the new individual is an
improvement otherwise recover an old individual;
end for
end
Procedure L3 :
begin
for each individual in J do
Create two strings as in L1 . A new individual is created from
the old one by randomly changing position of each element of
the Zj string. Accept if the new individual is an improvement
otherwise recover an old individual;
end for
end
Procedure DECODE(ch):

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 571 — #17

31-572 Handbook of Bioinspired Algorithms and Applications

begin
ai ’ ← a i ;
bj ’ ← bj ;
for k=1 to mn do
(p,q) ← g(chk )
if allocation (p,q) is feasible then
xpq ← min(ap ’, bq ’);
ap ’ ← ap ’ - xpq ;
bq ’ ← bq ’ - xpq ;
end if
end for
Store non-allocated goods into external ((n+1)–th) compartment;
end

To validate the approach, the performance of the proposed PLA implementation has been compared
with that of the “classic” evolutionary algorithm. The respective experiment involved two randomly
generated data sets of 10 instances each with number of goods and a number of compartments from
U [5,20] denoting the discrete uniform distribution between 5 and 20 inclusively. In data set 1, the internal
storage costs cij were randomly drawn from U [10,19], the external storage costs ci,n+1 from U [20,24].
In data set 2, the internal storage costs cij were randomly drawn from U [100,199], the external storage
costs ci,n+1 from U [200,249]. Quantities of cargos, ai and capacities of compartments, bj were randomly
generated from U [1,9] in both data sets. Elements of segregation matrices S and CS were generated from
U [0,5] and U [0,4], respectively.
All generated problem instances have been solved by EA and PLA. Optimum solutions have been also
obtained using a CPLEX solver. Mean relative error from optimal solution in case of the PLA is 2.17% (with
2.61% in case of EA). Mean relative computational effort is 91 and 100 for the PLA and EA, respectively.
The next example involves applying the population learning algorithm to solve the resource constrained
project scheduling problem (RCPSP) with makespan minimization as the objective function. A single-
mode RCPSP is considered. In a single-mode case a project consists of a set of activities, where each
activity has to be processed in a single, prescribed way (mode). Each activity requires some resources,
availability of which is constrained. The discussed problem is computationally difficult and belongs to
the NP-hard class. Because of its practical importance RCPSP has attracted a lot of attention and many
exact and heuristic methods have been proposed for solving it (see e.g., Christofides et al. [1987]). Exact
algorithms seem suitable for solving smaller instances of RCPSP. On the other hand, heuristic approaches,
used for solving its larger instances, can only be evaluated experimentally using benchmark datasets with
known optimal solutions or upper bounds.
A project consists of a set of n activities, where each activity has to be processed without interruption to
complete the project. The dummy activities 1 and n represent the beginning and end of the project. The
duration of an activity j is denoted by dj , where d1 = dn = 0. There are R renewable resource types. The
availability of each resource type k in each time period is Rk units, k = 1, . . . , r. Each activity j requires rjk
units of resource k during each period of its duration where r lk = rnk = 0, k = 1, . . . , r. All parameters
are nonnegative integers. There are precedence relations of the finish-start type with a zero parameter
value (i.e., FS = 0) defined between the activities. In other words activity i precedes activity j if j cannot
start until i has been completed. The structure of a project can be represented as an activity-on-node
network G = (V , A), where V is the set of activities and A is the set of precedence relationships. Sj (Pj )
is the set of successors (predecessors) of activity j. It is further assumed that 1 ∈ Pj , j = 2, . . . , n, and
n ∈ Sj , j = 1, . . . , n − 1. The objective is to find a schedule S of activities, that is, a set of starting times
(s1 , . . . , sn ) where s1 = 0 and resource constraints are satisfied, such that the schedule duration T (S) = sn
is minimized.
The respective PLA implementation involves three learning and improvement stages. The value of the
goal function is directly used as a measure of quality of individuals and, hence, as a selection criterion. An

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 572 — #18

Population Learning Algorithm and Its Applications 31-573

individual is represented as a vector of activities (schedule) S = [a1 , . . . , an ], each activity aj is an object

consisting of: starting time, sj , duration, dj , set of required units of resources, set of predecessors, SPj , and
set of successors, SSj .
The algorithm requires also that values of the following parameters are set:

• p — a multiplier used to calculate the size of the initial population.

• xi1 — a coefficient used to calculate the number of iterations at the first learning and improvement
stage.
• xm1 — a coefficient used to calculate a frequency of calls to a selection procedure within the first
learning and improvement stage.
• xi2 — a coefficient used to define a number of iterations at the second learning and improvement
stage.

Values of the above control parameters are set at the algorithm fine-tuning phase. All random moves
within the algorithm are drawn from the uniform distribution. All parameter values used in the following
pseudo-code have been set by trials and errors during the fine-tuning phase. In the pseudo-code shown
below P denotes the population, and |P| its size.

Procedure PLA:
begin
Set size of P, |P|= p*n;
Set values of xi1, xm1, xi2;
Generate the initial population;
{-- the first learning stage}
for it=1 to xi1*n do
for i=1 to 0.4*|P| do
Crossover(Random(P), Random(P));
end for
for i=1 to 0.05*|P|do
Mutation(Random(P));
end for
for i=1 to 0.05*|P| do
LSA(Random(P),2,6,10);
end for
if it mod (xm1*n)=0 then
Selection(medium makespan);
end if
end for
Selection(medium makespan);
{-- the second learning stage}
for it=1 to xi2*n do
for i=1 to 0.4*|P|do
Crossover(Random(P), Random(P));
end for
for 2 best solutions S∈P do
EPTA(S,6,4);
end for
for i=1 to 0.2*|P|do
LSA(Random(P),2,6,10);
end for
end for
Selection(medium makespan);

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 573 — #19

31-574 Handbook of Bioinspired Algorithms and Applications

{-- the third learning stage}

for every solution S∈P do
EPTA(S,6,2);
LSA(S,10,2,10_);
end for
end

The algorithm creates an initial population by producing four individuals using simple construction
heuristics and generating randomly the remaining ones. Heuristics are based on the following rules:
shortest duration first, shortest duration last, longest duration first, and longest duration last. The first
learning stage uses evolutionary operators and a simple local search algorithm (LSA). Three procedures,
such as Crossover, Mutation, and LSA are repeated xi1∗ n times. The Random(P) function chooses
randomly an individual from the population P.

Procedure Crossover(S1 ,S2 ):

begin
One-point crossover on two randomly chosen individuals S1 ,S2 ∈P;
end
Procedure Mutation(S):
begin
Move an activity from randomly chosen position in the schedule S
to another randomly chosen position such that no successor of this
activity can be found before it;
end
Procedure LSA(S, itNumber, iStep, fStep):
begin
for it=1 to itNumber do
for s= iStep to fStep do
for j=1 to scheduleLength-s do
Exchange activities at positions j and j+s;
if not the new solution is better
then recall the exchange;
end for
end for
end for
end

The LSA procedure requires four variables. The first (S) denotes an individual, the second (itNumber)
defines a number of iterations within the procedure. The last two (iStep, fStep) indicate the distance
between activities under exchange.
At the second learning stage a crossover and two heuristics, EPTA (exact precedence tree algorithm)
and LSA are used. EPTA is based on the precedence tree approach. It finds an optimum solution by
enumeration for a part of the schedule consisting of a sequence of activities, with the number of activities
in such sequence denoted as partExtent. In the following pseudo-code the variable S denotes an individual,
and |S| its current size (a number of activities). The step variable denotes a distance between starting points
of the considered partitions.

Procedure EPTA(S, partExtent, step):

begin
i=1;
while i*step + partExtent < |S|do
Find an optimal solution for a part of the schedule beginning

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 574 — #20

Population Learning Algorithm and Its Applications 31-575

from activity at position i*step and ending with activity at

position i*step + partExtent;
end while
end

Finally, at the third learning stage two heuristics EPTA and LSA are again used with settings resulting
in more iterations and higher granularity of the neighborhood explored as compared with the previous
stages.
To validate the proposed approach a computational experiment has been carried based on 1440
benchmark instances of the single-mode RCPSP. The benchmark data set used in the reported
experiments includes 480 instances for each out of the three problem sizes (30, 60, 90, and 120
activities, respectively). The benchmark data set together with known upper bounds can be found at
http://www.wior.uni-karlsruhe.de/rcpsp
The fine-tuning phase of the experiment has been devoted to finding values of the PLA parameters
assuring acceptable compromise between computation time and quality of solutions. This search has been
carried using the subset of available benchmark instances consisting of all RCPSP with 30 activities. The
resulting setting is shown in Table 31.10.
The experiment involved solving all benchmark instances twice. The results were evaluated in terms
of mean and maximum relative errors (MRE, max RE), percent of solutions equal to respective upper
bounds (eqUB) as well as mean computation time (MCT) needed to solve a single instance. Relative errors
have been calculated considering the available upper bounds. Experiment results are shown in Tables 31.11
and 31.12. The proposed PLA implementation has contributed to finding an improved upper bound for
one benchmark instance with 90 activities. The experiment was carried out on a PC computer with the
AMD XP 1600+ processor and 256 MB RAM.
Application of the population learning algorithm to solving several difficult combinatorial problems
proves that it is a useful tool extending the range of available techniques. Although the PLA does not seem
to be able to produce very good solution quickly it is a technique providing consistently a good to very
good solution in reasonable time, leading in many instances to improving upper bounds on most difficult
combinatorial problems.

TABLE 31.10 Six Variants of

the Parameter Settings for the
Computational Experiment

Variant p xi1 xm1 xi2

a 3 0.5 0.333 0.333

b 3 0.5 0.333 0.5
c 3 0.5 0.5 0.5
d 4 0.5 0.333 0.333
e 4 0.5 0.333 0.5
f 4 0.5 0.5 0.5

TABLE 31.11 The Experiment Results (30 and 60 Activities)

Variant 30 activities 60 activities

MRE% max RE% eqUB% MTC [s.] MRE% max RE% eqUB% MTC [s.]

a 0.11 4.35 94.17 1.32 0.53 6.06 76.25 14.75

b 0.11 3.45 93.96 1.57 0.47 4.55 75.28 18.25
c 0.13 5.17 93.54 1.55 0.48 5.05 77.08 18.28
d 0.10 3.95 94.38 1.66 0.51 5.05 76.25 18.71
e 0.10 3.95 94.38 1.97 0.43 5.05 78.13 23.07
f 0.11 3.95 93.96 1.97 0.46 5.05 77.08 23.14

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 575 — #21

31-576 Handbook of Bioinspired Algorithms and Applications

TABLE 31.12 The experiment results (90 and 120 activities)

Variant 90 activities 120 activities

MRE% max RE% eqUB% MTC [s.] MRE% max RE% eqUB% MTC [s.]

a 0.64 6.42 75.83 73.46 1.05 8.09 74.92 217

b 0.58 5.65 76.04 91.93 0.93 6.85 74.60 242
c 0.58 5.50 76.04 91.02 0.92 8.22 75.42 462
d 0.63 6.42 76.04 93.15 0.91 6.86 77.08 384
e 0.54 5.50 76.46 116.92 0.87 6.92 78.04 487
f 0.52 5.45 76.88 115.05 0.86 6.69 78.02 437

31.4 Example Applications of the PLA to Machine Learning

31.4.1 Application of the Population Learning Algorithm to Training
Feed-Forward Artificial Neural Networks (ANN)
Artificial neural networks (ANN) are, nowadays, being used for solving a wide variety of real-life problems
like, for example, pattern recognition, prediction, control, combinatorial optimization, or classification.
Main advantages of the approach include the ability to tolerate imprecision and uncertainty, while still
retaining the possibility of achieving tractability, robustness, and low cost in practical applications. Since
training a neural network for practical application is often very time-consuming, an extensive research
work is being carried out in order to accelerate this process. Another problem with ANN training methods
is danger of being caught in a local optimum. Hence, researchers look not only for algorithms that train
neural networks quickly but rather for quick algorithms that are not likely, or less likely, to get trapped in
a local optima.
In this section an application of the population learning algorithm to ANN training is investigated. The
idea of applying various implementations of the population learning algorithm to ANN training has been
suggested in Czarnowski and Jedrzejowicz (2002a, 2002b). Several implementations of the PLA have been
proposed and applied to solving variety of benchmark problems. Initial results were promising showing
a good performance of the PLA as a tool for ANN training.
The proposed implementations are based on the following assumptions:

• An individual is a vector of real numbers from the predefined interval, each representing a value of
weight of the respective link between neurons in the considered ANN.
• The initial population of individuals is generated randomly.
• There are five learning/improvement procedures used — standard mutation, local search,
nonuniform mutation, gradient mutation, and gradient adjustment.
• There is a common selection criterion for all stages. At each stage, individuals with fitness below
the current average are rejected.

All five learning and improvement procedure L(1) to L(5) used for ANN training are shown in the
following pseudo-code:

Procedure L(1): { standard mutation}

begin
for i=1 to J do
for i=1 to k do {k -- number of iteration}
Select randomly two different elements x1 and x2 within individual i;
Generate new random value of x1 and x2 producing new individual;
Calculate the fitness of a new individual;
if new fitness < old fitness then accept changes;
end for

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 576 — #22

Population Learning Algorithm and Its Applications 31-577

end for
end
Procedure L(2): {local search}
begin
for i=1 to J do
for i=1 to k do {k -- number of iterations}
Select randomly two different elements x1 and x2 within individual i;
Exchange values between x1 and x2 producing a new individual;
Calculate the fitness of a new individual;
if new fitness < old fitness then accept changes;
end for
end for
end
Procedure L(3): {non-uniform mutation}
begin
for i=1 to J do
for t=1 to T do
Select random point within the individual i;
Apply non-uniform mutation;
Calculate the fitness of a new individual;
if new fitness < old fitness then accept change;
end for
end for
end
Procedure L(4): {gradient mutation}
begin
for i=1 to J do
for i=1 to k do {k -- number of iteration}
Select randomly two different elements x1 and x2 within
individual i;
Generate a random binary digit;
Change values x1 and x2 to x1 + ξ and x2 + ξ if a random digit
is 0 and to x1 -- ξ and x2 -- ξ otherwise, producing a new
individual;
Calculate the fitness of a new individual;
if new fitness < old fitness then accept changes;
end for
end for
end
Procedure L(5): {gradient adjustment operator}
begin
for i=1 to J do
α =1;
while α > 0 do
Apply gradient adjustment operator;
Calculate fitness of the new individual;
if new fitness < old fitness then accept changes and break;
α = α − 0.02;
end while
end for
end

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 577 — #23

31-578 Handbook of Bioinspired Algorithms and Applications

The above-proposed procedures require some additional comments. The first procedure, standard
mutation, modifies an individual by generating new values of its two randomly selected elements. If the
operation improves the fitness function value then the change will be accepted. The second learning and
improvement procedure involves exchange of values between the two randomly selected elements within
an individual. If the operation improves the fitness function value then the change will be accepted. The
third procedure, a nonuniform mutation, involves modifying an individual by repeatedly adjusting value
of the randomly selected element (in this case a real number) until the fitness function value has improved
or until a number of the consecutive improvements have been attempted unsuccessfully. The value of the
adjustment is calculated as:
 
(t , y) = y 1 − r (1−(t /T ))·r ,

where r is the uniformly distributed real number from (0, 1], T is equal to the length of the vector
representing an individual and t is a current number of adjustment. The fourth learning and improve-
ment procedure, a gradient mutation, changes two randomly selected elements within an individual by
incrementing or decrementing their values. Direction of change (increment/decrement) is random and
has identical probabilities equal to 0.5. The value of change is proportional to the gradient of an indi-
vidual. If the fitness function value of an individual has improved then the change is accepted. Number
of iterations for learning and improvement procedures 1, 2, and 4 has to be set at the fine-tuning phase.
Finally, the fifth procedure adjusts the value of each element of the individual by a constant value delta
(), proportional to its gradient. Delta is calculated as  = α · ξ , where α is the factor determining a size
of the step in direction of ξ , known as a momentum. Factor α takes values from (0, 1]. In the proposed
algorithm its value iterates starting from 1 with the step equal to 0.02. Here, ξ is a vector determining the
direction of search and is equal to the gradient of an individual.
Population learning algorithm for ANN training has been implemented in two versions — sequential
and parallel. A general structure of the sequential implementation is shown in the following pseudo-code:

Procedure Sequential_PLA
begin
Generate initial population;
P =: initial population;
for i=1 to 4 do
Procedure L(i);
P = Select(P);
end for
Procedure L(5);
Consider best individual in P as a solution;
end

The parallel PLA implementation is based on the cooperation between the master worker (server)
whose task is to manage computations and a number of slave workers, who act in parallel, performing
computations as requested by the master. The approach allows a lot of freedom in designing population-
learning process. The master worker is managing communication flow during the population learning. It
allocates computational tasks in terms of the required population size and the number of iterations and
also controls information exchange between slaves. The latter task involves upgrading, at various learning
stages, current populations maintained by all slaves with globally best individuals. Communication flow
between the master and slave workers is shown in Figure 31.4.
The parallel PLA implementation for ANN training denoted Parallel_PLA is based on the following
rules:
• Master worker defines the number of slave workers and the size of the initial population for each
of them.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 578 — #24

Population Learning Algorithm and Its Applications 31-579

Slave
worker-1

Master Slave
worker worker-2

Slave
worker-N

FIGURE 31.4 Communication flow between the master and slave workers.

• Each slave worker uses the same, described earlier, learning and improvement procedures.
• Master worker activates parallel processing.
• After completing each stage, workers inform master about the best solution found so far.
• Master worker compares the received values and sends out the best solution to all the workers
replacing their current worst individual.
• Master worker can stop computations if the desired quality level of the objective function has been
achieved. This level is defined at the beginning of computations through setting the desired value of
the mean squared error on a given set of training patterns. Alternatively, computations are stopped
after the predefined number of iterations at each stage has been executed.
• Slave workers can also stop computations if the above condition has been met.

A variant of the above approach, denoted Parallel-PLAs, differs with respect to the strategy of using
the best solution sent to slaves from the master. Now, the best solution is used to produce offspring by
applying two types of the crossover operators — a single point crossover and a position crossover. Both are
applied in turn to generate an offspring from all of the current individuals for each slave. Each individual
in each slave is coupled with the best current solution forwarded from the master to produce an offspring.
This takes place after completing each stage.
The proposed implementations of the PLA have been used to train several neural networks applied to
solving popular benchmarking classification problems — two spirals, 10-parity, Wisconsin breast cancer,
Cleveland heart disease, and credit approval.
Artificial neural network algorithms solving the two spirals problem has two real-valued inputs corres-
ponding to the x and y coordinates of the point, and one binary target output that classifies the point as
belonging to either of the two spirals coiling three times around the origin. The two spirals are constructed
from 97 points given by the respective coordinates from the training set. Solving the two spirals problem
appears to be a very difficult task for back-propagation networks (Shang and Wah, 1996). The topology of
the network for this problem has been adopted from (Fahlman and Lebiere, 1990), where also an efficient
constructive training algorithm called CASCOR was suggested. This so-called “shortcut” topology has 4
hidden units and 25 links with 25 weights.
In 10-parity problem ANN has to be trained to produce the Boolean “Exclusive OR” function of ten
variables. A two-layer neural network with 10, 10 and 1 neurons in layers 1, 2, and 3, respectively is used.
Diagnosis of breast cancer involves classifying a tumor as either benign or malignant based on
cell descriptions gathered by microscopic examination. Breast cancer databases were obtained from
Dr. William H. Wolberg, University of Wisconsin Hospitals, Madison (see [Mangasarian and Wolberg,
1990]). It includes 699 examples, 9 inputs and 2 outputs each. The corresponding ANN has, 9, 9, and 1
neuron in layers 1, 2, and 3, respectively.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 579 — #25

31-580 Handbook of Bioinspired Algorithms and Applications

TABLE 31.13 Correct Classification Ratios (%), Mean Errors and Mean Training Times for the PLA
Implementations

TTV 10CV Training

Problem Approach mean max min mean max min MRE MSE time [m.]

Two spirals Sequential-PLA 71.6 82.0 58.9 74.7 83.2 65.0 0.29 0.25 45.0
Two spirals Parallel-PLA 89.2 98.4 70.2 88.4 96.6 74.0 0.07 0.04 20.0
Two spirals Parallel-PLAs 91.7 98.0 77.0 92.0 98.2 80.0 0.07 0.04 18.0
10 parity Sequential-PLA 88.1 92.0 72.0 88.0 97.4 70.0 0.12 0.11 3.50
10 parity Parallel-PLA 93.6 98.0 80.5 92.2 96.0 75.0 0.06 0.04 2.00
Breast Sequential-PLA 95.4 98.5 82.2 96.4 98.0 78.0 0.02 0.01 2.50
Breast Parallel-PLAs 96.6 98.0 82.2 96.6 98.0 80.8 0.02 0.01 1.50
Heart Sequential-PLA 81.0 88.0 65.0 81.3 89.0 62.3 0.21 0.17 2.70
Heart Parallel-PLAs 85.7 90.0 68.0 86.5 89.2 70.0 0.11 0.10 1.50
Credit Sequential-PLA 82.0 86.4 63.0 82.6 87.4 63.0 0.20 0.18 3.40
Credit Parallel-PLAs 88.1 91.9 74.0 86.6 90.4 79.2 0.11 0.11 2.50

Cleveland heart disease problem involves predicting heart disease, that is, deciding whether at least one
of four major vessels is reduced in diameter by more than 50% . The binary decision is made based on
personal data. The data set includes 303 examples with 13 inputs and 2 outputs and the corresponding
ANN has 13, 13, and 1 neuron in layers 1, 2, and 3, respectively.
Credit card approval involves predicting the approval or rejection of a credit card request. The data set
consists of 690 examples with 15 inputs and 2 outputs with a good mix of attributes. The corresponding
ANN to be trained has 15, 15, and 1 neuron in layers 1, 2, and 3, respectively.
All data sets for the above problems are available at the UCI repository (Merz and Murphy, 1998) and are
often used to compare various techniques for ANN training. Data set for the 2-spirals problem is available
and has been obtained from the home page http://www.cae.wisc.edu/∼ece602/data/CMUbenchmark
Computational experiment designed to validate the proposed PLA implementations have been based
on the following quality measures:

• Correct classification ratio for the “10 cross validation” (10CV) approach.
• Correct classification ratio for the training, test, and validation sets constructed using the
50–25–25% (TTV) principle (Prechelt and Proben, 1994).
• Mean squared classification error, MSE.
• Mean relative classification error, MRE.

All artificial neural networks used in the experiment have sigmoid activation function with the sigmoid
gain equal to 1. The initial population size in all PLA implementations has been set to 200. A maximum
number of repetitions for each learning and improvement procedure has been set to 20 except for the
2-spirals problem in both, sequential and parallel versions, where it has been set to 40. Each benchmarking
problem has been solved 50 times and the reported values of the quality measures have been averaged
over these 50 runs. All computations have been carried on the Sun Challenge R4400 and, additionally, for
the Parallel_PLA and Parallel_PLAs using the PVM environment. Experiment results are shown in
Table 31.12. In Table 31.13 the performance of the PLA implementations is compared with some reported
results from the literature.
It can be observed that PLA implementations guarantee good or even competitive level of training
quality. It seems that the approach is also interesting in terms of computational time requirements. The
author has not been able to find comparable data on time performance of the alternative approaches.
Some conclusions can be drawn from the information on time performance of the Novel algorithm,
which required 900 min to train ANN solving 2-spirals problem and 35 min to train ANN solving
10-parity problem (Shang and Wah, 1996).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 580 — #26

Population Learning Algorithm and Its Applications 31-581

31.4.2 An Approach to Instance Reduction in Supervised Learning

As it has been observed by Wilson and Martinez (2000) in the supervised learning, a machine-learning
algorithm is shown a training set, which is a collection of training examples called instances. Each instance
has an input vector and an output value. After learning from the training set, the learning algorithm is
presented with additional input vectors, and the algorithm must generalize, that is, decide what the output
value should be. It is well known that in order to avoid excessive storage and time complexity, and possibly,
to improve generalization accuracy by avoiding noise and overfitting it is often necessary to reduce original
training set by removing some instances before learning phase or to modify the instances using a new
representation.
In the instance reduction approach a subset S of instances to keep from the original training set T
can be obtained using incremental, decremental, or batch search. An incremental search begins with an
empty subset S and adds an instance to S if it fulfills some criteria. The decremental search begins with
S = T , and then searches for instances to remove from S. Finally, in a batch search mode, all instances
are evaluated using some removal criteria. Then, all those that do meet these criteria are removed in a
single step.
This section presents a collection of heuristic algorithms IRA1 to IRA4 used to decide whether an
instance should be kept or removed from the training set. The algorithms were originally proposed in
Czarnowski and Jedrzejowicz (2003).
Let N denote a number of instances in T and n be a number of attributes. Total length of each instance
(i.e., training example) is equal to n + 1, where element numbered n + 1 contains the output value. Let
also X = {xij}(i = 1, . . . , N , j = 1, . . . , n + 1) denote a matrix of n + 1 columns and N rows containing
values of all instances from T . Each of the proposed algorithms groups instances from input set T into a
number of clusters with identical value of identity factor, which is used to indicate similar instances.
IRA1 involves the following steps:

Step 1. Transform X normalizing value of each xij into interval [0, 1] and then rounding it to the
nearest integer, that is 0 or 1.
Step 2. Calculate for each instance from the original training set the value of its identity factor Ii :


n+1
Ii = xij sj , i = 1, . . . , N ,
j=1

where

N
Sj = xij , j = 1, . . . , n + 1.
i=1

Step 3. Map input vectors (i.e., rows from X ) into t clusters denoted as Yv , v = 1, . . . , t . Each
cluster contains input vectors with identical value of the identity factor Ii , where t is a number of
different values of Ii .
Step 4. Set value of the representation level K , which denotes the maximum number of input
vectors to be retained in each of t clusters defined in step 3. Value of K is set arbitrarily by the user.
Step 5. Select input vectors to be retained in each cluster. Let yv denote a number of input vectors
in cluster v, v = 1, . . . , t . Then the following rules for selecting input vectors apply:
• If yv ≤ K then S = S ∪ Yv .
• If yv > K then the order of input vectors in Yv is randomized and the cluster is partitioned
into q = yv /K subsets denoted as Dvj , j = 1, . . . , q. Generalization accuracy of each subset
denoted as Avj is calculated carrying out the leave-(q-1)-out test with X = X − Yv + Dvj as a
training set. Subset of input vectors from cluster v maximizing value of Avj is kept in the reduced
training set.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 581 — #27

31-582 Handbook of Bioinspired Algorithms and Applications

IRA2 uses two additional parameters: pr, precision and ml, multiple. The first one determines a number
of digits after the decimal point. The second determines a way the identity factor value is rounded. The
algorithm proceeds as follows:

Step 1. Set value of the representation level K , which is the maximum number of input vectors to
be retained in each cluster.
Step 2. Set values pr = 0 and ml = 1.
Step 3. Transform X normalizing value of each xij into interval [0, 1].
Step 4. Round xij to the nearest value of 0 or 1 with precision pr digits after the decimal point.
Step 5. Calculate for each instance from the original training set the value of its identity factor Ii .
Step 6. Round Ii to the nearest multiple of ml.
Step 7. Map instances (i.e., rows from X ) into t clusters Yv , v = 1, . . . , t . Each cluster contains
instances with identical value of the identity factor Ii , and t is a number of such clusters.
Step 8. Select instances to be retained in each cluster. Let yv denote a number of input vectors in
cluster v, v = 1, . . . , t . Then the following rules for selecting input vectors apply:
• If yv ≤ K then S = S ∪ Yv .
• If yv > K and K = 1 then the order of input vectors in Yv is randomized. Generalization
accuracy of each Yv is calculated carrying out the leave-(q-1)-out test. An instance from cluster
v maximizing the generalization accuracy is kept in the reduced training set.
• If yv > K and K > 1 set pr to 1 and then for instances in Yv and for several arbitrary values
of ml (i.e., ml = {10, 20, 30}) repeat steps 4 to 8 calculating new values of identity factor and
creating q subsets denoted as Dvj , j = 1, . . . , q, until a number of elements in each subset is
at least equal to K . Generalization accuracy of each subset Avj is calculated carrying out the
leave-(q-1)-out test with X = X − Yv + Dvj as a training set. The subset of instances from
cluster v maximizing the value of Avj is kept in the reduced training set.

IRA3 uses the population-learning algorithm for the selection of instances to be kept. Steps 1 to 4 in
IRA3 are identical as in IRA1 but step 5 differs:
Step 5. Select instances to be retained in each cluster:

• If yv ≤ K then S = S ∪ Yv .
• If yv > K and K = 1 then S = S ∪ {x }, where xv is a selected reference instance from the cluster
v

n
yv
Yv , where the distance d(xv , µv ) = i=1 (xi − µi ) ) is minimal and µ = (1/yv )
v v 2 v
j=1 xv is
the mean vector of the cluster Yv .
• If yv > K and K > 1 then S = S ∪ {xjv }, where {xjv }, j = 1, . . . , K ) are reference instances from
the cluster Yv selected by applying the PLA.

The PLA implemented for the selection of reference instances maps instances xv from Yv into K subsets
Dvj (j = 1, . . . , K ), such that the sum of the squared Euclidean distances between each instance xv ∈ Dvj
and the mean vector µj of the subset Dvj is minimal. Vectors with minimal distance to the mean vector in
each subset are selected as K reference vectors. This selection method can be associated with one of the
clustering technique known as k-means algorithm (Likas et al., 2001).
A potential solution p is coded as the (K + yv )-element vector. Its first K positions inform how many
subsequent elements from all yv elements are contained in each Dvj (j = 1, . . . , K ). The next yv positions
of the potential solution represent input vector numbers from Yv .
The fitness of an individual p ∈ P where P is a population of individuals can be evaluated as:


K 
J (p) =  p[z] − µj 2 ,
j=1 z∈T

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 582 — #28

Population Learning Algorithm and Its Applications 31-583

where

(K , K + p[j]) if j = 1
T =
j−1
j
(K + i=1 p[i], K + i=1 p[i]) otherwise

and p[z] is an instance number stored at position z of individual p.

The remaining assumptions for the discussed PLA implementation can be summarized as follows:
• The initial population is generated randomly.
• Four learning and improvement procedures are used — random local search, partially-mapped
crossover (PMX), steepest ascent local search and tabu search.
• There is a common selection criterion for all stages — individuals with fitness below the current
average are rejected.
Random local search involves selecting randomly two elements of an individual each representing a
number of the input vector. Both numbers are then exchanged, providing such action improves value of
the fitness function. The procedure is run for a predefined number of iterations.
PMX crossover operates on two parents (individuals) and includes the following steps (Michalewicz,
1994) — make a copy of the second parent, choose an arbitrary part from the first parent, and make
minimal changes in the offspring necessary to achieve the chosen pattern. A crossover is performed on
each individual in turn. A partner is always selected randomly. After the operation is performed the best
out of parents and offspring is selected.
The steepest ascent local search procedure involves selecting randomly an element representing a
number of the input vector and allocating it to a cluster, minimizing the Euclidean distance to the mean
vector, providing such an allocation improves value of the fitness function. The procedure is run for a
predefined number of iterations.
The tabu-search procedure is based on a pair-wise exchange of elements representing numbers of
the input vectors. A standard tabu-search arrangements including tabu list and a short and long term
memories are used (Glover, 1990). The procedure is run for a predefined number of iterations.
IRA 4 is identical to IRA 3 except that for the selection of reference instances the PLA with a real number
representation instead of the permutation one is used.
To validate the proposed instance reduction algorithms it has been decided to use generalization
accuracy as a criterion. A set of artificial neural networks (ANN) was used as learning machines. The
machines were presented with the reduced sets of instances during the supervised learning stage and the
results were compared with those obtained without reducing the respective training sets. To train artificial
neural networks an implementation of the population learning algorithm, was used. Neural network
trained using PLA is further on referred to as the PLAANN. Possibility of applying population learning
algorithms to train ANN was discussed in the previous section. In order to increase efficiency of the
approach it has been decided to use a PVM parallel computing environment.
The experiment involved five datasets from UCI Machine Learning Repository (Merz and Murphy,
1998), which are Wisconsin breast cancer (699 input vectors), Cleveland heart disease (303 input vectors),
credit approval (690 input vectors), and thyroid disease (3772 input vectors). Additional problem is
customer intelligence in banking provided under the EUNITE World Competition (12000 input vectors)
(EUNITE, 2002). All the above problems require classification based on input vectors with both —
continuous and binary attributes.
All ANN used during the experiment have had the MLP structure with 3 layers — input, hidden, and
output. The range of weights has been set to [−10, 10] and the sigmoid activation (transfer) function has
had the sigmoid gain value set to 1.0. Number of neurons in layers 1, 2, and 3, respectively, has been set to
the following values:
• Wisconsin breast cancer — 9, 9, and 1
• Cleveland heart disease — 13, 13, and 1

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 583 — #29

31-584 Handbook of Bioinspired Algorithms and Applications

TABLE 31.14 PLA Versus the Reported Accuracy of Training Algorithms

(source for the literature reported accuracy:
http://www.phys.uni.torun.pl/kmk/projects/datasets.html)

Literature reported Best achieved PLA

Problem Approach accuracy (%) accuracy (%)

Two spirals Novel 94.0 92.0

Two spirals Simulated annealing 84.8 92.0
Two spirals CASCOR 62.0 92.0
Two spirals Back propagation 52.0 92.0
Breast Back propagation 96.7 96.6
Breast C 4.5 94.7 96.6
Heart Back propagation 81.3 86.5
Credit Back propagation 91.0 89.0

TABLE 31.15 Mean Percentages

of Input Vectors Retained by IRA1
to IRA4

Problem Representation level (%)

K =1 K =4 K = 10
Credit 28 55 69
Heart 34 55 67
Breast 37 63 81
Thyroid 9 18 30
CI 1 3 5

• Credit approval — 15, 15, and 1

• Thyroid disease — 21, 21, and 3
• Customer intelligence in banking (CI) — 36, 15, and 1

The above proposed instance reduction algorithms have been used to generate training sets for all con-
sidered problems. For each problem five reduced instance sets have been generated with the representation
levels varied between 1 and 10. In the computational experiment “10 cross-validation” approach was used.
Each dataset was divided into 10 partitions and each reduction technique was applied to a training set T
consisting of 9 of the partitions, from which it reduced a subset S. The ANN was trained using only the
instances in S. Ten such trials were run for each dataset with each reduction algorithm, using a different
partitions as the test set in each trial. Mean percentages of input vectors retained are shown in Table 31.14.
The PLAANN classifier has been run 20 times for each representation level and for each training set
generated by IRA1 to IRA4. Results of thus obtained classifications, averaged over 20 runs, are shown in
Table 31.15. The column “Original set” in Table 31.15 shows results obtained by applying the PLAANN
to the original, non-reduced training set using the “10 cross-validation” approach. Best results for each of
the 5 benchmark problems and for each of the representation levels tried are shown in bold. Best overall
results are underlined.
Applying the proposed variants of IRA clearly results in a substantial reduction of the training set
size as compared with an original data set. Overall performance of the PLAANN classifier seems quite
satisfactory. It is also clear that in majority of cases increasing the representation level leads to a better
performance in terms of the classifier quality at a cost of higher requirements in terms of the computation
time. For the credit, heart and breast problems the PLAANN classifier trained using the reduced training
set, performs with the accuracy of classification comparable to the accuracy achieved on the original
training set. The respective training time decreases, on average, 4 to 6 times (for K = 10). It might be
worth noting that in case of the Customer Intelligence problem the PLAANN applied to the original set of
instances has not been able to find any satisfactory solution in a reasonable time and in fact classification

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 584 — #30

Population Learning Algorithm and Its Applications 31-585

TABLE 31.16 Mean Accuracy (%) of PLAANN in the Classification

Experiment

Reduction algorithm Problem Representation level

K =1 K =4 K = 10 Original set
Credit 74.00 79.87 82.70 85.70
Heart 85.34 78.31 85.10 88.10
IRA1 Breast 93.04 93.68 95.84 96.60
Thyroid 93.45 94.71 95.69 93.10
CI 67.44 69.45 71.80 58.70
Credit 74.00 80.04 83.15 85.70
Heart 85.34 83.45 85.87 88.10
IRA2 Breast 93.04 94.21 94.81 96.60
Thyroid 93.45 94.50 94.68 93.10
CI 67.44 75.31 76.48 58.70
Credit 76.00 81.32 84.33 85.70
Heart 80.10 87.20 87.28 88.10
IRA3 Breast 93.80 94.92 96.20 96.60
Thyroid 93.45 94.80 96.60 93.10
CI 70.00 79.50 80.50 58.70
Credit 76.00 81.45 83.73 85.70
Heart 80.10 87.10 88.10 88.10
IRA4 Breast 93.80 95.10 96.15 96.60
Thyroid 93.45 98.21 98.43 93.10
CI 70.00 80.50 83.40 58.70

TABLE 31.17 Performance Comparison of Different Instance Reduction Algorithm

Breast Heart Credit

Algorithm Accuracy % Retained Accuracy % Retained Accuracy % Retained
IRA(K = 1) 93.80 37.34 76.00 33.66 85.34 28.32
IRA(K = 10) 96.20 81.00 84.33 67.00 88.10 69.00
CNN 95.71 7.09 73.95 30.84 77.68 24.22
SNN 93.85 8.35 76.25 33.89 81.31 28.38
IB2 95.71 7.09 73.96 30.29 78.26 24.15
IB3 96.57 3.47 81.16 11.11 85.22 4.78
DROP3 96.14 3.58 80.84 12.76 83.91 5.96
kNN 96.28 100.00 81.19 100.00 84.78 100.00
PLAANN 96.60 100.00 85.70 100.00 88.10 100.00

process has been stopped. Accuracy of classification obtained with the original set was then only 58% .
However the PLAANN trained on the reduced set guarantees accuracy better then 80% already 120 sec of
computation time.
For the thyroid problem the PLAANN trained using the original data set produces accuracy of classi-
fication at the level of about 92%, which is a standard performance. The accuracy of classification for the
discussed problem has grown substantially with the reduction of the original dataset size.
Comparison of the proposed IRAs with other approaches to instance reduction is shown in Table 31.16.
The column “Retained” in Table 31.16 shows a percentage of training vectors from the original training
set that have been retained by the respective instance reduction algorithm. All the results other than
these produced by the proposed IRAs were reported in Wilson and Martinez (2000). Acronyms used in
Table 31.16 stand for k-Nearest-Neighbor (kNN), Condensed Nearest Neighbor (CNN), Selective Nearest
Neighbor (SNN), Instance Based (IB), and Decremental Reduction Optimization Procedure (DROP).
Computational experiment was carried on a SGI Challenge R4400 workstation with 12 processors.
A number of slave workers used by the master varied in different runs from 5 to 15 and in each run were

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 585 — #31

31-586 Handbook of Bioinspired Algorithms and Applications

chosen randomly. A size of the initial population in the IRA3 implementation of the PLA was set to 100.
A maximum number of repetitions for each learning and improvement procedure was set to 500. In the
IRA4 implementation the respective parameters were set to 50 for the initial population size and 500 for
the number of iterations.
The proposed and provisionally validated simple, PLA-based heuristic instance reduction algorithms,
can be used to increase efficiency of the supervised learning. Computational experiment results support
the claim that reducing training set size still preserves basic features of the analyzed data and can be even
beneficial to a classifier accuracy. The approach extends a range of available instance reduction algorithms.
Moreover, it is shown that the proposed algorithm can be, for some problems, competitive in comparison
with the existing techniques. Possible extension of the approach could focus on establishing decision rules
for finding a representation level suitable for each cluster thus allowing a variable representation level for
different clusters.

References
Barbucha D. and Jȩdrzejowicz, P. A population learning algorithm for solving the generalized segregated
storage problem. In P. Sincak, J. Vascak, V. Kvasnicka, and R. Mesiar (Eds.), The State of the Art. In
Computational Intelligence”, Advances in Soft Computing, Physica-Verlag, Heidelberg, New York,
2000, pp. 355–360.
Chung, C-J. and Reynolds, R.G. CAEP: An evolution-based tool for real-valued function optimization
using cultural algorithms. Journal of Artificial Intelligence Tools, 7, 1998, 239–292.
Colorni, A., Dorigo, M., and Maniezzo, V. An investigation of some properties of an ant algorithm. In
R. Manner and B. Manderick (Eds.), Proceedings of the 2nd European Conference on Parallel Problem
Solving from Nature, Elsevier, Amsterdam, 1992, pp. 509–520.
Christofides, N., Alvarez-Valdes, R., and Tamarit, J.M. Project scheduling with resource constraints:
A branch and bound approach. European Journal of Operational Research, 29, 1987, 262–273.
Czarnowski, I., and Jȩdrzejowicz, P. Application of the parallel population learning algorithm to training
feed-forward ANN. In P. Sincak et al. (Eds.), Intelligent Technologies–Theory and Applications, IOS
Press, Amsterdam, 2002a, pp. 10–16.
Czarnowski, I. and Jȩdrzejowicz, P. An approach to artificial neural network training. In M. Bremer,
A. Preece, F. Coenen (Eds.), Research and Development in Intelligent Systems XIX, Springer-Verlag,
Heidelberg, 2002b, pp. 149–160.
Czarnowski, I. and Jȩdrzejowicz, P. An approach to instance reduction in supervised learning. In F. Coenen,
A. Preece, and A.L. Macintosh (Eds.), Research and Development in Intelligent Systems XX, Springer,
London, 2003, pp. 267–280.
Czarnowski, I., Gutjahr, W.J., Jȩdrzejowicz, P., Ratajczak, E., Skakowski, A., and Wierzbowska, I. Scheduling
multiprocessor tasks in presence of correlated failures. Central European Journal of Operations
Research, 11, 2003, 163–182.
Davis, L. Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York, 1991.
Dawkins, R. The Selfish Gene, Oxford University Press, Oxford, 1976.
Dorigo, M., and Di Caro, G. The ant colony optimization meta-heuristic. In D. Corne, M. Dorigo, and
F. Glover, (Eds.), New Ideas in Optimization, McGraw-Hill, New York, 1999, pp. 11–32.
The European Network of Excellence on Intelligent Technologies for Smart Adaptive Systems (EUNITE) —
EUNITE World competition in domain of Intelligent Technologies — 0, 2002.
Fahlman, S.E. and Lebiere, C. The cascade-correlation learning architecture. In E. Touretzky (Ed.),
Advances in Neural Information Processing II, Morgan Kauffman, San Mateo, CA, 1990,
pp. 524–532.
Feo, T.A. and Rosende, M.G.C. Greedy randomized adaptive search procedures. Journal of Global
Optimization, 6, 1995, 109–133.
Glover, F. Heuristics for integer programming using surrogate constraints. Decision Sciences, 8, 1977,
156–166.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 586 — #32

Population Learning Algorithm and Its Applications 31-587

Glover, F. Tabu search: A tutorial. Interfaces, 20, 1990, 74–94.

Goldberg, D.E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley,
Reading, MA, 1989.
Golden, B.L., Laporte, G., and Taillard, E.D. An adaptive memory heuristic for a class of vehicle routing
problems with minmax objective. Computers & Operations Research, 24, 1997, 445–452.
Gordon, V. and Whitley, D. Serial and parallel genetic algorithms as function optimizers. In S. Forrest (Ed.),
Proceedings of the 5th Internation Conference on Genetic Algorithms, Morgan Kaufman Publishers,
San Mateo, CA, 1993, pp. 177–183.
Hertz, A. and Kobler, D. A framework for the description of population based methods, 16 European
Conferencee on Operational Research, Tutorials and Research Reviews, 1998, 1–21.
Holland, J.H. Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor,
MI, 1975.
Jȩdrzejowicz, P. Social learning algorithm as a tool for solving some difficult scheduling problems.
Foundation of Computing and Decision Sciences, 24, 1999, 51–66.
Jedrzejowicz, J., Jȩdrzejowicz, P. PLA-based permutation scheduling, Foundation of Computing and
Decision Sciences, 28, 2003, 159–177.
Jedrzejowicz, J. and Jȩdrzejowicz, P. Population-based approach to multiprocessor task scheduling in
multistage hybrid flowshops. In V. Palade, R.J. Howlett, and L. Jain (Eds.), Knowledge-Based Intelli-
gent Information and Engineering Systems, Lecture Notes in Artificial Intelligence, No. 2773, Springer,
Berlin, 2003, pp. 279–286.
Jȩdrzejowicz, P., Rosicka, L., and Wierzbowska, I. Population learning algorithms for optimizing system
structure under multiple objectives. In Proceedings of the XVIII European Conference on System
Dependability and Safety – ESREL’02, Lyon, 2002, pp. 420–424.
Jȩdrzejowicz, P. and Ratajczak, E. Scheduling multiprocessor tasks with correlated failures using popula-
tion learning algorithm. V. Kurkova, (Ed.), Artificial Neural Nets and Genetic Algorithms, Springer
Computer Science, Wien-New York, 2001, pp. 296–300.
Jȩdrzejowicz, P. and Ratajczak, E. Population learning algorithm for resource-constrained project schedul-
ing. In D.W. Pearson, N.C. Steele, and R.F. Albrecht (Eds.), Artificial Neural Nets and Genetic
Algorithms, Springer Computer Science, Springer, Wien, 2003, pp. 223-228.
Jȩdrzejowicz, P. and Wierzbowska, I. Scheduling multiple variant programs under hard real time
constraints, European Journal of Operational Research, 127, 2000, 458–465.
Kennedy, J. and Eberhart, R.C. Particle swarm optimisation. In Proceedings of IEEE International
Conference on Neural Networks, Piscataway, NJ, 1995, pp. 1942–1948.
Kirkpatrick, S., Gelatt, C.D., Jr., and Vecchi, M.P. Optimization by simulated annealing, Science, 220, 1983,
671–680.
Likas, A., Vlassis, N., and Verbeek, J.J. The Global k-Means Clustering Algorithm, Technical report,
Computer Science Institute, University of Amsterdam, IAS-UVA-01-02, 2001.
Mangasarian, O.L. and Wolberg, W.H. Cancer diagnosis via linear programming, SIAM News, 23, 1990,
1–18.
Merz, C.J. and Murphy, M. UCI Repository of machine learning databases, (http://www.ics.uci.edu/
∼mlearn/MLRepository.html), University of California, Department of Information and Computer
Science, Irvine CA, 1998.
Michalewicz, Z. Genetic Algorithms + Data Structures = Evolution Programs, 2nd Extended edition,
Springer-Verlag, Berlin, Heidelberg, New York, 1994.
Norman M.G. and Moscato P. A Competitive and Cooperative Approach to Complex Combinatorial
Search, Caltech Concurrent Computation Program, C3P report 790, 1989.
Prechelt, L. and Proben, I. A Set of Benchmark and Benchmarking Rules for Neural
Network Training Algorithm, Fakultät für Informatik, Universität Karlsruhe, Anonym-
ous /pub/papers/techraports/1994/1994-21.ps.z. on ftp.ira.uka.de, Technical report no. 21/94,
Karlsruhe, 1994.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 587 — #33

31-588 Handbook of Bioinspired Algorithms and Applications

Reynolds, R.G. An introduction to Cultural Algorithms. In A.V. Sebald, L.J. Fogel (Eds.), Proceedings of
the 3rd Annual Conference on Evolutionary Programming, World Scientific, River Edge NJ, 1994,
pp. 131–139.
Shang, Yi. and Wah, B.W. A global optimization method for neural network training, Conference of
Neural Networks. IEEE Computer, 29, 1996, 45–54.
Taillard, E. Benchmarks for basic scheduling instances. European Journal of Operational Research, 64, 1993,
278–285.
Wilson, D.R. and Martinez, T.R. Reduction techniques for instance-based learning algorithm. Machine
Learning, 38, 2000, 257–286.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C031” — 2005/8/6 — 14:13 — page 588 — #34

 32
Biology-Derived
Algorithms in
Engineering
Optimization

32.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32-589

32.2 Biology-Derived Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . .32-590
Genetic Algorithms • Photosynthetic Algorithms •
Neural Networks • Cellular Automata • Optimization
32.3 Engineering Optimization and Applications . . . . . . . . . .32-594
Function and Multilevel Optimizations • Shape Design and
Optimization • Finite Element Inverse Analysis •
Inverse IVBV Optimization
Xin-She Yang References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32-599

32.1 Introduction
Biology-derived algorithms are an important part of computational sciences, which are essential to many
scientific disciplines and engineering applications. Many computational methods are derived from or
based on the analogy to natural evolution and biological activities, and these biologically inspired compu-
tations include genetic algorithms, neural networks, cellular automata, and other algorithms. However, a
substantial amount of computations today are still using conventional methods such as finite difference,
finite element, and finite volume methods. New algorithms are often developed in the form of a hybrid
combination of biology-derived algorithms and conventional methods, and this is especially true in the
field of engineering optimizations. Engineering problems with optimization objectives are often difficult
and time consuming, and the application of nature or biology-inspired algorithms in combination with
the conventional optimization methods has been very successful in the last several decades.
There are five paradigms of nature-inspired evolutionary computations: genetic algorithms, evolution-
ary programming, evolutionary strategies, genetic programming, and classifier systems (Holland, 1975;
Goldberg, 1989; Mitchell, 1996; Flake, 1998). Genetic algorithm (GA), developed by John Holland and
his collaborators in the 1960s and 1970s, is a model or abstraction of biological evolution, which includes
the following operators: crossover, mutation, inversion, and selection. This is done by the representation
within a computer of a population of individuals corresponding to chromosomes in terms of a set of

32-589

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 589 — #1

32-590 Handbook of Bioinspired Algorithms and Applications

character strings, and the individuals in the population then evolve through the crossover and mutation
of the string from parents, and the selection or survival according to their fitness. Evolutionary pro-
gramming (EP), first developed by Lawrence J. Fogel in 1960, is a stochastic optimization strategy similar
to GAs. But it differs from GAs in that there is no constraint on the representation of solutions in EP
and the representation often follows the problem. In addition, the EPs do not attempt to model genetic
operations closely in the sense that the crossover operation is not used in EPs. The mutation operation
simply changes aspects of the solution according to a statistical distribution, such as multivariate Gaussian
perturbations, instead of bit-flopping, which is often done in GAs. As the global optimum is approached,
the rate of mutation is often reduced. Evolutionary strategies (ESs) were conceived by Ingo Rechenberg
and Hans-Paul Schwefel in 1963, later joined by Peter Bienert, to solve technical optimization problems
(Rechenberg, 1973). Although they were developed independently of one another, both ESs and EPs have
many similarities in implementations. Typically, they both operate on real-values to solve real-valued
function optimization in contrast with the encoding in Gas. Multivariate Gaussian mutation with zero
mean are used for each parent population and appropriate selection criteria are used to determine which
solution to keep or remove. However, EPs often use stochastic selection via a tournament and the selec-
tion eliminates those solutions with the least wins, while the ESs use deterministic selection criterion that
removes the worst solutions directly based on the evaluations of certain functions (Heitkotter and Beasley,
2000). In addition, recombination is possible in an ES as it is an abstraction of evolution at the level of
individual behavior in contrast to the abstraction of evolution at the level of reproductive populations
and no recombination mechanisms in EPs.
The aforementioned three areas have the most impact in the development of evolutionary computations,
and, in fact, evolutionary computation has been chosen as the general term that encompasses all these
areas and some new areas. In recent years, two more paradigms in evolutionary computation have
attracted substantial attention: Genetic programming and classifier systems. Genetic programming (GP)
was introduced in the early 1990s by John Koza (1992), and it extends GAs using parse trees to represent
functions and programs. The programs in the population consist of elements from the function sets,
rather than fixed-length character strings, selected appropriately to be the solutions to the problems.
The crossover operation is done through randomly selected subtrees in the individuals according to their
fitness; the mutation operator is not used in GP. On the other hand, a classifier system (CFS), another
invention by John Holland, is an adaptive system that combines many methods of adaptation with learning
and evolution. Such hybrid systems can adapt behaviors toward a changing environment by using GAs
with adding capacities such as memory, recursion, or iterations. In fact, we can essentially consider the
CFSs as general-purpose computing machines that are modified by both environmental feedback and the
underlying GAs (Holland, 1975, 1995; Michaelewicz, 1996; Flake, 1998).
Biology-derived algorithms are applicable to a wide variety of optimization problems. For example,
optimization functions can have discrete, continuous, or even mixed parameters without any a priori
assumptions about their continuity and differentiability. Thus, evolutionary algorithms are particularly
suitable for parameter search and optimization problems. In addition, they are easy for parallel implement-
ation. However, evolutionary algorithms are usually computationally intensive, and there is no absolute
guarantee for the quality of the global optimizations. Besides, the tuning of the parameters can be very
difficult for any given algorithms. Furthermore, there are many evolutionary algorithms with different
suitabilities and the best choice of a particular algorithm depends on the type and characteristics of the
problems concerned. However, great progress has been made in the last several decades in the application
of evolutionary algorithms in engineering optimizations. In this chapter, we will focus on some of the
important areas of the application of GAs in engineering optimizations.

32.2 Biology-Derived Algorithms

There are many biology-derived algorithms that are popular in evolutionary computations. For engin-
eering applications in particular, four types of algorithms are very useful and hence relevant. They are

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 590 — #2

Biology-Derived Algorithms in Engineering Optimization 32-591

GAs, photosynthetic algorithms (PAs), neural networks, and cellular automata. We will briefly discuss
these algorithms in this section, and we will focus on the application of GAs and PAs in engineering
optimizations in Section 32.3.

32.2.1 Genetic Algorithms

The essence of GAs involves the encoding of an optimization function as arrays of bits or character
strings to represent the chromosomes, the manipulation operations of strings by genetic operators, and
the selection according to their fitness to find a solution to the problem concerned. This is often done
by the following procedure: (1) encoding of the objectives or optimization functions; (2) defining a
fitness function or selection criterion; (3) creating a population of individuals; (4) evolution cycle or
iterations by evaluating the fitness of all the individuals in the population, creating a new population by
performing crossover, mutation, and inversion, fitness-proportionate reproduction, etc., and replacing
the old population and iterating using the new population; (5) decoding the results to obtain the solution
to the problem.
One iteration of creating new populations is called a generation. Fixed-length character strings are used
in most GAs during each generation although there is substantial research on variable-length string and
coding structures. The coding of objective functions is usually in the form of binary arrays or real-valued
arrays in adaptive GAs. For simplicity, we use the binary string for coding for describing genetic operators.
Genetic operators include crossover, mutation, inversion, and selection. The crossover of two parent
strings is the main operator with highest probability pc (usually, 0.6 to 1.0) and is carried out by switching
one segment of one string with the corresponding segment on another string at a random position (see
Figure 32.1). The crossover carried out in this way is a single-point crossover. Crossover at multiple
points is also used in many GAs to increase the efficiency of the algorithms. The mutation operation is
achieved by the flopping of randomly selected bits, and the mutation probability pm is usually small (say,
0.001 to 0.05), while the inversion of some part of a string is done by interchanging 0 and 1. The selection
of an individual in a population is carried out by the evaluation of its fitness, and it can remain in the
new generation if a certain threshold of fitness is reached or the reproduction of a population is fitness-
proportionate. One of the key parts is the formulation or choice of fitness functions that determines the
selection criterion in a particular problem.
Further, just as crossover can be carried out at multiple points, mutations can also occur at multiple
sites. More complex and adaptive GAs are being actively researched and there is a vast literature on this
topic (Chipperfield et al. 1994; Pohlheim, 1999). In Section 32.3, we will give examples of GAs and their
applications in engineering optimizations.

Parent 1 1 1 0 1 1 0 1 0 1 1 0 1 1 1 1 0 Child 1

Parent 2 1 1 0 1 0 1 1 0 1 1 0 1 0 0 1 0 Child 2

Crossover point

1 1 0 1 1 0 1 0 Mutation 1 1 0 0 1 0 1 0

FIGURE 32.1 Diagram of crossover and mutation in a GA.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 591 — #3

32-592 Handbook of Bioinspired Algorithms and Applications

32.2.2 Photosynthetic Algorithms

The PA was first introduced by Murase (2000) to optimize the parameter estimation in finite element
inverse analysis. The PA is a good example of biology-derived algorithms in the sense that its computational
procedure corresponds well to the real photosynthesis process in green plants. Photosynthesis uses water
and CO2 to produce glucose and oxygen when there is light and in the presence of chloroplasts. The
overall reaction

light and green plants

6CO2 + 12H2 O −−−−−−−−−−−−−−−→ C6 H12 O6 + 6O2 + 6H2 O

is just a simple version of a complicated process. Other factors, such as temperature, concentration of CO2 ,
water content, etc., being equal, the reaction efficiency depends largely on light intensity. The important
part of photosynthetic reactions is the dark reactions that consist of a biological process including two
cycles: the Benson–Calvin cycle and photorespiration cycle. The balance between these two cycles can be
considered as a natural optimization procedure that maximizes the efficiency of sugar production under
the continuous variations of light energy input (Murase, 2000).
Murase’s PA uses the rules governing the conversion of carbon molecules in the Benson–Calvin cycle
(with a product or feedback from dihydroxyacetone phosphate or DHAP) and photorespiration reactions.
The product DHAP serves as the knowledge strings of the algorithm and optimization is reached when
the quality or the fitness of the products no longer improves. An interesting feature of such algorithms
is that the stimulation is a function of light intensity that is randomly changing and affects the rate of
photorespiration. The ratio of O2 to CO2 concentration determines the ratio of the Benson–Calvin and
photorespiration cycles. A PA consists of the following steps: (1) the coding of optimization functions in
terms of fixed-length DHAP strings (16-bit in Murase’s PA) and a random generation of light intensity (L);
(2) the CO2 fixation rate r is then evaluated by the following equation: r = Vmax /(1 + A/L), where Vmax
is the maximum fixation rate of CO2 and A is its affinity constant; (3) either the Benson–Calvin cycle or
photorespiration cycle is chosen for the next step, depending on the CO2 fixation rate, and the 16-bit strings
are shuffled in both cycles according to the rule of carbon molecule combination in photosynthetic path-
ways; (4) after some iterations, the fitness of the intermediate strings is evaluated, and the best fit remains
as a DHAP, then the results are decoded into the solution of the optimization problem (see Figure 32.2).
In the next section, we will present an example of parametric inversion and optimization using PA in
finite element inverse analysis.

Light (random)

O2/CO2 concentration

Benson–Calvin Photorespiration
cycle

Fitness remove
evaluation if poor

iterations
DHAP Strings and results

FIGURE 32.2 Scheme of Murase’s PAs.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 592 — #4

Biology-Derived Algorithms in Engineering Optimization 32-593

u1(t )

wi1 Σ−Θi ui (t +1)

wi2

u2(t )

Inputs Hidden layer Outputs

FIGURE 32.3 Diagram of a McCulloch–Pitts neuron (left) and neural networks (right).

32.2.3 Neural Networks

Neural networks, and the associated machine-learning algorithms, is one more type of biology-inspired
algorithms, which uses a network of interconnected neurons with the intention of imitating the neural
activities in human brains. These neurons have high interconnectivity and feedback, and their connectiv-
ity can be weakened or enhanced during the learning process. The simplest model for neurons is the
McCulloch–Pitts model (see Figure 32.3) for multiple inputs u1 (t ), . . . , un (t ), and the output ui (t ). The
activation in a neuron or a neuron’s state is determined by
 

n
ui (t + 1) = H  wij ui (t ) − i  ,
j=1

where H (x) is the Heaviside unit step function that H (x) = 1 if x ≥ 0 and H (x) = 0 otherwise. The
weight coefficient wij is considered as the synaptic strength of the connection from neuron j to neuron i.
For each neuron, it can be activated only if its threshold i is reached. One can consider a single neuron
as a simple computer that gives the output 1 or yes if the weighted sum of incoming signals is greater than
the threshold, otherwise it outputs 0 or no.
Real power comes from the combination of nonlinear activation functions with multiple neurons
(McCulloch and Pitts, 1943; Flake, 1998). Figure 32.3 also shows an example of feed-forward neural
networks. The key element of an artificial neural network (ANN) is the novel structure of such an
information processing system that consists of a large number of interconnected processing neurons.
These neurons work together to solve specific problems by adaptive learning through examples and self-
organization. A trained neural network in a given category can solve and answer what-if type questions
to a particular problem when new situations of interest are given. Due to the real-time capability and
parallel architecture as well as adaptive learning, neural networks have been applied to solve many real-
world problems such as industrial process control, data validation, pattern recognition, and other systems
of artificial intelligence such as drug design and diagnosis of cardiovascular conditions. Optimization is
just one possibility of such applications, and often the optimization functions can change with time as
is the case in industrial process control, target marketing, and business forecasting (Haykins, 1994). On
the other hand, the training of a network may take considerable time and a good training database or
examples that are specific to a particular problem are required. However, neural networks will, gradually,
come to play an important role in engineering applications because of its flexibility and adaptability in
learning.

32.2.4 Cellular Automata

Cellular automata (CA) were also inspired by biological evolution. On a regular grid of cells, each cell has
finite number of states. Their states are updated according to certain local rules that are functions of the
states of neighbor cells and the current state of the cell concerned. The states of the cells evolve with time

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 593 — #5

32-594 Handbook of Bioinspired Algorithms and Applications

in a discrete manner, and complex characteristics can be observed and studied. For more details on this
topic, readers can refer to Chapters 1 and 18 in this handbook.
There is some similarity between finite state CA and conventional numerical methods such as finite
difference methods. If one considers the finite different method as real-valued CA, and the real-values are
always converted to finite discrete values due to the round-off in the implementation on a computer, then
there is no substantial difference between a finite difference method and a finite state CA.
However, CA are easier to parallelize and more numerically stable. In addition, finite difference schemes
are based on differential equations and it is sometimes straightforward to formulate a CA from the
corresponding partial differential equations via appropriate finite differencing procedure; however, it is
usually very difficult to conversely obtain a differential equation for a given CA (see Chapter 18 in this
handbook).
An optimization problem can be solved using CA if the objective functions can be coded to be associated
with the states of the CA and the parameters are properly associated with automaton rules. This is an
area under active research. One of the advantages of CA is that it can simulate many processes such as
reaction–diffusion, fluid flow, phase transition, percolation, waves, and biological evolution. Artificial
intelligence also uses CA intensively.

32.2.5 Optimization
Many problems in engineering and other disciplines involve optimizations that depend on a number of
parameters, and the choice of these parameters affects the performance or objectives of the system con-
cerned. The optimization target is often measured in terms of objective or fitness functions in qualitative
models. Engineering design and testing often require an iteration process with parameter adjustment.
Optimization functions are generally formulated as:

Optimize: f (x),
Subject to: gi (x) ≥ 0, i = 1, 2, . . . , N ; hj (x) = 0, j = 1, 2, . . . , M .
where x = (x1 , x2 , . . . , xn ), x ∈ (parameter space).

Optimization can be expressed either as maximization or more often as minimization (Deb, 1995,
2000). As parameter variations are usually very large, systematic adaptive searching or optimization
procedures are required. In the past several decades, researchers have developed many optimization
algorithms. Examples of conventional methods are hill climbing, gradient methods, random walk, simu-
lated annealing, heuristic methods, etc. Examples of evolutionary or biology-inspired algorithms are GAs,
photosynthetic methods, neural network, and many others.
The methods used to solve a particular problem depend largely on the type and characteristics of
the optimization problem itself. There is no universal method that works for all problems, and there is
generally no guarantee to find the optimal solution in global optimizations. In general, we can emphasize
on the best estimate or suboptimal solutions under the given conditions. Knowledge about the particular
problem concerned is always helpful to make the appropriate choice of the best or most efficient methods
for the optimization procedure. In this chapter, however, we focus mainly on biology-inspired algorithms
and their applications in engineering optimizations.

32.3 Engineering Optimization and Applications

Biology-derived algorithms such as GAs and PAs have many applications in engineering optimizations.
However, as we mentioned earlier, the choice of methods for optimization in a particular problem depends
on the nature of the problem and the quality of solutions concerned. We will now discuss optimization
problems and related issues in various engineering applications.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 594 — #6

Biology-Derived Algorithms in Engineering Optimization 32-595

32.3.1 Function and Multilevel Optimizations

For optimization of a function using GAs, one way is to use the simplest GA with a fitness function:
F = A − y with A being the large constant and y = f (x), thus the objective is to maximize the fitness
function and subsequently minimize the objective function f (x). However, there are many different ways
of defining a fitness function. For example, we can use the individual fitness assignment relative to the
whole population
f (xi )
F (xi ) = N ,
i=1 f (xi )

where xi is the phenotypic value of individual i, and N is the population size. For the generalized
De Jong’s (1975) test function


n
f (x) = xi2α , |x| ≤ r, α = 1, 2, . . . , m,
i=1

where α is a positive integer and r is the half-length of the domain. This function has a minimum of
f (x) = 0 at x = 0. For the values of α = 3, r = 256, and n = 40, the results of optimization of this test
function are shown in Figure 32.4 using GAs.
The function we just discussed is relatively simple in the sense that it is single-peaked. In reality, many
functions are multi-peaked and the optimization is thus multileveled. Keane (1995) studied the following
bumby function in a multi-peaked and multileveled optimization problem

sin2 (x − y) sin2 (x + y)
f (x, y) =  , 0 < x, y < 10.
x2 + y2

The optimization problem is to find (x, y) starting (5, 5) to maximize the function f (x, y) subject to:
x + y ≤ 15 and xy ≥ 3/4. In this problem, optimization is difficult because it is nearly symmetrical about
x = y, and while the peaks occur in pairs one is bigger than the other. In addition, the true maximum
is f (1.593, 0.471) = 0.365, which is defined by a constraint boundary. Figure 32.5 shows the surface
variation of the multi-peaked bumpy function.
Although the properties of this bumpy function make it difficult for most optimizers and algorithms,
GAs and other evolutionary algorithms perform well for this function and it has been widely used as a test

6 6
Best estimate = 0.046346 Best estimate = 0.23486
5 5

4
4
3
3
log[f(x)]

log[f(x)]

2
2
1
1
0

–1 0

–2 –1
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Generation (t) Generation (t)

FIGURE 32.4 Function optimization using GAs. Two runs will give slightly different results due to the stochastic
nature of GAs, but they produce better estimates: f (x) → 0 as the generation increases.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 595 — #7

32-596 Handbook of Bioinspired Algorithms and Applications

0.6

0.4

f (x,y)
0.2

0
0

x 5
10

5
10 y
0

FIGURE 32.5 Surface of the multi-peaked bumpy function.

Th Ts

R R

FIGURE 32.6 Diagram of the pressure vessel.

function in GAs for comparative studies of various evolutionary algorithms or in multilevel optimization
environments (Jenkins, 1997; El-Beltagy and Keane, 1999).

32.3.2 Shape Design and Optimization

Most engineering design problems, especially in shape design, aim to reduce the cost, weight, and volume
and increase the performance and quality of the products. The optimization process starts with the
transformation of design specification and descriptions into optimization functions and constraints. The
structure and parameters of a product depend on the functionality and manufacturability, and thus
considerable effort has been put into the modeling of the design process and search technique to find
the optimal solution in the search space, which comprises the set of all designs with all allowable values
of design parameters (Renner and Ekart, 2003). Genetic algorithms have been applied in many areas of
engineering design such as conceptual design, shape optimization, data fitting, and robot path design.
A well-studied example is the design of a pressure vessel (Kannan and Kramer, 1994; Coello, 2000) using
different algorithms such as augmented Lagrangian multiplier and GAs. Figure 32.6 shows the diagram
of the parameter notations of the pressure vessel. The vessel is cylindrical and capped at both ends by
hemispherical heads with four design variables: thickness of the shell Ts , thickness of the head Th , inner
radius R, and the length of the cylindrical part L. The objective of the design is to minimize the total
cost including that of the material, forming, as well as welding. Using the notation given by Kannan and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 596 — #8

Biology-Derived Algorithms in Engineering Optimization 32-597

Kramer, the optimization problem can be expressed as:

Minimize: f (x) = 0.6224x1 x2 x3 + 1.7781x2 x32 + 3.1611x12 x4 + 19.84x12 x3 .

x = (x1 , x2 , x3 , x4 )T = (Ts , Th , R, L)T ,
Subject to: g1 (x) = −x1 + 0.0193x3 ≤ 0, g2 (x) = −x2 + 0.00954x3 ≤ 0,
g3 (x) = −π x32 x4 − 4πx33 /3 + 1296000 ≤ 0, g4 (x) = x4 − 240 ≤ 0.

The values for x1 , x2 should be considered as integer multiples of 0.0625. Using the same constraints as
given in Coello (2000), the variables are in the ranges: 1 ≤ x1 , x2 ≤ 99, 10.0000 ≤ x3 . x4 ≤ 100.0000
(with a four-decimal precision). By coding the GAs with a population of 44-bit strings for each indi-
vidual (4-bits for x1 · x2 ; 18-bits for x3 · x4 ), similar to that by Wu and Chow (1994), we can solve
the optimization problem for the pressure vessel. After several runs, the best solution obtained is
x∗ = (1.125, 0.625, 58.2906, 43.6926) with f (x) = 7197.9912$, which is compared with the results
x∗ = (1.125, 0.625, 28.291, 43.690) and f (x) = 7198.0428$ obtained by Kannan and Kramer (1994).

32.3.3 Finite Element Inverse Analysis

The usage and efficiency of Murase’s PA described in Section 32.3.4 can be demonstrated in the application
of finite element inverse analysis. Finite element analysis (FEA) in structural engineering is forward
modeling as the aims are to calculate the displacements at various positions for given loading conditions
and material properties such as Young’s modulus (E) and Poisson’s ratio (ν). This forward FEA is widely
used in engineering design and applications. Sometimes, inverse problems need to be solved. For a given
structure with known loading conditions and measured displacement, the objective is to find or invert the
material properties E and ν, which may be required for testing new materials and design optimizations. It
is well known that inverse problems are usually very difficult, and this is especially true for finite element
inverse analysis.
To show how it works, we use a test example similar to that proposed by Murase (2000). A simple beam
system of 5 unit length × 10 unit length (see Figure 32.7) consists of five nodes and four elements whose
Young’s modulus and Poisson’s ratio may be different. Nodes 1 and 2 are fixed, and a unit vertical load is
applied at node 4 while the other nodes deform freely. Let us denote the displacement vector

U = (u1 , v1 , u2 , v2 , u3 , v3 , u4 , v4 , u5 , v5 )T ,

where u1 = v1 = u2 = v2 = 0 (fixed). Measurements are made for other displacements. By using the PA
with the values of the CO2 affinity A = 10000, light intensity L = 104 to 5 × 104 lx, and maximum CO2

f =1
1
4
E4, n4
5
E1, n1 E3, n3

E2, n2
3
2

FIGURE 32.7 Beam system of finite element inverse analysis.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 597 — #9

32-598 Handbook of Bioinspired Algorithms and Applications

fixation speed Vmax = 30, each of eight elastic modulus (Ei , νi )(i = 1, 2, 3, 4) is coded as a 16-bit DHAP
molecule string. For a target vector

Y = (E1 , ν1 , E2 , ν2 , E3 , ν3 , E4 , ν4 ) = (600, 0.25, 400, 0.35, 450, 0.30, 350, 0.32),

and measured displacements U = (0, 0, 0, 0, −0.0066, −0.0246, 0.0828, −0.2606, 0.0002, −0.0110), the
best estimates after 500 iterations from the optimization by the PA are Y = (580, 0.24, 400, 0.31,
460, 0.29, 346, 0.26).

32.3.4 Inverse Initial-Value, Boundary-Value Problem Optimization

Inverse initial-value, boundary-value problem (IVBV) is an optimization paradigm in which GAs have
been used successfully (Karr et al., 2000). Some conventional algorithms for solving such search optim-
izations are the trial-and-error iteration methods that usually start with a guessed solution, and the
substitution into the partial differential equations and associated boundary conditions to calculate the
errors between predicted values and measured or known values at various locations, then the new guessed
or improved solutions are obtained by corrections according to the errors. The aim is to minimize the
difference or errors, and the procedure stops once the given precision or tolerance criterion is satisfied. In
this way, the inverse problem is actually transformed into an optimization problem.
We now use the heat equation and inverse procedure discussed by Karr et al. (2000) as an example to
illustrate the IVBV optimization. On a square plate of unit dimensions, the diffusivity κ(x, y) varies with
locations (x, y). The heat equation and its boundary conditions can be written as:

∂u
∂t = ∇ · [κ(x, y)∇u], 0 < x, y < 1, t > 0,
u(x, y, 0) = 1, u(x, 0, t ) = u(x, 1, t ) = u(0, y, t ) = u(1, y, t ) = 0.

The domain is discretized as an N × N grid, and the measurements of values at (xi , yj , tn ), (i, j =
1, 2, . . . , N ; n = 1, 2, 3). The data set consists of the measured value at N 2 points at three different times
t1 , t2 , t3 . The objective is to inverse or estimate the N 2 diffusivity values at the N 2 distinct locations. The
Karr’s error metrics are defined as
   
N N  measured computed  N N  known predicted 
u − u κ − κ
i=1 j=1  i,j i,j  i=1 j=1  i,j i,j 
Eu = A N N  measured  , Eκ = A N N  known  ,
i=1 j=1 ui,j  i=1 j=1 κi,j 

where A = 100 is just a constant.

The floating-point GA proposed by Karr et al. for the inverse IVBV optimization can be summarized
as the following procedure: (1) Generate randomly a population containing N solutions to the IVBV
problem; the potential solutions are represented in a vector due to the variation of diffusivity κ with
locations; (2) The error metric is computed for each of the potential solution vectors; (3) N new potential
solution vectors are generated by genetic operators such as crossover and mutations in GAs. Selection of
solutions depends on the required quality of the solutions with the aim to minimize the error metric and
thereby remove solutions with large errors; (4) the iteration continues until the best acceptable solution
is found.
On a grid of 16 × 16 points with a target matrix of diffusivity, after 40,000 random κ(i, j) matrices
were generated, the best value of the error metrics Eu = 4.6050 and Eκ = 1.50 × 10−2 . Figure 32.8 shows
the error metric Eκ associated with the best solution determined by the GA. The small values in the error
metric imply that the inverse diffusivity matrix is very close to the true diffusivity matrix.
With some modifications, this type of GA can also be applied to the inverse analysis of other problems
such as the inverse parametric estimation of Poisson equation and wave equations. In addition, nonlinear
problems can also be studied using GAs.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 598 — #10

Biology-Derived Algorithms in Engineering Optimization 32-599

× 10–3
2

1.5

Ek
1

0.5

0
15
10 15
5 10
node ( j ) 5
node (i)

FIGURE 32.8 Error metric Eκ associated with the best solution obtained by the GA algorithm.

The optimization methods using biology-derived algorithms and their engineering applications have
been summarized. We used four examples to show how GAs and PAs can be applied to solve optimization
problems in multilevel function optimization, shape design of pressure vessels, finite element inverse
analysis of material properties, and the inversion of diffusivity matrix as an IVBV problem. Biology-
inspired algorithms have many advantages over traditional optimization methods such as hill-climbing
and calculus-based techniques due to parallelism and the ability to locate the best approximate solu-
tions in very large search spaces. Furthermore, more powerful and flexible new generation algorithms
can be formulated by combining existing and new evolutionary algorithms with classical optimization
methods.

References
Chipperfield, A.J., Fleming, P.J., and Fonseca, C.M. Genetic algorithm tools for control systems engineer-
ing. In Proceedings of Adaptive Computing in Engineering Design and Control, Plymouth, pp. 128–133
(1994).
Coello, C.A. Use of a self-adaptive penalty approach for engineering optimization problems. Computers
in Industry, 41 (2000) 113–127.
De Jong, K. Analysis of the Behaviour of a Class of Genetic Adaptive Systems, Ph.D. thesis, University of
Michigan, Ann Arbor, MI (1975).
Deb, K. Optimization for Engineering Design: Algorithms and Examples, Prentice-Hall, New Delhi (1995).
Deb, K. An efficient constraint handling method for genetic algorithms. Computer Methods and
Applications of Mechanical Engineering, 186 (2000) 311–338.
El-Beltagy, M.A. and Keane, A.J. A comparison of various optimization algorithms on a multilevel
problem. Engineering Applications of Artificial Intelligence, 12 (1999) 639–654.
Flake, G.W. The Computational Beauty of Nature: Computer Explorations of Fractals, Chaos, Complex
Systems, and Adaptation, MIT Press, Cambridge, MA (1998).
Goldberg, D.E. Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley,
Reading, MA (1989).
Haykins, S. Neural Networks: A Comprehensive Foundation, MacMillan, New York (1994).
Heitkotter, J. and Beasley, D. The Hitch-Hiker’s Guide to Evolutionary Computation: A
List of Frequently Asked Questions (FAQ), (2000) Available via anonymous FTP from
ftp://rtfm.mit.edu/pub/usenet/news.answers/ai-faq/genetic/ About 110 pages.
Holland, J. Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann Anbor, (1975).
Holland, J.H. Hidden Order: How Adaptation Builds Complexity, Addison-Wesley, Reading, MA (1995).
Jenkins, W.M. On the applications of natural algorithms to structural design optimization. Engineering
Structures, 19 (1997) 302–308.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 599 — #11

32-600 Handbook of Bioinspired Algorithms and Applications

Kannan, B.K. and Kramer, S.N. An augmented Lagrange multiplier based method for mixed integer
discrete continuous optimization and its application to mechanical design. Journal Mechanical
Design, Transaction of ASME, 116 (1994) 318–320.
Karr, C.L., Yakushin, I., and Nicolosi, K. Solving inverse initial-value, boundary-valued problems via
genetic algorithms. Engineering Applications of Artificial Intelligence, 13 (2000) 625–633.
Keane, A.J. Genetic algorithm optimization of multi-peak problems: Studies in convergence and
robustness, Artificial Intelligence in Engineering, 9 (1995) 75–83.
Koza, J.R. Genetic Programming: On the Programming of Computers by Natural Selection, MIT Press,
Cambridge, MA (1992).
McCulloch, W.S. and Pitts, W. A logical calculus of the idea immanent in nervous activity, Bulletin of
Mathematical Biophysics, 5 (1943) 115–133.
Michaelewicz, Z. Genetic Algorithm + Data Structure=Evolution Programming, Springer-Verlag,
New York (1996).
Mitchell, M. An Introduction to Genetic Algorithms, Cambridge, MIT Press, MA (1996).
Murase, H. Finite element inverse analysis using a photosynthetic algorithm. Computers and Electronics
in Agriculture, 29 (2000) 115–123.
Pohlheim, H. Genetic and Evolutionary Algorithm Toolbox for Matlab (geatbx.com) (1999).
Rechenberg, I. Evolutionsstrategie: Optimerrung Technischer Systeme nach Prinzipien der Biologischen
Evolution, FrommannHolzboog, Stuttgart (1973).
Renner, G. and Ekart, A. Genetic algorithms in computer aided design. Computer-Aided Design, 35 (2003)
709–726.
Wu, S.Y. and Chow, P. T. Genetic algorithms for solving mixed-discrete optimization problems. Journal of
the Franklin Institute, 331 (1994) 381–401.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C032” — 2005/7/28 — 19:59 — page 600 — #12

 33
Biomimetic Models
for Wireless Sensor
Networks

33.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-601

33.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-602
33.3 Characteristics of a WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-605
33.4 The Current Status of WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-607
33.5 Biological Organizations and Behavior . . . . . . . . . . . . . . . .33-608
Biological Organizations • Biological Behavior
33.6 Application of Biomimetics to WSNs . . . . . . . . . . . . . . . . . .33-615
Kennie H. Jones Potential Applications
Kenneth N. Lodding 33.7 Biomimetic Implementation Techniques . . . . . . . . . . . . . .33-617
Stephan Olariu 33.8 Engineering for Emergent Behavior . . . . . . . . . . . . . . . . . . . .33-619
Ashraf Wadaa 33.9 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-620
Larry Wilson References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-620
Mohamed Eltoweissy Appendix A — Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .33-622

33.1 Introduction
With the advent of Micro-Electrical-Mechanical Systems (MEMSs) technology, complex and ubiquitous
control of the physical environment by machines will facilitate the diversity of mechanization and auto-
mation, long promised by visionaries. Wireless sensor networks have appeared as the first generation of
this revolutionary technology. The small size and low cost of sensor devices will enable deployment of
massive numbers, but initially place severe limitations on the computing, communicating, and power
capabilities of these devices. With these constraints, research efforts have concentrated on developing
techniques for executing simple tasks with minimal energy expense. But, as MEMS evolve, computing
and communicating capabilities are expected to improve at an accelerating rate and new techniques for
supplying energy will significantly reduce the low power constraint. Increased capabilities will be possible,
and it is predicted that societies of machines will evolve to be autonomous, cooperative, fault-tolerant, self-
regulating, and self-healing. Improvements in biomimetic software and evolvable hardware will lead to
self-sustaining communities of machines with emergent behavior that autonomously operate and adapt to
changes in the environment. The main goal of this chapter is to investigate biomimetic models in relation
to their potential application to the evolution of these systems, thus providing a framework that guides

33-601

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 601 — #1

33-602 Handbook of Bioinspired Algorithms and Applications

the evolution from the current primitive organizations of sensor nodes to pervasive societies of intelligent
electromechanical systems.

33.2 Background
The dream of ubiquitous machinery was a result of the Industrial Revolution: that machinery could
eventually be developed to service man’s needs, particularly replacing man’s direct participation in physical
work for provision of food, clothing, housing, and other necessities. The realization of this dream has
progressed for several centuries but has taken a revolutionary leap through the invention of electronic
computers. Computers have vastly improved the performance of mechanical machines by allowing more
sophisticated “thought” processes for control of the mechanics (e.g., capabilities of industrial robots
used in manufacturing are far more autonomous and complex, automobile engines are much more
efficiently controlled, etc.). Futurists have long predicted that machines would eventually communicate
and cooperate with each other to accomplish extraordinarily complex tasks without human intervention.
The rapidly accelerating improvement of computing capabilities over the last 50 years has contributed
to the realization of this dream. Four major factors are
1. Increasing computational performance: Moore’s Law has reliably predicted that computing
performance doubles every 18 months.
2. Reduction in physical size: Computer technology has gone from vacuum tubes to transistors, to
integrated circuits. Soon nano-technology will create another revolution by further reducing the
physical size of circuitry.
3. Decreasing cost of production: Improved technology has contributed, but commodity pricing has
had a greater affect.
4. Reduction in power requirements and improvements in power sources: Advances in technology com-
bined with reductions in size will continue to reduce power requirements. However, the next
paradigm shift will come from techniques making it possible to scavenge power from the ambient
environment (from various sources including vibration, heat, light, and background radio noise).
While futurists long dreamed of machines working with other machines, a giant step toward the
realization of this dream may be credited to a DARPA-sponsored program, SmartDust, originated in
1999 (Kahn et al., 1999). The title of the program creatively described its goal: to make machines, with
self-contained sensing, computing, transmitting, and powering capabilities, so small and inexpensive
that they could be released into the environment in massive numbers. Whether intended to be mobile
(e.g., small enough to be cast into the wind to stay aloft for extended periods) or immediately stationary
(e.g., deployed from an airplane over a large geographical area for ground surveillance), the sensors have
the formidable task of self-organizing into a network that can transmit information to a user while being
severely constrained by the onboard energy supply. As they were funded by the Department of Defense,
much of the research concerned surveillance of battlefield scenarios, but it was immediately apparent that
many peace-time applications could benefit from wireless sensor networks (WSNs).
Observations of the SmartDust project suggests the following definition of WSNs:

A wireless sensor network is a deployment of massive numbers of small, inexpensive, self-powered

devices that can sense, compute, and communicate with other devices for the purpose of gathering
local information to make global decisions about a physical environment.

The National Research Council’s (NRC 2001) Committee on Networked Systems of Embedded Computers
published a report expanding this definition. They defined the concept of the embedded network, EmNet,
as a network of heterogeneous computing devices pervasively embedded in the environment of interest.
Their stated objective was to “develop a research agenda that could guide federal programs related to
computing research and inform the research community (in industry, universities, and government)
about the challenging needs” of research in EmNets. They recognized a difference between EmNets and
traditional computer networks in that the former will be “more tightly integrated with their physical

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 602 — #2

Biomimetic Models for Wireless Sensor Networks 33-603

environment, more autonomous, and more constrained in terms of space, power, and other resources.
They will also need to operate, communicate, and adapt in real time, often unattended.” The Committee
enlarged the scope for SmartDust to paint a picture of the eventual embedding of sensor (and effector)
nodes into every aspect of our world: “computing and communications technologies will be embedded into
everyday objects of all kinds to allow objects to sense and react to their changing environments. Networks
comprising thousands or millions of nodes could monitor the environment, the battlefield, or the factory
floor; smart spaces containing hundreds of smart surfaces and intelligent appliances could provide access
to computational resources.” A subtle but important prediction is made here: these networks will not
only gather information about the environment but will affect environmental conditions or effect new actions.
Thus, a feedback mechanism is instantiated where input to the network may be affected by its own actions.
Diverse applications are anticipated:

EmNets will be implemented as a kind of digital nervous system to enable instrumentation of all sorts
of spaces, ranging from in situ environmental monitoring to surveillance of battlespace conditions;
EmNets will be employed in personal monitoring strategies (both defense related and civilian),
combining information from nodes on and within a person with information from laboratory
tests and other sources; and EmNets will dramatically affect scientific data collection capabilities,
ranging from new techniques for precision agriculture and biotechnological research to detailed
environmental and pollution monitoring.

When this point is reached, one could argue that the universe becomes one gigantic EmNet. The
Massachusetts Institute of Technology’s (MIT) Amorphous Computing group sees nanoscale computers
“combining microsensors, actuators and communications devices integrated on the same chip to produce
particles that could be mixed with bulk materials, such as paints, gels, and concrete” (Abelson et al., 1999).
These groups envision an evolution from SmartDust’s passive observation of the environment to active
manipulation of the environment, driven by the coordinated action of massive numbers of sensors and
actuators, coupled with vast computing resources. Clearly, the term wireless sensor network describes only
a small portion of this vision.
To accomplish this vision, research challenges abound. Despite constraints in power, memory, band-
width, etc., these devices will be embedded into systems designed to last for long periods of time.
While SmartDust predicted massive numbers of nodes per network, many issues resulting from density
and scale remain unsolved. EmNets will be self-configuring upon deployment and adaptive to changes in
both the network and environment. Trust and fault tolerance models will have to be developed to solve
problems well beyond those posed by conventional networks. EmNets will control real-time processes
where great costs will be incurred or life is at risk upon failure making reliability, security, and quality
of service the major research issues to be considered. There are also nontechnical issues. Ubiquitous and
pervasive devices constantly monitoring and controlling everything present many potential legal, ethical,
and policy controversies regarding privacy, security, reliability, intellectual property rights, etc. There are
other issues concerning production standards, commercialization, business models for implementation
and coordination, etc. For the sake of making some progress, many of these requirements are being
ignored for current prototypes and present significant challenges for future research.
The NRC expanded the definition of wireless sensor networks:

Wireless sensor networks are massive numbers of small, inexpensive, self-powered devices pervasive
throughout electrical and mechanical systems and ubiquitous throughout the environment that monitor
(i.e., sense) and control (i.e., effect) most aspects of our physical world.

While the NRC envisions ubiquitous use of WSNs in the next few years, some visionaries look beyond
that horizon to predict even greater developments. Kurzweil (Richards et al., 2002) predicts that within the
next 20 years, a $1000 computer (typical price of today’s PC) will have computational power that matches
the human brain. He states, “This level of processing power is necessary but not sufficient for achieving
human-level intelligence in a machine. Organizing these resources — the ‘software’ of intelligence — will
take us to 2029, by which time your average personal computer will be equivalent to a thousand brains.”

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 603 — #3

33-604 Handbook of Bioinspired Algorithms and Applications

Unlike humans who must share knowledge through speech or vicariously through print or other media,
computers can instantly transmit all of their knowledge to other computers in a relative instant. Science
fiction writers and Hollywood often portray such an intelligent machine as a walking, talking robot (e.g.,
Arnold Schwartzenegger’s character in the Terminator), but a more likely scenario is a distributed network
of intelligent machines. Kurzweil further predicts that humans will be directly connected to these net-
works via neural implants. Some of these implants are available today as “standalone” devices to counteract
Parkinson’s disease, for hearing (cochlear implants), and, soon, visual implants for the blind. Wearable
devices currently monitor a person’s health and transmit this information to a network of resources for
analysis and action. Kurzweil envisions these implants will progress to allow humans to immerse them-
selves in the network of intelligent machines (e.g., to enter virtual reality environments). While his focus
is on the expectation of vast improvements in the intelligence of machines, it is a reasonable assumption
that similar scales of improvements will be obtained for the size of components and power requirements.
Moravec (1988) believes that machines will not only exceed the capabilities of humans in a similar time
period but will eventually become self-sufficient. He sees a future in which, “the human race has been
swept away by the tide of cultural change, usurped by its own artificial progeny.” Thus, when machines
are more intelligent than humans, autonomous, adaptable, self-sustaining, and self-replicating, they will
have no further use for mankind.
These visionaries further extend the definition of WSNs predicting their destiny:

The ever increasing capabilities of pervasive and ubiquitous wireless sensor networks will improve
the intelligence, autonomy, and adaptability of electrical and mechanical systems such that they will
soon converge with and surpass the capabilities of humans.

Kurzweil’s and Moravec’s predictions do not go unopposed. Their vision of man–machine conver-
gence, titled Strong AI (Artificial Intelligence), is challenged by many (Richards et al., 2002). One would
expect theologians to contest these predictions, but many metaphysicians as well argue that there are char-
acteristics of humanity that cannot be created or duplicated through a programmed series of chemical
and physical reactions. These include consciousness, curiosity, creativity, compassion, freedom, etc. The
debate rages on as to whether machines could ever identify with humans. But, even if they never do, and
Kurzweil and Moravec are off by orders of magnitude in their projections, capabilities will improve to the
extent that networked machines in the near future will go far beyond those of today.
Throughout this sequence of definitions, it is apparent that the term wireless sensor network is inadequate
to describe this increasing complexity. EmNets is neither very descriptive, nor often used. Sensor/effector
network may be more complete but is still not comprehensive. As the original term is well understood
and accepted by the research community, it will be retained for the remainder of the discussion using the
abbreviation, WSN. The individual sensor device is often called a sensor mote (or mote), but in the context
of the network, a mote is referred to as a sensor node (or node).
The progression from simple machines to complex arrangements of machines that are autonomous,
self-regulating, self-healing, and even self-reproducing will require increasing sophistication not only of
individual nodes and the network but also of the supporting software. As will be shown, the evolution
needed is not unlike that of life from simple unicellular organisms to multicellular societies. Section 33.2
presents a set of defining characteristics of WSNs that differentiate them from conventional computing
networks by describing their functionality and organization. Section 33.3 describes the current status
of WSN implementations. Section 33.4 describes the evolution of living organisms in terms of their
organization and behavior. Of prime importance are

• The diversity of life and how such diversity is required of future WSNs.
• The differences between innate behavior and learned behavior and their effect on emergent behavior
from individual organisms acting as a group.

The key to the application of these principles to WSNs is a new discipline of engineering, engineering for
emergent behavior, which will be defined in later sections of this chapter. Section 33.5 presents our vision

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 604 — #4

Biomimetic Models for Wireless Sensor Networks 33-605

that current WSN implementations parallel early life forms and describes what types of biomimetics
will apply to future WSNs to build an ecosystem of WSNs. This section also identifies research techniques
currently used in Artificial Intelligence (AI) and Artificial Life (Alife) that apply to engineering for emergent
behavior of WSNs. Section 33.6 presents a philosophical discussion of the importance of engineering for
emergent behavior to the future of WSNs. Section 33.7 offers the concluding remarks.

33.3 Characteristics of a WSN

Current WSN research was largely spawned by SmartDust and the challenges it raised. A necessary first step
for any implementation is to design a network that can be constructed ad hoc to enable communication
amongst randomly distributed sensor nodes. Operating with a tight energy constraint presents particular
demands, and much current work focuses on the completion of work in an energy efficient manner.
Radio, the most common communication mechanism for WSNs, is known to expend a lot of power.
To conserve power, the technique typically used is to restrict the nodes to a relatively short transmission
distance. They must then form a multi-hop network where a source node broadcasts a message, which is
received by a neighboring node, and rebroadcast by that node. This continues until the message arrives at
its destination. Ad hoc networks present new issues for the traditional routing problem and the resource
constraints of WSNs add even more issues. For WSNs, research abounds at all levels of the network model
along with such related issues as security, quality of service, reliability, etc. Reflecting this concentration,
Tilak et al. (2002) proposed a taxonomy for WSNs “from a communication protocol perspective.” They
first categorize by data the delivery requirements of applications and define the categories of continuous,
event driven, observer-initiated, and a hybrid of these. They further categorize by network dynamics or
routing in this dynamic environment.
Providing an efficient networking infrastructure is a necessary first step in the evolution of WSNs.
However, assuming such low-level problems will be solved soon, this discussion will focus on modeling
a WSN in terms of the relative complexity of its organization, its functionality, and the effect of its
behavior. The universal network will be compared to an ecosystem, and the subsystem networks will
become organisms in that ecosystem. The questions to be addressed in this section are how does one
identify a subsystem and to what subsystems of WSNs might biomimetic models apply?
The conundrum of defining a WSN is reminiscent of a famous statement from United States Supreme
Court Justice Potter Stewart in 1973 concerning a Supreme Court decision. Placed in the context of WSNs,
it would read, “I shall not today attempt further to define [a WSN]; and perhaps I could never succeed in
intelligibly doing so. But I know it when I see it . . .” Although a precise definition may not be possible,
the following characteristics are common to all WSNs:
• Close ties to the physical world: The purpose of WSNs is to interact with the physical world. Although
their name denotes sensing, actuating may be as important as sensing. The Internet is dissimilar
from a WSN as most nodes function either as information repositories (servers) or information
users (clients) with virtual locations that can be physically moved with no consequence.
• Granularity of action: Small size and low cost allow massive numbers of sensor nodes to be deployed
over a geographical region instead of a small number of larger, more costly devices each of which
must act upon a larger space.
• Collection of local action results in global effects: It is not necessary for any sensor node to know
or understand the function of the network, but, like a soldier in a battalion, if each sensor node
does its job, the network will succeed. The global view will emerge from an aggregation of the local
information.
The following list of characteristics will be useful in categorizing WSNs:
• Random distribution of sensor nodes: Sensor nodes are deployed by means that preclude a pre-
determined spatial topography (e.g., dumped from an airplane). Distribution density guarantees
functionality and fault-tolerance, as failure of some nodes will not cause the network to fail. Because

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 605 — #5

33-606 Handbook of Bioinspired Algorithms and Applications

the WSN is closely tied to the physical environment, node location must be determined. Wadaa
et al. (2004a) devised a scalable energy-efficient training protocol to provide locations for nodes
that are initially anonymous, asynchronous, and unaware of their locations.
• Large numbers of sensor nodes: Many of the challenging and new issues of WSNs are concerned
with coordinating massive numbers of sensor nodes into a functioning network.
• Small physical size of sensor nodes: Initially, reduction in size is expensive. However, as manufacturing
improves, smaller size and increased volume should contribute to reduced prices. Also, smaller sized
nodes will widen the range of applications.
• No pre-assigned network topology: If the location of the sensor nodes is not predetermined, neither
can the network topology. Modest power budgets may prevent nodes from transmitting directly to
a common destination, therefore an ad hoc process will be applied to form a multi-hop network.
• Anonymous network: Scalability of massive numbers of nodes within a WSN precludes the assig-
nment of unique node identifiers during deployment. Anonymity can be useful in securing a WSN.
Wadaa et al. (2004b) describe how anonymity can be used to prevent denial of service attacks.
• Wireless communication: Random distribution dictates wireless communication.
• Limited onboard energy supply: Current battery technology ties capacity to physical size. Therefore,
the need for small sensor nodes competes with the need for more onboard energy. If technology
improves sufficiently, this will no longer be a major issue.
• Node mobility: Motion of sensor nodes may be:
◦ Static: Once deployed, the sensor nodes do not move.
◦ Relative: The sensors move, but as a group (e.g., a group of satellites may be placed in parallel
orbits but remain fixed relative to each other; a “satellite constellation” ).
◦ Fully mobile with independent movement : For example, a swarm of sensing robots.
• Application code distributed across nodes: WSNs range from those containing homogeneous nodes
reporting simple observations to heterogeneous nodes with differing (and perhaps changing)
functions.
• Data aggregation: In addition to individual sensor nodes reporting to a single destination,
sensor nodes may form clusters whereby local preprocessing of sensed data improves efficiency
by transmitting only the aggregated data. In fact, a hierarchy of clusters and aggregation may
be used.
• Performance and reliability: Because of the interaction with the physical world and the potential
for damage to it in many applications, new demands will be placed on these networks for real-time
performance and reliability.
• Security: As stated, these networks must perform and be reliable; therefore, nodes must be trusted
network members. Jones et al. (2003) proposed a new security paradigm in which security is based
upon parameterized frequency hopping and cryptographic keys in a unified framework.
Many of these characteristics may not apply to networks resembling WSNs. Consider for example a
“smart skin”where many sensors are attached to a surface (e.g., an airplane wing). Sensor and actuators may
be pervasive throughout a larger structure, serving not only as surface sensors to sample environmental
input but internally as “muscles” to move an object. Such a network may be massive but fixed such that
each sensor’s position and identity may be predetermined, its network connectivity may be predetermined
and need not be wireless, and its supply of power may not be limited. But it still has large numbers of
small, spatially distributed nodes that must communicate efficiently and aggregate large amounts of
locally collected information for global decisions. Because of its close ties to the physical world, real-time
performance, reliability, and security are extremely important.
WSNs have many close cousins that go by a variety of names such as:
• Heterogeneous WSNs: This is the nomenclature of Intel’s EcoSense research project, that is,“tackling
a difficult challenge: how to network large numbers of inexpensive wireless sensor nodes while
maintaining a high level of network performance” (Intel Research — Exploratory Research — Deep
Networking — Heterogeneous Sensor, 2004).
• Deep networking : Intel defines Deep Networking as

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 606 — #6

Biomimetic Models for Wireless Sensor Networks 33-607

Locally networking billions of embedded nodes, driving computing deeper into the infra-
structure that surrounds us. . . .Small, inexpensive, low-powered sensors and actuators, deeply
embedded into the physical environment can be [deployed] in large numbers, interacting and
forming networks to communicate, adapt, and coordinate high-level tasks. . . .As these micro
devices are networked, the Internet will be pushed not just into different locations but deep
into the embedded platforms within each location. This will enable a hundredfold increase
in the size of the Internet beyond the growth we are already anticipating. New and different
methods of networking devices to one another and to the Internet must be developed. (Intel
Research — Exploratory Research — Deep Networking, 2004)

• Amorphous computing : This is MIT’s term for “the development of organizational principles and
programming languages for obtaining coherent behavior from the cooperation of myriads of
unreliable parts that are interconnected in unknown, irregular, and time-varying ways.” (Abelson
et al., 2000)
• Sensor webs: NASA’s Jet Propulsion Laboratory (JPL) describes a sensor web as “an independent
network of wireless, intra-communicating sensor pods, deployed to monitor and explore a limitless
range of environments. This adaptable instrument can be tailored to whatever conditions it is sent
to observe.” (Delin, 2004)
• Mesh networks: This term describes a type of Wi-Fi network where nodes do not communicate
through a central controller, or access point, but rather, mobile ad hoc peers in the network form
a mesh topology to transmit data from source to destination via multiple hops throughout the
network. This technique improves on the reliability and efficiency of the centralized approach.
Although designed for connection of conventional computers, this technology shares many issues
with WSNs.
• Sensor constellation: This term is usually used to describe multiple satellites cooperating in a space
science experiment (Leveraging the Infosphere: Surveillance and Reconnaissance in 2020, 1995).
• Pervasive computing : The Centre for Pervasive Computing (2004) defines this as,
the next generation computing environments with information & communication technology
everywhere, for everyone, at all times. Information and communication technology will be an
integrated part of our environments: from toys, milk cartons and desktops to cars, factories
and whole city areas — with integrated processors, sensors, and actuators connected via
high-speed networks and combined with new visualization devices ranging from projections
directly into the eye to large panorama displays. . . .Pervasive computing goes beyond the
traditional user interfaces, on the one hand imploding them into small devices and appliances,
and on the other hand exploding them onto large scale walls, buildings and furniture.
• Ubiquitous computing : This term was coined by Mark Weiser (1996) whose goal is to “Activate the
world. Provide hundreds of wireless computing devices per person per office, of all scales (from 1
displays to wall sized). This has required new work in operating systems, user interfaces, networks,
wireless, displays, and many other areas.”
• Invisible computing : An ACM SIGGRAPH conference in 2000 explored “the tiny, cheap, special-
purpose devices that experts expect to diffuse into our lives over the next two decades. Though
these devices themselves may be visible, our common goal is to make them so comfortable to use
that they seem not to be computers at all — the computing power they use stays invisible” (Invisible
Computing: Scope, 2000).

33.4 The Current Status of WSNs

Prototypes such as SmartDust have begun the implementation of futurist dreams. In March 2001, a
team of researchers led by Pister and Culler of the University of California at Berkeley (UCB) held a
demonstration at the Marine Corps Air/Ground Combat Center, Twentynine Palms, CA (UCB/MLB 29

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 607 — #7

33-608 Handbook of Bioinspired Algorithms and Applications

Palms UAV-Dropped Sensor Network Demo). They dropped six sensor motes from an Unmanned Aerial
Vehicle (UAV) along a road. These motes self-organized by synchronizing their clocks and forming a
multi-hop network. They magnetically detected vehicles passing and reported the time of the passing.
Using information collected from all motes, velocity of passing vehicles was computed.
On August 27, 2001, researchers from UCB and the Intel Berkeley Research Lab demonstrated a self-
organizing WSN to those attending the kickoff keynote of the Intel Developers Forum. Several students
each brought on stage a wireless sensor mote and activated it at different times. As the motes were
initiated, their icons appeared on a display with lines connecting all motes that could “hear” each other and
highlighted lines depicting a multi-hop routing structure over which sensor data could be transmitted to a
central collector, a PC. The network grew as the nodes were initiated and adapted to changing conditions.
Additionally, color cues on the display indicated changing lighting conditions on the stage as sensors
detected these changes. As the students left the stage, the network disintegrated. In a second large-scale
demonstration of the networking capability, the quarter-sized motes were hidden under 800 chairs in
the presentation hall and simultaneously initiated forming what Culler described as, “the biggest ad hoc
network ever to be demonstrated” (Lammers, 2001).
In the spring of 2002, Culler’s group (Mainwaring et al., 2002) collaborated with the College of the
Atlantic in Bar Harbor to instate a WSN on Great Duck Island, Maine. The initial application was to
monitor the microclimates of nesting burrows in the Leach’s Storm Petrel, and, by disseminating the
data worldwide, to enable researchers anywhere to nonintrusively monitor sensitive wildlife habitats. The
sensor motes were placed in the habitat and formed a multi-hop network to pass messages in an energy-
efficient manner back to a laptop base station. The data was intermediately stored at the base station
and eventually passed by satellite to servers in Berkeley, CA, where it was distributed via the Internet to
any interested viewer. The sensors measured temperature, humidity, barometric pressure, and mid-range
infrared by periodically sensing and relaying the sensed data to the base station. The largest deployment
had 190 nodes with the most distant placement over 1000 feet from the nearest base station.
The FireBug system (Chen et al., 2003) is a network of GPS-enabled, wireless thermal sensors motes,
communicating through a control layer for processing sensor data, and a command center for interactive
monitoring and control of the WSN. The FireBug network self-organizes into clusters in which cluster
leader motes act as base stations, receiving sample data from cluster members and brokering commands
to these members. The controller is a personal computer running the Apache web server interfaced with
MySQL using PHP. The FireBug Command Center allows user interaction for controlling the FireBug
network and displays real-time changes in the network.

33.5 Biological Organizations and Behavior

Life began as an organization of entities performing functions. The entities and their functions survived
or became extinct based on the sustainability of those functions. Proteins catalyzed chemical reactions.
Aggregates of proteins resulted in more complex reactions eventually evolving RNA and then DNA that
direct protein production. Further aggregation eventually produced cells. Cells aggregated in symbiotic
relationships to form more complex cells and then to form tissues and organs composing multicellular
organisms. Throughout this progression, organisms developed more complex forms and functionality.
Successful species adapt to an ecological niche providing sustenance, and live to reproduce at a rate that
keeps them in existence. Critical to understanding this process is the realization that evolution occurs not
because complexity is better but because, while simple, successful organisms remain, random mutations
result in increasing complexities that are sustainable in some niche (i.e., one did not develop eyes so one
can see; rather one sees because one developed eyes — and this enhances success). Detrimental mutations
result in unsuccessful species. Development of WSNs will also follow such an evolution but with two very
important distinctions:
• While biological evolution occurs randomly with resulting successes and failures over long periods
of time, initially, electromechanical systems can be directly designed and these designs evaluated

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 608 — #8

Biomimetic Models for Wireless Sensor Networks 33-609

and improved rapidly. If Kurzweil and Moravec are correct, eventually electromechanical systems
may design themselves. Again, random mutation and selection for sustainability may prevail but
at a much faster pace than ever occurred in life.
• Functionality, as are all other characteristics of living organisms, is only important as it affects
sustainability. While humans are in control, functionality will be most important for a WSN,
whereas sustainability only affects issues such as cost, efficiency, longevity, etc. Evolution of WSNs
can be directed for increasing functionality with sustainability only a secondary issue. Again,
Kurzweil and Moravec predict that this may change back if machines become autonomous giving
primary importance to sustainability.

Organization models and behavior influence functionality. In general, with more complexity comes
more functionality and more diverse behavior. It is important to note that whereas man continually
strives to divide living systems into distinct categories organisms have developed as a continuum of both
organizations and behavior in which there are typically exceptions to any rule applied for distinction.
Development of machines can be perceived as such a continuum, but it is necessary to adopt a model
in which clear distinctions are made. A distinction will be made between the capabilities of unicellular
organisms compared with those of multicellular organisms and, furthermore between organisms surviving
alone and those organized into groups (colonies, swarms, societies, etc.).
Behaviors will be categorized as those that result from responses to stimuli that ultimately are determined
genetically (i.e., innate) and those that are cognitive. Cognition is defined as the mental process by
which knowledge is acquired, which is a result of awareness, perception, intuition, reasoning, memory, and
judgment. Thus, a behavior is said to be cognitive if its effect is known and understood by the effector.
Cognitive behavior begets new cognitive behavior that is not genetically encoded. For this discussion, new
behavior that results from prior experience defines learning. A simple response to stimuli is not considered
learning. An interesting effect of cognition in living systems is that it may introduce selection factors other
than sustainability.

33.5.1 Biological Organizations

Life can be represented hierarchically with the most complexity at the root of the tree and the simplest
organization at the leaves. Levels of organization are defined by degree of complexity:

• Simple organization, cells: Although cells are composed of identifiable structures, and some
independent but subcellular components function in a capacity that arguably demonstrates char-
acteristics of life, cells are the basic unit of life for the purpose of this discussion. All cells are
contained by a plasma membrane in which the chemical “soup” of life operates. All cells have a
genetic structure, DNA, controlling protein production that, in turn, catalyzes all metabolic pro-
cesses within the cell. Cells are at least potentially capable of self-replication. A cell may serve many
functions and exist independently as a distinct organism or have a very specialized function while
cooperating symbiotically with many other adjacent cells forming a multicellular organism.
• Building blocks of complexity, tissues, and organs: Tissues are contiguous, homogeneous, highly spe-
cialized cells, each serving a function where the collective result of their actions yields a cumulative
result. For example, muscle cells contract, applying a pulling force, when stimulated. While the
power of individual cells is small, the cumulative power of millions of cells contracting simulta-
neously produces a significant force. Unlike unicellular organisms that must forage for food, cells
of tissues are not self-sustaining as they are provided nourishment by the multicellular organism
to which they belong and cannot survive independently. Organs are a heterogeneous organization
composed of many kinds of tissues to serve a cumulative function greater than the individual
functions of their component tissues (e.g., a stomach has muscle tissues and tissues to secrete acid
and mucus; it functions to digest food). Organs may combine as components in an organ system to
perform a higher function (e.g., the digestive system has many organs working together to intake,
digest, and distribute food while eliminating waste).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 609 — #9

33-610 Handbook of Bioinspired Algorithms and Applications

• Complex organization, multicellular organisms: Multicellular organisms are composed of tissues and,
usually, organs. They range from the simplest of these composed of few tissues (e.g., Coelenterates
such as jellyfish) to plants with simple organs to mobile animals with complex motor and nervous
systems. Of particular interest to this discussion are the latter. Because of these complex systems,
they can not only respond to stimuli but can learn from these stimuli to alter their behavior.
Multicellular organisms reproduce as a unit.
• Organizations of organisms: Colony, population, community, and society are nearly synonymous
words describing a group of organisms living or growing together. Although they are often used
interchangeably, colony, population, and society usually denote a homogeneous group of the same
kind of animals, plants, or unicellular organisms, while community usually indicates all life in a
given habitat.
• Add the environment to comprise the ecosystem: An ecosystem is the community of living organisms
in a habitat together with all nonliving (abiotic) components of the environment with which the
community interacts.
• The root of the tree is the biosphere: On Earth, the biosphere is all of the area of Earth from the
highest altitude in the atmosphere to the lowest depths of the oceans and land where life exists.
Thus, life on Earth can be represented as a tree with the biosphere the root, which is composed of
ecosystems, which are composed of communities and the abiotic environment, etc. The beauty of
the evolutionary process is that the biosphere comprises all levels of organizations and behavior:
if an organism fills a niche for sustainability, it survives. Life did not begin with the simple organisms
and, as it evolved to more complex organisms, discard the simple organisms. There is a delicate
balance within an ecosystem between all of the interdependent organisms together with their
environmental resource requirements. With the obvious competition for food, it may appear that
life is a constant fight for survival. While this is true for an individual organism, when the biosphere
is considered, life is what is sustained. Margulis and Sagan (1986) stated, “Life did not take over the
globe by combat, but by networking.” The biosphere as a whole may be viewed as a giant symbiotic
relationship.
Applying the ecosystem model to future WSNs encourages one to plan a diverse collection of interacting
(i.e., networked) sensor/effector systems of all sizes and complexities to fulfil every required niche.

33.5.2 Biological Behavior

Biological behavior ranges along a continuum from responses to stimuli resulting from genetic encoding to
complex, learned behavior that build upon experience, to social organizations where the society exhibits
emergent behavior as the result of individual actions of its members. Organisms exhibiting cognitive
behavior must have the ability to learn and often form societies.
33.5.2.1 Non-Cognitive Behavior of Singletons
What distinguishes unicellular organisms from multicellular organisms is their relative independence
from all other cells. Although many may cohabitate spatially, they remain relatively independent having
simple, independent functions that result in their sustainability. They reproduce asexually with mainly
two functions: eat and reproduce. Although they may exhibit other functionality, these two are necessary
for sustainability. While their metabolism may greatly affect the ecosystem (e.g., bacteria metabolize dead
organisms returning chemicals to the ecosystem that are used to sustain other organisms), these are side
effects unknown and not understood by the organism. Dusenbery (1996) describes a diversity of behavior in
microbes such as locomotion, navigation, migration, habituation, and communication. He gives examples
of how they can anticipate the future (e.g., circadian responses). For habituation, he gives examples in
which protozoa appear to have memory as they will “learn” to ignore a stimulus for the duration of a
few hours. He concludes that “most microbes are capable of only the most specialized forms of learning:
sensory adaptation, setting internal clocks, and crude habituation. Many biologists and most psychologists
would not consider any of these to be ‘real’ learning.” Primio, et al. (2000) argue that life and cognition

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 610 — #10

Biomimetic Models for Wireless Sensor Networks 33-611

are synonymous, “Bacteria and other unicellular organisms are autonomous and social beings showing
(the lowest levels of) cognition. They have the fundamental cognitive abilities to identify elements of
the environment and to differentiate between them (and self), to choose among alternatives, to adapt to
changes, to coordinate their behavior in groups, to act purposefully.” While these cells respond to external
stimuli, their response is ultimately preprogrammed within and limited by their DNA. In view of previous
definitions of cognition and learning, these behaviors are not cognitive. Although improvements occur
through generations of randomly mutated cells with mutations selected for sustainability, this is also not
considered learning.
Multicellular organisms often act as singletons. Other than to mate, mosquitoes normally function
independently. Mosquitoes cohabitate in massive numbers, often appearing to swarm and are described
as such. In northern ecosystems, these swarms congregate in such density as to inhibit breathing (Conniff,
1996) and have been known to drain enough blood to cause the death of both cows and caribou (Budiansky,
2002). However, as defined below, this differs from the behavior exhibited by swarms of bees or ants. This
action is the result of massive singletons simultaneously attacking the same food source; they are not
cooperating but are competing. Again, for this discussion, these organisms do not exhibit cognitive
behavior.

33.5.2.2 Non-Cognitive Behavior of Swarms

Swarms are organisms at any level that form a group in which relatively simple and individual behavior
combine to result in what is known as emergent behavior (i.e., the behavior of the group emerges from
individual behavior). The emergent behavior of the group is neither individually possible nor understood
by the individuals. From a view outside of the colony, insects such as ants, termites, and bees exhibit
behavioral characteristics that could appear to be intelligent. Bonabeau et al. (1999) catalogue a list of
behaviors often demonstrated by social insects that seem to indicate cognitive thought: optimal route
finding, division of labor and task allocation, cemetery organization, and brood sorting. For example,
consider that to maintain a healthy beehive specific gas-exchange requirements must be met. When
necessary, honeybees will actively ventilate the hive by positioning workers at the hive entrance to fan their
wings and control the flow of gases through the hive. This wing fanning event is not a simple operation,
as the bees must synchronize their fanning actions to form an on–off pattern that allows the hive to inhale
and exhale: the hive breathes. In the context of a colony, this behavior has been termed social homeostasis
(Turner, 2000).
A decade ago, Kevin Kelly (1994) in his landmark book, Out of Control, wrote of what he termed the
hive mind:

The marvel of the “hive-mind” is that no one is in control, and yet an invisible hand governs, a
hand that emerges from very dumb members. The marvel is that more is different. To generate a
colony organism from a bug organism requires only that the bugs be multiplied so that there are
many, many more of them, and that they communicate with each other. At some stage the level of
complexity reaches a point where new categories like “colony” can emerge from simple categories
of “bug.” Colony is inherent in bugness, implies this marvel. Thus, there is nothing to be found in
a beehive that is not submerged in a bee. And yet you can search a bee forever with cyclotron and
fluoroscope, and you will never find the hive.”

Social homeostasis, hive-minds, swarming, and similar behavior associated with large groups of social
insects share a common thread: the intelligence that these behaviors seem to exhibit is not the result of a
cognitive process. The intelligence of the swarm is emergent and results from the interactions of many
thousands of basically dumb, autonomous individuals. Each individual follows its own set of rules and
reacts to local state information. The rules are primarily a result of a manifestation of genetic encoding in
the presence of environmental stimuli. There is no central control command issuing orders. Individuals
are highly connected within their immediate neighborhood and can share information, but they do not
have a central server with which to communicate. Control and management of the swarm is distributed

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 611 — #11

33-612 Handbook of Bioinspired Algorithms and Applications

throughout and within the swarm members. It is distributed control with no single point of command
and no single point of failure.
What level of intelligence (i.e., the extent of their cognition) is required of bees to perform such
behavior? There is evidence that some degree of memory is required. Gould and Gould (1988) believed a
bee’s ability for spatial recollection not only provided for storage of many mental maps to food locations but
also the ability to calculate shortcut routes between visited sites. However, Dyer and Seeley (1991) found
that bees are limited to storing route maps comprising of memorized landmarks (i.e., a much condensed
representation of spatial topology) and found no evidence that bees can form spatial relationships between
different routes. Bees also use visions of solar movement together with some internal circadian indicator
to relate landmarks for directional cues.
Gregory (1997) stated,

One’s first thought might be that insects are mere automata, lacking ability to learn, but this is not so.
Bees can learn flower patterns and complex routes to food. They can navigate, and communicate in
many ways including the famous dance-language discovered by Karl von Frisch. The waggle dance . . .
is related to the position of the Sun, indicating to other workers direction and distance and food
quality of flowers. No doubt there is always a danger of reading too much intelligence into innate
behavior, but it seems impossible to see this as completely stereotyped, for it is adjusted to particular
conditions and needs. . . .Insect learning may be limited to immediate adaptive uses; but insects
do have associative learning by Pavlovian conditioning with rewards from previously meaningless
stimuli. . . .They do show latent learning, though probably without “insight.” They cannot reorganize
their memories for a new situation (at least not in laboratory conditions), and there is little transfer
learning from one situation to another. But it remains remarkable that such small brains can do
so much.

Reinhard et al. (2004) demonstrated a “Pavlov’s dog” response in bees by exposing a hive to two food
sources each laced with a different fragrances: rose and lemon. Later, all food and scent was removed
from the stations. When exposed to a scent, the bees returned to the empty feeding station that previously
contained that scent. While this demonstrates memory and learning, they do not give evidence for long
duration memory in bees.
Remarkable as the behavior is, Gregory (1997) sees no evidence of concept formation and insight in
bees. Gregory agreed with Gould and Gould (1988) when they said,

(bee) communication and navigation are the most complex known among invertebrates and, except-
ing humans, quite likely among vertebrates as well. . . .Far and away the most complex animal
behaviors we know of usually are innate. . . Orb-weaving spiders make their characteristic webs in
total darkness with no previous experience or learning. . . .In our opinion, selection has operated to
make complex behavior innate for the simple reason that, if it were not, animals could not hope to
discover it by trial and error or to learn it by observation in time to be able to perform it. In the end,
complexity of behavior is one of the worst guides of all to intelligence.

While these behaviors demonstrate the innate ability for efficient calculations using simple memories of
environmental cues such as landmarks and the movement of the sun, this does not compare with the
cognitive abilities of higher animals.

33.5.2.3 Cognitive Behavior of Societies

The term “social” is often applied to colonies of organisms described as swarms: a society of bees, social
ants, etc. In this discussion, societies will be limited to those that are cognitive. A society is similar to a
swarm in that individuals combine their actions for an aggregate result that could not be accomplished
individually. It differs from a swarm in that the behavior of individuals are cognitive: they understand and
exploit the cumulative power of the group and depend upon each other for support. A pack of wolves
exemplifies a cognitive society. An individual wolf is able to track down and kill prey unassisted. But it

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 612 — #12

Biomimetic Models for Wireless Sensor Networks 33-613

understands that the chance of success is greater in a pack and, therefore, prefers to participate as a member
of the pack. In this case, individual abilities are similar although differing duties may be assumed. Wolves
have behaviors for greeting pack members, keeping the pack together, and identifying social ranking. The
entire pack participates in raising the young: they feed the mother and the young, protect them from
harm, and nurse the young when the mother is away. Human societies are much more complex with
more diverse duties of individuals contributing to the good of the community. In a successful society, a
symbiotic relationship develops among individuals, where actions that are good for the society are also
good for the individual.
Most will agree that cognitive organisms require a higher level of intelligence. But that does not add to
understanding without a definition of intelligence. Pfeifer and Scheier (2000) summarized the opinions
of major experts in the field of psychology in 1921 on the definition of intelligence:

• The ability to carry on abstract thinking.

• Having learned or having the ability to learn to adjust oneself to the environment.
• The ability to adapt oneself adequately to relatively new situations in life.
• A biological mechanism by which the effects of a complexity of stimuli are brought together and
given a somewhat unified effect in behavior.
• The capacity to acquire capacity.
• The capacity to learn or profit by experience.
They point out interesting consistencies and inconsistencies. Opinions vary at the level in which intel-
ligence operates (“abstract thinking” versus “biological mechanism”). Some mention interaction with
the environment; some do not. Most indicate the ability to change behavior (“adjust”, “adapt,” “effect in
behavior,” “learn . . . by experience”).
For this discussion, both interacting with the environment and adapting to changes in the environment
are apropos. It has been stated that even unicellular organisms do both, but lack the ability to creatively
develop new behavior for adaptation. Albert Einstein said, “Insanity is doing the same thing over and
over again and expecting a different result.” In this context, a restatement is, “The unintelligent do the
same thing over and over and get the same result.” Conversely, “The intelligent create new behavior to
improve the result.” What do intelligent animals have that lower life forms lack? The ability to learn is
paramount.
33.5.2.4 Learning Models
A necessary but not sufficient component for learning is memory. In higher animals, there are different
types of memory for different functions (short term, long term, etc.). Learning requires more than just
memory: learning requires relationships among the memories so that new behavior can be conceived.
Long-term memory is critical for cognitive behavior (i.e., if you cannot collect memories over long periods
of time, the development of relationships is limited). As described, paramecia “remember” for minutes or
hours, bees perhaps for weeks. However, after memory is lost, the experience must be relearned.
Various learning models have been popularized. In the early years of the 20th century, two models
for behavioral modification gained wide acceptance. Pavlov developed the respondent conditioning model
informally known as “Pavlov’s dog” response. In this model, a subject can be trained to associate an
unrelated stimulus with a desired result (e.g., when a bell rings, food is presented and the dog salivates).
After enough exposure, the dog will salivate if a bell rings and no food is present. Skinner proposed the
operant conditioning model: behavior can be changed by applying rewards when a desired behavior is
observed (reinforcement is a reward to continue a desired behavior; negative reinforcement is a reward
when an undesirable behavior ceases) and punishment to extinguish an undesirable behavior. Clark
(1997) added an interesting observation concerning these models of learning: “Biological cognition is
highly selective and it can sensitize an organism to whatever (often simple) parameters reliably specify
states of affairs that matter to the specific life-form.”
Cognitive science was influenced greatly by the development of electronic computers. Behaviorists
came to believe that all behavior can be understood by the language of information processing. Lachman

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 613 — #13

33-614 Handbook of Bioinspired Algorithms and Applications

et al. (1979) stated that behavior can be modeled by “how people take in information, how they recode
and remember it, how they make decisions, how they transform their internal knowledge states, and
how they translate these states into behavioral outputs.” Pfeifer and Scheier (2000) described this as
cognitivistic paradigm or functionalism, which formed the basis of informational processing psychology.
This conveniently aligned with computer simulation, where thinking could be modeled as a computer
program having input, data structures, computation, and output.
While this model served well for simulating problem solving (culminating with the chess championship
of IBM’s “Deep Blue” computer over the reigning world champion chess player), it proved less than
adequate for modeling simple behavior even a child can perform, particularly those requiring interaction
with the environment (e.g., identification of colors and shapes, depth perception, navigation, etc.). Pfeifer
and Scheier (2000) promote a new model of learning they call embodied cognitive science (alias new
artificial intelligence or behavior-based artificial intelligence). Summarizing the view of Rodney Brooks
of the MIT Artificial Intelligence Laboratory, one of the founders of this new model, they agree that
the cognitivistic paradigm’s model of thinking, logic, and problem solving is fundamentally flawed by
our own introspection of how we see ourselves. Brooks suggested we abandon this approach and focus
on interaction with the real world. Intelligence emerges from the interaction of an organism with its
environment: intelligence must operate within a body. This effort has overcome some of the problems
encountered by use of the cognitivistic paradigm, and developed the approach of designing for emergent
behavior, producing some unexpected though desirable results.
Pfeifer and Scheier (2000) specify two additional concepts that pertain to this discussion: First, “The
essence of learning is that the [organism] can use its own experience to improve its behavior.” Because
experience is nondeterministic, predicting changes in behavior is difficult. Second, adaptive behavior
requires two components, compliance with existing “tried and true” rules and applying diverse, new
rules. They call this diversity–compliance tradeoff. It is the application of diverse, new behavior related to
previously learned behavior (rules) that has the best chance for improvement.
Cognitive behavior described above is defined as intra-generational; each individual of a generation
learns based on its own experience. Response to stimuli is evaluated by a fitness function (e.g., was lifting
my hand quickly from a hot stove good or bad?); behavior is either reinforced or altered based on the
result. In higher animals, stimuli and the resulting chosen behavior are remembered, organized, related,
and prioritized. Thus, when the same stimuli are encountered again, the best behavior is exhibited. But
inter-generational learning may be man’s greatest achievement. All organisms benefit from information
that is inherited (i.e., “learned”) from the previous generation through genetic codes. Genetic codes are
passed through DNA in reproduction and determine behavior as responses to stimuli (e.g., phototropism
in plants — plants “know how” to grow toward a light source). But as we have defined, this is not cognitive
behavior. In humans, this inter-generational transfer is augmented as information is acquired from parents
and other members of previous generations. In most animals this transfer of knowledge occurs only by
direct interaction with parents or contemporaries (e.g., other animals in the pack). Man however has
developed the ability to permanently capture knowledge through books and other media and indirectly
pass it on to future generations. The power of this technique cannot be overemphasized.

33.5.2.5 Social Models

In lower organisms, behavior is learned either through inheritance or by experience. However, higher
animals may control behavior through social means. These means may be represented as a continuum
between two extremes, centralized and distributed control:

• Centralized: Behavior of the society is dictated by a central authority. This approach has the
weakness of a single point of failure.
• Federated: Behavior of the society is dictated by a federation of authorities. This approach mitigates
the single point of failure but increases the complication of decision making.
• Distributed: Behavior of the society can emerge from actions of individuals on the basis of rewards,
extinction, and punishment or other learning models.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 614 — #14

Biomimetic Models for Wireless Sensor Networks 33-615

Imagine a town is newly formed. Who decides how many tailors the town needs? How many merchants?
How many carpenters? Obviously, a leader could emerge that would make those decisions for all. But the
leader could be wrong. He could fail and chaos would develop until new leadership emerges. He could
misappropriate resources for personal gain, but to the detriment of the society. Alternatively, individuals
could decide their own role in the society. A balance would naturally evolve: if too many people decide
to be a tailor, there will not be enough work and some will starve or be forced to change. It is likely that
the better tailors will get what work is available and survive. When individuals are free to make their
own choices of behavior, the choices can be detrimental to the society. The most successful societies have
been those in which individuals have been free to work for individual rewards within certain bounds
(i.e., individuals cannot seek personal rewards to the detriment of the society).

33.6 Application of Biomimetics to WSNs

The current WSN implementations described in Section 33.4 represent pioneering work that demonstrate
the enormous promise of the WSN concept. They share a number of similarities. Most of the motes
used are similar in capability to those used in the UCB-Intel large-scale demonstration (Lammers, 2001):
a 4 MHz microprocessor with 16 KB of flash instruction memory, 512 B of SRAM, a 256 KB EEPROM for
secondary storage, and a 10 Kbpsec radio transmitter. Each network has homogeneous motes although
the motes may be multifunctional having multiple sensing capabilities and serving other functions such
as routers, cluster leaders, etc. PCs serve as base stations for interfacing with the world outside the WSN.
However, all are passive, sensing systems (i.e., no effectors) observing the environment and reporting
these observations to a central authority where decisions will be made often by human observers. These
demonstrations used between 6 and 800 motes. Will the techniques used scale to massive numbers? In these
designs, behavior is predetermined, its results collected, and otherwise managed by a central authority.
Motes appear to be working together, though only in simple ways: multi-hop networks are formed and
clocks are synchronized. An examination of the method used for clock synchronization in the Twentynine
Palms demonstration (UCB/MLB 29 Palms UAV-Dropped Sensor Network Demo, 2001) reveals that it
closely models singleton behavior. Each mote periodically broadcasts its current time. Each mote that
receives the broadcast updates its time if the broadcast value is greater than its own. Though effective,
are the motes really working together? Or are they singletons following simple rules: they periodically
broadcast their time and they listen for the broadcast of packets containing time values. Do they know or
care that other motes exist? From the individual application of these simple rules, a simple behavior for
the group emerges: their times are synchronized. However, neither the individual nor the group behavior
can be described as cognitive.
Thus, it can be argued that much of the behavior of current WSNs parallels the behavior of simple
singleton organisms. The singleton model with simple emergent behavior will be insufficient as sens-
ing increases and diversifies, effecting becomes a player, and, thus, environmental influences become
nondeterministic. To reach full potential, cognitive WSNs must evolve to apply the diversity–compliance
tradeoff: deciding when to apply established rules and when to try new behavior based on previous learning
from experience. The more new behavior is tried, the more interaction with the environment becomes
nondeterministic and, therefore, unpredictable. Cognition becomes a great advantage.
As with biological societies, the WSN is expected to operate for a long time, although it is both acceptable
and expected that individual motes will not. For longevity of such a cognitive network, inheritance is
required to preserve knowledge between generations of motes. Consider the changing of a guard in a
combat scenario. A guard is deployed with the assignment of watching for the appearance of a threat.
If the guard was a sensor mote of today, he may be asked to notify the command center if a light, a noise, or a
magnetic disturbance is detected. The command center would take the observations and decide whether
or not to act. But a human guard would represent a cognitive mote. He would apply his cumulative
experience and training to his observation of the environment. He would apply known rules (e.g., he
would report seeing a tank), but he would exhibit new behavior when presented with new experience.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 615 — #15

33-616 Handbook of Bioinspired Algorithms and Applications

If he sees a light or hears a noise, he may not report the incident but store that information for further
consideration. He may pay more attention to that area in the future. He may also rely on inter-generational
learning: when faced with a new situation, he may consult the command center for more information; he
may consult a manual. What happens during the changing of the guard (i.e., generational replacement)?
The guard could tell the central command all information to be transferred to the new guard, let the
central command inform the new guard, and pass the new guard without saying a word. Perhaps it would
be more efficient for the current guard to directly discuss the information with the new guard allowing
interaction; the new guard could ask questions for clarification. The current guard could do a complete
“data dump” to the new guard telling him every incident that happened so that the new guard had all
experience that happened during the previous shift. Perhaps it would be more efficient for the current
guard to use his awareness, perception, intuition, reasoning, memory, and judgment (i.e., cognition) to
decide what knowledge to impart to the new guard. Similarly, cognitive nodes can significantly improve
the efficiency and effectiveness of many WSN implementations.

33.6.1 Potential Applications

Lessons from life that should be applied to future WSNs include aspects of both organizational and
behavioral models. Organization characteristics directly affect the sustainability and the functionality
of WSNs:

• Genetic material: The “genes” of sensor nodes are represented by characteristics endowed in their
creation. These can be both software and hardware. Flexibility requires that this genetic material
be mutable. Current nodes can change their software “genes” via their wireless connection but
hardware is constrained to the genes granted to them at “birth” (i.e., their manufacture). Efforts
are currently underway to design hardware that is evolvable such that it can redesign itself for new
functions.
• Metabolism: The metabolism of a sensor node is represented by electromechanical processes. The
source of “food” to energize these processes is the power source. Currently, power in WSNs is
provided by an onboard battery. The lifetime of a sensor node can be extended by providing it with
more powerful batteries, and technological improvements will surely provide more power in smaller
sized batteries. But this approach, no matter how potent, is still a finite power source. As the NRC
predicts, WSNs will be deployed requiring a long lifetime. A better approach is to design sensor
nodes that “eat”: mobile nodes forage for energy as do animals; stationary nodes must acquire
energy from a nearby, renewable source as do plants (photosynthesis using sunlight and water).
Research in this area is active, discovering techniques to make it possible to scavenge power from
the ambient environment. Potential sources include vibration, heat, light, and background radio
noise.
• Self-healing : Just as a multicellular animal does not fail because of malfunction or the death of
individual cells, a WSN must be able to compensate for malfunction or death of nodes.
• Symbiosis: Some organisms have developed symbiotic relationships in which each benefit from a
direct relationship with others. WSNs must be designed to organize in cooperation; complimenting
each other rather than competing.

Behavioral characteristics also directly affect the sustainability and the functionality of WSNs:

• Adaptability: The ability to adapt to a changing environment is paramount for any sustainable
system. If WSNs are to be autonomous, cooperative, fault-tolerant, self-regulating, and self-healing,
they must be able to adapt to changes in their environment. As with living organisms, some
adaptability can be innate, but to achieve maximum flexibility, cognition is required.
• Functional mobility: Obviously, mobile nodes can carry their functionality to different locations.
However, functionality can also migrate using stationary nodes. This migration is facilitated by the
reassignment of individual node duties.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 616 — #16

Biomimetic Models for Wireless Sensor Networks 33-617

• Sharing knowledge: Cognitive organisms pass information among their community in order to
reduce collectively the cost of learning. By sharing knowledge amongst nodes, WSNs can get
universal benefits for the experiences of a single node.
• Generational learning : As with communities of organisms, individual nodes will be “born”
(i.e., manufactured and deployed) and die (i.e., cease to function). New nodes coming into an exist-
ing WSN represent a new generation. WSNs must be designed such that knowledge is preserved
and made available to newer generations.
• Policing the society: It must be expected that nodes will not only fail, they may “misbehave.” In a
self-sustaining WSN, such rogue nodes must be discovered and dealt with to prevent damage to
the functioning network (i.e., the WSN must police itself). The ideal condition would be that all
damage is prevented, but, as with human societies, this may not be possible. However, the policing
must keep damage at an acceptable level.
• Cooperative sustenance: In animal societies, weak or sick individuals are cared for by the healthy and
strong members. In a well-regulated WSN, duties may be assumed by a stronger node, or perhaps
they may “nurse” the weaker node by providing sustenance (e.g., energy) or repairs (e.g., software
patches).
This is not meant to be an exhaustive list but to merely inspire thought on how models of living systems can
be directly applied to the improvement of WSNs. Many of these examples require extensive cooperation
among nodes in a network and among WSNs. To avoid the requirement (and all of its weaknesses) of
a centralized controller, the WSNs must be designed for the emergence of desirable behavior from this
cooperation.

33.7 Biomimetic Implementation Techniques

Cognitive societies of WSNs will require hardware more capable and less constrained than what is available
today. As stated, these improvements are to come soon enough. As Kurzweil recognized (Richards et al.,
2002), designing software systems to fully utilize these improvements for building cognitive societies is a
much more difficult problem. Therefore, the design of software should not wait until improved hardware
is available. Software design for the advanced WSNs can and should begin now. Tools and techniques
from the Alife communities directly apply.
A major building block for the embodied cognitive scientists is the intelligent agent. Although definitions
abound for the intelligent agent, Noor Malone (1999) offers a definition that is pertinent to this discussion:
“All agent categories have the common feature that they, to a large extent, independently perform tasks on
behalf of their contracting party (or user) for which specialized knowledge is needed, or which consist
of many time-intensive individual steps. . . .a software/hardware agent resides in an environment, uses
sensors to identify certain aspects of the environment, and executes commands that affect the environment.
Intelligent . . . agents act on behalf of people, take initiatives, and make suggestions.”
Agents do not have to be intelligent to be effective. Like bees and ants, individual agents can apply a
relatively small set of known rules in the presence of environmental stimuli. The actions of many such
agents can result in the emergence of more complex behavior. Consider John Conway’s “Game of Life”
(Gardner, 1970). The environment is a cellular automaton: a rectangular grid of square blocks or cells.
The grid really depicts a torus in that movement off the grid results in the reappearance on the opposite
side. Agents exist as the cells that are initially in one of two states, light or dark, and apply the following
rules for each change in state of the automaton:
• If a cell is lighted:
◦ If it has one or no lighted neighbors, it dies (darkens), as if by loneliness.
◦ If it has four or more lighted neighbors, it dies (darkens), as if by overpopulation.
◦ If it has two or three lighted neighbors, it survives (remains lighted).
• If a cell is dark:
◦ If it has three lighted neighbors, it lightens.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 617 — #17

33-618 Handbook of Bioinspired Algorithms and Applications

The behavior of the system emerges not only by the application of these rules but by the chosen initial state
of the environment (i.e., the initial position of the agents: which cells are lighted). When the changing
states are displayed as an animation, objects appear to form and move. Groups of cells consistently lighted
relative to each other begin to glide across the grid (called gliders). Sometimes, objects will appear to
remain still (called blinkers). Sometimes, blinkers will appear to emit gliders (called glider guns). The
behavior seems surprising and unpredictable.
Epstein and Axtell (1996) designed a similar system they called Sugarscape. Using a similar cellular
automaton, they endowed the environment with simple rules: each cell was assigned an initial amount of
sugar, a rate of replenishment when sugar was consumed by agents, and no two agents could simultaneously
occupy a single cell. Agents had two genetic endowments:

• A vision whereby they could see a given distance in cells (e.g., some could see two cells ahead; some
three cells, etc.) in each of the four horizontal and vertical directions.
• A metabolism rate: the rate at which they consume sugar as they move from cell to cell.

Agents were given one simple rule: look as far as you can see in all directions, find the cell that has the
most sugar, go there, and consume the sugar. Sugar was dispersed in the environment with concentrations
at certain places. Four hundred agents were initially assigned locations and the rules applied in a series of
states. Not surprisingly, at the end of the simulation, agents with high metabolism rates and poor vision
died (they consumed all of their sugar before they found nourishment) and those with better vision and
lower metabolism survived to locate the cells richest in sugar: the system demonstrated a living Darwinian
ecosystem. They added more and different rules to eventually simulate other living attributes such as
commerce, combat, and sexual reproduction.
The question is, can agent-environment systems be engineered to elicit desired behavior and prevent
undesired behavior? That is, can Kelly be correct that one cannot find the hive in the bee; or is it that,
if one knew enough about the bee, the hive could be seen in it. And, can this technique be applied in the
design of WSNs?
Fuzzy logic and artificial neural networks (ANN) can be applied to the development of cognitive
WSNs. Introduced by Zadeh (1965), fuzzy logic was conceived to define partial truth; that is, truth values
that are not completely true and not completely false. ANNs are based on the concept of an artificial
animal neuron, the Threshold Logic Unit (TLU), proposed by McCulloch and Pitts (1943). Gurney (1997)
defines an ANN as, “an interconnected assembly of simple processing elements, units or nodes, whose
functionality is loosely based on the animal neuron. The processing ability of the network is stored in
the inter-unit connection strengths, or weights, obtained by a process of adaptation to, or learning from,
a set of training patterns.” The strength of the ANN is that they can be trained when the weights of the
connections are adjusted on the basis of input experienced (i.e., the ANN learns). Hashem et al. (1995)
demonstrated that neural networks benefit real-time data analysis by WSNs. For the task of identifying
contaminants in the environment, they demonstrated that using neural network-based analysis of data
from an array of heterogeneous sensors, selectivity of the array is significantly improved using sensors that
are not individually selective. The combination of a heterogeneous sensor array with an automated analysis
system is described as an artificial or electronic nose and has been demonstrated for use in monitoring food
and beverage odors, analyzing fuel mixtures, and environmental monitoring. The array is designed such
that each sensor measures a different property of the sensed sample. Each chemical composition presented
to the array produces a characteristic signature through the fusion of sensor readings. By presenting
many different compositions to the array, a database of signatures can be recorded. Conventionally, the
number of sensors must be at least as great as the number of analytes. The quantity and complexity
of data collected can make analysis of the data intricate. Using ANN to analyze the data for pattern
recognition can not only reduce the necessary computation, but can reduce the required number and type
of sensors in the array. Can ANNs and/or fuzzy logic be used in other applications of WSNs to facilitate
learning ?

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 618 — #18

Biomimetic Models for Wireless Sensor Networks 33-619

33.8 Engineering for Emergent Behavior

There exists a good deal of parallelism between the evolutionary development of life as presented and what
will be the evolutionary development of the WSN. The evolutionary roots of WSNs may be taken as having
begun with the development of the microprocessor, which allowed the development of the initial WSNs:
simple sensor nodes receive environmental stimuli as input (e.g., a command to take a sensory reading,
the sensed value, etc.), and respond as output (e.g., transmit the sensed data). A central processor of some
kind is required to eventually aggregate the sensed data from all nodes and make some global decisions
based on those results. As WSNs evolve, the sensor nodes will become more intelligent, allowing greater
amounts of processing to occur locally. The architectural requirement for the network to have a central,
controlling processor will be greatly reduced or eliminated (Lodding, 2004a). For some purposes, nodes
will be given the ability to affect their local environment through the structural coupling of sensor and
effector technologies. One example of such an application is aircraft smartskin, in which a sensor-affector
network built from a large number of smart sensors and affectors will be directly embedded into the wing
structure to dynamically alter the shape of an aircraft wing, morphing it into the most efficient shape for
the current aerodynamic conditions. With a biomimetic architecture there will be no need for a central,
controlling computer as the wing can think for itself.
Technological evolution, like its biological twin, is an ongoing process. That the WSN will over time
evolve from dumb sensor nodes to more intelligent and capable sensor-affector nodes is largely self-
evident. A less obvious change is that for anything but the simplest of dedicated WSNs, intelligence will
migrate away from a central, controlling computer, and distribute itself fully into the network, living
within and across the collection of ever more intelligent and flexible nodes. The WSN is moving from a
unicellular to a multicellular architecture (Lodding, 2004b).
Sensor nodes that are capable of altering both their behavior and their relationships with other nodes
will be developed. As node intelligence, the number of nodes, and the number of inter-node connections
increase, the stage becomes set for a phenomenon called circular causation to possibly arise. This is when
the actions of the parts cause the overall behavior of the organism (i.e., emergent behavior) and the overall
behavior guides the action of the parts making up the organism (i.e., self-organization). A possible mani-
festation of this behavior might take the form of network nodes forming local assemblies, or hierarchies,
of embedded micro sensor-affector networks: organs. The formerly limited WSN will now evolve into a
responsive, smart, adaptable, self-healing sensor-affector network, what can be described as a cognitive
network. The primary question is: will this evolution occur in an accidental, haphazard manner, or will it be
engineered?
Today, the primary focus of WSN research has been on the unicellular level of evolutionary development.
Difficult questions concerning the hardware and software aspects of implementing real-world WSNs are
being addressed. Concerns about electrical power requirements, communications issues, and network
topology discovery are some of the basic problems that are currently being explored. While these parts
of the puzzle are both necessary and appropriate for research, there is a larger, missing piece of the puzzle
that is being overlooked.
The missing piece of the puzzle is how do we engineer emergent behavior? How do we program individual,
unreliable, autonomous units that make up a larger system to exhibit specific, desired behavior? How do we
predict the high-level behavior of the group from the individual programs executing in the individuals?
And just as importantly, how do we inhibit the emergence of unwanted behavior from the collective
interactions of numerous individuals?
There are people outside of the sensor-network research community who have been looking at just these
types of problems. The majority of these researchers are in the fields of Artificial Intelligence and Artificial
Life. Mathematicians have looked at the complexities of self-organizing systems, but with little, if any,
direct influence in software algorithm development. Understanding how to design and build emergent
systems is for all practical purposes an unknown and unexplored problem domain. If we want to control
the development of cognitive networks, into which it is believed the WSN is predestined to evolve, then
it is critical to understand and apply this new engineering science. Furthermore, the development and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 619 — #19

33-620 Handbook of Bioinspired Algorithms and Applications

application of this science needs to begin now, while the focus is on the unicellular stage of evolution.
Rather than wait until nodes have the expected increased capability to begin application, it is far better
to begin now to understand and plan with this new method of engineering. When these networks are
endowed with the qualities and abilities that are expected, it must be guaranteed that a command such as,
“Please open the pod bay door, HAL.,” produces the desired result.

33.9 Concluding Remarks

WSNs are destined to become pervasive in machines and ubiquitous throughout the Earth (and beyond)
as they evolve into sensor/effector systems that not only monitor the environment but affect it as well.
They are also going to rapidly expand their functionality. It has been demonstrated that the evolution
of these networks has, parallels in the evolution of life. Current WSNs represent the beginning stage
of their “life,” where they function as a community of unicellular singletons receiving stimuli (input)
and reacting to those stimuli (output). A centralized authority may provide some of that input and will
aggregate their output for its own gain, but this is unknown or understood by the sensor nodes. The
observation here is that their evolution can be accelerated and improved by engineering for emergent
behavior. This engineering should not only be developed for achieving desired emergent behavior from
large numbers of increasingly functional nodes (bees and ants), but should be preparing now for nodes
that are cognitive (i.e., can learn). The goal for the design of future WSNs should be a “biosphere” of
WSNs interacting with each other and their environment “symbiotically” such that they are autonomous,
cooperative, fault-tolerant, self-regulating, self-healing, self-sustained and exhibit only desirable behavior.

References
Abelson, H., Allen, D., Coore, D., Hanson, C., Homsy, G., Knight, T., Nagpal, R., Rauch, E., Sussman, G.,
and Weiss, R. (1999). Amorphous Computing. MIT Artificial Intelligence AI Memo 1665. Retrieved
April 5, 2004, from http://www.swiss.ai.mit.edu/projects/amorphous/papers/ aim1665.pdf.
Abelson, H., Allen, D., Coore, D., Hanson, C., Rauch, E., Sussman, G., and Weiss, R. (2000).
Amorphous computing. Communications of the ACM, 43, Retrieved April 5, 2004, from http://
www.swiss.ai.mit.edu/projects/amorphous/cacm-2000.html.
Bonabeau, E., Dorigo, M., and Theraulaz, G. (1999). Swarm Intelligence: From Natural to Artificial Systems.
Oxford University Press, Oxford.
Budiansky, S. (2002). Creatures of our own making. Science, 298: 80–86.
Centre for Pervasive Computing (2004). Retrieved April 5, 2004, from http://www.pervasive.dk.
Chen, M., Majidi, C., Doolin, D., Glaser, S., and Sitar, N. (2003). Design and Construction of a Wildfire
Instrumentation System using Networked Sensors (Poster). Network Embedded Systems Technology
(NEST) Retreat, Oakland, CA. Retrieved April 5, 2004, from http://firebug.sourceforge.net.
Clark, A. (1997). Being There: Putting Brain, Body, and the World Together Again. MIT Press,
Cambridge, MA.
Conniff, R. (1996). Spineless Wonders. Henry Holt and Company, New York.
Delin, K. (2004). NASA/JPL Sensor Webs Project. Retrieved April 5, 2004, from http://sensorwebs.jpl.
nasa.gov.
Dusenbery, D. (1996). Life at Small Scale: The Behavior of Microbes. Scientific American Library Series
No. 61.
Dyer, F. and Seeley, T. (1991). Dance dialects and foraging range in three Asian honey bee species.
Behavioral Ecology and Sociobiology, 28: 227–233.
Epstein, J. and Axtell, R. (1996). Growing Artificial Societies: Social Science from the Bottom Up. The
Brookings Institution, Washington, D.C.
Gardner, M. (1970). MATHEMATICAL GAMES: The fantastic combinations of John Conway’s new solit-
aire game “life.” Scientific American, 223: 120–123. Retrieved April 5, 2004, from http://ddi.cs.uni-
potsdam.de/HyFISCH/Produzieren/lis_projekt/proj_gamelife/ConwayScientificAmerican.htm.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 620 — #20

Biomimetic Models for Wireless Sensor Networks 33-621

Gould, J., and Gould, C. (1988). The Honey Bee. Scientific American Library, New York.
Gregory, R. (1997). Editorial: Brains of ants and elephants. Perception, 26. Retrieved April 5, 2004, from
http://www.perceptionweb.com/perc0397/editorial.html.
Gurney, K. (1997). Introduction to Neural Networks. Routledge, an imprint of Taylor and Francis Books
Lt, London.
Hashem, S., Keller, P., Kouzes, R., and Kangas, L. (1995). Neural network based data analysis for
chemical sensor arrays. In Proceedings of International Society for Optical Engineering (SPIE)
AeroSense Conference, Orlando, FL (April 17–21, 1995), in Applications and Science of Artifi-
cial Neural Networks, Vol. 2492, Paper #2492–05, pp. 33–40. Retrieved April 5, 2004, from
http://citeseer.ist.psu.edu/519919.html.
Intel Research — Exploratory Research — Deep Networking (2004). Retrieved April 5, 2004, from
http://www.intel.com/research/exploratory/deep_networking.htm.
Intel Research — Exploratory Research — Deep Networking — Heterogeneous Sensor Networks (2004).
Retrieved April 5, 2004, from http://www.intel.com/research/exploratory/heterogeneous.htm
Invisible Computing: Scope (2000). Retrieved April 5, 2004, from http://invisiblecomputing.org/
scope.html.
Jones, K., Wadaa, A., Olariu, S., Wilson, L., and Eltoweissy, M. (2003). Towards a new paradigm for
securing wireless sensor networks. In Proceedings New Security Paradigms Workshop 2003, Ascona,
Switzerland. August 18–21, 2003, pp. 115–122.
Kahn, J., Katz, R., and Pister, K. (1999). Next century challenges: Mobile networking for
“Smart Dust”. In ACM MOBICOM Conference, Seattle, WA. Retrieved April 5, 2004, from
http://www.cs.berkeley.edu/∼randy/Papers/mobicom99.pdf.
Kelly, K. (1994). Out of Control: The New Biology of Machines, Social Systems, and the Economic World.
Perseus Books.
Lachman, R., Lachman, J., and Butterfield, E. (1979). Cognitive Psychology and Information Processing.
Lawrence Erlbaum Assoc, Hillsdale, NJ.
Lammers, D. (2001). Embedded projects take a share of Intel’s research dollars. EE Times. Retrieved
April 5, 2004, from http://today.cs.berkeley.edu/800demo/eetimes.html.
Leveraging the Infosphere: Surveillance and Reconnaissance in 2020 (1995). Airpower Journal — Summer
1995, A SPACECAST 2020 White paper. Retrieved April 5, 2004, from http://www.airpower.
maxwell.af.mil/airchronicles/apj/spacast1.html.
Lodding, K.N. (2004a). Hitchhikers Guide to Biomorphic Software. ACM Queue.
Lodding, K.N. (2004b). Multi-agent organisms for persistent computing. In Proceedings of the 3rd Inter-
national Joint Conference an Autonomous Agents and Multi Agent Systems (AAMASO4). New York,
NY, July 19–23.
Mainwaring, A., Polastre, J., Szewczyk, R., and Culler, D. (2002). Wireless sensor networks for
habitat monitoring. (Intel Research, IRB-TR-02–006, June 10, 2002) In ACM International Work-
shop on Wireless Sensor Networks and Applications. Retrieved April 5, 2004, from http://www.
greatduckisland.net.
Margulis, L. and Sagan, D. (1986). Microcosmos: Four Billion Years of Microbial Evolution. Simon and
Schuster, New York, p. 15.
McCulloch, W. and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin
of Mathematical Biophysics, 7: 115–133.
Moravec, H. (1988). Mind Children. Harvard University Press.
National Research Council (2001). Embedded, Everywhere: A Research Agenda for Systems of Embedded
Computers, Committee on Networked Systems of Embedded Computers, for the Computer Science
and Telecommunications Board, Division on Engineering and Physical Sciences, Washington, DC.
Retrieved April 5, 2004, from http://www.nap.edu/catalog/10193.html.
Noor, A. and Malone, J. (1999). Intelligent Agents and Their Potential for Future Design and Synthesis
Environment. NASA CP-1999-208986.
Pfeifer, R. and Scheier, C. (2000). Understanding Intelligence. MIT Press, Cambridge, MA.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 621 — #21

33-622 Handbook of Bioinspired Algorithms and Applications

Primio, F., Müller, B., Lengeler. (2000). Minimal cognition in unicellular organisms. In SAB2000
Proceedings Supplement, International Society for Adaptive Behavior. Honolulu, HI. pp. 3–12,
Retrieved April 5, 2004, from http://www.ais.fraunhofer.de/BAR/papers/diprimio-mincog.pdf.
Reinhard, J., Srinivasan, M., and Zhang, S. (2004). Olfaction: scent-triggered navigation in honeybees.
Nature, 427: 411.
Richards, J., Gilder, G., Kurzweil, R., Searle, J., Dembski, W., Denton, M., and Ray, T. (2002). Are We
Spiritual Machines? Discovery Institute, Seattle, WA.
Tilak, S., Abu-Ghazaleh, N., and Heinzelman, W. (2002). A taxonomy of wireless micro-sensor
network models. ACM Mobile Computing and Communications Review (MC2R), 6. Retrieved
April 5, 2004, from http://www.cs.colorado.edu/∼rhan/CSCI_7143_001_Fall_2002/Papers/
Tilak2002_p28-tilak.pdf.
Turner, J. (2000). The Extended Organism: The Physiology of Animal-Built Structures. Harvard University
Press.
UCB/MLB 29 Palms UAV-Dropped Sensor Network Demo (2001). University of California, Berkeley, CA.
Retrieved April 5, 2004, from http://robotics.eecs.berkeley.edu/∼pister/29Palms0103.
Wadaa, A., Olariu, S., Wilson, L., Eltoweissy, M., Jones, K., and Sundaram, P. (2004a). Training a sensor
network. In Special Issue of MObile NETwork (MONET) on Algorithmic Solutions for Wireless,
Mobile, Ad Hoc and Sensor Networks, Bar-Noy, A., Bertossi, A., Pinotti, M., and Raghavendra, C.
Eds. January 2004.
Wadaa, A., Olariu, S., Wilson, L., Eltoweissy, M., and Jones, K. (2004b). On providing anonymity
in wireless sensor networks. In Proceedings of the 10th International Conference on Parallel and
Distributed Systems, (ICPADS-2004). Newport Beach, CA. July 2004.
Weiser, M. (1991). The computer for the Twenty-first century. Scientific American, 94–10. September 1991.
Retrieved April 5, 2004, from http://www.ubiq.com/hypertext/weiser/ UbiHome.html.
Weiser, M. (1996). Ubiquitous Computing. Retrieved April 5, 2004, from http://www.ubiq.com/
hypertext/weiser/UbiHome.html.
Zadeh, L. (1965). Fuzzy sets. Information and Control, 8: 338–353.

Appendix A — Glossary
The following terms are pertinent to this topic:
Analyte: The object of measurement in an analytical procedure (in this case a chemical property).
Artificial intelligence: The science of simulating intelligence in a creation of humans (i.e., not natural).
Artificial life: Beyond intelligence, the science of simulating a living organism or some property thereof
in a creation of humans (i.e., not natural).
Biomimetics: The study of the origin, structure, or function of biological mechanisms, processes, and
materials as models for the design of artificial constructs.
Circadian: Relating to approximately a 24 h period (from Latin meaning “around the day”).
Effector: A device that, in response to a stimulus (input or command), initiates an effect on or affects
its environment.
Emergent Behavior: Behavior that results from:

• Innate behavioral rules applied to unpredictable environmental stimuli.

• Behavior of individuals interacting and combining to result in behavior of a group.
• Learning (i.e., in cognitive systems, learning from experiences of behavior may initiate new
behavior).

Evolvable hardware: Hardware designed by the application of evolution to automate its creation and
adaptation. A goal is to use these techniques in situ producing hardware that can adapt to unpredicted
environmental conditions, thus improving its survivability for long durations in unknown and changing
environments.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 622 — #22

Biomimetic Models for Wireless Sensor Networks 33-623

Homeostasis: The state of a relatively constant internal environment. The physical and chemical states
that an organism must maintain to allow proper functioning, in maximum efficiency, of its components:
cells, tissues, organs, and organ systems.
Innate behavior: Behavior that results from genetic encoding. Such behavior does not require learning
and changes little in response to environmental stimuli.
Learned behavior: Behavior that results from the cognitive assessment of prior experiences to determine
a new response.
Metabolism: The chemical processes (breaking down substances to provide energy or synthesis of new
substances) within a living organism that are necessary for life.
Mote: Defined as a small particle, here it species the self-powered, physically independent device that
contains hardware and software for sensing or effecting, computing, and communication in a wireless
sensor network.
Nanoscale: Measurement on a scale of nanometers (dimensions under ∼100 nm).
Nano-technology: The process used to design and build electronic circuits and devices from atoms and
molecules.
Node: In computer science, this is either a terminal or hop point in a communications network. Here
it specifies a mote in the context of the network.
Pervasive: Defined as the quality of being present throughout; to permeate. Pervasive computing
and ubiquitous computing seem to be synonymous in computer science literature, but here, pervasive is
used to describe presence throughout a system (e.g., sensors are pervasive in an automobile if they are
installed and function throughout the automobile).
Photosynthesis: The process in some organisms (plants and some microbes) by which carbohydrates are
synthesized from carbon dioxide and water using light as an energy source. These carbohydrates serve as
energy storage and are later metabolized resulting in the release of carbon dioxide, water, and energy used
by the organism.
Protozoa: Plural for protozoan collectively naming any of a large group of single-celled, micro-
scopic organisms (e.g., amoeba, sporozoans, ciliates, flagellates, etc.). Protozoans differ from bacteria
(prokaryotic) in that they, like higher plants and animals, contain cellular constructs such as nucleus,
mitochondria, etc. (eukaryotic).
Sensor: A device that receives and reports a signal or stimulus.
Society: In biology, a society is defined as a colony of organisms, usually of the same species such as a
society of ants. Here, a society is limited to an association of cognitive animals such as wolves or humans.
Symbiosis: A relationship between organisms that is mutually beneficial and which, over time, forms a
dependence.
Ubiquitous: Defined as the quality of being or seeming to be everywhere at the same time. Pervasive
computing and ubiquitous computing seem to be synonymous in computer science literature, but here,
ubiquitous is used to describe presence within all systems (e.g., sensors are ubiquitous in automobiles if
they are installed and function in all automobiles).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C033” — 2005/8/6 — 14:13 — page 623 — #23

 34
A Cooperative Parallel
Metaheuristic Applied
to the Graph Coloring
Problem

34.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34-625

34.2 A Cooperative Metaheuristic. . . . . . . . . . . . . . . . . . . . . . . . . . . .34-626
First Step Search • Memorizing during the Search •
Cooperative Search
34.3 Graph Coloring Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34-628
Applications • Formalization • Used Methods • Reference
Graphs
34.4 Basic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34-630
IG Operator • Break Operator • Operators Efficiency
34.5 A Cooperative Metaheuristic. . . . . . . . . . . . . . . . . . . . . . . . . . . .34-632
Information to Memorize • Partial Coloring Building •
Parallel Implementation
34.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34-635
Preliminary Tests • Global Architecture
Benjamin Weinberg 34.7 Conclusion and Future Works . . . . . . . . . . . . . . . . . . . . . . . . . .34-635
El-Ghazali Talbi References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34-637

34.1 Introduction
Combinatorial optimization often needs a large amount of computation resource. This is particularly the
case for NP-hard problems, for which no efficient algorithm is known. Parallel computers may supply
this resource. However, it may be interesting to use those platforms in another way than to parallelize the
computation of the objective function of the visited configurations.
Coevolutionary computation [1] leads to algorithms that can be implemented in a distributed way.
We call this kind of implementation cooperative algorithms. Such an algorithm presents a low coupling
between its components. In this chapter, we present a distributed design and implementation of the
metaheuristic COSEARCH [2,3]. This metaheuristic is a very general model which can easily be imple-
mented in several ways. We propose a view of this algorithm bearing in mind the need to explicitly balance

34-625

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 625 — #1

34-626 Handbook of Bioinspired Algorithms and Applications

diversification and intensification during the search. We evaluate our approach on the graph coloring
problem.
Nevertheless, the implementation of COSEARCH needs the design of mechanisms dedicated to the
studied problem. In our case, we use two new operators which are the break and the Xρ operators. The
former is an operator modifying a unique coloring as mutation in genetic algorithms. The latter combines
two colorings as crossover.
This chapter is organized as follows. In Section 34.2, we present an overview of the parallel cooper-
ative COSEARCH method: concepts and implementations of suggestions. In Section 34.3, we present
the studied problem, which is the graph coloring problem (MGCP). In Section 34.4, we present some
preliminary works on the search operators. In Section 34.5, we present an implementation of COSEARCH
for MGCP. In Section 34.6, we have given some experimental results. Finally, we conclude and present
some perspectives in Section 34.7.

34.2 A Cooperative Metaheuristic

In this section, we present a brief overview of a parallel metaheuristic, called COSEARCH. Some work
on this metaheuristic was undertaken in Reference [2] for the quadratic assignment problem (QAP) and
in reference [3] for the frequency assignment problem (FAP). This presentation is very general and this
metaheuristic may be instantiated in many ways. The head of the algorithm is the use of a memory
gathering information on the overall search.
This section is organized as follows: In Section 34.2.1, we present the main idea to visit the search space.
In Section 34.2.2, we present the challenge to memorize information during the search. In Section 34.2.3,
we present an overview of our approach.

34.2.1 First Step Search

Optimizing an NP-hard problem leads in general to the use of heuristic approaches. One of the most
used approach is the local search. Those algorithms modify configurations (candidate solution) by local
transformations. Generally, it is said, that we move from a configuration to one of its neighbors by a small
modification.
Iterating successive modification leads to a configuration which is better than the initial configuration.
But generally, the configuration found is only a local optimum; that is we cannot improve the configuration
by any small modifications, we use. This is typically the case in the hill climbing algorithm. In such an
algorithm, an elementary operation is called a move.
That is why the additional search mechanism is often used to escape from the local optimum and
continue the search. Then, the best visited local optimum is returned at the end of the algorithm. There
are plenty of such mechanisms more or less dedicated to the studied problem. We can cite the three
following examples:

• Random mechanisms using biased probabilities of acceptance of the worst configurations as in

simulated annealing [4].
• Memorizing of the last visited configurations or the last made moves to prevent visiting again a
previously visited local optimum as in a tabu search [5].
• Multistart methods consist in the searching of the local optimum from different initial
configurations [6].

Such methods may provide good results. However, in some cases, these methods may be insufficient to
find the optimum. We propose to use the local search algorithm as a basic component of our algorithm.
Several parallel/distributed local search algorithms may work in a cooperative way to optimize a problem.
In our approach, such an algorithm is called a (search) agent.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 626 — #2

A Cooperative Parallel Metaheuristic 34-627

34.2.2 Memorizing during the Search

Search agents work from different configurations of the search space. We suggest to iterate their application
in order to cover the whole search space (as in multistart approach). However, for a better guidance, we
have to synchronize the different agents around an adaptive memory.
The adaptive memory gathers information on the visited configurations. The main difficulty in apply-
ing this algorithm is to know what is pertinent to keep in this memory. Indeed, the search may take
a long time and it would be unhelpful and inapplicable to keep every visited configuration in the
memory.
The second difficulty is to answer the following question: “what do we use with this memory?” A search
agent needs an initial configuration to search. The memory has to supply interesting initial configurations.
For that reason, we may be able to build configurations with new or insufficiently explored characters
with regard to the memory.
So, in answer to the second point, the way to surmount the main difficulties, in our view, is that the
memory must keep some good configurations and notice the unexplored area of the search space.

34.2.3 Cooperative Search

Our approach consists in cooperation of several search algorithms (called agent) through the adaptive
memory (see Figure 34.1).
Typically, an agent is a local search strategy (due to its efficiency), but it is not a strict rule. So, we can
imagine to select any algorithm as an agent.
The different searches made by agents are totally independent of each other. So, each agent can be
executed in parallel.
Furthermore, we can use heterogeneous agents. As the coupling is very low, several algorithms can be
used in a complementary manner:

• Local searches with different operators

• Different metaheuristics as simulated annealing or tabu search
• Complex algorithms using two or more configurations as genetic algorithms

adaptive memory

deduce unexplored characters

good configurations provide

provide
generate new start points
update
provide
Input/output
agent agent

agent
agent

agent agent
agent

FIGURE 34.1 Global overview of our approach.

Note: In this figure, we use ellipses to indicate the presence of dynamic code and rectangles for static data.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 627 — #3

34-628 Handbook of Bioinspired Algorithms and Applications

34.3 Graph Coloring Problem

This section is about recalling the definition of the academic generic problem which is graph coloring
problem (MGCP). In Section 34.3.1, we present some applications of the graph coloring problem. Then
in Section 34.3.2, we present the formalizations that we use. In Section 34.3.3, we present an overview of
the heuristic methods applied to this problem.

34.3.1 Applications
Graph coloring problem models a lot of real-life problems. Those problems consist in distributing objects
into different groups in order to separate some pairs of objects, for example, the Register Allocation
Problem (RAP) [7]. In this problem, the objects are the variables of a source code and the groups are the
registers of a computer. For each variable, we must assign a register number. Two variables, having the
same register number, cannot exist at the same time. Thus, we obtain a set of binary constraints of mutual
exclusion on variables. The smaller the number of used registers, the more efficient is the program. The
objective function is obviously to minimize the number of registers.
There are a lot of other problems, which are direct applications of MGCP. In this category, we can
find:
• Frequency assignment problem [8]. In this problem, the objects are the channels of the transceiv-
ers and the groups are the available frequencies. Two channels covering near by areas have to
communicate using different frequencies.
• Timetabling problem [9]. A course combines the constraints from three elements: teacher, class,
and location; knowing that a teacher cannot make two lectures at the same time, idem for class and
location.
• Some other problems such as pattern matching [10], air route conception [11], etc.

34.3.2 Formalization
All problems presented above can be modeled using a graph G(V , E). The set of vertices V is the set of
object to color (distribute or assign) and the set of edge E symbolizes the constraints between vertices:
vertices connected by an edge must have different colors (be assigned with different values).
We use the following formalization for a given graph G(V , E).
Definition 34.1 We call coloring (of G) any mapping from V into the set of natural numbers {1 . . . |V |}.
Definition 34.2 We call color of v with regards to a coloring C , the image of v by C (i.e., C (v)).
Definition 34.3 We call ith class of a coloring C , the set of vertices colored into i (i.e., C −1 (i)).

Technically, a coloring may be trivially encoded by an array of integers indexed by the vertices. Never-
theless, for some algorithms it will be helpful to have the vertices gathered by classes. We choose a mixed
encoding (see Figure 34.2). This encoding allows the visit of vertices of a class without investigating the
other vertices. An array allows us to change quickly (in O(1)) the color of one vertex.
Definition 34.4 We call violation of constraint (in a coloring C ), an edge whose extremities is colored by the
same color (i.e., C −1 (i)).

Definition 34.5 We define g the mapping defined by:

g : {colorings} → N
1

C  → g (C ) = × δC (i),C (j) × IE (i, j)
2
i,j,∈V

where δi,j is the Kronecker’s symbol (i.e., δi,j = 1 if i = j; 0 otherwise) and IE (i, j) is the characteristic
function of E (i.e., IE (i, j) = 1 if (i, j) ∈ E; 0 otherwise).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 628 — #4

A Cooperative Parallel Metaheuristic 34-629

belonging to a group composition of a part

O1 C1 C1 O1
O2 Ci O1 O5 O6 O8 O2

Ci
O2 O4

Ck
Om Ck O7 O9 Om Om

FIGURE 34.2 Example of encoding.

Note: Technically the number χ (C ) of non-empty parts is stored and only the first χ (C ) labels are used.

Definition 34.6 We say that a coloring C is a proper coloring if there is not any violation (i.e., g (C ) = 0).

Definition 34.7 A coloring is a k-coloring if the used colors are less than or equal to k.

Definition 34.8 The chromatic number of graph is the smaller number such that the graph allows a proper
k-coloring. For a given graph G, we denote χ (G) as its chromatic number.

Definition 34.9 Let k be a positive number. We call graph k-coloring problem (GkCP), the problem consists
in finding a k-coloring such that g (C ) is minimal.

Definition 34.10 We call Minimal graph coloring problem (MGCP), the problem consisting in finding a
coloring with the less used colors. So, we minimize the following function:

 
f : proper colorings → N
C  → f (C ) = max C (v)
v∈V

This is this second objective function, we want to optimize.

Both problems are known to be NP-hard [12].

In this problem, there is a symmetry. Indeed, we do not care about the effective value assigned to a
vertex. The unique interest resides in the fact that vertices share or do not share the same color. So, the
graph coloring problem is a partitioning problem (also called grouping problem).
In Reference [13], we propose some tools to operate on the space of partitioning without the problems
due to the symmetry. We provide a format to compute in linear time if two colorings represent the same
partitioning.
We also present how to compute the distance between two partitioning independently of the choice of
their representations. Our study is based on the use of what we call a refining matrix. The refining matrix
of two colorings A and B is a matrix MAB defined by:

mijAB = card(A(i) ∩ B(j)). (34.1)

This matrix will be helpful for the design of search operators.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 629 — #5

34-630 Handbook of Bioinspired Algorithms and Applications

34.3.3 Used Methods

Minimal graph coloring problem is an old and deeply studied problem. As the problem is NP-hard, the
heuristic approach is often used. There are two great families of algorithms for MGCP. The first one
operates on proper coloring:

• Greedy algorithm DSATUR [14]

• Recursive Largest First RLF [15]
• Local search algorithm using the Kempe chain [16]
• Iterated Greedy, proposed by J. Culberson [17] is a kind of local search
• Memetic Algorithm using a special encoding of a coloring [18]
• Ants colony [19]
• Constraint Programming [11]

The second one operates on k-colorings. The search is decomposed into steps. At each step, the algorithm
works with a fixed k and tries to find a proper k-coloring in the whole k-colorings. Then, k is decreased for
a new step. When a step fails (does not find a proper coloring), the previously found proper k + 1-coloring
is returned. This strategy was used in many heuristics:

• Local search (hill climbing, simulated annealing) [16] with the change of one vertex as operator
• Local search with using “permutation-neighborhood” operator [20]
• Tabu search [21–23]
• Addition of several diversification techniques to local search procedures [24]
• Local search with variable neighborhood search [25]
• Modification of bias of neighborhood operator [26,27]
• Evolutionary Algorithm with dedicated operators [23,28,29]
• A multi-level approach [30]

Nevertheless, Clerc [31] studied a method using a mix of both strategies. He used a method working
on valid colorings and k-colorings in particle swarm optimization algorithm.
In this study, our attention was focused on the first strategy.

34.3.4 Reference Graphs

We evaluate our algorithm on several graphs from the site of Trick.1 Those graphs are issued from the files
of the second DIMACS challenge. Table 34.1 presents the benchmark and some of their features.

34.4 Basic Algorithm

This study is based on the use of two algorithms modifying a coloring. Both of them change proper
coloring into another proper one where the number of colors decreases or remains the same. In this
section, we present those operators. Then, we experiment upon their efficiencies.

34.4.1 IG Operator
Culberson [17] proposes a manner to generate a coloring C , where the number of colors is less than
or equal to the number of colors of the genitor C . The new coloring is built as follow: First, all the
vertices are uncolored in C . Second, all (uncolored in C ) vertices of class C −1 (1) are placed into the first
class of C . Then, the class is filled with other (uncolored in C ) vertices (i.e., adding vertices until the
class does not accept any one without inducing constraint violations). This step is repeated to build the

1 http://mat.gsia.cmu.edu/COLOR/color.html

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 630 — #6

A Cooperative Parallel Metaheuristic 34-631

TABLE 34.1 Used Benchmarks

Name nb vertices Density

DSJC125.1 125 0.1

DSJC125.9 125 0.9
DSJC500.9 500 0.9
DSJC125.5 125 0.5
DSJC250.9 250 0.9
DSJR500.1c 500 0.1
le450_5c 450 0.1
le450_5d 450 0.1
queen8_12 96 0.3
queen9_9 81 0.3
queen10_10 100 0.3
queen11_11 121 0.3
school1 385 0.3
school1_nsh 352 0.2

second color and so on. The visit of the class of C are made in a disorder by manner to generate different
colorings. Technically, this operator can be implemented using a unique coloring. So, we can apply again
this technique on the obtained coloring.
Algorithm 34.1 presents this idea. In line 5, we complete the class without regarding the colorings C .
The uncolored vertices are shuffled to increase the diversity of the search.

Algorithm 34.1 IG Operator

Require: C current coloring
1: C is uncolored;
2: shuffle the class of C ;
3: for all class i of C do
 
4: C −1 (i) = C −1 (i) uncolored vertices of C ;
5: Complete C −1 (i) with uncolored vertices of C ;
6: end for
7: C ← C ;

34.4.2 Break Operator

This operator operates as opposed to the previous one. Instead of filling the classes we try to empty them.
For each class, we distribute its vertices into the others (nonempty) classes. In this case too, the number
of used colors cannot increase.
The Algorithm 34.2 explains how we proceed to break a coloring. We notice that we use the same order
to visit the classes of the coloring in line 1 and line 3.

34.4.3 Operators Efficiency

Both methods usually change the current coloring to a coloring with the same number of colors. To evaluate
their performance, we iterate them at most |V | times without improvement. Table 34.2 summarizes the
results, we obtain.
We notice that the number of colors needed by iteration of break operator are always less than those
obtained by IG. We can state that the break operator is better than IG when they are used alone.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 631 — #7

34-632 Handbook of Bioinspired Algorithms and Applications

Algorithm 34.2 Break Operator

Require: C the current coloring
1: shuffle the class of C ;
2: for all class i of C do
3: for all vertices s of C −1 (i) do
4: for all class j of C = i do
5: if s is colorable in j then
6: color s into j;
7: break;
8: end if
9: end for
10: end for
11: end for

TABLE 34.2 Operatory Efficiency

IG Break

Name min max avg std dev. min max avg std dev.
DSJC125.1 6 7 6.90 0.31 6 7 6.4 0.5
DSJC125.5 19 21 20.64 0.59 19 21 20.35 0.59
DSJC125.9 45 47 46.4 0.60 44 47 45.8 0.77
DSJC250.9 78 81 79.15 0.88 76 79 77.2 0.89
DSJC500.9 141 145 143.2 1.24 137 141 138.85 1.46
DSJR500.1c 85 85 85 0 85 85 85 0
le450_5c 6 8 6.8 0.52 5 7 6.40 0.6
le450_5d 5 8 6.7 0.66 5 7 6.05 0.69
queen10_10 12 14 12.95 0.39 12 13 12.95 0.22
queen11_11 14 15 14.1 0.31 14 15 14.5 0.51
queen8_12 13 14 13.05 0.22 13 13 13 0
queen9_9 12 12 12 0 11 12 11.8 0.41
school1 14 19 14.7 1.13 14 15 14.25 0.44
school1_nsh 14 18 15.6 1.05 14 16 14.6 0.6

34.5 A Cooperative Metaheuristic

Our goal is to provide cooperative heuristic coloring graphs. In this section, we present the implementation
we made of COSEARCH to MGCP.

34.5.1 Information to Memorize

In the adaptive memory, we store a set of N colorings, called elite. This elite present no double, for that
we use the techniques presented in Reference [13]. When a coloring is received, the new coloring replaces
the worst coloring (if the new coloring is better).
We also use four specific sets: one for each type of agents. More precisely, when a new coloring is obtained
from an agent a, we try to add it to the elite set. If this fails, we try to add it to the set corresponding to a.
The replacement strategy keeps only the best colorings in both cases.
The adaptive memory also records some additional information: for each pair of vertices, we notice the
last iteration where they were in the same class, and the last iteration they were placed in different classes.
Using this data, we can infer if some area of the search space is insufficiently explored.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 632 — #8

A Cooperative Parallel Metaheuristic 34-633

The adaptive memory takes decision from the following list:

• Intensify: The memory randomly picks a coloring in elite and sends it to an agent IG or a break
agent. Indeed those agents can easily modify a coloring and a better coloring can be reached by
several applications of those algorithms. We call such an agent an I-agent.
• Force a same class: The memory randomly picks one of the best colorings. It chooses the pair of
nonconnected vertices that are not colored by the same color for a long time. The memory forces
both vertices to be in the same class. This may increase the number of colors. Finally, the new
coloring is sent to a break agent. The agent respects always that both vertices are in the same class.
It behaves as both chosen vertices form a new one. The set of neighbors of the new vertex is the
union of the set of neighbors of the old ones. We call such an agent an FS-agent.
• Force different classes: The memory picks randomly one of the best colorings. It chooses the pair
of nonconnected vertices that are always colored by the same color for a long time. The memory
forces both vertices to be in different classes. This usually increases the number of colors. Finally,
the new coloring is sent to a break agent. The agent respects always that the two vertices are in
different classes. It behaves as if there is an additional edge between the two chosen vertices. We
call such an agent an FD-agent.
• Greedy crossover: Memory extracts from two colorings the commun part. This partial coloring is
sent to an agent to be completed. This completion is made in a greedy manner. Then, the break
operator is applied few times. Several compressions are tested before returning the best one. We
call such an agent an Xρ -agent.

34.5.2 Partial Coloring Building

To construct a partial coloring C representing the commun part of two colorings, we use the refining
matrix defined as below.
For two colorings A and B, and their associated refining matrix MAB , we build a mapping
AB .
τ : {1, . . . , χ(A)} → {1, . . . , χ(B)} defined by τ (i) = arg max j mi,j
With τ , we build C as follows:

∀j ∈ {1, . . . , χ(B)} C (j) = A(i). (34.2)
i∈τ −1 (j)

Typically, this partial coloring C has few nonempty classes than B.

We remark also that such a scheme presents no constraints violations as soon as the coloring B is proper.
Indeed classes of C are part of classes of B.

34.5.3 Parallel Implementation

Our algorithm is implemented using the C-PVM environment [32]. For the synchronization of the
different tasks, we use three (kind of) automata: one to control the distributed agents (see Figure 34.3),
one to control a manager of a fleet of several agents of the same type2 (see Figure 34.4) and the last one
to control the adaptive memory (see Figure 34.5).
A fleet manager works as an asynchronous gate. A new job is provided with workers because of the
manager since one of its agents is available.
The automaton can be in one of the three main states:

• Every worker is working, so the adaptive memory waits for one result.
• There are some computational available resources, so a policy is chosen with regard to the
availabilities.

2 Scheme, break, force different or force same.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 633 — #9

34-634 Handbook of Bioinspired Algorithms and Applications

Initialisation

end Signal
Wait for work End

work
new
Work

Return result

FIGURE 34.3 Automaton for the searching agents.

Initialisation

end Signal
Receipt End
ava
ilab
work
re sult new ility?

New result New work Answer for availability

FIGURE 34.4 Automaton for the fleet managers.

Initialisation

Ask of availabilities

Receipt
ble as
ila yn
available

va ch
agent

na ron
u ou
ent sr
ag es
all ult
Wait result Choose policy New result

Intensification policy
Treat result Force same class policy
Force different class policy
Greedy crossover policy
End

FIGURE 34.5 Automaton for the adaptive memory.

• A new result is asynchronously obtained, so the memory incorporates the result before any choice
of a policy.

As suggested in Section 34.2.3, the automaton of the adaptive memory and the agents communicate
with the fleet managers. A fleet manager corresponds to the Input/Output interface of the memory.
The end of the algorithm is decided by the adaptive memory. The end is chosen as a policy but it
happens after a fixed number of iterations. The other policies of the memory is randomly picked from
the available agents with fixed biases. For example, we may apply 60% of I-agent, 10% of FD-agent,

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 634 — #10

A Cooperative Parallel Metaheuristic 34-635

10% of FS-agent, and 10% of Xρ -agent. At a given moment, if all intensification agents are working then
the policy cannot be an intensification.

34.6 Experiments
Our algorithm is composed of many components. It is possible that some presented agents do not
contribute well to the global quality of the search. So, we have to tune the use of each agent. This is
complicated by the fact that all agents are based on the break operator that is very efficient due to its
capacity to diversify. Thus, we cannot detect if we have to continue to use I-agent or operate some radical
changes.
In a first time, we evaluate each agent with the I-agent. Finally, we present results of the global
architecture.

34.6.1 Preliminary Tests

For our experiments, we use the following parameters:

• Number of iterations of the global cycle: At each iteration, adaptive memory collects a new coloring
and chooses a policy. We bound experimentally this parameter to 500.
• Number of iterations for each type of agent : The number of application of the break operator by the
I-agents, FS-agents and FD-agents is equal to the number of vertices. The Xρ agents recombine
two colorings and halve the number of vertices applications of break operator on each obtained
colorings.
• Number of available agent for each type of agent : We use ten I-agents, four FS-agents, four FD-agents,
and four Xρ -agents. For those tests, we use six diversification agents (FS, FD, Xρ ) in parallel with
ten intensification agents (break).

Table 34.3 summarizes the obtained results. For each combination, we made four runs on different
hardware platforms using different architectures: Linux PCs (between 2.7 and 3 Ghz , 512 Mo and 1 Go
RAM) and an IBM SP3 (with 16 processors Power3 NH2 per node running at 375 Mhz, 16 Go RAM).
We can see that the combination allows us to go further in the search. Indeed, the qualities of the overall
solutions are better for any combination than for the break operator.
Incidentally, we observe that the different combinations perform more or less well on different graphs.
I+FS provides the best results on DSJC250.9, I+FD provides the best results on Queen9_9 and I+X always
found the same quality of results for Queen8_12.

34.6.2 Global Architecture

To profit from different diversification mechanism, place into concurrence all the agents. Results using
four instances of each diversification agents and ten intensification agents are gathered in Table 34.3.
We observe that the complete combination provides results at least as good as the worst combination
previously seen. But for some graphs (DSJC125.9 and DSJC250.9) results are not abreast of the best
combination. Nevertheless, for one graph (DSJC125.5) interactions between components provide the
best colorings than those seen before (Table 34.4).

34.7 Conclusion and Future Works

In this chapter, we present a cooperative metaheuristic called COSEARCH makes an explicit balancing
of intensification and diversification in the search space. This metaheuristic is intrinsically parallel. As
the model is very general, it can be implemented in many ways. We propose a manner to implement
it in parallel. This implementation is hierarchical and is based on three levels. At the top the adaptive

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 635 — #11

34-636 Handbook of Bioinspired Algorithms and Applications

TABLE 34.3 Summarize Experiments with one Diversification Agent and the Break Agent

I+FS I+FD I+X

Name min max avg std dev. min max avg std dev. min max avg std dev.

DSJC125.1 6 6 6 0 6 6 6 0 6 6 6 0
DSJC125.5 19 19 19 0 19 19 19 0 19 19 19 0
DSJC125.9 44 44 44 0 44 44 44 0 44 44 44 0
DSJC250.9 73 75 74 0.81 74 74 74 0 74 74 74 0
DSJC500.9 131 133 132 0.81 131 132 131.75 0.5 133 134 133.5 0.57
DSJR500.1c 85 85 85 0 85 85 85 0 85 85 85 0
le450_5c 5 5 5 0 5 5 5 0 5 5 5 0
le450_5d 5 5 5 5 5 5 5 0 5 5 5 0
queen10_10 12 12 12 0 12 12 12 0 12 12 12 0
queen11_11 13 14 13.5 0.57 13 14 13.75 0.5 14 14 14 0
queen8_12 12 13 12.25 0.5 12 13 12.5 0.57 12 12 12 0
queen9_9 11 11 11 0 10 11 10.75 0.5 11 11 11 0
school1 14 14 14 0 14 14 14 0 14 14 14 0
school1_nsh 14 14 14 0 14 14 14 0 14 14 14 0

Note: We use the following notation:

• I+FS for combination of break agent and FS-agent
• I+FD for combination of break agent and FD-agent
• I+X for combination of break agent and Xρ -agent

TABLE 34.4 Used Benchmarks Presentation

I+FS+FD+X

Name min max avg std dev.

DSJC125.1 6 6 6 0
DSJC125.5 18 19 18.6 0.54
DSJC125.9 44 44 44 0
DSJC250.9 74 75 74.4 0.54
DSJC500.9 132 133 132.4 0.54
DSJR500.1c 85 85 85 0
le450_5c 5 5 5 0
le450_5d 5 5 5 0
queen10_10 12 12 12 0
queen11_11 13 14 13.25 0.5
queen8_12 12 13 12.25 0.5
queen9_9 11 11 11 0
school1 14 14 14 0
school1_nsh 14 14 14 0

memory summarizes information and decides the direction of the search: intensification or a type of
diversification. Below the memory, specific agents allow to manage an asynchronous fleet of agents of the
same type. At the lowest level, the workers called agents apply strategies dedicated to the problem.
We apply such a strategy to the graph coloring problem. For that, we develop dedicated operators:
Break operator and Xρ . We propose several ways to diversify the search forcing the presence of characters
in colorings used by searching agents.
Combinations of components was tested. We saw that that diversification leads to obtain better results
but no diversification worth better than all others.
The combination of all diversification techniques improves results on a graph but pays back its
performance on two other graphs. This may be due to the too heavy diversification made in this case.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 636 — #12

A Cooperative Parallel Metaheuristic 34-637

We are now working on new search agents as agents working with k-colorings and a way to tune to each
diversification technique.

References
[1] Jan Paredis. Coevolutionary computation. Artificial Life, 2: 355–375, 1995.
[2] V. Bachelet. Métaheuristique parallèle hybride: application au problème d’affectation quadratique.
Ph.D. thesis, Université des sciences et technologies de Lille, cité scientifique Villeneuve d’Ascq
59655, December 1999.
[3] B. Weinberg, V. Bachelet, and E.-G. Talbi. A co-evolutionnist meta-heuristic for the assignment of
the frequencies in cellular networks. In First European workshop on Evolutionary Computation in
Combinatorial Optimization (EvoCOP), Como, Italy, 2001. Springer-Verlag, pp. 140–149.
[4] S. Kirkpatrick, D.C. Gelatt, and M.P. Vecchi. Optimization by simulated annealing. Science, 220:
671–680, 1983.
[5] F. Glover and M. Laguna. Tabu search. In C. Reeves, Ed., Modern Heuristic Techniques for
Combinatorial Problems, Blackwell Scientific Publishing, Oxford, England, 1993.
[6] T. Feo and M. Resende. Greedy randomized adaptive search procedures. Journal of Global
Optimization, 6: 109–133, 1995.
[7] G.J. Chaitin, M. Auslander, A.K. Chandra, J. Cocke, M.E. Hopkins, and P. Markstein. Computer
Languages, chapter Register allocation via graph coloring, IBM. T.J. Watson Research Center, 1981,
pp. 47–57.
[8] A. Gamst. Some lower bounds for a class of frequency assignment problems. IEEE Transactions of
Vehicular Technology, 1(35): 8–14, 1986.
[9] S. Miner, S. Elmohamed, and H. Yau. Optimizing timetabling solutions using graph coloring. In
NPAC REU Program, NPAC, Syracuse University, NY, 1995.
[10] H. Ogawa. Labeled point pattern matching by delaunay triangulation and maximal cliques. In
Pattern Recognition, number 1 in 19, 1986, pp. 35–40.
[11] N. Barnier and P. Brisset. Coloriage de graphe en programmation par contraintes. In ROADEF
2003, ROADéF, February 2003, pp. 348–349.
[12] R.M. Karp. Complexity of Computer Computations, chapter Reducibility among combinatorial
problems, Plenm Press, New York, 1972, pp. 85–103.
[13] B. Weinberg and E.-G. Talbi. On symmetry of partitionning problems. In J. Gottlieb and G. Raidl
Eds., EvoCOP, LNCS, 2004. To appear.
[14] D. Brelaz. New methods to color the vertices of a graph. Communications of the ACM, 22: 251–256,
1979.
[15] F.T. Leighton. A graph coulouring algorithm for large scheduling problems. J. Res Natl Bur.
Standards, 84: 489–506, 1979.
[16] D.S. Johnson, C.R. Aragon, L.A. McGeoch, and C. Schevon. Optimization by simulated annealing:
An experimental evaluation; part II, graph coloring and number partitioning. Operations Research,
39: 378–406, 1991.
[17] J. Culberson. Iterated Greedy Graph Coloring and the Difficulty Landscape. Technical report,
University of Alberta, June 1992.
[18] S. Hurley, D. Smith, and C. Valenzuela. A permutation based genetic algorithm for minimum span
frequency assignment. In T. Baeck, A. Eiben, M. Schoenauer, and H. Schwefel, Eds, PPSN V: Pro-
ceedings of the Fifth International Conference on Parallel Problem Solving from Nature, Vol. 1498 of
Lecture Notes in Computer Science, Amsterdam, The Netherlands, September 1998. Springer-Verlag
Publication, pp. 907–916.
[19] D. Costa and A. Hertz. Ants can colour graphs. Journal of the Operational Research Society, 48:
295–305, 1997.
[20] C.A. Glass and A. Prügel-Bennett. A Polynomially Searchable Exponential Neighbourhood for
Graph Colouring. Technical report, Departement of Electronics and Computer Science, University
of Southampton, 1998.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 637 — #13

34-638 Handbook of Bioinspired Algorithms and Applications

[21] A. Hertz and D. de Werra. Using tabu search techniques for graph coloring. Computing, 39:
345–351, 1987.
[22] R. Dorne. Étude des méthodes heuristiques pour la coloration, la T-coloration et l’affectation de
fréquence. Ph.D. thesis, Université de Montpellier II Science et Technique, May 1998.
[23] P. Galinier and J-K. Hao. Hybrid evolutionary algorithms for graph coloring. Journal of
Combinatorial Optimization, 3: 379–397, 1999.
[24] L. Paquete and T. Stützle. An experimental investigation of iterated local search for coloring graphs.
In S. Cagnoni, J. Gottlieb, E. Hart, M. Middendorf, and G. Raidl, Eds, Applications of Evolutionary
Computing, Proceedings of Evo Workshops2002: EvoCOP, EvoIASP, EvoSTim, Vol. 2279, Kinsale,
Ireland, 3–4 Springer-Verlag, 2002, pp. 121–130.
[25] C. Avanthay, A. Hertz, and N. Zufferey. A variable neighborhood search for graph coloring.
European Journal of Operational Research, 151: 379–388, 2003.
[26] A. Vesel and J. Zerovnik. How good can ants color graphs? Journal of computing and Information
Technology, 8: 131–136, 2000.
[27] A. Petford and D. Welsh. A randomised 3-colouring algorithm. Discrete Mathematics, 74: 253–261,
1989.
[28] J-P. Hamiez and J-K. Hao. Scatter search for graph coloring. In Artificial Evolution, Le Creusot,
France, October 2001, pp. 267–278.
[29] D. Fotakis, S. Likothanassis, and S. Stefanakos. An evolutionary annealing approach to graph
coloring. In Proceedings of Applications of Evolutionary Computing, Vol. 2037 of Lecture Notes in
Computer Science, Evo Workshops 2001, Springer-Verlag, April 2001, pp. 120–129.
[30] C. Walshaw and M.G. Everett. Multilevel Landscapes in Combinatorial Optimisation. Technical
Report 02/IM/93, Comp. Math. Sci., Univ. Greenwich, London SE10 9LS, UK, April 2002.
[31] M. Clerc. Optimisation par essaim particulaire et coloriage de graphe. Technical report, France
Télécom, 2001.
[32] A. Geist, A. Beguelin, J. Dongarra, W. Jiang, R. Manchek, and V. Sunderam. PVM: Parallel
Virtual Machine, A Users’ Guide and a Tutorial Networked Parallel Computing. The MIT Press,
Cambridge, MA, 1994.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C034” — 2005/8/6 — 14:14 — page 638 — #14

 35
Frameworks for the
Design of Reusable
Parallel and
Distributed
Metaheuristics

35.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35-639

35.2 Parallel and Distributed Local Searches (PDLS) . . . . . . .35-640
Principles of LS • Parallel and Distributed Models of LS
35.3 Parallel and Distributed Evolutionary Algorithms
(PDEAs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35-642
35.4 Hybrid Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35-642
35.5 Frameworks for the Design of PDM . . . . . . . . . . . . . . . . . . .35-643
Why Using Frameworks? • Design Requirements and
N. Melab Objectives • An Overview of Existing Frameworks
El-Ghazali Talbi 35.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35-646
S. Cahon References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35-646

35.1 Introduction
Real-world optimization problems are often NP-hard, complex, and CPU time-consuming. Moreover,
their modeling evolves continuously in terms of constraints and objectives. Therefore, their resolution
requires the use of parallel/distributed hybrid metaheuristics. Unlike exact methods, metaheuristics allow
to find sub-optimal solutions in a reasonable execution time. They allow to meet the resolution delays
often imposed in the industrial field.
Metaheuristics fall in two categories: single solution-oriented or local search (LS) methods, and
population-based or evolutionary algorithms (EAs). An LS starts with a single initial solution. At each step
of the search the current solution is replaced by another (often the best) solution found in its neighborhood.

This work is a part of the current national joint grid computing project ACI-GRID DOC-G (Défis en Optimisation
Combinatoire sur Grilles). It includes research teams from different laboratories: OPAC from LIFL, OPALE from
PRISM and O2 and P3-ID from IMAG. The project is supported by the French government.

35-639

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 639 — #1

35-640 Handbook of Bioinspired Algorithms and Applications

Very often, LS methods allow to find a local optimal solution, and so are called exploitation-oriented
methods. On the other hand, evolutionary algorithms work on a randomly generated population of solu-
tions. The initial population is enhanced through a natural evolution process. At each generation of the
process, the whole or a part of the population is replaced by the newly generated individuals (often the
best ones). EAs are often called exploration-oriented methods.
Although their time complexity is polynomial, metaheuristics remain insufficient for large-size prob-
lems. Therefore, parallel/distributed and concurrency tools are necessary to tackle these problems.
Different parallel/distributed models have been proposed for each class of methods. They are detailed
in Sections 35.2 and 35.3. In order to take benefit from the exploitation power of the LS methods and the
exploration merit of EAs, their hybridization is recommended [1]. Hybrid metaheuristics allow to deliver
high quality and robust solutions.
Several parallel and distributed metaheuristics and their implementations have been proposed in the
literature. Most of them are available on the Internet, and can be reused and adapted to its own problems.
Reusability may be defined as the ability of software components to build many different applications [2].
However, one has to rewrite the problem-specific sections of the code. Such task is tedious, error prone,
energy and time-consuming. Moreover, the new developed code is harder to maintain. A better way to
reuse the code of existing parallel and distributed metaheuristics is the use of libraries [3]. Their objective
is twofold: they are reliable as they are often well tested and documented. In addition, they allow a better
maintainability and efficiency. However, libraries do not allow the reuse of design. A better approach to
reuse the design and the code at the same time is the framework-based reuse [4].
In the literature, very often the authors do not make a clear of difference between a library and a
framework. In a framework, the provided code calls the user defined one according to the Hollywood
property: “do not call us, we call you.” Therefore, frameworks provide the full control structure of
the invariant part of the algorithms, and the user has to only supply the problem-specific details. This
chapter focuses on the frameworks, and aims at removing the ambiguity on their use, highlighting their
characteristics, requirements, and objectives.
Most of existing frameworks related to metaheuristics for discrete optimization problems are Object-
Oriented (OO) [5–14]. They include a set of classes that embody an abstract design of solution methods
of a family of related problems [2]. They are based on a strong conceptual separation of the invariant
(generic) part of PDM and their problem-specific part. Such characteristic allows the PDM programmer
to redo very little code.
The frameworks focus only on either EA [5–9] or LS [10,11]. Only few frameworks are dedicated
on the design of both EA and LS, and their hybridization [12–14]. All these frameworks are described,
summarized, and compared in this chapter. The comparison is mainly based on the class of provided
solution methods, the parallel/distributed models they implement, the hybridization mechanisms they
allow, and some implementation choices, mainly the programming language and the communication and
concurrency API. The presented overview will help the user to choose the framework corresponding to
his/her needs. At our best known, such overview has never been proposed in the literature.
The rest of the chapter is organized as follows: Sections 35.2 and 35.3 present the working principles
and their major parallel/distributed models of, respectively, LS methods and EA. In Section 35.4, we
present the main hybrid mechanisms of metaheuristics. In Section 35.5, we propose an overview of the
major frameworks dedicated to the LS methods, to the EA, and both of them. The main characterist-
ics of each of these frameworks are summarized. Section 35.6 ends the chapter with some concluding
remarks.

35.2 Parallel and Distributed Local Searches (PDLS)

In this section we first present the main existing LS methods and their working principles. Afterward, we
describe the main existing parallel/distributed models of their design and implementation.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 640 — #2

Design of Reusable Parallel and Distributed Metaheuristics 35-641

35.2.1 Principles of LS
Local search are metaheuristics dedicated to the improvement of a single solution. They are generally
based on the concept of the neighborhood. They start from a solution randomly generated or provided by
another metaheuristic. This solution is then updated, systematically, by replacing the current solution by
another found in the neighborhood. The specific features of LS are mainly: the heuristic internal memory,
the strategy used to choose the initial solution, the generator of candidate solutions, and the selection
policy of the candidate moves. Three major LS stand out: Hill Climbing (HC) [15], Simulated Annealing
(SA) [16], and Tabu Search (TS) [17].
A serial LS is composed of generic and specific features. Generic features include the initializing of a
given movement, the exploration strategy of the neighborhood, and the computation of the fitness value
of a given solution corresponding to a given movement. Specific features, such as the Tabu list involved in
the TS method, allow to differentiate LS.

35.2.2 Parallel and Distributed Models of LS

There exist mainly three parallel distributed models of LS: the parallel/distributed exploration of the
neighboring model, the multi-start model, and the parallel/distributed evaluation of each solution. The
first two models are illustrated in Figure 35.1.
• Parallel/distributed exploration of the neighboring: This model is a kind of Farmer–Worker model.
At the beginning of each iteration, the farmer copies the current solution on the workers. Each
worker explores some neighboring candidates, and returns the results to the farmer. The model
could be efficient if the evaluation of each solution is time consuming and there is a large number
of candidate neighbors to evaluate. Such a model is not well suited to SA since only one candidate
solution is evaluated at each iteration. On the other hand, the efficiency of the model for HC is
not always guaranteed, as the number of neighboring solutions to process before finding one that
improves the current objective function may be highly variable.
• Multi-start model: The model consists of launching in parallel several independent
homo/heterogeneous LS. It allows to enhance the robustness of the execution. On the other hand,
the model is well suited for low-speed networks of workstations. Moreover, its exploitation is
natural and easy for the user.
• Parallel/distributed evaluation of a single solution: The fitness of each solution is evaluated in a
parallel centralized way. That model could be efficient when the evaluation function is CPU time

Parallel and independent

walks starting from
different solutions

Copy

Parallel
exploration
of the neighborhood
An iteration

FIGURE 35.1 The first two parallel models for LS.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 641 — #3

35-642 Handbook of Bioinspired Algorithms and Applications

Aggregation
E.A. E.A. of fitness
Solutions values
E.A. Solutions

Fitness Partial
values fitnesses
Migrations

Full evaluating
nodes Partial evaluating
nodes

FIGURE 35.2 Three major parallel distributed models for EA.

consuming and IO intensive. The function could be viewed as an aggregation of a set of partial
functions. A reduction operation is performed on the results returned by the partial functions.
Consequently, for this model the user has to indicate a set of partial functions and an aggregation
operator of these.

35.3 Parallel and Distributed Evolutionary Algorithms (PDEAs)

The EAs [18] are population-based metaheuristics. They improve an initial randomly generated popu-
lation of solutions during a certain number of generations. At each generation, individuals are selected,
paired, and recombined in order to generate new solutions that replace others. The major characteristics of
EA are the way the population is initialized, the selection strategy (deterministic/stochastic) by fostering
“good” solutions, the replacement strategy that withdrawn individuals, and the continuation/stopping
criterion, evaluated at the end of each generation.
Three major parallel and distributed models for EA stand out in the literature (Figure 35.2): the
island (a)synchronous cooperative model, the parallel/distributed evaluation of the population model,
and the parallel/distributed evaluation of a single solution model. The last model is the one presented in
Section 35.2.2.
In the Island model, several homo/heterogeneous EA run simultaneously and cooperate to compute
better and robust solutions. They exchange genetic stuff to improve the diversity of the search. The model
aims at delaying the global convergence, especially when the EA are heterogeneous regarding to the
variation operators. The migration of individuals follow a policy defined by the following parameters: the
migration decision criterion, the exchange topology, the number of emigrants, the emigrants selection
policy, and the replacement/integration policy.
The evaluation of the population is necessary because it is, in general, the most time-consuming step of
an EA. The parallel evaluation is farmer/worker oriented. The farmer applies the selection, transformation,
and replacement operations as they require a global management of the population. At each generation,
it distributes the new solutions among the workers, which evaluate them and return their fitness values.
An efficient execution is often obtained particularly when the evaluation of each solution is costly.

35.4 Hybrid Metaheuristics

Hybridization allows to combine different metaheuristics and make them cooperate in order to solve the
same optimization problem. Nowadays, hybrid metaheuristics have gained great interest [1]. They allow
to solve many practical and academic problems, and to provide the best-known solutions. In Reference 1,
two levels (low and high) and two modes (relay and cooperative) of hybridization have been distinguished
(Figure 35.3).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 642 — #4

Design of Reusable Parallel and Distributed Metaheuristics 35-643

Hybrid metaheuristic

High level Low level

Coevolution Relay

FIGURE 35.3 Hierarchical taxonomy of hybrid metaheuristics.

The low-level hybridization consists in changing the internal component of the metaheuristic on which
it is performed. A given function of a given metaheuristic is replaced by another metaheuristic. For
instance, the mutation operator of a given GA could be replaced by an LS method. In order to make easier
this kind of hybridization the semantics of the internal function must be the same as the metaheuristic. The
low-level hybridization requires to examine the internal working of the metaheuristic [14]. Conversely,
in high-level hybrid algorithms the combined metaheuristics are self-containing, meaning no direct
relationship to their internal working is considered.
On the other hand, in the relay hybridization mode the metaheuristics are applied in a pipeline way.
The output of a given metaheuristic (except the last) is the input of its successor (except the first) in
the pipeline. At the contrary, coevolutionist hybridization is a cooperative optimization model. Each
metaheuristic performs a search in a solution space, and exchange good solutions with others.

35.5 Frameworks for the Design of PDM

In this section we first explain why it is important to use frameworks for the reusable design of PDM, their
design requirements and objectives, and an overview and a comparison of existing frameworks.

35.5.1 Why Using Frameworks?

There are mainly three ways to reuse existing PDM: from scratch or no reuse, only code reuse, and both
design and code reuse. The from scratch-oriented approach is attractive due to the apparent simplicity of
the metaheuristics code. Indeed, some programmers are tempted to develop themselves their code rather
than reusing a code developed by a third party. However, several problems arise: the coding effort is time
and energy consuming, error prone, and difficult to maintain, etc. Only code reuse is the reuse of the third-
party code available on the Internet either as free single programs or as libraries. Old third-party code
has usually application-dependent sections that must be extracted before the new application-dependent
code can be inserted. Changing these sections is often time consuming and error prone. The code reuse
through libraries [3] is obviously better because these are often well tried, tested, and documented, thus,
more reliable. However, libraries allow to reuse the code but they do not make easier the reuse of the
complete invariant part of algorithms, particularly their whole design.
The both code and design reuse approach is devoted to overcome this problem, that is, to redo as little
code as possible each time a new optimization problem is dealt with. The basic idea is to capture into
special components the recurring (or invariant) part of solution methods to standard problems belonging
to a specific domain. Frameworks provide the full control structure of the invariant part of the algorithms,
and the user only has to supply the problem-specific details. In order to meet this property the design
of a framework must be based on a clear conceptual separation between the solution methods and the
problems they tackle.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 643 — #5

35-644 Handbook of Bioinspired Algorithms and Applications

35.5.2 Design Requirements and Objectives

A framework is normally intended to be exploited by as many users as possible. Therefore, its exploitation
could be successful only if some important user criteria are satisfied. The following criteria are the major
of them:

• Maximum design and code reuse: The framework must provide for the user a whole architecture
(design) of his/her solution method. Moreover, the programmer may redo as little code as possible.
This objective requires a clear and maximum conceptual separation between the solution methods
and the problems to be solved, and thus a deep problem domain analysis. The user might, therefore,
develop only the minimal problem-specific code.
• Utility and extendibility: The framework must allow the user to cover a broad range of
metaheuristics, problems, parallel distributed models, hybridization mechanisms, etc. It must
be possible for the user to easily add new features/metaheuristics or change existing ones
without implicating other components. Furthermore, as in practice existing problems evolve
and new others arise these have to be tackled by specializing/adapting to them the framework
components.
• Transparent use of parallel/distributed models and hybridization mechanisms: In order to facilitate
its use it is implemented so that the user can deploy his/her parallel algorithms in a transparent
manner. Moreover, the execution of the algorithms must be robust to guarantee the reliability
and the quality of the results. The hybridization mechanism allows to obtain robust and better
solutions.
• Portability: In order to satisfy a large number of users the framework must support different
material architectures and their associated operating systems.

35.5.3 An Overview of Existing Frameworks

It is quoted that a framework is mainly based on the conceptual separation between the solution methods
and the problems they deal with. This separation requires a deep understanding of the application domain.
In Reference 19, OO composition rather than inheritance is recommended to perform the conceptual
separation. Indeed, classes are easier to reuse than individual methods. The modeling of the application
domain results in a set of reusable classes with some constant and variable aspects. The implementation
of the constant part is provided by the framework, and encapsulated into generic/abstract classes or
skeletons [12]. The variable part is problem specific, and is specified and fixed in the framework, but it
is implemented by the user. This part is a set of holes or hot spots [20] that serve to fill the skeletons
provided by the framework when building specific applications.
According to the extendibility requirement, two types of frameworks can be distinguished: white- or
glass-box frameworks and black-box frameworks. In black-box frameworks one can reuse components
by plugging them together through static parameterization and composition, unaware of their internal
working [4]. In contrast, white-box frameworks require an understanding of the internal working of
the classes so that correct subclasses (inheritance-based) can be developed. Therefore, they allow more
extendibility of the provided classes.
Several white-box frameworks for the reusable design of PDM are available on the Internet. This
table summarizes the most important frameworks. The different frameworks are classified according to
some criteria: the classes of methods they are dedicated to (LS or EA), the kind of hybridization mech-
anisms they provide, the parallel/distributed models they implement, the language they are developed
with, and the API used for communication and concurrency. These criteria allow to evaluate the differ-
ent frameworks regarding the quality and design requirements and objectives of frameworks quoted in
Section 35.5.2.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 644 — #6

Design of Reusable Parallel and Distributed Metaheuristics 35-645

LS EA Hybrid Parallel/Distributed Language Communication/

Concurrency
ECJ — + EA/EA Island Java Java threads
TCP/IP
Sockets
D. Beagle — + EA/EA M/S C++ TCP/IP
Sockets
Jdeal — + EA/EA M/S Java TCP/IP
Sockets
Easylocal ++ + — LS/LS — C++ —
Localizer ++ + — LS/LS — C++ —
MAFRA — + EA/EA — Java —
DREAM — + EA/EA Island Java Java threads
TCP/IP
Sockets
MALLBA + + EA/LS all C++ Netstream
EA/EA MPI
LS/LS
ParadisEO + + EA/LS all C++ MPI
LS/LS PVM
EA/EA PThreads

According to the first two criteria, the frameworks fall in three categories that distinguish those,
respectively, dedicated to only EA, only LS, and both of them. Some of the frameworks limited to only
EA are the following: DREAM1 [5], ECJ2 [6], JDEAL3 [7], and Distributed BEAGLE4 [9]. These software
are frameworks as they are based on a clear OO conceptual separation. They are portable as they are
developed in the Java language except the last system, which is programmed in C++. However, they are
limited regarding the parallel distributed models. Indeed, in DREAM and ECJ only the island model is
implemented using Java threads and TCP/IP sockets. DREAM is particularly deployable on peer-to-peer
platforms. Furthermore, JDEAL and Distributed BEAGLE provide only the Master–Slave (M/S) model
using TCP/IP sockets. The latter also implements the synchronous migration-based island model, but
deployable on only one processor.
In the LS domain, most of existing frameworks [10,11] do not allow parallel distributed implement-
ations. Those enabling parallelism/distribution are often dedicated to only one solution method. For
instance, Reference 21 provides parallel skeletons for the TS method. Two skeletons are provided and
implemented in C++/MPI: independent runs (multi-start) model with search strategies, and a Master–
Slave model with neighborhood partition. The two models can be exploited by the user in a transparent
way.
In practice, only few frameworks available on the Internet are devoted to both PDEA and PDLS,
and their hybridization. MALLBA5 [12], MAFRA6 [13], and ParadisEO7 [14] are good examples of
such frameworks. MAFRA is developed in Java using design patterns [22]. It is strongly hybridization
oriented; however, it is very limited regarding parallelism and distribution. MALLBA and ParadisEO have
numerous common characteristics. They are C++/MPI open source frameworks. They provide all the

1 Distributed Resource Evolutionary Algorithm Machine: http://www.world-wide-dream.org

2 Java Evolutionary Computation: http://www.cs.umd.edu/projects/plus/ec/ecj/
3 Java Distributed Evolutionary Algorithms Library: http://laseeb.isr.ist.utl.pt/sw/jdeal/
4 Distributed Beagle Engine Advanced Genetic Learning Environment: http://www.gel.ulaval.ca/∼ beagle
5 MAlaga+La Laguna+BArcelona: http://neo.lcc.uma.es/mallba/mallba.html
6 Java Mimetic Algorithms Framework: http://www.cs.nott.ac.uk/∼nxk/MAFRA/MAFRA.html
7 Parallel and distributed Evolving Objects: http://www.lifl.fr/∼cahon/paradisEO

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 645 — #7

35-646 Handbook of Bioinspired Algorithms and Applications

previously presented parallel/distributed models, and the different hybridization mechanisms. However,
they are quite different as ParadisEO seems to be more flexible because the granularity of its classes is
finer. Moreover, ParadisEO (an extended EO [23]) provides also the PVM-based communication layer
and PThreads-based concurrency. On the other hand, MALLBA is deployable on wide area networks [12].
Communications are based on NetStream, an ad hoc flexible and OOP message passing service upon MPI.
Furthermore, MALLBA allows the cooperation between metaheuristics and exact methods.

35.6 Concluding Remarks

In this chapter, we presented the different commonly used parallel/distributed models and hybridization
mechanisms of metaheuristics. Most of these models and mechanisms have been implemented and
are available on the Internet. We believe that the best way to reuse them is to encapsulate them into
frameworks. Frameworks provide the invariant part of the metaheuristics; therefore the user has to just
supply the specific part of his/her problem. We have also proposed an overview of existing white-box
frameworks, including those dedicated either only to LS, only to EA, or to both.
The different frameworks have been compared according to some objectives: the code and design reuse,
their utility and extendibility, the transparent use of parallelism/distribution and hybridization, and the
portability. These objectives are evaluated according to the following criteria: the families of methods the
frameworks are dedicated to, the parallel/distributed models and hybridization mechanisms they provide
and to be exploited transparently by the user, the language and communication and concurrency API they
are developed with.
All the presented frameworks are object oriented, and are based on a clear conceptual separation between
solution methods and problems. Therefore, they allow a maximum code and design reuse. Moreover,
they allow a portable deployment as they are developed in either C++ or Java. In the literature, only
few frameworks provide the major parallel/distributed models. ECJ and DREAM implement the island
model for cooperative EA, and Distributed Beable and Jdeal provide the Farmer/Worker model. Finally,
ParadisEO and MALLBA are particularly outstanding. Indeed, they allow to achieve all the objectives
quoted above. They include LSs as well as EAs. They provide all the major parallel/distributed models and
the hybridization mechanisms.

References
[1] E.-G. Talbi. A taxonomy of hybrid metaheuristics. Journal of Heuristics, 8: 541–564, 2002.
[2] A. Fink, S. Vo, and D. Woodruff. Building reusable software components for heuristc search.
In P. Kall and H.-J. Luthi (Eds.), Operations Research Proc., 1998, Springer-verlag, Berlin, 1999,
pp. 210–219.
[3] M. Wall. GAlib: A C++ library of genetic algorithm components. http://lancet.mit.edu/ga/.
[4] R. Johnson and B. Foote. Designing reusable classes. Journal of Object-Oriented Programming,
1: 22–35, 1988.
[5] M.G. Arenas, P. Collet, A.E. Eiben, M. Jelasity, J.J. Merelo, B. Paechter, M. Preuß and M. Schoenauer.
A framework for distributed evolutionary algorithms. In Proceedings of PPSN VII, September 2002.
[6] S. Luke, L. Panait, J. Bassett, R. Hubley, C. Balan, and A. Chircop. ECJ: a Java-based
evolutionary computation and genetic programming research system. http://www.cs.umd.edu/
projects/plus/ec/ecj/.
[7] J. Costa, N. Lopes, and P. Silva. JDEAL: The Java Distributed Evolutionary Algorithms Library.
http://laseeb.isr.ist.utl.pt/sw/jdeal/home.html.
[8] E. Goodman. An Introduction to GALOPPS — The “Genetic Algorithm Optimized for Portability
and Parallelism” System. Technical report, Intelligent Systems Laboratory and Case Center for
Computer-Aided Engineering and Manufacturing, Michigan State University, November 1994.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 646 — #8

Design of Reusable Parallel and Distributed Metaheuristics 35-647

[9] C. Gagné, M. Parizeau, and M. Dubreuil. Distributed BEAGLE: An environment for parallel and
distributed evolutionary computations. In Proceedings of the 17th Annual International Symposium
on High Performance Computing Systems and Applications (HPCS) 2003, May 11–14, 2003.
[10] L. Di Gaspero and A. Schaerf. Easylocal++: An object-oriented framework for the design of local
search algorithms and metaheuristics. In MIC ’2001 4th Metaheuristics International Conference,
Porto, Portugal, July 2001, pp. 287–292.
[11] L. Michel and P. Van Hentenryck. Localizer++: An Open Library for Local Search. Technical report
CS-01-02, Brown University, Computer Science, 2001.
[12] E. Alba and the MALLBA Group. MALLBA: A library of skeletons for combinatorial optimization.
In R.F.B. Monien, Ed., Proceedings of the Euro-Par, Vol. 2400 of Lecture Notes in Computer Science
Paderborn, Springer-Verlag, Heidelberg, 2002, pp. 927–932.
[13] N. Krasnogor and J. Smith. MAFRA: A java memetic algorithms framework. In Alex A. Freitas,
William Hart, Natalio Krasnogor, and Jim Smith, Eds, Data Mining with Evolutionary Algorithms,
Las Vegas, Nevada, USA, August 2000, pp. 125–131.
[14] S. Cahon, N. Melab, and E.-G. Talbi. ParadisEO: A framework for the reusable design of parallel
and distributed metaheuristics. Journal of Heuristics, 10(3): 357–380, 2004.
[15] C.H. Papadimitriou. The Complexity of Combinatorial Optimization Problems. Master’s thesis,
Princeton University, 1976.
[16] S. Kirkpatrick, C.D. Gelatt, and M.P. Vecchi. Optimization by simulated annealing. Science,
220: 671–680, 1983.
[17] F. Glover. Tabu search, part I. ORSA. Journal of Computing, 1: 190–206, 1989.
[18] J.H. Holland. Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann
Arbor, MI, 1975.
[19] D. Roberts and R. Johnson. Evolving frameworks. A pattern language for developing object-
oriented frameworks. In Proceedings of the Third Conference on Pattern Languages and Programming
(PLoP ‘96), Allerton Park, Illinois, September 4–6, 1996.
[20] W. Pree, G. Pomberger, A. Schappert, and P. Sommerlad. Active guidance of framework
development. Software — Concepts and Tools, 16: 94–103, 1995.
[21] M.J. Blesa, Ll. Hernandez, and F. Xhafa. Parallel skeletons for tabu search method. In 8th Interna-
tional Conference on Parallel and Distributed Systems (ICPADS’01), IEEE Computer Society Press,
Kyongju City, Korea, 2001, pp. 23–28.
[22] E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design Patterns, Elements of Reusable Object-
Oriented Software, Addison-Wesley, Reading, MA, 1994.
[23] M. Keijzer, J.J. Morelo, G. Romero, and M. Schoenauer. Evolving objects: A general purpose
evolutionary computation library. In Proceedings of the 5th International Conference on Artificial
Evolution (EA ’01), Le Creusot, France, October 2001.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C035” — 2005/8/8 — 22:03 — page 647 — #9

 36
Parallel Hybrid
Multiobjective
Metaheuristics on
P2P Systems

36.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .36-649

36.2 Parallel Hybrid MOMs and P2P Computing . . . . . . . . . .36-650
Multiobjective Optimization • Parallelism and
Hybridization • P2P Computing for Parallel MO
Optimization
36.3 A Model for P2P Coordination . . . . . . . . . . . . . . . . . . . . . . . . .36-652
Model Description • Implementation on Top of XtremWeb
36.4 Application to BPFSP and Experimentation . . . . . . . . . . .36-655
Problem Formulation • A Genetic–Mimetic Algorithm for
Solving BPFSP • Parallel Hybrid AGMA • Deployment and
N. Melab Experimentation • Deployment and Fault Tolerance •
El-Ghazali Talbi Experimental Results
M. Mezmaz 36.5 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . .36-661
B. Wei References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .36-663

36.1 Introduction
Metaheuristics allow to provide near-optimal solutions of NP-hard complex problems in a reasonable time.
They fall into two complementary categories: evolutionary algorithms (EAs) that have a good exploration
power, and local searches (LSs) characterized by better intensification capabilities. The hybridization of
the two categories permits to improve the effectiveness (quality of provided solutions) and the robustness
of the metaheuristics [11]. Nevertheless, as it is CPU time consuming it is not often fully exploited in
practice. Indeed, experiments with hybrid metaheuristics are often stopped before the convergence is
reached. Nowadays, Peer-to-Peer (P2P) computing [8] and grid computing [5] are two powerful ways
to achieve high performance on long-running scientific applications. Parallel hybrid metaheuristics used
for solving real-world multiobjective problems (MOPs) are good challenges for P2P and grid computing.
However, to the best of our knowledge no research work has been published on that topic.

36-649

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 649 — #1

36-650 Handbook of Bioinspired Algorithms and Applications

In this chapter, we contribute with the first results on parallel hybrid multiobjective metaheuristics
on P2P systems. The design and deployment of these optimization methods require a middleware that
allows cooperation between parallel tasks. In addition, the traditional parallel models and hybridization
mechanisms have to be re-thinked and adapted to be scaled up. Moreover, these require to be fault-tolerant
to allow long-running problem resolutions. We particularly focus here on the island model and the
multistart model.
Recently, few middlewares [1,4,13] allowing to exploit P2P systems have emerged. These middlewares
are well suited for embarrassingly parallel applications such as multi parameter simulations. However,
they are limited regarding the parallelism as they do not allow direct cross-peer (or cross-task) commu-
nication. Our contribution is to propose a Linda-like [7] coordination model and its implementation
on top of XtremWeb [4]. This is a Dispatcher/Worker oriented middleware, in which the Dispatcher
distributes application tasks submitted by clients to volunteer worker peers at their request. In addition,
the considered middleware provides fault-tolerance mechanisms that are costly in a highly volatile P2P
environment. Indeed, a work unit is re-started from scratch each time it fails. Another contribution of this
chapter is to deal with the fault-tolerance issue at application level. We propose a check-pointing approach
for the two parallel models quoted above.
To be validated the proposed approaches have been experimented on the Bi-criterion Permutation
Flow-Shop Problem (BPFSP) [12]. The problem consists roughly to find a schedule of a set of jobs
on a set of machines that minimizes the makespan and the total tardiness. Jobs must be scheduled in
the same order on all machines, and each machine cannot be simultaneously assigned to two jobs.
In Reference 2, a hybrid MultiObjective Metaheuristic (MOM) has been proposed to solve this problem.
In this chapter, we extend this work with two P2P-based fault-tolerant parallel models: the island and
multistart models. Our extended version allows to fully exploit the hybridization and provides clearly
better results. This constitutes another contribution of this chapter.
This chapter is organized as follows: Section 36.2 presents briefly parallel hybrid multiobjective optimiz-
ation (MOO). Section 36.3 highlights the requirements of MOO and describes the proposed coordination
model and its implementation on top of XtremWeb. Section 36.4 presents the experimentation of the
model and its implementation through a parallel hybrid metaheuristic applied to the BPFSP, and analyzes
the preliminary experimental results. Finally, Section 36.5 concludes the chapter.

36.2 Parallel Hybrid MOMs and P2P Computing

36.2.1 Multiobjective Optimization
An MOP consists generally in optimizing a vector of nbobj objective functions F (x) = (f1 (x), . . . , fnbobj (x)),
where x is an d-dimensional decision vector x = (x1 , . . . , xd ) from some universe called decision space.
The space the objective vector belongs to is called the objective space. F can be defined as a cost function
from the decision space to the objective space that evaluates the quality of each solution (x1 , . . . , xd ) by
assigning it an objective vector (y1 , . . . , ynbobj ), called the fitness.
Unlike single-objective optimization problems, an MOP may have a set of solutions known as the Pareto
optimal set rather than a unique optimal solution. The image of this set in the objective space is denoted as
Pareto front. Graphically, a solution x is Pareto optimal if there is no other solution x  such that the point
F (x  ) is in the dominance cone of F (x). This dominance cone is the box defined by F (x), its projections
on the axes and the origin (Figure 36.1).

36.2.2 Parallelism and Hybridization

In Reference 3, different parallel models have been distinguished: those associated with LSs and those
dedicated to EAs. Three major parallel models for EAs are presented: the island (a)synchronous cooperative
model, the parallel evaluation of the population, and the distributed evaluation of a single solution.
The parallel models for LSs are mainly: the parallel exploration of neighboring candidate solutions and

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 650 — #2

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-651

f2
Pareto solution
Dominated solution

f1

FIGURE 36.1 Example of nondominated solutions.

the multistart model. In this chapter, we focus only on the coarse-grained models: the island model and the
multistart model. Due to the communication delays, fine-grained models are often inefficient when they
are deployed in a large-scale network.
In the island (a)synchronous cooperative model, different EAs are simultaneously deployed and cooper-
ate for computing better and robust solutions. They exchange, in an asynchronous way, the genetic stuff
to diversify the search. The objective is to allow to delay the global convergence, especially when the EAs
are heterogeneous regarding the variation operators. The migration of individuals follows a policy defined
by few parameters: the migration decision criterion, the exchange topology, the number of emigrants, the
emigrants selection policy, and the replacement/integration policy.
The multistart model consists in simultaneously launching several local searches. They may be hetero-
geneous, but no information is exchanged between them. The results would be identical as if the algorithms
were sequentially run. Very often deterministic algorithms differ by the supplied initial solution and/or
some other parameters. This trivial model is convenient for low-speed networks of workstations.
Combinations of different metaheuristics often provide very powerful search methods. In Reference 11,
two levels and two modes of hybridization are distinguished: Low and High levels, and Relay and Cooper-
ative modes. The low-level hybridization consists in replacing an internal function (e.g., an operator) of
a given metaheuristic by another metaheuristic. In high-level hybrid algorithms, the different metaheur-
istics are self-containing, meaning no direct relationship to their internal working is considered. Relay
hybridization means a set of metaheuristics is applied in a pipeline way. The output of a metaheuristic
(except the last) is the input of the following one (except the first). Conversely, teamwork hybridization is a
cooperative optimization model. Each metaheuristic performs a search in a solution space, and exchanges
solutions with others. In this chapter, we address the high-level hybridization mechanism in the relay and
cooperative modes.

36.2.3 P2P Computing for Parallel MO Optimization

In this chapter, we focus on Dispatcher/Worker-oriented P2P middlewares such as XtremWeb [4] and
SETI@Home [1]. In such systems, clients can submit their jobs to the Dispatcher. A set of volatile workers
(peers) request the jobs from the Dispatcher according to the cycle stealing model. Then, they execute the
jobs and return the results to the Dispatcher to be collected by the clients. In these middlewares, even a
central server (the Dispatcher) is required for controlling the peers (workers) they are considered as P2P
software environments. Indeed, an important part of these systems is executed on these peers with a high
autonomy.
One of the major limitations of P2P computing environments is that they are well suited for embarrass-
ingly parallel (e.g., multiparameter) applications with independent tasks. In this case, no communication
is required between the tasks, and thus peers. The deployment of parallel hybrid metaheuristics that
need cross-peer/task communication is not straightforward. The programmer has the burden to manage

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 651 — #3

36-652 Handbook of Bioinspired Algorithms and Applications

and control the complex coordination between the workers. To deal with such problem existing middle-
wares must be extended with a software layer that implements a coordination model. Several interesting
coordination models have been proposed in the literature [6,9]. In this chapter, we focus only on one of
the most popular of them, that is, Linda [7], as our proposed model is an extension of this model.
In the Linda model, the coordination is performed through generative communications. Processes
share a virtual memory space called a tuple-space (set of tuples). The fundamental data unit, a tuple,
is an ordered vector of typed values. Processes communicate by reading, writing, and consuming these
tuples. The “eval” operation is particularly useful in a P2P environment as it allows to spawn tasks to be
executed on volunteer peers. A small set of four simple operations allows highly complex communication
and synchronization schemes:

• Out(tuple): puts tuple into tuple-space.

• In(pattern): removes a (often the first) tuple matching pattern from tuple-space.
• rd(pattern): is the same as in(pattern), but does not remove the tuple from tuple-space.
• Eval(expression): puts expression in tuple-space for evaluation. The evaluation result is a tuple left
in tuple-space.

Nevertheless, Linda has several limitations regarding the design and deployment of parallel hybrid
metaheuristics for P2P systems. First, it does not allow rewriting operations on the tuple space. Due to
the high communication delays in a P2P system, tuple rewriting is very important as it allows to reduce
the number of communications and the synchronization cost. Indeed, in Linda a rewriting operation
is performed as an “in” or “rd” operation followed by a local modification and an “out” operation.
The operations “in”/“rd” and “out” involve two communications and a heavy synchronization. Therefore,
the model needs to be extended with a rewriting operation. Furthermore, the model does not support
group operations that are useful for efficiently writing/reading Pareto sets in/from the tuple-space. Finally,
nonblocking operations that are very important in a P2P context are not supported in Linda. In the next
section, we propose an extension of the Linda model that allows to meet these requirements.

36.3 A Model for P2P Coordination

36.3.1 Model Description
Designing a coordination model for parallel MOO requires the specification of the content of the tuple-
space, a set of coordination operations and a pattern matching mechanism. The tuple-space may be
composed of a set of Pareto optimal solutions and their corresponding solutions in the objective space.
For the parallel island model of the multiobjective metaheuristics, the tuple-space contains a collection of
(parts of) Pareto optimal sets deposited by the islands for migration. The mathematical formulation of
the tuple-space (Pareto Space or PS) is the following:


PS = PO, with PO = {(x, F (x)), x is Pareto optimal}.

In addition to the operations provided in Linda, parallel P2P multiobjective optimization needs other
operations. These operations fall into two categories: group operations and nonblocking operations. Group
operations are useful to manage multiple Pareto optimal solutions. Nonblocking operations are necessary
to take into account the volatile nature of P2P systems. In our model, the coordination primitives are
defined as follows:

• in, rd, out and eval: These operations are the same as those of Linda defined in Section 36.2.3.
• ing(pattern): Withdraws from PS all the solutions matching the specified pattern.
• rdg(pattern): Reads from PS a copy of all the solutions matching the specified pattern.
• outg(setOfSolutions): Inserts multiple solutions in PS.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 652 — #4

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-653

• update(pattern, expression): Updates all the solutions matching the specified pattern by the solutions
resulting from the evaluation of expression.
• inIfExist, rdIfExist, ingIfExist, and rdgIfExist : These operations have the same syntax than
respectively in, rd, ing, and rdg but they are non blocking probe operations.

The update operation allows to locally update the PS, and so to reduce the communication and syn-
chronization cost. The pattern matching mechanism depends strongly on how the model is implemented,
and in particular on how the tuple-space is stored and accessed. For instance, if the tuple-space is stored
in a database the mechanism can be the request mechanism used by the database management system.
More details on the pattern matching mechanism of our model are given in the next section.

36.3.2 Implementation on Top of XtremWeb

XtremWeb [4] is a Java P2P project developed at Paris-Sud University. It is intended to distribute applica-
tions over a set of peers, and is dedicated to multiparameter applications that have to be computed several
times with different inputs. XtremWeb manages tasks following the Dispatcher/Worker paradigm (see
Figure 36.2). Tasks are scheduled by the Dispatcher to workers only on their specific demand since they
may adaptively appear (connect to the Dispatcher) and disappear (disconnect from the Dispatcher). The
tasks are submitted by either a client or a worker, and in the latter case, the tasks are dynamically generated
for parallel execution. The final or intermediate results returned by the workers are stored in a MySQL
database. These results can be requested later by either the clients or the workers. The database stores also
different information related to the workers and the deployed application tasks.
XtremWeb is well suited for embarrassingly parallel applications where no cross-peer communication
occurs between workers, and these can only communicate with the Dispatcher. Yet, many parallel dis-
tributed applications particularly parallel MOMs need cooperation between workers. In order to free
the user from the burden of managing himself/herself such cooperation we propose an extension of the
middleware with a software layer.
The software layer is an implementation of the proposed model composed of two parts (see Figure. 36.3):
a coordination API and its implementation at the worker level and a coordination request broker (CRB).
The PS is a part of the MySQL database associated with the Dispatcher. Each tuple or solution of the PS
is stored as a record in the database.

XtremWeb clients

...

Get results

Submit work
Internet
Send results
Get a work unit

... XtremWeb dispatcher

XtremWeb workers

FIGURE 36.2 Global architecture of XtremWeb.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 653 — #5

36-654 Handbook of Bioinspired Algorithms and Applications

XtremWeb dispatcher
switch(OP) { Pareto
... Space
...
case ING: Manager
switch(){}
local call of ...
ing(pattern); ...
... ing(pattern){}
CRB_Skeleton ...
};

XtremWeb worker
Work unit RMI call SELECT * FROM PS
of ing WHERE pattern is
...
matched
s=ing(pattern);
... CRB_Stub

switch() Pareto
{} space (PS)

execlp(..., CRB_Stub, switch(OP) {

ING, ARGS_FILE, ...); ...
case ING: XtremWeb database
RMI call of ing(pattern); information
...
};

FIGURE 36.3 Implementation of the coordination model on top of XtremWeb.

From the worker side the coordination API is implemented in Java and in C/C++. The C/C++
version allows the deployment and execution of C/C++ applications with XtremWeb (written in Java).
The coordination library must be included in these programmer applications. From the Dispatcher side,
the coordination API is implemented in Java as a PS manager. The CRB is a software broker allowing the
workers to transport their coordination operations calls to the Dispatcher, and has two components: one
for the worker (CRB stub) and another for the Dispatcher (CRB skeleton). The role of the CRB stub is to
transform the local calls to the coordination operations performed by the tasks executed by the worker into
RMI calls. The role of the CRB skeleton is to transform these RMI calls into local calls to the coordination
operations performed by the PS Manager. These local calls are translated into MySQL requests addressed
to the PS.
To illustrate the implementation of the coordination layer on top of XtremWeb, let us consider the scen-
ario presented in Figure 36.3. The work unit performed by an XtremWeb worker calls the ing (template)
coordination operation. In the C++ version of the coordination API, the implementation of each coordin-
ation operation makes the system call execlp() with appropriate parameters to plug in the CRB_Stub Java
object. In our scenario, the major parameters are the number ING designating the operation and the file
ARGS_FILE containing the arguments specified in the template parameter. CRB_Stub translates the ing
local call into an RMI call to the CRB_Skeleton Java object. This latter translates the RMI call into a local
call to the ing operation implemented in the PS Manager class. The implementation of the coordination
operation consists in a MySQL select request addressed to the PS part of the XtremWeb information
database.
Note that the method declarations for the coordination operations in the PS Manager class contain the
Java synchronized keyword. Hence, the system associates a unique lock with the instance of the PS Manager
class. Whenever control enters a synchronized coordination operation, other calls to a synchronized
cooperation method are blocked until the PS Manager object is unlocked. In the next section, the proposed
coordination model is applied to parallel hybrid MOMs.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 654 — #6

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-655

M1 J2 J4 J5 J1 J6 J3

M2 J2 J4 J5 J1 J6 J3

M3 J2 J4 J5 J1 J6 J3

FIGURE 36.4 Example of permutation flow-shop with 6 jobs and 3 machines.

36.4 Application to BPFSP and Experimentation

36.4.1 Problem Formulation
The Flow-Shop problem is a scheduling problem [12] that has received a great attention given its import-
ance in many industrial areas. The problem can be formulated as a set of N jobs J1 , J2 , . . . , JN to be
scheduled on M machines. The machines are critical resources as each machine cannot be simultaneously
assigned to two jobs. Each job Ji is composed of M consecutive tasks ti1 , . . . , tiM , where tij represents the
jth task of the job Ji requiring the machine mj . To each task tij is associated a processing time pij , and each
job Ji must be achieved before a due date di .
In this chapter, we focus on the BPFSP where jobs must be scheduled in the same order on all the
machines (see Figure 36.4). Therefore, two objectives have to be minimized:

• Cmax : Makespan (Total completion time)

• T : Total tardiness

The task tij being scheduled at time sij , the two objectives can be formulated as follows:

f1 = Cmax = max{siM + piM |i ∈ [1, . . . , N ]},

N
f2 = T = [max(0, siM + piM − di )].
i=1

The Pareto front PF associated with BPFSP may be formulated as follows:

∀y, ∃x ∈ PF , (m(x) ≤ m(y)) or (t (x) ≤ t (y)),

where x and y are solutions of the MOP, and m(x) (respectively t (x)) is the value of x corresponding to
the makespan (respectively tardiness) criterion.

36.4.2 A Genetic–Mimetic Algorithm for Solving BPFSP

In single objective optimization, it is well known that GAs provide better results when they are hybridized
with LS algorithms. Indeed, the GA convergence is too slow to be really effective without any hybridization
[10]. In Reference 2, a hybrid Genetic–Mimetic algorithm named AGMA has been proposed for solving
BPFSP. The simplified pseudocode of the algorithm is presented in Algorithm 36.1 and illustrated in
Figure 36.5.
AGMA combines a genetic algorithm (GA) and a mimetic algorithm (MA). In this chapter, we do not
give the details and parameters of the two algorithms, and if needs be, the reader is referred to Reference 2.
The GA uses mainly two parameters: an archive (Pareto Front) PO ∗ of nondominated solutions, and a
progression ratio PPO ∗ of PO ∗ . At each generation, these two parameters are updated. If no significant
progression is noticed (PPO ∗ < α, where α is a fixed threshold), an intensified search process is triggered.
The intensification consists in applying MA (see Algorithm 36.2) to the current population during one
generation. The application of MA returns a Pareto Front PO ∗ that serves to update the Pareto Front PO ∗
of the GA.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 655 — #7

36-656 Handbook of Bioinspired Algorithms and Applications

Algorithm 36.1 AGMA Algorithm

Create an initial population
while run time not reached do
Perform a GA generation with adaptive mutation
Update PO∗ and PPO∗
if PPO∗ < α then
Perform a generation of MA on the population (Algorithm 2)
Update PO∗ and PPO∗
end if
Update selection probability of each mutation operator
end while

Genetic algorithm Mimetic algorithm

Ppo* < a
POP POP
Crossover
PO* POP⬘

PO*⬘

Neighbors

FIGURE 36.5 Illustration of AGMA.

Algorithm 36.2 MA algorithm

while MA run time not reached do
Select randomly a set P of solutions from the current population
Apply the crossover on P to generate a set P  of new solutions
Compute the nondominated set PO ∗ from P 
while New solutions found do
Create the neighborhood N of each solution of PO 

Let PO ∗ be the nondominated set of N PO ∗
end while
end while

Mimetic algorithm consists in selecting randomly a set of solutions from the current population of
the GA. A crossover operator is then applied to these solutions and new solutions are generated. Among
these new solutions only nondominated ones are maintained to constitute a new Pareto Front PO ∗ . An
LS is then applied to each solution of PO ∗ to compute its neighborhood. The nondominated solutions
belonging to the neighborhood are inserted into PO ∗ .

36.4.3 Parallel Hybrid AGMA

Different parallel models have been sketched and analyzed in Section 36.2. The fine-grained parallel models
could not be exploited efficiently in a P2P environment due to the communication delays. In BPFSP, the
model based on parallel evaluation of each solution is fine-grained and is not likely to lead to better

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 656 — #8

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-657

performance. Indeed, the evaluation of each objective has a low cost. Therefore, it is useless to evaluate
in parallel the two objectives and evaluate each of them in parallel. Conversely, it is useful to exploit the
following parallel models: (1) the island model that consists in performing in parallel several cooperative
AGMAs; (2) the parallel evaluation of the population of each AGMA; (3) the multistart model that consists
in applying in parallel an LS on each solution of the Pareto Front PO ∗ in MA. The parallel evaluation of
the neighborhood of each solution could not be efficient for the same reason as the parallel evaluation of
each solution.
We have limited our implementation to the coarse-grained parallel models, that is, the island model
and the multistart model. Figure 36.6 illustrates the parallel hybrid AGMA exploiting these two models.

• The island model: Due to its exorbitant cost in terms of CPU time on large-size instances of BPFSP,
the island model has not been exploited in Reference 2. Indeed, the exploitation of the model on
large-size BPFSP is possible only on large-scale P2P networks or grids. In our implementation
(see Figure 36.6), the parameters of the model are the following: The different cooperative AGMA

Island
model

Genetic
Mimetic
algorithm
algorithm

Ppo* < a
POP POP
Crossover
PO* POP⬘

PO*⬘

Neighbors

XtremWeb
interface PO*

Multistart
model

XtremWeb workers

FIGURE 36.6 Illustration of parallel AGMA.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 657 — #9

36-658 Handbook of Bioinspired Algorithms and Applications

exchange their whole archives PO ∗ , and the number of emigrants is dynamic. At its arrival, the
immigrant archive is merged with the local one. Migrations occur periodically (each a fixed number
of iterations). The migration topology is the random one, meaning the destination island is selected
randomly.
• The multistart model: The multistart model is exploited during the execution of MA. Each solution
of the Pareto Front PO ∗ computed by the algorithm represents the initial solution of an LS method
that calculates its neighborhood. The different LSs are executed in parallel according to the Master–
Slave model. The master, that is, the algorithm MA merges with PO ∗ the neighborhoods returned
by the different slaves and computes the new PO ∗ that contains the nondominated solutions.

36.4.4 Deployment and Experimentation

In this section, we present the deployment scheme of the different parallel hybrid models on the XtremWeb
architecture, and the preliminary experimental results obtained on the application presented above.

36.4.5 Deployment and Fault Tolerance

A deployment scheme may be defined as a function that consists in embedding the different components
of the parallel models on the different components of the P2P architecture. Different deployment schema
of the island and multistart models on XtremWeb are possible. Indeed, the AGMA algorithms of the island
model can be deployed either as XtremWeb clients or workers. For the multistart model, the master can
be either a client or a worker, and the slaves are necessarily deployed as workers. For our experimentation,
the deployment scheme is illustrated in Figure 36.7.
The island model is deployed on three XtremWeb clients, and each client runs the AGMA algorithm.
During the hybridization phase (execution of MA), the LSs initiated on the Pareto Front PO ∗ are submitted
as tasks to the Dispatcher that launches them on the workers at their request. The multistart model is thus
deployed on a client and a set of workers.
In XtremWeb, the fault tolerance issue is tackled at Worker and Dispatcher levels. When a worker fails
the work unit being executed is re-started from scratch. If the Dispatcher crashes it is re-started using
its information database. The problem with such solution is that in a highly volatile environment a large

Worker1 Worker2 WorkerN

LS LS
Neighbors Neighbors

LS11 LS25 ... LSm4

Dispatcher

MySQL
PO*
Neighbors LS PO*

Client1 Client2 Clientm

FIGURE 36.7 Deployment schema of parallel hybrid AGMA on top of XtremWeb.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 658 — #10

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-659

amount of CPU time is wasted as the system spends its time in re-starting work units performed by the
workers. Therefore, we propose a check-pointing approach at the client level that allows to solve more
efficiently the fault-tolerance problem. Indeed, the problem data and intermediate results are periodically
stored. If the Dispatcher fails the application is restored and re-started from the last checkpoint. In case
of worker failure the work unit is re-started using the intermediate results. The check-pointing operation
(storing) is performed after each LS and/or every 100 generations. The second condition is necessary when
no LS has been launched during the last 100 generations. This means that a significant progression of the
Pareto Front has been observed at each generation, what excludes any resort to the hybridization.

36.4.6 Experimental Results

In our experiments, we consider the BPFSP instance 200 jobs on 10 machines. The parameters of the
island model are fixed as the following: migrations occur every 10 generations, the number of emigrants
is fixed at each migration operation to 20 the size of the archive PO ∗ is upper than 20, and the whole
archive otherwise, and the population size of each AGMA is 100.
The application has been deployed during working days (nondedicated environment) on the education
network of the Polytech’Lille engineering school. The experimentation hardware platform is composed of
120 heterogeneous Linux Debian PCs. The characteristics of these PCs are presented in Table 36.1.
Three parallel hybrid versions of AGMA are experimented, evaluated and compared:

• Version 1 is that proposed in Reference 2, and exploits only the multistart model. This means
that the GA is executed on a single machine and the hybridization phase is deployed on a parallel
machine (IBM-SP2) according to the Master–Slave model (Push mode, i.e., work distribution is
initiated by the master).
• Version 2 is the same as Version 1 except that the hybridization is deployed in a distributed way
on a set of XtremWeb workers according to the cycle stealing paradigm (Pull mode, i.e., work
distribution is initiated by the workers).
• Version 3 is not considered in Reference 2, and is a combination of the multistart and island models.
As illustrated in Figure 36.7, three AGMA algorithms are deployed on client machines and cooperate
according to the island model. Each AGMA is an implementation of Version 2.

Figure 36.8 illustrates the Pareto Fronts obtained with the versions 1 and 2 after 80 LSs. The two fronts
are approximately the same, but Version 2 has the advantage to be fault tolerant.
The execution of Version 1 is stopped after 80 LSs as it is not fault tolerant. Conversely, with Version 2
long-lasting executions are possible. For instance, Figure 36.9 shows that the execution goes on up to
350 LSs. The execution lasted one week, and 10 failures have been observed and as many check-pointing
operations have been performed. As a result, the Pareto Front obtained with 350 LSs is clearly better than
that obtained with 80 LSs using Version 1 or Version 2. One has to note that such results are possible only
with a scalable and fault tolerant version of the algorithm.

TABLE 36.1 Experimentation Hardware Platform

Processor Number

AMD Duron(tm) Processor 14

Celero (Coppermine) 14
Intel(R) Celeron(R) CPU 2.00 GHz 8
Intel(R) Celeron(R) CPU 2.20 GHz 28
Intel(R) Celeron(R) CPU 2.40 GHz 21
Intel(R) Celeron(R) CPU 1400 MHz 7
Pentium III (Katmai) 28

Total 120

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 659 — #11

36-660 Handbook of Bioinspired Algorithms and Applications

11250

11200 Version 1 (80 LS)

Version 2 (80 LS)

11150

11100
Makespan

11050

11000

10950

10900

10850
40000 45000 50000 55000 60000 65000 70000
Tardiness

FIGURE 36.8 Pareto Fronts obtained with Version 1 and Version 2 (80 LSs).

11350

11300 Version 1 (80 LS)

Version 2 (350 LS)
11250

11200

11150
Makespan

11100

11050

11000

10950

10900

10850
35000 40000 45000 50000 55000 60000 65000 70000
Tardiness

FIGURE 36.9 Pareto Fronts with Version 1 (80 LSs) and Version 2 (350 LSs).

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 660 — #12

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-661

11250

11200 Version 2 (80 LS)

Version 3 (80 LS)

11150

11100
Makespan

11050

11000

10950

10900

10850
42000 44000 46000 48000 50000 52000 54000 56000 58000 60000 62000 64000
Tardiness

FIGURE 36.10 Pareto Fronts with Version 2 (80 LSs) and Version 3 (80 LSs).

Figure 36.10 allows to compare the Pareto Fronts obtained with Version 2 and Version 3 and to
demonstrate the contribution of the island model to the effectiveness. With 80 LSs, the Pareto Front
obtained using Version 3 is better than that obtained using Version 2. More experiments with more LSs
are in progress.
Figure 36.11 (Part A) shows the oscillation between GA and MA (or LS) over time obtained with
Version 3. Figure 36.11 (Part B) is a zoom up of Figure 36.11 (Part A) on the 50,000 first time units. It
shows that the MA (thus LS) is frequently solicited and this lasts long. Figure 36.11 (Part C) illustrates the
evolution in time of the number of deployed workers at the beginning of the execution (zoom out on the
first 350 time-units). The maximum number of workers is 60 because during the starting phase the Pareto
Front contains a small number of solutions. The number of workers decrease to 0 when the GA succeed
to improve the Pareto Front without calling MA.
One has to note that the spectrum blackens with the time (from left to right). This means that the GA
solicits more and more the MA, that is, the LS because it never enhances again the Pareto Front, in other
words the GA converges. On the other hand, the local search lasts less and less time. Therefore, even the
intensification (by LS) does not contribute to enhance the effectiveness, meaning that the AGMA con-
verges. Through this experimentation, we have learned more on the convergence of the AGMA algorithm.
Therefore, one can note that P2P computing allows to“push far”the limits in terms of computing resources
to better evaluate the contribution of the hybridization but also its limitations.

36.5 Conclusion and Future Work

The hybridization of metaheuristics having complementary behaviors allows to enhance the effectiveness
and robustness in combinatorial optimization [11]. However, its exploitation on industrial applications is
possible only by using a great computing power. Large-scale parallelism based on the use of computational
grids and/or P2P systems is recently revealed to be a good way to get at hand such computing power and
exploit hybridization. To our best of knowledge, no research work has been published on parallel hybrid

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 661 — #13

36-662 Handbook of Bioinspired Algorithms and Applications

60 .
(Part C)
50

40
Number of workers
30

20

10

0
0 50 100 150 200 250 300 350
MA Time
(LS)

GA
(Part A)

0 000 300000 350000 400000 450000 500000

mps Time
(Part B)
5000 10000 15000 20000 25000 30000 35000 40000 45000 50000

Time

FIGURE 36.11 Oscillation spectrum between the GA and MA (thus LS).

metaheuristics on P2P systems. Nowadays, existing P2P computing middlewares are inadequate for the
deployment of parallel cooperative applications. Indeed, these need to be extended with a software layer
to support the cooperation. In this chapter, we have proposed a Linda-like cooperation model that has
been implemented on top of XtremWeb.
In Reference 2, a hybrid metaheuristic (AGMA) has been proposed and experimented on BPFSP.
The performed experiments on large-size instances such as 200 jobs on 10 machines are often stopped
without the convergence is reached. The full exploitation of the hybridization needs a large amount of
computational resources and the management of the fault-tolerance issue. We have proposed a fault-
tolerant hybrid parallel design of the AGMA combining two parallel models: the multistart model and the
island model. The algorithm has been implemented on our extended version of XtremWeb.
The first experiments have been performed on the education network of the Polytech’Lille engineering
school. The network is composed of 120 heterogeneous Linux PCs. The preliminary results, obtained
after several execution days, demonstrate that the use of P2P computing allows to fully exploit the benefits
of hybridization. Indeed, the obtained Pareto Front is clearly better than that obtained in Reference 2.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 662 — #14

Parallel Hybrid Multiobjective Metaheuristics on P2P Systems 36-663

On the other hand, the deployment of the island model allows to improve the effectiveness. Beyond the
improvement of the effectiveness, the parallelism on P2P systems allows to push far the limits in terms
of computational resources. As a consequence, it permits to better evaluate the benefits and limitations
of the hybridization. Such result has to be confirmed again on a larger P2P network and larger instances
of the problem.

References
[1] D.P. Anderson, J. Cobb, E. Korpela, M. Lepofsky, and D. Werthimer. SETI@home: An experiment
in public-resource computing. Communications of the ACM, 45: 56–61, 2002.
[2] M. Basseur, F. Seynhaeve, and E.-G. Talbi. Adaptive mechanisms for multi-objective evolution-
ary algorithms. In Congress on Engineering in System Application CESA ’03, Lille, France, 2003,
pp. 72–86.
[3] S. Cahon, N. Melab, and E.-G. Talbi. ParadisEO: A framework for the reusable design of parallel
and distributed metaheuristics. Journal of Heuristics, 10: 353–376, 2004.
[4] G. Fedak, C. Germain, V. Neri, and F. Cappello. XtremWeb: Building an experimental platform
for Global Computing. Workshop on Global Computing on Personal Devices (CCGRID2001), IEEE
Press, May 2001.
[5] I. Foster and C. Kesselman. The Grid: Blueprint for a New Computing Infrastructure. Morgan
Kaufmann, San Fransisco, CA, 1999.
[6] D. Gelernter and N. Carriero. Coordination languages and their significance. Communications of
the ACM, 35: 97–107, 1992.
[7] D. Gelernter. Generative communication in Linda. ACM Transactions on Programming Languages
and Systems, 7: 80–112, 1985.
[8] A. Oram. Peer-to-Peer: Harnessing the Power of Disruptive Technologies. O’Reilly & Associates, 2001.
[9] G.A. Papadopoulos and F. Arbab. Coordination models and languages. Advances in Computers:
The Engineering of Large Systems, Academic Press, 1998, p. 46.
[10] E.-G. Talbi, M. Rahoual, M.-H. Mabed, and C. Dhaenens. A hybrid evolutionary approach
for multicriteria optimization problems: Application to the Flow Shop. In E. Zitzler et al.,
Eds., Evolutionary Multi-Criterion Optimization, Vol. 1993 of Lecture Notes in Computer Science,
Springer-Verlag, Heidelberg, 2001, pp. 416–428.
[11] E.-G. Talbi. A Taxonomy of Hybrid Metaheuristics. Journal of Heuristics, 8: 541–564, 2002.
[12] V. T’kindt and J.-C. Billaut. Multicriteria Scheduling — Theory, Models and Algorithms. Springer-
Verlag, Heidelberg, 2002.
[13] J. Verbeke, N. Nadgir, G. Ruetsch, and I. Sharapov. Framework for peer-to-peer distributed com-
puting in a heterogeneous, decentralized environment. In Proceedings of the Third International
Workshop on Grid Computing (GRID ’2002), Baltimore, MD, January 2002, pp. 1–12.

© 2006 by Taylor & Francis Group, LLC

CHAPMAN: “C4754_C036” — 2005/8/6 — 14:14 — page 663 — #15