Artificial Intelligence (AI)

Artificial Intelligence (AI): Elementary to Advanced Practices

Series Editors: Vijender Kumar Solanki, Zhongyu (Joan) Lu, and Valentina E Balas
In the emerging smart city technology and industries, the role of artificial intelligence is getting
more prominent. This AI book series will aim to cover the latest AI work, which will help the naïve
user to get support to solve existing problems and for the experienced AI practitioners, it will assist
to shedding light for new avenues in the AI domains. The series will cover the recent work car-
ried out in AI and its associated domains, it will cover Logics, Pattern Recognition, NLP, Expert
Systems, Machine Learning, Block-Chain, and Big Data. The work domain of AI is quite deep, so
it will be covering the latest trends which are evolving with the concepts of AI and it will be help-
ing those new to the fiield, practitioners, students, as well as researchers to gain some new insights.
Cyber Defense Mechanisms
Security, Privacy, and Challenges
Gautam Kumar, Dinesh Kumar Saini, and Nguyen Ha Huy Cuong
Artificial Intelligence Trends for Data Analytics Using Machine Learning and Deep
Learning Approaches
K. Gayathri Devi, Mamata Rath and Nguyen Thi Dieu Linh
Transforming Management Using Artificial Intelligence Techniques
Vikas Garg and Rashmi Agrawal
AI and Deep Learning in Biometric Security
Trends, Potential, and Challenges
Gaurav Jaswal, Vivek Kanhangad, and Raghavendra Ramachandra
Enabling Technologies for Next Generation Wireless Communications
Edited by Mohammed Usman, Mohd Wajid, and Mohd Dilshad Ansari
Artificial Intelligence (AI)
Recent Trends and Applications
Edited by S. Kanimozhi Suguna, M. Dhivya, and Sara Paiva
For more information on this series, please visit: https​:/​/ww​​w​.rou​​tledg​​e​.com​​/Arti​​ficia​​l​-
Int​​ellig​​ence-​​AI​-El​​ement​​a ry​-t​​o​-Adv​​anced​​-Prac​​tices​​/​book​​-seri​​es​/CR​​CAIEA​P
Artificial Intelligence (AI)
Recent Trends and Applications

Edited by
S. Kanimozhi Suguna, M. Dhivya, and Sara Paiva
First edition published 2021
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742

and by CRC Press

2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

© 2021 Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, LLC

The right of S. Kanimozhi Suguna, M. Dhivya, and Sara Paiva to be identified as the authors of the editorial
material, and of the authors for their individual chapters, has been asserted in accordance with sections 77 and
78 of the Copyright, Designs and Patents Act 1988.

Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and publish-
ers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has not
been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, 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, access www​.copyright​.com or con-
tact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For
works that are not available on CCC please contact mpkbookspermissions​@tandf​.co​​.uk

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

Names: Suguna, S. Kanimozhi, editor.

Title: Artificial intelligence (AI) : recent trends and applications /
edited by S. Kanimozhi Suguna, M. Dhivya, and Sara Paiva.
Description: First edition. | Boca Raton : CRC Press, 2021. | Series:
Artificial intelligence (ai): elementary to advanced practices |
Includes bibliographical references and index.
Identifiers: LCCN 2020048919 (print) | LCCN 2020048920 (ebook) | ISBN
9780367431365 (hardback) | ISBN 9781003005629 (ebook)
Subjects: LCSH: Artificial intelligence.
Classification: LCC Q335 .A8584 2021 (print) | LCC Q335 (ebook) | DDC
006.3--dc23
LC record available at https://lccn.loc.gov/2020048919
LC ebook record available at https://lccn.loc.gov/2020048920

ISBN: 978-0-367-43136-5 (hbk)

ISBN: 978-0-367-75969-8 (pbk)
ISBN: 978-1-003-00562-9 (ebk)

Typeset in Times
by Deanta Global Publishing Services, Chennai, India
Contents
Preface......................................................................................................................vii
Acknowledgements....................................................................................................ix
Editor biographies......................................................................................................xi
List of Contributors................................................................................................. xiii

Chapter 1 Advances in Large-Scale Systems Simulation Modelling

Using Multi-agent Architectures Optimized with Artificial
Intelligence Techniques for Improved Concurrency-Supported
Scheduling Mechanisms with Application to Wireless
Systems Simulation............................................................................... 1
P.M. Papazoglou and Dimitrios Alexios Karras

Chapter 2 Let’s Find Out: Why Do Users React Differently to

Applications Infused with AI Algorithms?......................................... 59
Jayden Khakurel, Indu Manimaran, and Jari Porras

Chapter 3 AI vs. Machine Learning vs. Deep Learning .................................... 73

R. Lalitha

Chapter 4 AI and Big Data: Ethical Reasoning and Responsibility.................... 83

Sweta Saraff

Chapter 5 Online Liquid Level Estimation in Dynamic Environments

Using Artificial Neural Network......................................................... 93
Thulasi M. Santhi and S. Sathiya

Chapter 6 Computer Vision Concepts and Applications................................... 111

Bettina O’Brien and V. Uma

Chapter 7 Generative Adversarial Network: Concepts, Variants, and

Applications....................................................................................... 131
K. Rakesh and V. Uma

v
vi Contents

Chapter 8 Detection and Classification of Power Quality Disturbances in

Smart Grids Using Artificial Intelligence Methods.......................... 149
Gökay Bayrak and Alper Yılmaz

Chapter 9 Robust Design of Artificial Neural Network Methodology to

Solve the Inverse Kinematics of a Manipulator of 6 DOF................ 171
Ma. del Rosario Martínez-Blanco,
Teodoro Ibarra-Pérez, Fernando Olivera-Domingo,
and José Manuel Ortiz-Rodríguez

Chapter 10 Generative Adversarial Network and Its Applications...................... 211

A. Vijayalakshmi and Deepa V. Jose

Chapter 11 Applications of Artificial Intelligence in Environmental Science.... 225

Praveen Kumar Gupta, Apoorva Saxena, Brahmanand
Dattaprakash, Ryna Shireen Sheriff,
Surabhi Hitendra Chaudhari, Varun Ullanat,
V. Chayapathy

Chapter 12 A Genetic Algorithm-based Artificial Intelligence Solution for

Optimizing E-Commerce Logistics Vehicle Routing....................... 241
Suresh Nanda Kumar and Ramasamy Panneerselvam

Chapter 13 Application of Machine Learning for Fault Detection and

Energy Efficiency Improvement in HVAC Application.................... 263
Umashankar Subramaniam, Sai Charan Bharadwaj, Nabanita
Dutta, and M. Venkateshkumar

Chapter 14 Smart City Using Artificial Intelligence Enabled by IoT.................. 279

P. Srividya and Sindhu Rajendran

Chapter 15 AI Emerging Communication and Computing................................. 293

N. Girija and T. Bhuvaneswari
Index.......................................................................................................................307
Preface
Artificial Intelligence (AI): Recent Trends and Applications covers AI technologies
and highlights the prospects of AI to problems in real-time day-to-day applications.
In short, AI is for everyone, everywhere, and in everything.
Artificial intelligence has embarked its niche in all fields of engineering and the
dissemination is seemingly important in interdisciplinary research. This book serves
as a background for covering the basic principles of AI; it elaborates the concepts of
artificial intelligence to counterfeit human intelligence, which requires learning and
reasoning.
The book encompasses both the theoretical aspects and recent trends of AI incor-
porated with deep learning strategies towards Smart city, Smart Grid, image process-
ing and computer vision, robotics, instrumentation and automation, etc. The readers
will benefit from the exhaustive range of applications in all fields of engineering
presented in this book.

vii
Acknowledgements
We would like to recognize CRC Press for providing us with the opportunity and
the professionalism shown by teams along the process. Namely, we would like to
thank Ms. Erin Harris and Keith Emmanual Arnold for all the support and leadership
­during the course of publishing this book.
A word of appreciation and recognition to all the authors who have contributed
to this book. Also, a special recognition and our gratitude to all the reviewers who
made an extraordinary effort to enhance the quality of the final chapters of this
book.
Kanimozhi Suguna S has a great privilege to thank the Vice Chancellor,
Dean – School of Computing and other fellow colleagues of SASTRA Deemed
University, Thanjavur, for their assistance in and encouragement in throughout the
progress of the publication. She would like to extend her gratitude to her husband,
Shyam Sundar M, son Dharun Krishna S, and parents Gunavathi K and Subramanian
G. She would also like to thank her brother Arun Prasad S and all his family mem-
bers for their continuous support.
Dhivya wishes to thank Dr. Mohan Manghnani, Chairman, New Horizon Edu­
cational Institutions and Dr. Manjunatha, Principal, New Horizon College of
Engineering, for their whole-hearted cooperation and great encouragement in all
endeavours. She also wishes to thank her parents for their continuous support and
motivation.
With great respect and honour, Sara Paiva would like to thank the Instituto
Politécnico de Viana do Castelo, Portugal and all her colleagues. Special thanks
go to her husband Rogério Paiva, children Diana and Leonardo Paiva, and also her
parents for all their support.
Obviously, a big word of appreciation goes to our families, our main source of
inspiration and strength, namely, in these difficult times we have lived in the last
year. Dedicated to The Almighty – The Supreme Power That Drives Us.

Dr. S. Kanimozhi Suguna

Dr. M. Dhivya
Dr. Sara Paiva

ix
Editor biographies
S. Kanimozhi Suguna received her B.Sc., M.C.A., M.Phil., and Ph.D., in 2005,
2008, 2009, and 2015, respectively. She has seven years of teaching experience and
is working as Assistant Professor in the Department of CSE, School of Computing,
SASTRA University, Thanjavur, India since 2017. Her research interests include
image processing, data mining, web mining, and wireless sensor networks. She
is an editorial member in IGI: Global and Editorial Advisory Board in various
International Conferences. She is a member of IEEE and CSI. She has published 4
book chapters, 12 international journals, and 5 conference presentations.

M. Dhivya received her B.E., M.E., and Ph.D. degree in Electrical Engineering in
2006, 2008, and 2013, respectively. At present, she is working with New Horizon
College of Engineering, Bangalore. She has worked as Associate Professor in the
Department of Electronics and Communication Engineering at Dr. N.G.P. Institute
of Technology. She has received a grant of Rs. 4,00,000 from Texas Instruments
for organizing four-day faculty empowerment programme on “Analog, Power,
Embedded Systems and Wireless (IoT).” She has guided 30 postgraduate scholars
in several disciplines: Applied Electronics, Electric Drives and Embedded Control,
Power Electronics and Drives and Control, and Instrumentation. Her research inter-
ests are wireless sensor networks, embedded systems, real-time systems, optimiza-
tion techniques, and machine learning.

Sara Paiva is Associate Professor at the Polytechnic Institute of Viana do Castelo.

She received her Ph.D. in Informatics Engineering from University of Vigo. She is
also a postdoctoral researcher at the University of Oviedo since January 2018; her
field of research is advanced driving assistants and urban mobility. Paiva coordi-
nates ARC4DigiT, the Applied Research Center for Digital Transformation (http://
arc4digit​.ipvc​.pt), created in January 2018. Her main line of research is urban mobil-
ity solutions. She has edited five books for Springer and IGI and other six are ongo-
ing. She has contributed more than 40 publications to several international journals,
conferences, and books. She is a frequent reviewer of journals and international con-
ferences and supervises several final projects at Bachelor’s and Master’s levels in her
main line of work.

xi
Contributors
Alper Yılmaz Indu Manimaran
Department of Electrical and Electronics Deaprtment of Computer Science
Bursa Technical University California State University
Turkey Long Beach, California

Bettina O’Brien Jari Porras

Department of Computer Science Department of Software Engineering
Pondicherry University LUT University
India Lappeenranta, Finland

Bhuvaneswari T. Jayden Khakurel

Department of CS&A Department of Child Psychiatry
Queen Mary’s College University of Turku
Chennai, India Turku, Finland
Dimitrios Alexios Karras
José Manuel Ortiz-Rodríguez
General Department
Laboratorio de Innovación y
National and Kapodistrian University of
DesarrolloTecnológicoenInteligencia
Athens
Artificial
Evia, Greece
Universidad Autónoma de Zacatecas
and
Zacatecas, México
Computer Engineering Department
Epoka University
Tirana, Albania Lalitha, R.
Department of Computer Science and
Fernando Olivera-Domingo Engineering
UnidadProfesionalInterdisciplinaria de Rajalakshmi Institute of Technology
Ingeniería Campus Zacatecas Chennai, India
InstitutoPolitécnicoNacional
Zacatecas, México Ma​.d​el Rosario Martínez-Blanco
Laboratorio de Innovación y
Girija Narasimhan DesarrolloTecnológicoenInteligencia
Department of Information Technology Artificial
Higher College of Technology Universidad Autónoma de Zacatecas
Muscat, Oman Zacatecas, México

Gökay Bayrak Nabanita Dutta

Department of Electrical and Department of Energy and Power
Electronics Electronics
Bursa Technical University VIT University
Turkey Vellore, India

xiii
xiv Contributors

Papazoglou, P. M. Teodoro Ibarra-Pérez

General Department Laboratorio de Innovación y
National and Kapodistrian University of DesarrolloTecnológicoenInteligencia
Athens Artificial
Evia, Greece Universidad Autónoma de Zacatecas
Zacatecas, México
Rakesh K.
Department of Computer Science Thulasi M. Santhi
Pondicherry University Deaparment of Instrumentation and
India Control Engineering
Dr. B R Ambedkar National Institute of
Sai Charan Bharadwaj Technology
Department of Energy and Power
Jalandhar, India
Electronics
VIT University
Vellore, India Udendhran Mudalyar
Department of Computer Science and
Sathiya S. Engineering
Department of Instrumentation and Bharathidasan University
Control Engineering Tiruchirappalli, India
Dr. B R Ambedkar National Institute of
Technology Uma V.
Jalandhar, India Department of Computer Science
Pondicherry University
Sindhu Rajendran India
Department of Electronics and
Communication Uma Shankar Subramaniam
R. V. College of Engineering Renewable Energy Lab
Bangalore, India Prince Sultan University Riyadh
Saudi Arabia
Srividya P.
Department of Electronics and
Communication Venkateshkumar M.
R. V. College of Engineering Department of EEE
Bangalore, India Amrita Vishwa Vidyapeetham
University
Sweta Saraff Chennai, India
Department of Behavioral Sciences
Amity Institute of Psychology and Vijayalakshmi A.
Allied Sciences Department of Computer Science
Amity University Kolkata Christ(Deemed to be University)
India Bangalore, India
 1 Advances in Large-Scale
Systems Simulation
Modelling Using Multi-
Agent Architectures
Optimized with
Artificial Intelligence
Techniques for
Improved Concurrency-
Supported Scheduling
Mechanisms with
Application to Wireless
Systems Simulation
P.M. Papazoglou and Dimitrios Alexios Karras

CONTENTS
1.1 Literature Review..............................................................................................3
1.1.1 Simulation Methodologies Applied in Wireless Communication
Systems (WCS)......................................................................................3
1.1.1.1 Simulation of WCS................................................................. 3
1.1.1.2 Discrete Event Simulation....................................................... 4
1.1.1.3 Event Scheduling....................................................................7
1.1.2 Channel Assignment in WCS.............................................................. 10
1.1.3 Multi-Agent Systems in WCS.............................................................. 12

1
2 Artificial Intelligence (AI)

1.1.3.1 Agent and Multi-Agent Systems........................................... 12

1.1.3.2 Multi-Agent Systems in WCS............................................... 13
1.1.4 The Concept of Cellular Network........................................................ 14
1.1.5 Simulation Languages (SLs)................................................................ 15
1.2 The Proposed Simulation Model..................................................................... 16
1.2.1 Network Structure............................................................................... 16
1.2.1.1 Operational Parameters........................................................ 17
1.2.2 Modelled Network Services and Channel Allocation......................... 17
1.2.2.1 Network Services.................................................................. 17
1.2.2.2 Channel Allocation............................................................... 18
1.2.2.3 Traffic Generation................................................................. 19
1.2.3 The Multi-Agent/Multilayered Model.................................................20
1.2.4 Theoretical Analysis of Agents Adapted to Modelled Network
Services................................................................................................ 22
1.2.4.1 Network Agent Definition..................................................... 22
1.2.4.2 Architecture of the Intelligent Network Agents................... 23
1.2.4.3 Network Agent Interface.......................................................25
1.2.4.4 Network Agents Which Maintain State................................ 27
1.2.4.5 Network Agent Utility Functions.......................................... 27
1.2.4.6 Multi-Agent Encounters........................................................28
1.2.5 Event Interleaving as Scheduling Technique Based on Real-
Time Scheduling Theory..................................................................... 29
1.2.5.1 Real-Time Scheduling Algorithms for Implementing
Synchronized Processes or Events........................................ 29
1.2.5.2 Process Life Span in a Real-Time Scheduling Set-Up......... 30
1.2.5.3 Scheduling Concurrent Events in WCS................................ 30
1.2.5.4 Response Time Analysis....................................................... 31
1.2.5.5 Pre-emptive Stationary Priority Scheduling (PSPS)............34
1.2.6 Supported DCA Variations..................................................................34
1.2.6.1 The Conventional Unbalanced Variation (Classical DCA)...34
1.2.6.2 The Conventional Balanced Variation (Min Cell Congestion)..35
1.2.6.3 The Conventional Best CNR Variation................................ 35
1.2.6.4 The Conventional Round Blocking Variation....................... 36
1.2.6.5 The Proposed Novel Artificial Intelligence Based
Balanced and Best CNR DCA Variation for
Concurrent Channel Assignment.......................................... 37
1.2.7 Implementation Architectures............................................................. 38
1.2.7.1 Conventional Model.............................................................. 38
1.2.7.2 Concurrent Models............................................................... 38
1.3 Simulation Model Evaluation.......................................................................... 41
1.3.1 Network Behaviour.............................................................................. 41
1.3.2 Monte Carlo Simulation Method......................................................... 42
1.3.3 Simulation Model Behaviour............................................................... 42
1.3.4 Results Accuracy................................................................................. 42
1.3.5 Reference Analysis Model Employing One Cell Only........................ 43
Large-Scale Multi-Agent Systems Simulation  3

1.4 Experimental Results....................................................................................... 45

1.4.1 Indicative Results Based on Five Days of Network Operation............ 45
1.4.2 Model Behaviour Based on Architectural Variations......................... 45
1.4.3 Scheduling Mechanism Comparison................................................... 45
1.4.4 Response Time Analysis Results......................................................... 45
1.5 Conclusions and Future Work.......................................................................... 53
References................................................................................................................. 53

1.1 LITERATURE REVIEW
1.1.1 Simulation Methodologies Applied in Wireless
Communication Systems (WCS)
1.1.1.1 Simulation of WCS
A real wireless network’s efficiency and behaviour can be tested using simulation
systems without the need for field experiments and prototype creation. The simula-
tion solutions give the opportunity to grow to a desired wireless network channel
allocation schemes, network architectures, etc. The simulation software development
approach becomes a very critical issue influencing the resulting network model and
efficiency, due to the complexity of real wireless networks. A big challenge for wire-
less network simulation is the discovery of a way to tackle the actual actions of the
network and not just speed up execution time using parallel machines. The simula-
tion model and environment structure affect the performance of simulated wireless
networks, and for this reason the design and development of such systems is studied
thoroughly. Modern simulation tools provide network engineers with the opportu-
nity to develop and test wireless communication systems at low cost very quickly.
There are three major simulation techniques (Chaturvedi, A., et al. 2001): discrete
event simulation (DES), system dynamics, and multi-agents. The most widely known
simulation tools are based on the DES concept and use various model architectures
to implement. A more accurate and reliable simulation environment can be developed
with the help of efficient model architectures (Chaturvedi, A., et al. 2001; Liu, W.,
et al. 1996; Zeng, X., et al. 1998; Bajaj, L., et al. 1999; Kelly, O.E., et al. 2000; Liu, et al.
2001; Boukerche, A., et al. 2001; Bononi, L. and D’Angelo, G., 2003) to speed up
execution times. Rapid development of parallel systems with multiple processors has
contributed to more effective simulation execution. On the other hand, the program-
ming technology, and especially the multithreading systems, provides an alternative
approach to simulation model implementation. Inside the simulation model, the actual
network activities must be realistically modelled. In the case of cellular networks,
such operations are the essential network functions such as new call entry, realloca-
tion (handoff) (Krishna, S., et al. 2016), device movement, and call termination. The
simulation model will include sub-models for three types of basic components:

• Facilities by network
• Network operating parameters (e.g. number of cells, base station locations,
allocation channel schemes, etc.)
4 Artificial Intelligence (AI)

• Mathematical network models (e.g. propagation models, statistical distribu-

tions, signal computations, etc.)

The sequential simulation is a very time-consuming method due to the high complex-
ity of the models used (Liu, W., et al. 1996) and thus the parallelization is a critical
issue particularly for wireless networks of large scale (Liu, W., et al. 1996; Zeng, X.,
et al. 1998; Bajaj, L., et al. 1999; Kelly, O.E., et al. 1999; Liu, et al. 2001; Boukerche,
A., et al. 2001; Bononi, L. and D’Angelo, G. 2003). The key goal of parallel simula-
tion is the minimization of the execution time (Zeng, X., et al. 1998; Bajaj, L., et al.
1999). In the process of parallelization, sets of network entities have to be mapped to
multiple processors to achieve load balance (Zeng, X., et al. 1998). Multiple-processor
load distribution is a key issue (Liu, W., et al. 1996; Zeng, X., et al. 1998; Bajaj, L., et
al. 1999). Several features, procedures, and entities, such as geographical area (Kelly,
O.E., et al. 1999; Boukerche, A., et al. 2001), radio channels (Kelly, O.E., et al. 1999;
Boukerche, A., et al. 2001), interference calculations (Kelly, O.E., et al. 1999), mobile
hosts (MH) (Boukerche, A., et al. 2001), and network cells (Boukerche, A., et al.
2001) can be parallelized in parallel implementations. There are important problems
that need to be tackled effectively in order to achieve optimum parallelization, such as
synchronization of processors (Zeng, X., et al. 1998; Bajaj, L., et al. 1999; Boukerche,
A., et al. 2001), load balancing control (Liu, W., et al. 1996; Zeng, X., et al. 1998), and
cycle mapping to processors (Liu, W., et al. 1996; Zeng, X., et al. 1998). The above-
mentioned studies have resulted in two basic points for further investigation:

• How to achieve full parallelization for sophisticated computer systems

• How to model the current network environment and behaviour more effec-
tively using advanced programming principles

Advanced computer and programming technologies must be used to build an effec-

tive simulation model, tailored to the actual network (Papazoglou et al. 2008–2016).

1.1.1.2 Discrete Event Simulation

1.1.1.2.1 The Main Concepts of Discrete Event Simulation
DES (Subramania, S., 2001; Tolk, A. 2012; Barr, R., 2004; Goh, R.S.M. and Thng, I.,
L-J, 2003; Reddy, D., et al. 2006; Schriber, T.J., and Brunner, D.T., 1997; Misra, J., 1986;
Tropper, C., 2002; Balakrishnan, V., et al 1997; Jefferson, D.R., and Barnes, P.D., 2017)
represents the most well-known methodology of simulation, in particular for commu-
nication systems. According to the principle of DES, events occur within the simula-
tion time at discrete points in time. Simulation time progresses based on sequence of
events. These events express the basic actions within the physical network, such as new
call arrival. Each event is created with an exclusive time stamp which is later involved
when running the event. A scheduler defines the processing of events over simulation
time (Mehta et al. 2010; Siow, R., et al. 2004; Chung, K., et al. 1993; Whitaker, P., 2001;
Naoumov, V., and Gross, T., 2003) that picks out events with a minimum time stamp
associated with maximum priority. Therefore, a priority queue is the foundation of the
entire scheduling process. Characterized DES systems can be sequential or parallel.
Large-Scale Multi-Agent Systems Simulation  5

Sequential DES systems are the most common in science community. In such systems,
in a three-step cycle, the scheduling mechanism can be analyzed:

• Dequeue: an incident of minimum time stamp is deleted from the queue

• Run: diagnosis of dequeued case
• Enqueue: entrance into the queue of a newly created event

ns-2, ns-3 (Fall, K., and Varadhan, K., 2015), widely accepted by the scientific com-
munity (Kurkowski, S., et al 2005), use this scheduling mechanism for event execu-
tion. In ns-2, ns-3 (Fall, K., and Varadhan, K., 2015), the scheduler selects the next
earliest incident, executes it until ending it up, and returns to run the next event of
highest priority or with least time stamp among the outstanding events. In ns-2, ns-3
(Fall, K., and Varadhan, K., 2015), only one operation can be executed at any given
time, and so it is a single-threaded simulator. If two or more events are scheduled to
take place (to execute) at the “same time,” their execution is performed in a first come
first scheduled scheme – first dispatched, based on the time stamp of each incidence
and thus, the simulation model’s adaptability to the actual network behaviour, includ-
ing all relevant network physical activities, is strongly based on the scheduling algo-
rithm. On the other hand, parallel implementation of DES systems was introduced in
the literature primarily to achieve speed up of execution. Due to the fact that the main
goal is load distribution and speed-up execution, the parallel DES systems do not
modify the principle of the event scheduling found in sequential DES. The system per-
formance is thus a critical point in terms of execution time (Liu, W., et al 1996; Zeng,
X., et al 1998; Bajaj, L., et al 1999; Kelly, O.E, et al 1999; Liu, et al. 2001; Boukerche,
A., et al 2001; Bononi, L. and D’Angelo, G., 2003). Parallel efficiency plays a major
role in the performance of the entire system (Liu, W., et al. 1996; Zeng, X., et al. 1998;
Kelly, O.E, et al. 1999; Boukerche, A., et al. 2001) and therefore synchronization of
multiple processors is important (Liu, W., et al. 1996; Zeng, X., et al. 1998).

1.1.1.2.2 Simulation Time Definition

The time can be defined in different ways with respect to the real time and simula-
tion time. There are three known definitions of time (Fujimoto, R.M. 2000):

• Physical time length

• Simulator period
• Official clock time

Physical time is the time in the physical (real) system; simulation time is the mod-
elled physical time in the simulation system; and wall-clock time is the simulation
program’s runtime.
The definition of simulation time is given in Fujimoto (2000) as follows:

«Simulation time is defined as a totally ordered set of values in which each value
represents an instant of time in the modelled physical system. Furthermore,
for any two simulation time values T1 representing physical time P1, and
6 Artificial Intelligence (AI)

T2 representing P2, if T1 < T2, then P1 occurs before P2, and (T2 − T1) is
equal to (P2 −P1) K for some constant K. If T1 < T2 then T1 is said to occur
before T2 and if T1>T2, then, obviously T1 happened after T2.»

1.1.1.2.3 Efficient Usage of Computer Technology for

Supporting Parallel DES Systems
The availability of computer technology and tools among academic and research
institutes led to the development of advanced simulation systems that effectively
confront the complexity and time of execution. Scientific community and industry
have a great interest in that field. In the case of speeding up the simulation execution
time, INTEL (Liu, W., et al. 1996) and DARPA (Zeng, X., et al. 1998; Bajaj, L., et
al 1999) supported several projects in educational institutions such as UCLA, but
according to Fujimoto et al. (2003), the effective simulation of large-scale networks
is still an unresolved problem. In general, parallel computing can be categorized
according to system features and capabilities (e.g. architecture). Multiprocessing,
computer cluster, parallel supercomputer, and distributed computing are the parallel
computing-based categories. In parallel implementation of DES systems, two dif-
ferent strong technologies were used: clusters and multiple processor machines. A
cluster usually consists of two or more workstations connected via a local network
(Baker M., et al 2000). It is possible to view this group of loosely coupled computers
as one single computer. The most common cluster types are high-availability clus-
ters and high-performance clusters. High-availability clusters are used to improve
a cluster’s availability of the services it offers. High-performance clusters are used
mainly by splitting a computational function into more than one processing unit
to provide improved performance. One of the most common cluster projects is the
Beowulf project used in over 50 educational institutes and research centres. The
parallel computing system forms a computer of multiple processors. The process
to solve a problem in this scheme can be divided into smaller tasks performed by
different processors. There are several critical issues in parallel computing such as
synchronization and parallel overhead which affect the performance of the whole
system. Many strategies for execution time minimization were applied based on the
model’s thorough decrease (Liu, W., et al. 1996). If the parallelization is sufficiently
effective, the reduction of details of the model is not necessary. Also, in a paral-
lel system with N = P (N = partitions of the layout, P = number of processors), there
must be efficient synchronization between processors. Effective parallelization can
be accomplished by eliminating global variables and unnecessary model synchro-
nizations, effective partitioning (locality of exploitation and balance of loads), etc.
A range of workstation processors with a total of 16 processors is used in simula-
tion experiments of (Liu, W., et al. 1996) and (Zeng, X., et al. 1998). The use of 16
processors for over 200 network nodes provided major acceleration. The partitioning
scheme plays a major role in the resulting speedup according to Zeng et al. (1998).
In Kelly et al. (1999), interference calculations are parallelized and sub-geographic
points and sub-matrix zones for path losses are mapped as entities in one processor
(Kelly, O.E., et al. 1999). Multiple channel zones are mapped to multiple processors
Large-Scale Multi-Agent Systems Simulation  7

according to Kelly et al. (1999), and significant speedup is achieved using up to eight
processors. For the effective simulation of wireless and mobile networks, a cluster
of workstations is used (Boukerche, A., et al. 2001). Quite impressive speedup is
accomplished using up to 16 processors based on Mobile Host partitioning. For par-
allel DES, a networked cluster of PCs use a set of physical execution units (Bononi,
L. and D’Angelo, G., 2003).
1.1.1.3 Event Scheduling
1.1.1.3.1 The Importance of Event Scheduling
Events, being the main entities of a DES system, express the physical activities of
a real-world wireless network. With respect to the aforementioned network facili-
ties, each service could be considered an incidence for a particular mobile user
(MU). During simulation time, an incidence generator creates events as, for instance,
new call arrivals in the system. Such an event is marked with a time stamp (within
simulation time) that specifies the starting execution point in time. A critical entity
within the simulation system, namely, the scheduler, is executing by selecting the
next earliest event, based on its associated time stamp, running it until ending it up,
and returning to execute the next event expressed by the next least time stamp. The
scheduling system is a standard tool within the actual wireless network for the event
service incidence sequence control. When an operation is not executed, the pending
operation set (PES) (Goh, R.S.M. and Thng, I., L-J, 2003; Tan, K.L., and Thng, L.-J.,
2000; Siangsukone, T., et al. 2003) is the set of all non-simulated and processed
tasks generated during simulation time. Consequently, PES corresponds to a priority
queue controlling the stream of events simulation based on the contemporary mini-
mum time stamp, that is the task of highest priority (Goh, R.S.M. and Thng, I., L-J,
2003). The selected scheduling method imposes how the actual network activities
that occurred can be reflected in the simulation model in a realistic way. Therefore,
scheduling could be viewed as a generalized mapping method within the simulation
time of the DES algorithm for the real network events-activities execution.

1.1.1.3.2 The Current Standard Event Scheduling

Mechanism: Calendar Queue (CQ)
The major application of priority queuing is the realization of the PES sets within
the DES simulation systems (Blackstone, J.H., et al 1981; Henriksen, J.O., 1977;
Kingston, J. H., 1984; McCormack, W.M. and Sargent, R.G., 1981). The CQ concept
was first introduced by Brown (1988), has been analyzed in Siangsukone et al. (2003),
and even nowadays is established as the most popular scheduling scheme among
the implemented DES systems such as ns-2, ns-3 (Berkeley, USA) (Fall, K., and
Varadhan, K., 2015), Ptolemy II (Berkeley, USA) (Whitaker P. 2001), Jist (Cornell
University, USA) (http:​/​/jis​​t​.ece​​.corn​​ell​.e​​du​/ja​​vadoc​​/inde​​x​.htm​​l​?jis​​t ​/run​​time/​​Sched​​u​
ler.​​Calen​​dar​.h​​tml). Numerous other variations of the CQ aiming at increasing per-
formance of the queuing system, as for instance by optimally resizing the queue,
have been proposed in the literature such as DSplay (Siow, R. et al. 2004), MList
(Goh, R.S.M. and Thng, I., L-J, 2003), Markov hold model (Chung, K., et al. 1993),
and SNOOPY CQ (Tan, K.L., and Thng, L.-J., 2000). The CQ concept is derived
8 Artificial Intelligence (AI)

from the common desktop calendar of one-page entry every day. Each incident is
scheduled on the relevant calendar page. Each event’s schedule timing defines its
priority. If an event on the calendar is enqueued, then this event is scheduled for
future running. The earliest event within the calendar is dequeued by searching it
within the relevant page of the current date and removing it from that page (Brown,
R. 1988).
A CQ scheme is considered to consist of an array of lists. Each such list queue
contains future incidences. According to the CQ principle, a large queue of N events
is partitioned to M shorter queues called Buckets. Each bucket is associated with a
specific time range corresponding to future events executions. Any incidence bound
to an occurrence time t(e) is associated with the mth bucket in year y (y = 0, 1, 2, …)
if and only if

(
ë )
t (e) Î é ( yM + m ) d , ( yM + m + 1) d ù (1.1)
û
In order to find the associated relevant bucket number m(e) where an incidence e will
occur at time t(e), the following type is involved:

ê t ( e) ú
m (e) = ê ú mod M (1.2)
ë d û

Consider that N = 10, δ = 1, M = 8, and t(e) = 3.52 (Figure 1.1) for a new incidence e.
Using Equation 1.2, the relevant bucket number for the incidence e is m(e) = 3.
An analytical study of the CQ scheme can be found in Erickson et al. (1994).

1.1.1.3.3 Application of CQ in Sequential DES

Figure 1.2 demonstrates a complete CQ operation in a sequential system of DES. The
CQ initially includes six events (step 0) that are to be executed at a later time unit.
The event e1 (NC – New Call), with time stamp 3.62, will be dequeued first as it has
the highest priority among the queue buckets, having acquired the minimum time
stamp. A novel incidence with time stamp 9.98 is enqueued in step (3), and in step (4)
the event e2 (RC – reallocation of a call) is dequeued. The entire CQ function follows
the Dequeue-Run-Enqueue protocol. In any step of the entire operation, this rule can
be evaluated (e.g. Dequeue{step(9)}-Run{step(10)}- Enqueue{step(11)}).

FIGURE 1.1  Illustration of a CQ implementation with eight buckets.

Large-Scale Multi-Agent Systems Simulation  9

FIGURE 1.2  Sample of a Sequential DES operation (events/facilities: NC – new call arrival,
RC – call reallocation/handoff, MC – movement of user running a call, FC – terminated call).

The generator of the incidences creates events with corresponding time signs. The
dequeued incidences are running either in rising priority order or in increasing time
stamp order (Figure 1.3). The final processes running is sequential oversimulation
time except for events of nearly identical time stamps.

1.1.1.3.4 Application of CQ in Parallel DES

Where a key goal is to speed simulation execution, parallel models and implementa-
tions are used.
Speeding up can be performed in a parallel DES system if the work load of the
total processing is spread over individual processors. Figure 1.4 shows a scenario
with a cellular network that is divided into four clusters. Each cluster contains seven
uniquely frequently used cells. A separate processor is assigned to the processing
load for simulation of all possible MU resources in each cluster.
In parallel DES, the implementation of just one central CQ is an incorrect selec-
tion since, given the multiple processor availability, the time stamps of the enqueued
events compel the entire system to be sequential. This is because all activities in any
cell must be carried out in a particular order based on their associative time stamps.
10 Artificial Intelligence (AI)

FIGURE 1.3  Event running following the dequeuing order of the processing unit (events/
facilities: NC – new call arrival, RC – call reallocation/handoff, MC – movement of user run-
ning a call, FC – terminated call).

The events belonging to various clusters can be executed in parallel according to

Figure 1.5. All the events are running serially within each processor and its associ-
ated cluster. Many more processors are required to obtain true competition. Parallel
sets of sequential events can run in parallel within this scheme. Therefore, only
selected rivalry can be accommodated within this framework, and only for services
that belong to different network partitions.

1.1.2 Channel Assignment in WCS

A cellular system’s capability may be defined in terms of the number of channels
available, or the number of users that the network can support at the same time. The
total number of channels made available to a system depends on each channel’s allot-
ted spectrum and bandwidth. The frequency spectrum available is limited and the
number of MUs is increasing day by day, therefore the channels have to be reused
as much as possible to increase the system capacity. One of the basic resource man-
agement issues in a mobile communication system is allocating channels to cells or
mobile phones. A channel allocation scheme has the function of allocating channels
to cells or MUs in a way that minimizes:

(A) the probability of discontinuing incoming calls;

(B) the probability of dropped calls, and
(C) the probability of any call carrier-to-interference ratio falling below the
prespecified value (Papazoglou et al. 2008–2016).
Large-Scale Multi-Agent Systems Simulation  11

FIGURE 1.4  Excerpt of spatial partitioning/parallelization.

Many channel allocation schemes were widely invested in literature with a goal
of maximizing reuse of frequency. Channel allocation schemes are classified into
three strategies: fixed channel allocation (FCA) (Zhang, M. and Yum, T.S., 1989;
Lai, W.K. and Coghill, G.C., 1996; MacDonald, V.H., 1979; Elnoubi, S.M., et al
1982; Xu, Z. and Mirchandani, P.B., 1982); dynamic channel allocation (DCA)
(Zhang, M. and Yum, T.S., 1989; Cimini L.J. and Foschini, G.J., 1992; Cox, D.C.
and Reudink, D.O., 1973; Del Re, E, Fantacci, R, and Giambene, G. 1996; Kahwa,
TJ, and Geor-gans, ND. 1978; Papazoglou, PM, and Karras, DA. 2016; Papazoglou,
PM, Karras, DA, and Papademetriou, RC. 2008a; Papazoglou, PM, Kar-ras, DA,
and Papademetriou, RC. 2008b; Papazoglou, PM, Karras, DA, and Papademetriou,
RC. 2009; Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2011; Papazoglou,
PM, Karras, DA, and Papademetriou, RC. 2016; Sivarajan, KN, McEliece, RJ, and
Ketchum, JW. 1990; Subramania Sharma, T, and Thazhuthaveetil, MJ. 2001) and
HCA (Hybrid Channel Allocation).
12 Artificial Intelligence (AI)

FIGURE 1.5  Excerpt of different CQs based on network spatial partitioning (events/facili-
ties: NC – new call arrival, RC – call reallocation/handoff, MC – movement of user running
a call, FC – terminated call).

In FCA, every cell is permanently assigned a set of channels based on a pre-

estimated traffic intensity. No permanent allocation of channels to cells occurs
in DCA. Rather, all the channels available are open to all the cells, and the
channels are allocated in a dynamic manner on a call-by-call basis. The FCA
scheme is simple, but does not adapt to changing traffic and user delivery condi-
tions. In addition, frequency planning in a microcellular environment becomes
more difficult because it is based on accurate knowledge of traffic and interfer-
ence conditions. DCA overcomes these shortcomings however; under heavy load
conditions, FCA outperforms most known DCA schemes. HCA combines the
features of both the FCA and DCA techniques to overcome the drawbacks of
FCA and DCA.

1.1.3 Multi-Agent Systems in WCS

1.1.3.1 Agent and Multi-Agent Systems
An agent is considered to be a computational machine that autonomously interacts
with its environment and works for the reason for which it was programmed (Maes,
P., 1995), (Papazoglou et al. 2008–2016). Agents may be approached as organizations
committed to a specific task and smaller than a standard application (Smith, D.C.,
et al. 1994). An agent also perceives the conditions of the world, behaves according
to these conditions, interprets perceptions, and addresses issues (Hayes-Roth, B.,
1995). Mobile agents are used in a distributed environment such as the internet and
offer flexible networks customized to MU needs, and use the available bandwidth
effectively. There are certain basic attributes which distinguish agents from other
programs. Such qualities of greatest significance are as follows:
Large-Scale Multi-Agent Systems Simulation  13

• Adaptability that reflects a learning and development potential centred

on growing experience. This attribute is, in other words, about the shift
of agent according to external or internal events (Splunter, S., et al. 2003;
Russell, S., and Norvig, P., 2002).
• Independence in the acts needed to achieve predefined goals. Growing
agent has the power over its own behaviour according to this characteristic
(Huhns, M., and Singh, M., Eds. 1998; Norman, T., and Long, D., 1995;
Ekdahl, B., 2001).
• Partnering with other stakeholders to reach mutual goals.
• Interactivity with surroundings.

According to Jennings (2000), a multi-agent system (MAS) has a number of agents

that interact with each other through communication. These agents function in an
environment and have different areas of environmental control. Many areas of influ-
ence can be coincided within the environment. In modern problems, the modelling
of all system aspects requires a number of agents, and so the MASs emerged. MASs
can be described as a loosely connected network of problem-solver entities (Sycara,
K., 1998) working together with the shared aim of solving the whole problem beyond
each individual entity’s solving capabilities. Agent technology has been used in the
management of telecommunications networks in many instances (Hayzelden, A.,
and Bigham, J., 1999; Bodanese, E.L, 2000; Iraqi, Y. and Boutaba, R., 2000; Jiang,
Y., and Jiang, J., 2008). In these cases, the architecture is straightforward. Modelling
of cellular systems and particularly the methodology of simulation have not been
presented in terms of MAS so far. Their simulation is based primarily on sequential
or parallel models (for faster execution) but not on models of the agents. According to
Wooldridge (2004), negotiation approaches are based on other components, such as
negotiation sets (possible agents’ proposals), protocols (definition of legal proposals),
tactics selection (usual private), etc.
Growing agent has its own view of the environment and also personal objec-
tive and resource claim in a competitive or collaborative environment. Negotiation
between the agents is a required re-requirement for achieving the goal and satisfy-
ing the specified con-stresses. Negotiations may be challenged as competitive or
cooperative (Zhang, X., et al. 2001). The negotiators are self-interested in successful
deals to achieve their own objectives. The negotiators are working together in joint
discussions to find a compromise for a shared target.

1.1.3.2  Multi-Agent Systems in WCS

For the issue of resource distribution, the MAS technology has been used in many
studies (Papazoglou et al. 2008–2016). Different network elements, such as Base
Stations (BSs), cells, etc. were modelled as agents within the established models. An
overview of communication systems agent technology is presented in Hayzelden and
Bigham (1999). This overview is focused on software agents used for the manage-
ment of communications. More specifically, agents can be used to deal with some
important issues, such as complexity of the network, mobility of the MU, and network
management. A MAS for the control of resources in multimedia wireless mobile
14 Artificial Intelligence (AI)

networks is provided in Iraqi and Boutaba (2000). The call dropping likelihood is
small based on the proposed MAS (Iraqi, Y. and Boutaba, R., 2000), although the
wireless network makes a high average usage of the bandwidth. According to Iraqi
and Boutaba (2000), the final decision to accept a call is based on neighbouring cells
participating. Therefore, every cell or BS has an agent working. A cooperative agree-
ment in a MAS is defined in Jiang and Jiang (2008) to help real-time load balancing
of a mobile cellular network. Within the current MAS, agents are used (a) for market
representation of different service providers and (b) for network operators handling
the radio capital of different network regions. According to Jiang and Jiang (2008),
the coordination of agents is done by messages, and negotiations are focused on the
final agreements. Negotiation is an intelligent control strategy in the above study to
dynamically change the shape and size of the cells and to manage the traffic load
over the network.
To suggest a distributed channel allocation scheme using intelligent software
agents, a detailed simulation model for wireless cellular networks was developed
by Bodanese (2000). In the above analysis, smart collaborative software agents
offer BSs autonomy, enhance network robustness, enable network resource negotia-
tion, and improve resource allocation. To this end, several aspects of infrastructure
and operation of the cellular network have been modelled. According to Bodanese
(2000), the basic features implemented in the model are call set-up, handoffs based
on signal strength and defined threshold, forced call termination emulation (due
to insufficient signal), Mobile Station self-location based on signal strength from
nearest BSs, Mobile Stations with arbitrary trajectories, cell boundaries, number of
Mobile Stations per unit. Four specific aspects of the configuration and function of
the entire network were modelled in Bodanese (2000). Those aspects include (a) the
network model defining the network layout (cell structure, BSs, MSs), (b) the mobile
station (signal calculations, call generation inter-arrival time, call length), (b1) call
server (SNR measurements), (b2) handoff, (b3) call dropping, (b4) call termination,
(b5) filter (sending packets from the radio link to the server), (b6) trajectory track-
ing, which is responsible for Mobile Station movement (random cyclical, driving on
a highway, etc.), (c) BS which includes SNR measurements, handoff (monitoring the
received voice packets), etc., (d) MTSO (Mobile Telephone Switching Office) which
offers services such as HLR (Home Location Register), VLR (Visitor Location
Register), etc., (d1) keeps track of each active MS. All the implemented network
models in Bodanese (2000) do not support multimedia real-time services. A novel
modelling methodology for supporting wireless network services is also described
in the current report, based on MAS technology.

1.1.4 The Concept of Cellular Network

The cellular principle divides the geographic area covered into a series of smaller
service areas called cells. That cell is fitted with a BS and a number of mobile ter-
minals (e.g. mobile phones, palms, laptops, etc.). The BS is fitted with reception and
radio transmission equipment. The mobile terminal inside a cell communicates with
the cell-related BS via wireless links. A number of BSs are linked via microwave
Large-Scale Multi-Agent Systems Simulation  15

links or dedicated leased lines to the Base Station Controller (BSC). The BSC pro-
vides logic for the management of the base stations’ radio resources under its control.
As a MU moves from cell to cell, it is also responsible for transferring an ongoing
call from one base station to another. A mobile terminal must first obtain a channel
from the BS in order to establish a communication with a BS. A channel consists of
a pair of signal transmission frequencies between the BS and MU.

1.1.5 Simulation Languages (SLs)

The great interest in simulation applications led to the development of specialized
Simulation Languages (SLs) that offered the possibility of a developed model being
simulated quickly. The biggest advantage of the SLs is the significant reduction in
time for programming the model. This advantage, contrasting with the program-
ming time, helps the developer to give more time for model analysis and develop-
ment. Modern SLs provide design guidance over production and study. The SLs most
known offer several useful features:

• Are process or event oriented

• Mechanism simulating time flow
• Generation of random numbers and various distributions
• Fit for debugging devices

On the other hand, SLs have several drawbacks:

• High procurement and maintenance costs

• Time cost for the initial language learning
• High resource requirements in contrast with conventional programming
languages (computer power, memory)

In order to pick every SL correctly, multiple variables need to be taken into account.
Those characteristics include:

• SL-accessibility as a (free) commercial or academic resource

• SL technical features such as (a) how the SL encounters incidents, activi-
ties, and processes, (b) problem-solving usability, (c) data collection and
analysis capabilities, (d) model extension support, (e) random number and
sample generation, (f) debugging procedures
• The programming effort required
• Computer time
• Specifications for software resources
• The mobile model

Most approaches to simulation models and applications have allowed us using gen-
eral-purpose languages, general SLs, and packages for special-purpose simulations.
Languages of general use have the following principal characteristics:
16 Artificial Intelligence (AI)

• Bottom cost
• Are available
• No supplementary training is required
• Every model starts from the ground up
• Hard to check
• Reusability of code is minimal
• The business process period is a long one

General SLs deliver uniform modelling features, shorter development cycle, and
simple verification. These languages have higher costs, restricted portability, and
require more instruction.
Languages such as Simula (Nierstrasz, O., 1989), GPSS (Karian, Z.A., and
Dudewicz, E.J., 1998), SimScript (Russell, E.C., 1999), Parsec (Subramania, S.,
2001; Zeng, X., et al. 1998; Takai, M., 1999; Reddy, D., et al. 2006; Abar, S., et al.
2017), modsim (West, J. and Mullarney, A., 1998), slam (Miller, J.O., et al. 1988),
GlomoSim (Zeng, X., et al. 1998), and other special purposes of high-flexibility
simulation devices. Other software tools such as MATLAB have plenty of scientific
libraries and capabilities for the creation of portable script code but suffer from low-
performance execution. A large-scale simulation model includes a scalable network
environment with advanced features such as portability, internetworking, and high
network activity adaptability.
The invention and development of new languages for simulation introduces
important drawbacks:

• The new languages are specific to the domain and rarely adapted by the
scientific community.
• The respective libraries require designers to adapt their applications to spe-
cific needs.
• Designers can’t achieve high simulation kernel adaptability in their
applications.

Java language is the most appropriate to develop scalable, portable, and high-perfor-
mance network applications, and is adopted by the majority of the scientific com-
munity. On the other hand, Java supports multiple threads that mean manipulation of
the concurrent tasks. This functionality is very critical for supporting network events
at the same time.

1.2 THE PROPOSED SIMULATION MODEL

1.2.1 Network Structure
The cellular network consists of N cells and each of the cells includes a BS in the
centre position (Figure 1.6).
Each cell is divided into a mesh of spots where users may exist from the start of
the call or because of handoff situations.
Large-Scale Multi-Agent Systems Simulation  17

FIGURE 1.6  Structure of cellular network model.

The basic parameter for mesh building is the fineness, which means the distance
between the spots.

1.2.1.1 Operational Parameters
A set of parameters determine the conditions of network operation, such as signal
propagation, signal rates, and MU behaviour. The parameter basics are: CNR (car-
rier-to-noise ratio) on cell edge (dB), CNIR (carrier-to-noise plus interference ratio)
threshold (dB), average call arrival rate (calls/hour, lambda), average call hold time
(seconds), simulation time, cell channels, path loss factor (alpha), standard shadow-
ing deviation (sigma), cell mesh fineness, etc.

1.2.2 Modelled Network Services and Channel Allocation

1.2.2.1 Network Services
1.2.2.1.1 New Call Arrival (NC)
The number or MUs is increasing, the calls by each MU are restricted, and so the
call arrivals can be considered as random and independent. In the simulation system,
18 Artificial Intelligence (AI)

the new calls result from a random or a Poisson distribution with regard to a pre-
defined regular pattern.

1.2.2.1.2 Call Termination (FC)

For each new MU, the call length is based on an exponential function. The call keep-
ing time for later review is applied to current simulation time. The related protocol
checks the phased call time for linked MUs.

1.2.2.1.3 Reallocation Management-Handoff (RC)

The computations are based on the intensity of the signal and how it affects adjacent
cells from other related MUs. If the CNIR threshold is not reached by a MU signal,
the procedure may attempt to find another suitable source. First, the algorithm calcu-
lates the strength of the signal between MU and BS and then calculates any interfer-
ence from connected MUs in later time. If an approved channel is identified, then the
new MU is allocated, otherwise the call will be dropped.

1.2.2.1.4 MU Movement (MC)

The algorithm locates the MUs which are related and adjusts their current positions.
A movement based on the Gaussian distribution is generated by the MU. This distri-
bution is also used in simulation models similar to that found in literature (Dixit et
al. 2001). The formula of Gaussian distribution probability density function is
1
e( )
2
- x - m / 2s 2
f ( x) = (1.3)
s 2p
where σ is the standard deviation and μ is the baseline expected value. The Gaussian
distribution applied has mean value 0 and standard deviation 1. Consequently, the
probability density function is

1 - x2 / 2
f ( x) = e (1.4)
2p
In the MU movement procedure, firstly a Gaussian number is generated, e.g. x1 and
the corresponding f(x1) whenx1 Î éë -0.5, 0.5ùû . If another Gaussian number, e.g. x2,
where x2 ³ f ( x1 ) , a MU move is generated. The new position of any MU is maxi-
mum two cells distance from the initiated cell.

1.2.2.2 Channel Allocation
1.2.2.2.1 Channel Allocation Criteria
Any channel allocation procedure is successful only if the following criteria are met:

• Channel availability
• Carrier strength (between MU and BS)
• CNR (signal purity)
• Signal to noise plus interference ratio CNIR (interference from other con-
nected MUs)
Large-Scale Multi-Agent Systems Simulation  19

1.2.2.2.2 Channel Allocation Procedure and Calculation

After a new call arrival, several actions take place in turn:

(a) Check if the maximum MU capacity in the cell neighbour has been reached.
(b) Calculate a random MU position in the mesh.
(c) Place the new MU according to the cell’s BS position and mesh spot.
(d) Calculate the signal strength between BS and new MU in the call-initiated
cell.

Firstly, the shadow attenuation (Molisch, A., 2012; Lee, W.C.Y., 1995) is derived.
Next

s ×n
sh = 10 10 (1.5)

where σ is the shadowing standard deviation and n is the associated number from the
normal distribution. The distance attenuation dw can be derived from the shadow
attenuation and distance between MU and BS. Through MU and BS, the CNR could
be computed (Molisch, A., 2012; Lee, W.C.Y., 1995):

cnedge
cn = 10 10 × dw (1.6)

where cnedge is the CNR on cell edge (dB).

(e) Calculate interference among the new MU and other MUs that use the same
channel.
(f) Check if the CNIR ratio is acceptable according to a predefined threshold.
(g) If CNIR is acceptable, establish the new call and update UR, otherwise use
any alternative selected DCA variation.

1.2.2.3 Traffic Generation
When a new call arrival occurs with regard to Poisson distribution, a random number
x is generated from Poisson distribution and the corresponding P(x) is calculated as
follows:

e-l × l x
P( x ) = (1.7)
x!
where λ represents the new call arrival rate. This rate is analogous to the time of the
day. A fixed set of simulation time steps represents a whole day. Each day is divided
into five zones according to traffic conditions. Table 1.1 illustrates the five zones with
corresponding λ parameter. Zone 1 represents the hours 12–9 a.m. (λ = 1) and zone
3 the hours 1–4 p.m. (λ = 5).
20 Artificial Intelligence (AI)

TABLE 1.1
Daily Zones
Zone Percentage
Zone (Approximately) λ Description
1 34% 1 Least busy zone
2 21% 4
3 12% 5 Most busy zone
4 17% 3
5 17% 2

1.2.3 The Multi-Agent/Multilayered Model

The basic events like new call delivery, user movement, call status control etc. will
arise concurrently in a real-world cellular network. As soon as a new user attempts
to be connected to the network through a call, another call could be terminated
at the matching time, another user is moved or reassigned to a new channel, etc.
Hence, these things need to be simultaneously modelled. The main idea is that each
basic event is independent of each other, acting independently, interacting with the
environment, and being able to be implemented as an agent. The entire structure is
divided into three layers in the proposed architecture (Figure 1.7).

FIGURE 1.7  The proposed layered agent architecture.

Large-Scale Multi-Agent Systems Simulation  21

FIGURE 1.8  Three agent sets for three individual MUs.

The fundamental layer (layer 1) embodies the structure of the cellular network,
where the basic events (new call, call rearrangement, user movement, call termina-
tion) take place. Layer 2 is made up of four agents that implement the synchronized
events and describe the behaviour of the network. The control agent (main agent) that
exists in the third layer activates the system with corresponding event agents. The set
of four agents is recurring for every individual MU (Figure 1.8).
The main agent includes a clock logic which is responsible for the layer 2 agents’
activation and ensures that the additional procedures are triggered in the correct
order. The agents communicate for the activation status through layers 2 and 3 and
exchange information (messages) etc. Event agents are independent entities that react
with the appropriate environment and are coordinated by the control agent. The cel-
lular network environment includes parameters such as information about the signal
power, user characteristics, network status, etc. Every agent is implemented as a Java
thread and is defined in relation to layered architecture by state, actions, and position.
Agents in layer 2 are concurrently executed according to signals from the control
node. The simulation time changes only when the layer 2 agents have completed the
concurrent acts. Layer 1 uses shared data objects that are protected from the event
agents in order to simultaneously access (read or write). Each event agent keeps the
22 Artificial Intelligence (AI)

control agent informed about the status of its execution. The control agent is active,
while the time for simulation is not over and so the rest of the agents. This agent has
a sequential step-value clock (e.g. 1,2,3, etc.). A corresponding action is turned on
in each value. All the acts (agents and supplementary procedures) are carried out as
threads. The requisite supplementary activities (preparation tasks) are triggered in
the first phase of the clock, while the other agents and procedures are disabled. In
the second step of the clock, event agents are turned on while other procedures are
disabled. Control agent prevents simultaneous activation within simulation time of
the event agents, initial and final simulation procedures.

1.2.4 Theoretical Analysis of Agents Adapted

to Modelled Network Services

1.2.4.1 Network Agent Definition

According to Wooldridge and Jennings (1995), “An agent is a computer device
located in any setting that is capable of autonomous action in this environment to
meet its design goals.”
An agent communicates with its environment, obtains input information from
it, and carries out a reasonable appropriate behaviour that affects this environment
(Figure 1.9).
NC and RC agents (NCA, RCA) interconnect with the cellular network (envi-
ronment), obtain input information (blocking the probability-network performance,
decreasing RCA probability), and perform certain actions (giving priority to new
calls or handovers for different MUs) affecting network efficiency (Figure 1.10).

FIGURE 1.9  Agent definition.

FIGURE 1.10  Network agent definition.

Large-Scale Multi-Agent Systems Simulation  23

The agents’ main features can be concisely outlined as follows:

• Responsiveness. An agent perceives its environment and reacts to fulfil

the design goals. NCA perceives network efficiency and (a) prioritizes new
calls and (b) negotiates the best deal with RCA to achieve its design goal
of reducing blocking risk (network performance optimization). RCA works
similarly but with probability dropping.
• Pro-activeness. Takes the initiative to demonstrate the goal-oriented
behaviour to achieve the design goals. To get performance benefits for the
network, network agents send messages to other MU agents (NCA to RCA
and vice versa) in order.
• Socio-power. Interaction with other MU agents to attain the concept goals.
NCA works with RCA to reach design goals.

1.2.4.2 Architecture of the Intelligent Network Agents

Let’s assume that the conceivable discrete states of the network setting can be des-
ignated by the set E as

E = {e, e¢,¼} (1.8)

Moreover, let’s assume that the probable discrete states of the wireless network set-
ting up can be defined by the set E as

E = {LL, LH , HL, HH} (1.9)

where the memberships of E signify the network performance, L signifies low level,
and H the high level. The pairs are linked to blocking and dropping probability,
respectively.
On the other hand, it is adopted that each agent has a set of possible interactions
with this set-up. These interactions change the environmental status and are deter-
mined by the set

Ac = {a k 0 , a k1,¼} (1.10)

In the case of network set-up, the matching actions are determined by the set

Ac = {INPNC , INPRC , DePNC , DePRC , DN NC , DN RC } (1.11)

where INP is the action “Increase Priority,” DeP represents the action “Decrease
Priority,” and, finally, DN is the action “Do Nothing.”
The external network set-up changes its state with regard to the above-mentioned
activities. A sequence of actions results in a sequence of environmental state changes.
Thus, a run r of an agent within the environment can be defined as

a0 a1 a2 a3 a u-1
r : e0 ® e1 ® e2 ® e3 ® ® eu (1.12)
24 Artificial Intelligence (AI)

A run r for the cellular network might develop

INPNC , DN RC DN NC , DN RC DN NC , INPRC DN NC , DN RC a u-1

r : HL ® LL ® LH ® LL ® ××× ® eu (1.13)

Let, also, consider the following sets:

R: Set of possible finite sequences (over E and Ac)

R AC: Subset of R that ends with an action
RE: Subset of R that ends with an environment state

A state transformer function has been made known by Fagin et al. (1995) in order to
depict the effect of an agent activity on the network set-up:

t : R Ac ® g ( E ) (1.14)

The previously mentioned function associates a run to a set of probable environmen-

tal states. Assuming that no successor state exists to r, τ(r) is developed as

t ( r ) = Æ (1.15)

The entire environment is defined (states, transformer function) as

Env = E, e0 ,t (1.16)

where Ε is the state set, eo is the initial state, and τ is the transformer function. For
modelling agents, it is adopted that an agent characterizes a function for mapping
runs to activities and thus:

Ag : R E ® Ac (1.17)

In other words, an agent decides about actions (to select an appropriate action to
perform) based on the past events of the system. For expressing the entire system
(agents, environment), a suitable set is developed:

R ( Ag, Env ) (1.18)

Finally, the sequence (e 0 ,a 0 , e1,a1, e 2 ,… )expresses a run of an agent Ag (in the net-
work set-up) Env = E, e0 ,t , if:

a 0 = Ag ( e0 ) (1.19)

A run of agent NCA or RCA is expressed for instance by the sequence

( HL < INPNC >, LL < DN >, LH <,…), if IPNC = Ag ( HL )
Large-Scale Multi-Agent Systems Simulation  25

and for u > 0,

eu Î t ( e0 , a 0 ,..., a u -1 ) (1.20)

a u = Ag ( e0 , a 0 ,..., eu ) (1.21)

In the case of network set-up and for u > 0

eu Î t ( HL < INPNC , DN RC >,¼, a u -1 ) (1.22)

a u = Ag ( HL < INPNC , DN RC >,¼, eu ) (1.23)

where Ag corresponds to NCA or RCA.

1.2.4.3 Network Agent Interface

As mentioned above, an agent perceives environment and acts on it. These two dis-
tinct actions are expressed by two functions separately (Figure 1.11).
Correspondingly, NCA and RCA, perceive network set-up and react on it. These
two distinct actions for NCA and RCA are expressed by two functions, respectively
(Figure 1.12).

FIGURE 1.11  Agents interface.

FIGURE 1.12  Network agents interface.

26 Artificial Intelligence (AI)

The “perceive” function maps environment states to observations and “activity”

maps sequences of perceptions to actions.
As a concrete example of the above framework, let x express the statement “met-
ric M1 is adequate” and let y represent the statement “metric M2 is adequate.” Thus,
the set E contains four combinations of x and y. This set can be expressed as follows:

E= {{x , y} ,{x , y} ,{x, y} ,{x, y}} (1.24)

with

e1 = { x , y } , e2 = { x , y} , e3 = { x, y } , e4 = { x, y} (1.25)

Network behaviour is evaluated through two basic statistical metrics which are the
blocking and dropping likelihoods. Thus, (1.24) and (1.25) will be expressed in terms
of the above metrics as follows:

E= {{B, D},{B, D},{B, D},{B, D}} (1.26)

with

e1 = {B, D} , e2 = {B, D} , e3 = {B, D} , e4 = {B, D} (1.27)

where B represents the statement “Blocking possibility is adequate” and D repre-

sents the statement “Dropping possibility is adequate.” Now, the set E contains four
mixtures of B and D.
With regard to (1.24) and (1.25), the “perceive” function of the agent, will have
two observations in its range, P1 and P2 that indicate if the metric M1 is acceptable
or not. The behaviour of the “perceive” function can be described as follows:

ì p1 if e = e1 or e = e2 ü
see(e) = í ý (1.28)
î p2 if e = e3 or e = e4 þ

According to (1.26) and (1.27), the “perceive function” of the NCA will have two
observations in its range P1 and P2 that indicate if the blocking possibility is accept-
able or not. The behaviour of the “perceive” function can be outlined as follows:

ì P1 if e = e1 or e = e2 bad ü
see(e) = í ý (1.29)
î P2 if e = e3 or e = e4 good þ

Identically for RCA, the perceive function see(e) is formulated as follows:

ì P1 if e = e1 or e = e3 bad ü
see(e) = í ý (1.30)
î P2 if e = e2 or e = e4 good þ
Large-Scale Multi-Agent Systems Simulation  27

With two given environmental states e Î E and e¢ Î E , then e~e’ can be written only
if see(e) = see(e’). An agent has perfect perception if the different environment states
are equal to distinct observations. In such case

~ = E (1.31)

In contrast with (1.15), when the perception of an agent does not exist,

~ = 1 (1.32)

1.2.4.4 Network Agents Which Maintain State

In many cases, an agent maintains state. In such a case, there are two additional
objects within the internal structure (Figure 1.13). In the proposed MA model, each
of the two basic agents (NCA, RCA) can be in one of the following states: (a) active
(in cooperation), (b) active (in competition), (c) inactive (Figure 1.14).
The agent (internal) state is updated by the “next” function. This function can be
expressed as next(i0,see(e)) and the corresponding action is action(next(i0,see(e))),
where i0 is the internal initial state.

1.2.4.5 Network Agent Utility Functions

A different utility function can be assigned to each agent (NCA, RCA) in order
to “measure” how good the corresponding outcome is (blocking/dropping). This

FIGURE 1.13  Agent with state.

FIGURE 1.14  Network agent with state.

28 Artificial Intelligence (AI)

function assigns a real number to each outcome, indicating how good the outcome
is for the selected agent. In other words, this utility function defines the preference
ordering over all the outcomes.
Let ω and ω’ be two possible outcomes in Ω set (where Ω = {ω1,ω2,…}), with
utility functions that give

ui (w ) ³ ui (w¢ ) (1.33)

Based on (1.16), agent i prefers better the ω outcome.

1.2.4.6 Multi-Agent Encounters
In a large-scale environment with more than one agent, except the environment
itself, the agent encounters also have to be modelled. Assume that there are two pos-
sible actions by two agents such as “C”, which means “Cooperate”, and “D”, which
means “Defect” and let the action set Ac = {C,D}. Based on the above, the environ-
ment behaviour can be modelled with the function

t : Aci ´ Ac j ® W (1.34)

where Aci and Acj represent the two agents i and j, respectively.
As an example, let the environment function

t ( D, D ) = w1, t ( D, C ) = w2 , t ( C , D ) = w3 , t ( C , C ) = w4 (1.35)

The two possible actions for NCA and RCA are the Cooperation (CR) and Competition
(CT) and so the action set becomes Ac = {CP, CT}. Now the expression (1.35) is as follows:

t ( CT , CT ) = w1,t ( CT , CR ) = w2 ,t ( CR, CT ) = w3 ,t ( CR, CR ) = w4 (1.36)

According to (1.35), each action is mapped to a different outcome.

Using utility functions for the ω in (1.35), let

ui (w1 ) = 1 ui (w2 ) = 1 ui (w3 ) = 4 ui (w4 ) = 4

(1.37)
u j (w1 ) = 1 u j (w2 ) = 4 u j (w3 ) = 1 u j (w4 ) = 4
The corresponding utility functions (NCA, RCA) in (1.18a) are

uNC (w1 ) = 1 uNC (w2 ) = 1 uNC (w3 ) = 4 uNC (w4 ) = 4

(1.38)
uRC (w1 ) = 1 uRC (w2 ) = 4 uRC (w3 ) = 1 uRC (w4 ) = 4
From (1.35) and (1.37), the outcomes for every action combination of the two agents
i,j can be expressed as follows:

ui ( D, D) = 1 ui ( D, C ) = 1 ui (C , D) = 4 ui (C , C ) = 4
(1.39)
u j ( D, D ) = 1 u j ( D, C ) = 4 u j (C , D) = 1 u j (C , C ) = 4
Large-Scale Multi-Agent Systems Simulation  29

Combining now (1.36) and (1.38) for NCA and RCA, the outcomes can be expressed
as follows:

uNC (CT , CT ) = 1 uNC (CT , CR) = 1 uNC (CR, CT ) = 4 uNC (CR, CR) = 4
(1.40)
uRC ( CT , CT ) = 1 uRC (CT , CR) = 4 uRC (CR, CT ) = 1 uRC (CR, CR) = 4

The only possible agent actions for the given outcomes can be rewritten (see (1.39)):

C , C ³i C , D >i D, C ³i D, D (1.41)

In the same way, the possible NCA, RCA actions based on (1.20a) are

CR, CR ³ NC CR, CT > NC CT , CR ³ NC CT , CT (1.42)

It is obvious from (1.39), (1.40), and (1.41) which action will be selected from each
agent, and thus each agent knows exactly what to do.

1.2.5 Event Interleaving as Scheduling Technique

Based on Real-Time Scheduling Theory
1.2.5.1 Real-Time Scheduling Algorithms for Implementing
Synchronized Processes or Events
Real-time scheduling is the most commonly recognized method for handling and
running simultaneous tasks where response time is a critical problem. The principle
of real time scheduling can be used to establish an alternative scheduling of events
compared to the state-of-the-art scheduling mechanism (CQ) used in DES systems
to accommodate concurrent events.
Real-time (RTS) systems are commonly used in a range of vital applications,
including robotics, avionics, telecommunications, process control, etc. In several
embedded systems, RTSs are also used. The common aspect of the above programs
is the tight deadlines for production tests. In other words, the timing of the results is
critical and in some cases even harmful. According to Burns and Wellings (2001),
“Real-time systems are those whose correctness depends not only on the rationality
of the computational results, but also on their precise timing.” According to Nahas
(2014), the interaction of the real-time control system with the environment requires
different components, such as sensors that read the external environment state. The
software architectures are common RTS implementation. Based on Nahas (2014)
and Locke (1992), these architectures can be classified into (a) cyclic executives, (b)
event-activated simultaneous task systems, (c) message-passing systems, and (d) cli-
ent–server systems. A specialized scheduler is required in concurrent task systems
(Nahas, M., 2014) in order to take scheduling decisions as the notion of a task is
retained in real time. The real-time scheduling theory is one way of predicting the
timing behaviour of complex multitasking software. Several developments have been
made since the initial formulation of the scheduling theory, such as rate monotonic
30 Artificial Intelligence (AI)

analysis in Liu and Layland (1973) and deadline monotonic analysis in Audsley et
al. (1992). These two improvements were incorporated into the scheduling theory
of fixed priorities (Audsley, et al. 1995). Some very important characteristics of the
RTSs found in Buttazzo (2011) are (a) repetitive contact with the environment, (b)
multiple actions concurrently transitioning rapidly to events and requiring a high
degree of competition, (c) competition for shared resources, (d) actions caused exter-
nally by events (within the system) or after time development, (e) must be stable in
overload, and (f) must be maintainable and of scalable open architecture.

1.2.5.2 Process Life Span in a Real-Time Scheduling Set-Up

The process life span in an RTS can be investigated by utilizing the single-queue
methodology combined with the pre-emption scheme.
An RTS consists of three basic components (Fidge, C.J., 2002): (a) a collection
containing the computational processes to be performed (typically subroutines with
a private control thread), (b) a run-time scheduler (e.g. dispatcher) specifying the
process execution sequence, and (c) a collection of common resources to be accessed.
Figure 1.15 demonstrates the standard life span of a job execution through a queue
from event to termination.
The processes that are facilitated by the system can be characterized by the time
features of arrival as (a) periodic (non-changing static arrival intervals) (Buttazzo,
G.C, 2011), (b) aperiodic (random arrivals) (Sprunt, B. 1990), and (c) infrequent (they
have only a non-changing minimum inter-arrival time) (Sprunt, B. 1990).

1.2.5.3 Scheduling Concurrent Events in WCS

In a wireless set-up, where events occur synchronously, the performance of the net-
work depends not only on the rationality of the computational results, but, also on
the correctness of the time sequence in which the results are encountered, implying
that the logical sequence of the various network procedures is also related to efficient
bandwidth management. Network operations such as new call entry, reallocation
(handoff), call termination, and so on can be regarded as tasks the network needs
to perform. A major role for the success of the resulting network is the scheduling
mechanism that manages the process under service. That is, this framework deter-
mines how, in terms of network performance metrics, the network can precisely
express the synchronized processes (incidents) in the most efficient way. The prog-
ress and actions of the event service may be seen by applying the response time
analysis as a first step. The sequential execution of an event is due to the existence of
one processor in conventional computers. The most common simulation tools, such

FIGURE 1.15  Task lifecycle.

Large-Scale Multi-Agent Systems Simulation  31

as ns-2/ns-3, are focused on the scheduling of the CQ form, which reflects the serial
logic. Within this conceptual framework the execution sequence of each generated
event is defined from the associated time stamps. CQ retains information about
activities for future executions. Such time stamps can also be used as goals. The
computational multi-processing approach can be implemented when the network
events are addressed as synchronized. In a simultaneous model, another MU is mov-
ing or attempting for reallocation (handoff) when a MU is under network processing.
Thus, the execution of the mission (event) has to be incomplete to accommodate the
waiting synchronized MUs. This reasoning leads to the inter-leaving scheme that
can normally be enforced using a multi-threading framework. Figure 1.16 demon-
strates the queuing models for the two approaches (consecutive and concurrent).
In a concurrent set-up, events can be viewed as multiple processes that must be
facilitated by the available computing unit (Figure 1.17).

1.2.5.4 Response Time Analysis

Owing to the frequency of tasks, the whole system can be interpreted as a multi-
processing device in real time when other tasks are being performed. That approach
leads to sophisticated control and interleaving of events. Multi-processing concepts

FIGURE 1.16  Queuing models for serial and synchronized processes.

FIGURE 1.17  Multiple processes (events) generators related to one server.

32 Artificial Intelligence (AI)

and an analysis of schedulability (Fidge, C.J. 2002), reflect events behaviour that
could be analyzed as follows:
Any event e, with e Î {NC , RC , MC , FC}has a release time r:

re = ( n × Te ) - J e (1.43)

where n is the nth event, Te is the event period, and Je is the associated Jitter. Je is
computed as

J e = re ( max ) - re ( min ) (1.44)

In the computations, Je = 0.

Computational time of an event is

C = (n + 1) × Ce (1.45)

Let Ri be the response time for process i, Ci be the worst-case calculation time, and Ii
be the interference, due to higher priority processes. The response time for process i is

Ri = Ci + I i (1.46)

Assuming that all priorities are unique, the interference for process i, is

é Ri ù
Ii = å êê T úú C (1.47)
jÎhp ( i ) j
j

where hp(i) is the set of all processes that have greater priority than i.
From Equations 1.46) and 1.47, the worst-case response time of process i can be
computed through the following formula:

é Ri ù
Ri = Ci + å êê T úúC (1.48)
jÎhp ( i ) j
j

Equation 1.48, can be solved iteratively (Audsley et al. 1993), commencing with the
initial estimate for Ri equal to 0. Knowing the xth value, the (x + 1) value can be
approached as follows:

é Rix ù
Rix +1 = Ci + å êê T
jÎhp ( i )
úC j (1.49)
j ú

The main and useful conclusion derived from Equation 1.49 is process i schedulabil-
ity, which means that the deadlines required can be met. For safe conclusions, the
above equation must be re-evaluated to see if it converges to an Ri value such that
Ri < Di (Di = the given process i time limit). Instead of the existence of specialized
Large-Scale Multi-Agent Systems Simulation  33

tools to evaluate the above equations, such tools do not necessarily give the pro-
grammer any insight into the dynamic conduct of each process (Fidge, C.J., 2002).
Simulation is the most effective method for demonstrating the timing actions of the
processes in consideration (Fidge, C.J., 2002).
The event goals must be set for conducting full calculations. Let PFC, PRC, PMC,
and PNC be the priorities for the event types FC, RC, MC, and NC with PFC > PRC,
PRC > PMC, PMC > PNC, respectively. The corresponding hp(j) sets involving the
definitions above can be defined as follows:

hp ( FC ) = {Æ} (1.50)

hp ( RC ) = {FC} (1.51)

hp ( MC ) = {RC , FC} (1.52)

hp ( NC ) = {MC , RC , FC} (1.53)

Equation 1.49 can be solved iteratively (Audsley et al. 1993). Thus, the response time
for next event n + 1 can be calculated as follows:

é Ren ù
Ren +1 = C + å ê ú × C j (1.54)
T
jÎhp ( e ) ê j ú

Formulating the above equations to network procedures, the resulted response times
are as follows:
Assuming that Re0 = 0 ,

é RFC
n ù
n +1
RFC = CFC + å ê
jÎhp ( FC ) ê
Tj ú
ú × C j Þ RFC = CFC (1.55)

Since hp ( FC ) = {Æ}

n +1 é Rn ù
RRC = C RC + ê RC ú × CFC (1.56)
ê TFC ú

n +1 é Rn ù é Rn ù
RMC = C MC + ê MC ú × C RC + ê MC ú × CFC (1.57)
ê TRC ú ê TFC ú

n +1 é Rn ù é Rn ù é Rn ù
RNC = C NC + ê NC ú × C MC + ê NC ú × C RC + ê NC ú ×CFC (1.58)
ê TMC ú ê TRC ú ê TFC ú
Based on the above rules, a pre-emptive event scheduling can be applied.
34 Artificial Intelligence (AI)

FIGURE 1.18  Pre-emptive stationary priority scheduling (tasks/events: NC – new call, RC

– call reallocation/handoff, MC – movement call, FC – finished call).

1.2.5.5 Pre-emptive Stationary Priority Scheduling (PSPS)

In PSPS scheduling, each task has a non-changing priority that is constant during
execution time. The main rewards of this strategy are (a) faster facilitations of pro-
cesses with higher priority (Fidge, C.J., 2002; Pont et al. 2007), (b) simpler realiza-
tion, (c) stability in overloads and spiking behaviour. In a real network set-up, while
an MU is under processing, another MU might be arrived.
Figure 1.18 illustrates four processes with different priorities (Low1 < Low2 <
High1 < High2) during a time window of execution.
For example, while an NC type of event is under processing, a new event (RC) is
arrived (Figure 1.18).

1.2.6 Supported DCA Variations

For the assessment procedure, we used various DCA variations:

• Unbalanced version reflecting traditional DCA

• A conventional balanced version based on a less congested neighbourhood
algorithm)
• A conventional Best CNR channel assignment algorithm that is based on
the best measured CNR between current MU and local or neighbouring BS
• A traditional round blocking algorithm where the adjacent channels are in
turn checked

All the above algorithms assume sequential search only within the neighbouring
cells between the current user and available channels.
The following proposed algorithm assigns channels inside neighbouring cells
simultaneously, provided that N agents/users emerge at a given time.
• Artificial intelligence-based variation controlled and good CNR DCA for
concurrent network assignment

1.2.6.1 The Conventional Unbalanced Variation (Classical DCA)

Unbalanced variation represents the classical DCA scheme where only one attempt (in
call initiating cell) for connection is done based on the CNIR criterion (Figure 1.19).
Large-Scale Multi-Agent Systems Simulation  35

Compute CNR from Current BS

Find available channel
Compute Interference from other users
CNIR > CNIR Threshold ?
YES : Success =1
NO : Success = 0
FIGURE 1.19  Unbalanced algorithm.
1.2.6.2 The Conventional Balanced Variation (Min Cell Congestion)
The unbalanced DCA algorithm assigns the first channel which meets the conditions
and the balanced DCA uses the least congestion algorithm to complete the channel
assignment.
The balanced approach uses the least congested algorithm, as we have already
stated. Initially we presume that the cell where the new call occurs is the least con-
gested. The algorithm checks all the neighbouring region’s cells sequentially and
identifies the least congested cell (Figure 1.20).

1.2.6.3 The Conventional Best CNR Variation

The best CNR method goes around the initiated cell’s six neighbour cells and cal-
culates the CNR between the base stations for the user and the neighbour cells. The
software tries to create a new channel from the corresponding cell until the strongest
CNR is identified. Initially, the algorithm assumes that the strongest CNR occurs

FIGURE 1.20  Min cell congestion algorithm.

36 Artificial Intelligence (AI)

FIGURE 1.21  Best CNR algorithm.

from the first neighbour cell and proceeds to measure CNR sequentially from the
rest of the neighbour cells until the strongest CNR is identified. Figure 1.21 illus-
trates the structure and logic of the algorithm, respectively.

1.2.6.4 The Conventional Round Blocking Variation

The round blocking algorithm checks in neighbouring cells, and the algorithm stops
when a successful channel assignment is made. If the call establishment in the last
neighbouring cell is not successful, then the call is blocked. The round blocking
algorithm and its rationale are shown in Figure 1.22, respectively.

FIGURE 1.22  Round blocking algorithm.

Large-Scale Multi-Agent Systems Simulation  37

1.2.6.5 The Proposed Novel Artificial Intelligence Based Balanced and

Best CNR DCA Variation for Concurrent Channel Assignment
Suppose at any given moment t, N agents/users emerge in a given neighbour cell
under consideration. Also suppose that C(j) is the number of available cells per
cell j in this neighbour. The problem is how each agent/user i out of the total N
agents/users should be assigned to a channel in an optimal way according to the
least congestion criterion and best CNR criterion. If this is done sequentially, then
it is not obviously optimal. In order that this assignment problem could be solved
in an optimal way, we should involve either dynamic programming or a global
optimization methodology. In this research, we have employed genetic algorithms
which can lead to an acceptable solution, especially when the problem is of fair
dimensionality.
The chromosome in the proposed algorithm is as follows

agent_1 agent_2 agent_3 … ….. …. ….. …. agent_N-1 agent_N

Ch(1) Ch(2) Ch(3) Ch(N-1) Ch(N)

Therefore, the chromosome comprises the possible channels (out of the available
ones) to be assigned to every agent/user. If there are no available channels for all N
agents/users we put zeros in the corresponding positions in the chromosome. An
exhaustive search of all possible combinations of C in total available channels in
æC ö
groups of N agents/users, ç ÷ , requires factorial time increase as it is known and
èNø
makes impossible the task in real time. Genetic algorithms in this case can provide
an affordable near optimal solution.
The fitness function used is comprised of the following parts:

all cells C ( j )

1. å åCNR (i, j ) should be maximized. This means that for all channels
j =1 i =1

C(j) of cell j assigned, let’s say, to C(j) agents/users out of N and for all cells
the total CNR should be maximized.

åå ( congestion (i ) - congestion ( j ))
2
2. should be minimized for all cell
i j

pairs (i,j).
3. Each available channel ch(i) assigned to agent i should be used only once.

å
4. ch ( i, j ) £ C ( j ) , that is, all channels of cell j assigned to agents/users i
i

should be less than or equal to the available channels C(j) of channels j.

The above fitness function and the chromosome architecture define completely the
genetic algorithm (GA) solution
38 Artificial Intelligence (AI)

1.2.7 Implementation Architectures
We have implemented and tested three different approaches of agent-oriented archi-
tectures using Java threads. Each approach is characterized by the parallelization
grade that influences the simulation efficiency and the reliability related to network
behaviour.

1.2.7.1 Conventional Model
The conventional model consists of one thread and the event processing is based on
the CQ mechanism (Figure 1.23). According to this model, any generated event is
served through a three-step procedure which includes enqueue, dequeue, and execu-
tion. Based on the type of the highest priority event, the corresponding network
service is activated.

1.2.7.2 Concurrent Models
As we mentioned before, in a real network, processes take place simultaneously. A
model that supports concurrent architectures approaches in a more reliable way the
real network. We have developed and tested two architectural models with different

FIGURE 1.23  Conventional model.

Large-Scale Multi-Agent Systems Simulation  39

FIGURE 1.24  Basic concurrency.

levels of concurrency. These models rely on the mentioned layered agent architecture
using Java multi-threading techniques for the implementation.

1.2.7.2.1 Basic Concurrency
The concurrent version of the model must prevent simultaneous access to the user
registry where critical data for the MUs are stored. Figure 1.24 illustrates the
structure of the implemented architecture. Level 0 represents the control agent
that synchronizes the execution of the concurrent tasks (New call, reallocation,
etc.) in combination with the thread scheduler and administers the simulation
time. Concurrent tasks are activated within the level 1 and share the connection
establishment methods. The main network processes like new call arrival, real-
location check, user movement, etc. execute common tasks such as access to user
registry and connection establishment. Connection establishment and user reg-
istry access (read, write) are implemented with three mutual excluded methods
(level 2).
These methods belong to the same thread. The main agents act simultaneously
and try to finish for each MU in each cell. A new connection based on reallo-
cation, user movement, or new call arrival requires signal calculations and other
DCA operations. These actions are implemented using the try_connection method.
All the threads work simultaneously except the above method and access to user
registry that is used by one thread at a time preventing the data corruption of user
registry.
40 Artificial Intelligence (AI)

FIGURE 1.25  Concurrency with multiple synchronized blocks.

1.2.7.2.2 Concurrency with Multiple Synchronized Blocks

In this model, every main agent (new call, reallocation, etc.) has a separate instance
of the try_connection method. Mutual exclusion is now moved to user registry access
(Figure 1.25). Each agent can complete its connection establishment independently
from the others. The resource competition between concurrent threads is restricted
to user registry access and not in try_connection method like the previous mentioned
model. Thus, we achieve a higher level of concurrency.

1.2.7.2.3 Thread Usage
The main implementation concept is based on thread logic. Each user/network
operation is implemented as a thread. The Java programs with regard to the above-
mentioned concurrent architectures consist of four basic threads (call termination,
reallocation, new call arrival, user move) and three additional (clock synchroniza-
tion, initial/final procedures). As the concurrency grade increases, the Java threads
become more competitive for accessing the user table (read/write tasks). Thus, the
possibility for deadlocks is rising. In order to face effectively the deadlock conditions,
synchronized blocks instead of synchronized methods can be used. Synchronized
blocks set locks for shorter periods than synchronized methods and so the deadlock
possibilities decrease. Table 1.2 shows the threads and synchronized methods/blocks
that are used in the proposed architectures.
Large-Scale Multi-Agent Systems Simulation  41

TABLE 1.2
Thread Usage by Architectural Model
Concurrency with Multiple
Concurrency Architecture Basic Concurrency Synchronized Blocks
No. of threads 7 7
Thread tasks Clock, initial loop procedures, final loop procedures, call termination,
reallocation check, new call arrival, user movement
No. of synchronized methods 3 0
Synchronized methods Read User Registry –
Write User Registry
Try_connection
No. of synchronized blocks 0 18 (Read/Write User registry)

In the first approach, the most of the access points for user registry exist within
the try_connection method that is synchronized. Due to that synchronization, the
competition for user registry access is low. Increasing concurrency, the try_connec-
tion is concurrent among the four main threads and so only the access methods
(read/write to user registry) are synchronized. With this approach, we achieve high
concurrency but the deadlocks seem to be a difficult situation. Using synchronized
blocks in contrast with synchronized methods, we faced successfully the deadlock
conditions. Extending the above models for avoiding deadlocks, conditional synchro-
nization can be applied. With conditional synchronization, the user registry access
is synchronized only when two or more threads try to access data for the same MU.

1.3 SIMULATION MODEL EVALUATION

1.3.1 Network Behaviour
One of the most important features for characterizing the performance of a cellular
network is the probability of blocking. When a new call arrival happens and a chan-
nel cannot be allocated by the network, then we say this call is blocked. The prob-
ability of blocking Pblocking is determined from the proportion.

number of blocked calls

Pblocking = (1.59)
number of calls
The risk of a call dropping is also an additional and a very critical cellular network
feature. If a call is ongoing and the required quality conditions are not met, then this
call is mandatorily led to termination. The probability of dropping Pfc is computed
from the following proportion:

number of forced calls

Pfc = (1.60)
number of calls - number of blocked calls
42 Artificial Intelligence (AI)

1.3.2 Monte Carlo Simulation Method

Monte Carlo (MC) methods are stochastic techniques, meaning they are based on
the use of random numbers and probability statistics to investigate problems. For MC
simulations, the processes are random; so each time it is run, it will come up with
slightly different results. This method gives us a way to model complex systems that
are often extremely hard to investigate with other types of techniques. Moreover,
MC method is used in order to estimate parameters of an unknown distribution (e.g.
network behaviour) by statistical simulation (Fishman, G.S., 1995).

1.3.3 Simulation Model Behaviour

The stability of the simulation model is derived by sequential Monte Carlo simulations
from the progressive standard deviation of selected network properties and metrics.
A further metric for evaluating the performance of the simulation model is peri-
odicity, a metric dependent on outliers. As we have already mentioned, simulation
time is based on a regular model. The periodicity is determined within the simula-
tion time using all available pairs of days. Assume that the likelihood of blockage
within a two-day set of 48 hours is expressed by the set a and set b matrices, respec-
tively. First, we calculate the BPDk blocking likelihood differences within a pair of
days for every hour:
:

BPDk [i ] = seta éëi ùû - setb éëi ùû (1.61)

where k is the selected pair and i is the i-element of each set. The difference seta[i]
− setb[i] is calculated only if the seta[i] and setb[i] values are not outliers; otherwise,
the data pairs are rejected. We then compute the expected value Mk for all possible
such successive pairs:

å BPD éëi ùû
k

M k [] = i =1
(1.62)
n
Third, we calculate periodicity by measuring the mean value between the mean Mk.
When the difference between elements tends to be zero between pairs of different days,
the pairs tend to be equal; so the factor of periodicity tends to be zero for the given
pair. When more pairs have a low periodicity factor, then the mean factor value of more
pairs is also low. Smaller periodicity factor means higher periodicity and simulation
model behaviour is more stable in simulation time as opposed to the regular model.

1.3.4 Results Accuracy
Outliers and other common statistical metrics are used to measure the accuracy
of results. Outliers are unusual data values which may distort the results obtained
Large-Scale Multi-Agent Systems Simulation  43

(Barnett, V. and Lewis, T., 1994). When out of an agreed area, we describe a value
as outlier. This region’s boundaries are defined as follows:

L = m ± 0.95stdX (1.63)

where stdX is the standard deviation of the likelihood of blocking or link dropping,
and m is the mean simulation time value.
To investigate how the system design and implementation affects the accuracy and
reliability of the test, we employed several overall statistical metrics with regard to
blocking and dropping probabilities throughout all simulation runs such as expected
value of standard deviation, mean value, variance, etc.

1.3.5 Reference Analysis Model Employing One Cell Only

Initially, a conceptual level reference model was developed to check the accuracy
of the event interleaving approach compared to established CQ approaches. Due to
the conceptual model’s simplicity, only one cell is presumed to support MUs within
the coverage area. Figure 1.26 shows the cell offering n channels to support new
call entries, reallocation, and movement of MUs. This model includes no advanced

FIGURE 1.26  MU service operation based on CQ approach.

44 Artificial Intelligence (AI)

mathematical model (e.g. for signal propagation) since it is based on how the chan-
nels express the clients (MUs).
Both models were initially tested for event-generation with deterministic data.
Figure 1.26 shows how some of the five events (NC1, NC2, RC1, MC1, RC2) per-
form. The scheduler restores the event with the least time stamp (highest priority)
for subsequent execution from the queue. When the event interleaving system man-
ages the same events queue (Figure 1.27), the execution sequence remains identical,
while the scheduler time slice (time slice width – TSW) is greater than or equal to
the event computational time (ECT). In other words, if TSW > ECT, then one slice
of time in the predetermined sequence is enough to complete execution of each gen-
erated event. And the two conceptual models yield similar results. When the TSW
is less than the ECT for finishing execution of the event, or the TSWs are assigned
asymmetrically to active threads, the results are totally different from the traditional
CQ approach. This is due to the competitiveness of running threads, associated with
individual MUs, for specific resources management and the relevant scheme of chan-
nel allocation at different time units.

FIGURE 1.27  MU service operation based on the Event Interleaving scheme.

Large-Scale Multi-Agent Systems Simulation  45

1.4 EXPERIMENTAL RESULTS
1.4.1 Indicative Results Based on Five Days of Network Operation
The simulation results are based on unbalanced and balanced DCA schemes where
the wireless network consists of 19 cells with capacity of 50 users per cell and 32
channels per cell. Figures 1.28–1.30 show a typical blocking probability graph for
the given network using the daily traffic model (five days of network operation) and
based on the proposed architectures.

1.4.2 Model Behaviour Based on Architectural Variations

It is obvious from Figures 1.31–1.36 that the selection of the concurrency affects
positively the simulation model behaviour.

1.4.3 Scheduling Mechanism Comparison

Figures 1.37 and 1.38

1.4.4 Response Time Analysis Results

Figures 1.39–1.42 show the time behaviour of two tested event interleaving schemes.
In Pre-emptive Event Interleaving (PEI) scheme, when a new event arrives with
higher priority, the current under processing event is pre-empted and so the PEI is

Conventional Model, BLOCKING

0.5

0.45

0.4
Blocking probability (%)

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Simulation time

FIGURE 1.28  Conventional model, blocking probability.

46 Artificial Intelligence (AI)

Basic Concurrency, BLOCKING

0.5

0.45

0.4
Blocking probability (%)

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Simulation time

FIGURE 1.29  Basic concurrency, blocking probability.

Concurrent model, Multiple Synch Blocks, BLOCKING

0.5

0.45

0.4
Blocking probability (%)

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Simulation time

FIGURE 1.30  Concurrency with multiple synchronized blocks, blocking probability.

Large-Scale Multi-Agent Systems Simulation  47

ROUND BLOCKING MEAN mean Progress, BLOCKING

0.208

0.206

0.204
Blocking Probability(%)

0.202

0.2 Conventional model

Basic concurrency
0.198 Concurrent, Multiple Synch Blocks

0.196

0.194

0.192

0.19

0.188
0 5 10 15
Executions (x2)

FIGURE 1.31  Proposed balanced DCA and best CNR based on GAs, blocking probability.

BALANCED MEAN mean Progress, BLOCKING

0.19
Conventional model
0.188 Basic concurrency
Concurrent, Multiple Synch Blocks

0.186
Blocking Probability(%)

0.184

0.182

0.18

0.178

0.176

0.174
0 5 10 15
Executions (x2)

FIGURE 1.32  Conventional balanced DCA, blocking probability.

48 Artificial Intelligence (AI)

ROUND BLOCKING MEAN mean Progress, BLOCKING

0.208

0.206

0.204
Blocking Probability(%)

0.202

0.2 Conventional model

Basic concurrency
0.198 Concurrent, Multiple Synch Blocks

0.196

0.194

0.192

0.19

0.188
0 5 10 15
Executions (x2)

FIGURE 1.33  Best CNR DCA, blocking probability.

BALANCED MEAN mean Progress, BLOCKING

0.19
Conventional model
0.188 Basic concurrency
Concurrent, Multiple Synch Blocks

0.186
Blocking Probability(%)

0.184

0.182

0.18

0.178

0.176

0.174
0 5 10 15
Executions (x2)

FIGURE 1.34  Proposed balanced DCA and best CNR based on GAs, dropping probability
variance.
Large-Scale Multi-Agent Systems Simulation  49

BEST CNR MEAN mean Progress, BLOCKING

0.077

0.076

0.075
Blocking Probability(%)

0.074

0.073

0.072

0.071

0.07

0.069
Conventional model
0.068 Basic concurrency
Concurrent, Multiple Synch Blocks
0.067
0 5 10 15
Executions (x2)

FIGURE 1.35  Unbalanced DCA, dropping probability periodicity.

BALANCED VAR mean Progress, DROPPING

0.028

0.026

0.024

0.022

0.02
Variance

0.018

0.016

0.014

0.012 Conventional model

Basic concurrency
0.01
Concurrent, Multiple Synch Blocks

0.008
0 5 10 15
Executions (x2)

FIGURE 1.36  Proposed balanced DCA and best CNR based on GAs, dropping probability
periodicity.
50 Artificial Intelligence (AI)

BLOCKING MEAN Mean Progress ONE-CELL REF MODEL

0.75

0.7
Blocking Probability (%)

0.65

0.6

0.55

0.5

0.45
Event Interleaving
Calendar Queue
0.4
0 5 10 15
Executions (x2)

FIGURE 1.37  CQ blocking probability and Event Interleaving techniques (MC runs, refer-
ence model of one cell).

DROPPING MEAN Mean Progress ONE-CELL REF MODEL

0.9

0.85
Dropping Probability (%)

0.8

0.75

0.7

Event Interleaving
Calendar Queue
0.65
0 5 10 15

Executions (x2)

FIGURE 1.38  CQ dropping probability and Event Interleaving techniques (MC runs, refer-
ence model of one cell).
Large-Scale Multi-Agent Systems Simulation  51

ARRIVAL-DEPARTURE MEAN Mean Progress

ONE-CELL REF model (Tnc=Trc=Tmc=Tfc)
25
PEI-NCDD
PEI-RCDD
PEI-MCDD
PEI-FCDD
20 FEI-NCDD
FEI-RCDD
FEI-MCDD
FEI-FCDD

15
Delay (sec)

10

0
0 5 10 15 20 25 30
Executions (x2)

FIGURE 1.39  Delay calculations between events arrival and departure time units (equal
arrival time units).

ARRIAVAL-DEPARTURE MEAN Mean Progress

ONE-CELL REF model (Tnc#Trc#Tmc#Tfc)
18

16
PEI-NCDD
PEI-RCDD
14 PEI-MCDD
PEI-FCDD
FEI-NCDD
12 FEI-RCDD
FEI-MCDD
Delay (sec)

FEI-FCDD
10

2
0 5 10 15 20 25 30
Executions (x2)

FIGURE 1.40  Delay calculations between events arrival and departure time units (aperi-
odic and random arrival times).
52 Artificial Intelligence (AI)

DEPARTURE-DEPARTURE MEAN Mean Progress

ONE-CELL REF model (Tnc=Trc=Tmc=Tfc)
30
PEI-NCDD
PEI-RCDD
PEI-MCDD
PEI-FCDD
25 FEI-NCDD
FEI-RCDD
FEI-MCDD
FEI-FCDD

20
Delay (sec)

15

10

0
0 5 10 15 20 25 30
Executions (x2)

FIGURE 1.41  Delay measurements between events departure times (equal arrival times).

DEPARTURE-DEPARTURE MEAN Mean Progress

ONE-CELL REF model (Tnc#Trc#Tmc#Tfc)
45
PEI-NCDD
40 PEI-RCDD
PEI-MCDD
PEI-FCDD
35 FEI-NCDD
FEI-RCDD
30 FEI-MCDD
FEI-FCDD
Delay (sec)

25

20

15

10

0
0 5 10 15 20 25 30
Executions (x2)

FIGURE 1.42  Delay measurements between events departure times (aperiodic arrival
times).
Large-Scale Multi-Agent Systems Simulation  53

based clearly on the real-time scheduling concept. On the other hand, a variation of
PEI called Fair Event Interleaving (FEI) is also tested. In FEI, each event is executed
in turn based on priority settings, and the system gives equal computational periods
for each event under processing.

1.5 CONCLUSIONS AND FUTURE WORK

This research work investigates scheduling issues in multi-agent systems modelling
wireless communications systems. Several contributions have been herein intro-
duced by analyzing conventional scheduling methods and the more suitable con-
current scheduling methodologies in case of synchronized agent models. Moreover,
a comprehensive simulation model for simulating large-scale wireless networks is
proposed based on multi-agent analysis. A novel modelling methodology of large-
scale network services and the corresponding theoretical analysis and implementa-
tion based on the multi-agent concept is initially presented. Based on this modelling,
an optimized channel allocation algorithm through artificial intelligence techniques,
namely, a genetic algorithm model, has been presented with success when concur-
rent events emerge in the system, which is the actual real-world case. Involving this
modelling approach, a novel and alternative to the state-of-the-art event scheduling
mechanism using real-time scheduling theory for supporting improved concurrent
network events modelling is also presented. The simulation results show the effec-
tiveness of the concurrent multi-agent-based models using artificial intelligence opti-
mization techniques as well as the efficiency of the event interleaving for concurrent
network events. However, such promising results should be verified in larger scale
networks, especially in wireless sensor networks, where the parameter of power con-
sumption plays a critical role together with congestion and best CNR control.

REFERENCES
Abar, S, Theodoropoulos, GK, and Lemarinier, P. 2017. Agent based modelling and simula-
tion tools: A review of the state-of-art software. Computer Science Review, 24, 13–33.
Audsley, NC, Burns, A, Richardson, MF, and Wellings, AJ. 1992. Deadline monotonic sched-
uling theory. In Proceedings 18th IFAC/IFIP Workshop on Real-Time Programming
(WRTP’92), June.
Audsley, N, Burns, A, Richardson, M, Tindell, K, and Wellings, A. 1993. Applying new
scheduling theory to static priority pre-emptive scheduling. Software Engineering
Journal, 8(5), 284–292.
Audsley, NC, Burns, A, Davis, RI, Tindell, KW, and Wellings, AJ. 1995. Fixed priority pre-
emptive scheduling: An historical perspective. Real-Time Systems, 8, 173–198.
Bajaj, L, Takai, M, Ahuja, R, and Bagrodia, R. 1999. Simulation of large-scale heteroge-
neous communication systems. In Proceedings of IEEE Military Communications
Conference (MILCOM’99).
Baker, Mark, Rajkumar Buyya, and Dan Hyde. “Cluster computing: A high-performance
contender.” arXiv preprint cs/0009020 (2000), pp. 79–83, https​:/​/ar​​xiv​.o​​rg​/ft​​p​/cs/​​paper​​
s​/000​​9​/​0 00​​9020.​​pdf
Balakrishnan, V, et al. 1997. A framework for performance analysis of parallel discrete event
simulators. In Proceedings of the 29th Conference on Winter Simulation.
54 Artificial Intelligence (AI)

Barr, R. 2004. An Efficient, Unifying Approach To Simulation Using Virtual Machines.

Cornell University, PhD Dissertation
Beowulf Project. http://www​.beowulf​.org.
Blackstone, JH, Hogg, CL, and Phillips, DT. 1981. A two-list synchronization procedure for
discrete event simulation. Communications of the ACM, 24(12), 825–629.
Bodanese, EL. 2000. A Distributed Channel Allocation Scheme for Cellular Networks Using
Intelligent Software Agents. PhD thesis, University of London.
Bononi, L, and D’Angelo, G. 2003. A novel approach for distributed simulation of wireless
mobile systems. PWC 2003, LNCS 2775, 829–834.
Boukerche, A, Das, SK, and Fabbri, A. 2001. SWiMNet: A scalable parallel simulation test-
bed for wireless and mobile networks. Wireless Networks, 7, 467–486.
Brown, R. 1988. Calendar queues: A fast O(1) priority queue implementation for the simula-
tion event set problem. Communications of the ACM, 31(10), 1220–1227.
Burns, A, and Wellings, AJ. 2001. Real-Time Systems and Programming Languages: Ada
95, Real-Time Java, and Real-Time POSIX. Pearson Education, 3rd edition, ISBN-
13: 978-0201729887, ISBN-10: 0201729881, https​:/​/ww​​w​.ama​​zon​.c​​om​/ Re​​al​-Ti​​me​-Sy​​
stems​​-Prog​​rammi​​ng​-La​​nguag​​es​/dp​​/0201​​72988​​1​/ref​​= sr​_1​​_1​?dchi​​ld​=1&​​keywo​​rds​=9​​
78020​​17298​​87​&li​​nkCod​​e​​=qs&​​qid​=1​​60848​​8674&​​s​=boo​​ks​&sr​​=1-1
Buttazzo, GC. 2011. Hard Real-Time Computing Systems: Predictable Scheduling Algorithms
and Applications, Vol. 24. Springer Science & Business Media, LLC 2011, New York,
doi: 10.1007/978-1-4614-0676-1.
Chaturvedi, A, Dickieson, J, Dolk, DR, and Scholl, J. 2001. Introduction to agent-based simu-
lation and system dynamics minitrack. In Proceedings of the 34th Hawaii International
Conference on System Sciences.
Chung, K, Sang, J, and Rego, V. 1993. A performance comparison of event calendar algo-
rithms: An empirical approach. Software—Practice and Experience, 23(10), 1107–1138.
Cimini, LJ, and Foschini, GJ. 1992. Distributed algorithms for dynamic channel allocation in
microcellular systems. In IEEE Vehicular Technology Conference, 641–644.
Cox, DC, and Reudink, DO. 1973. Increasing channel occupancy in large scale mobile
radio systems: Dynamic channel reassignment. IEEE Transanctions on Vehicular
Technology, VT-22, 218–222.
Del Re, E, Fantacci, R, and Giambene, G. 1996. A dynamic channel alloca-tion technique
based on hopfield neural networks. IEEE Transanctions on Vehicular Technology, VT-
45(1), 26–32.
Dixit, S, Yile Guo, Antoniou Z, Resource management and quality of service in third genera-
tion wireless networks, IEEE Communications Magazine ( Volume: 39, Issue: 2, Feb.
2001)
Ekdahl, B. 2001. How autonomous is an autonomous agent? In Proceedings of the 5th
Conference on Systemic, Cybernetics and Informatics (SCI 2001), July 22–25, Orlando,
USA.
Elnoubi, SM, Singh, R, and Gupta, SC. 1982. A new frequency channel assignment algorithm
in high capacity mobile communication systems. IEEE Transactions on Vehicular
Technology, VT-21(3), 125–131.
Erickson, KB, Ladner, RE, and LaMarca, A. 1994. Optimizing static calendar queue. Annual
IEEE Symposium on Foundations of Computer Science, 35, 732–743.
Fagin, R. et al. 1995. Reasoning About Knowledge. MIT Press, Cambridge, MA.
Fall, K, and Varadhan, K. 2015. The ns manual (2011). http://www​.isi​.edu​/nsnam​/ns​/doc​/ns
doc. pdf.
Fidge, CJ. 2002. Real-Time Scheduling Theory, Technical Report. University of Queensland,
No. 02-19, https​:/​/es​​pace.​​libra​​r y​.uq​​.edu.​​au​/da​​ta​/ UQ​​_1054​​3​/svr​​c​_02_​​19​.pd​​f​?dsi​​_vers​​
ion​=5​​3e60a​​4 4db4​​5e980​​2571f​​fd4a1​​6a4c3​​3​&Exp​​i res=​​16084​​90365​​&Key-​​Pair-​​Id​=AP​​
Large-Scale Multi-Agent Systems Simulation  55

KAJKN​​BJ4MJ​​BJNC6​​N LQ​&S​​ignat​​u re​=R​​QTLaQ​​~jI7v​​5b6gl​​xW3O3​​ctcm0​​P Gv​~h​​

78x5y​​uyTLL​​D8bhK​​V8Imu​​zsy7O​​P1Usy​​h8​-yU​​9EwkQ​​b0​-hn​​zDkzY​​vpZEJ​​1gzJb​​N RlJU​​
YwFmO​​R9fgO​​qIPjr​​8UJnL​​RjgWV​​6n2fb​​x Mwwr​​PXi0I​​o dLe~​​vQsrs​​M ZXRF​​g AAe2​​
Dq3vw​​F Ib​-2​​bS7fd​​1v3Zl​​~tsuj​​Y Uk0J​​ZXCqe​​9eT9f​​Uz6XH​​0q​-dm​​nyORh​​O0WgE​​LFnj6​​
ett​-I​​H Xq7K​​b eTXA​​FnCMu​​0 ​- Ls6​​I XKAq​​Fs0pR​​​t XTcb​​qDpj5​​X1oa3​​jHypG​​DBvPQ​​
2TCvZ​​0jlRu​​J PRmh​​2d0Oi​​EBJVw​​-JJsP​​k hCM6​​A KALn ​​NmZA8​​D r2Ze ​​eBf07​​2​~UZb​​
Q__
Fishman, GS. 1996. Monte Carlo: Concepts, Algorithms, and Applications. New York:
Springer-Verlag, Springer Science + Business Media.
Fujimoto, RM. 2000. Parallel and Distributed Simulation Systems. (Vol. 300). New York:
Wiley-Interscience.
Fujimoto, RM, et al. 2003. Large-scale network simulation: How big? How fast? In IEEE,
International Symposium on Modeling, Analysis, and Simulation of Computer and
Telecommunications Systems MASCOTS.
Goh, RSM and Thng, IL-J. 2003. MLIST: An efficient pending event set structure for discrete
event simulation. International Journal of Simulation, 4(5–6), 66–77.
Hayes-Roth, B. 1995. An architecture for adaptive intelligent systems. Artificial Intelligence:
Special Issue on Agents and Interactivity, 72, 329–365.
Hayzelden, A. and Bigham, J. 1999. Software Agents for Future Communications Systems.
Springer-Verlag, Berlin.
Henriksen, JO. 1977. An improved events list algorithm. In Proceedings of the 1977 Winter
Simulation Conference (Gaithersburg. Md., Dec. 5–7). IEEE, Piscataway, NJ, 547–557.
http:​/​/jis​​t​.ece​​.corn​​ell​.e​​du​/ja​​vadoc​​/jist​​/runt​​ime​/S​​chedu​​ler​.C​​​alend​​a r​.ht​​ml.
Huhns, M, and Singh, M. Eds. 1998. Agents and Multi Agent Systems: Themes, Approaches,
and Challenges. Readings in Agents, Chapter 1, Morgan Kaufmann Publishers, USA,
1–23.
Iraqi, Y, and Boutaba, R. 2000. A multi-agent system for resource management. In Wireless
Mobile Multimedia Networks, LNCS 1960. Springer-Verlag, Berlin Heidelberg,
218–229.
Jefferson, DR, and Barnes, PD. 2017. Virtual time III: Unification of conservative and opti-
mistic synchronization in parallel discrete event simulation. In 2017 Winter Simulation
Conference (WSC). IEEE.
Jennings, NR. 2000. On agent-base software engineering. Artificial Intelligence, 117,
277–296.
Jiang, Y, and Jiang, J. 2008. Contextual resource negotiation-based task allocation and load
balancing in complex software systems. IEEE Transactions on Parallel and Distributed
Systems, 20(5), 641–653.
Kahwa, TJ, and Georgans, ND. 1978. A hybrid channel assignment schemes in large-
scale, cellular structured mobile communication systems. IEEE Transactions on
Communications, 26, 432–438.
Karian, ZA, and Dudewicz, EJ (1998). Modern Statistical, Systems, and GPSS Simulation,
2nd edition, CRC Press.
Kelly, OE, Lai, J, Mandayam, NB, Ogielski, AT, Panchal, J, and Yates, RD. 2000. Scalable
parallel simulations of wireless networks with WiPPET: Modeling of radio propagation,
mobility and protocols. Mobile Networks and Applications, 5(3), 199–208, Springer.
Kingston, JH. 1984. Analysis of Algorithms for the Simulation Event List. Ph.D. thesis, Basser
Dept. of Computer Science, Univ. of Sydney, Australia. July.
Krishna, S, Goldsmith, A, and Carlton, M. Seamless handoff, offload, and load balancing
in integrated Wi-Fi/small cell systems. US Patent 9,510,256, 2016 - Google Patents.
Kurkowski, S, Camp, T and Colagrosso, M. MANET Simulation Studies: The Current State
and New Simulation Tools, 2005.
56 Artificial Intelligence (AI)

Lai, WK, and Coghill, GC. 1996. Channel assignment through evolutionary optimization.
IEEE Transactions on Vehicular Technology, 45(1), 91–96.
Lee, WCY. 1995. Mobile Cellular Telecommunications. McGraw-Hill Professional.
ISBN:978-0-07-038089-9, https​:/​/dl​​.acm.​​org​/d​​oi​/ bo​​ok​/10​​.555​5​​/5557​​61
Liu, CL, and Layland, JW. 1973. Scheduling algorithms for multiprogramming in a hard real-
time environment. Journal of the ACM, 20(1), 46–61.
Liu, W, Chiang, CC, Wu, HK, Jha, V, Gerla, M, and Bagrodia, R. 1996. Parallel simula-
tion environment for mobile wireless networks. In Proceedings of the 1996 Winter
Simulation Conference, WSC’96, 605–612.
Liu, J, Perrone, LF, Nicol, DM, Liljenstam, M, Elliott, C, and Pearson, D, 2001, June.
Simulation Modeling of Large-Scale ad-hoc Sensor Networks. In European Simulation
Interoperability Workshop (Vol. 200, No. 1).
Locke, CD. 1992. Software architecture for hard real-time applications: Cyclic executives vs.
fixed priority executives. The Journal of Real-Time Systems, 4, 37–53.
MacDonald, VH. 1979. The cellular concepts. The Bell System Technical, Journal, 58,
15–42.
Maes, P. 1995. Artificial life meets entertainment: Life like autonomous agents.
Communications of the ACM, 38(11), 108–114.
McCormack, WM, and Sargent, RG. 1981. Analysis of future event-set algorithms for dis-
crete event simulation. Communications of the ACM, 24(12), 801–812.
Mehta, S, Kwak, KS, and Najnin, S. 2010. Network and System Simulation Tools for Next
Generation Networks: A Case Study. Modelling, Simulation and Identification, Azah
Mohamed (Ed.), ISBN: 978-953-307-136-7, InTech, Available from: http:​/​/www​​.inte​​
chope​​n​.com​​/ book​​s​/mod​​ellin​​g-​-si​​mulat​​ion​-a​​ndide​​ntifi​​catio​​n ​/net​​work-​​and​-s​​ystem​​-simu​​
latio​​n​-too​​ls​-fo​​r​-wir​​​eless​​-netw​​orks-​​a​-cas​​e​-stu​​dy
Miller, J, Weyrich, O, and Suen, D. 1988. A software engineering oriented compari-son of
simulation languages. In Proceedings of the 1988 Eastern Simulation Conferences:
Tools for the Simulationists. Orlando, FL, April.
Misra, J. 1986. Distributed discrete-event simulation, ACM. Computing Surveys, (CSUR),
18(1), 39-65, March.1
Molisch, AF. 2012. Wireless Communications, Vol. 34. John Wiley & Sons, UK, 2nd edition,
ISBN: 1118355687, 9781118355688
Nahas, M. 2014. Studying the impact of scheduler implementation on task jitter in real-
time resource-constrained embedded systems. Journal of Embedded Systems, 2(3),
39–52.
Naoumov, V, and Gross, T. 2003. Simulation of Large Ad Hoc Networks, MSWiM’03. In
Proceedings of the 6th ACM international workshop on Modeling analysis and simula-
tion of wireless and mobile systems (pp. 50–57), September, San Diego, CA.
Nierstrasz, O. 1989. A survey of object-oriented concepts. In W Kim and F Lochovsk, editors,
Object- Oriented Concepts, Databases and Applications. ACM Press, Addison Wesley,
3–21.
Norman, T, and Long, D. 1995. Goal creation in motivated agents. In Proceedings of the
Workshop on Agent Theories, Architectures, and Languages on Intelligent Agents.
Papazoglou, PM, and Karras, DA. 2016. A conceptual multi-agent modeling of dynamic
scheduling in wireless sensor networks. Lecture Notes in Electrical Engineering, 348,
385–398. Publisher Springer-Verlag, ISSN:1876-1100.
Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2008a. An improved multi-agent
simulation methodology for modelling and evaluating wireless communication sys-
tems resource allocation algorithms. Journal of Universal Computer Science, 14(7),
1061–1079.
Large-Scale Multi-Agent Systems Simulation  57

Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2008b. On a new generation of event
scheduling algorithms and evaluation techniques for efficient simulation modelling
of large scale cellular networks bandwidth management based on multitasking the-
ory. WSEAS Transactions on Communications, 7(10), 1024–1034, Publisher: World
Scientific and Engineering Academy and Society (WSEAS) Press, ISSN: 1109-2742.
Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2009. On the efficient implementa-
tion of a high performance multi-agent simulation system for modeling cellular com-
munications involving a novel event scheduling. International Journal of Simulation
Systems, Science & Technology, Special Issue on: Performance Engineering, 10(1).
Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2011. Evaluating novel DCA varia-
tions for efficient channel assignment in cellular communications through a generic
Java simulation system. In Proceedings of 2011 Intern. Conference on Computer
Science and Network Technology. Harbin, December 24–26.
Papazoglou, PM, Karras, DA, and Papademetriou, RC. 2016. On integrating natural comput-
ing based optimization with channel assignment mining and decision making towards
efficient spectrum reuse in cellular networks modelled through multi-agent system
schemes. Lecture Notes in Electrical Engineering, 348, 783–798. Publisher Springer-
Verlag, ISSN:1876-1100.
Pincus, R. (1995). Barnett, V., and Lewis T.: Outliers in Statistical Data. J. Wiley & Sons
1994, XVII. 582 pp.,£ 49.95. Biometrical Journal, 37(2), 256–256.
Pont, MJ, Kurian, S, Wang, H, and Phatrapornnant, T. 2007. Selecting an appropriate sched-
uler for use with time-triggered embedded systems. In EuroPLoP, 595–618.
Reddy, D, et al. 2006. Measuring and explaining differences in wireless simulation models.
In 14th IEEE International Symposium on Modeling, Analysis, and Simulation. IEEE.
Russell, EC. 1999. Building Simulation Models with Simscript II.5. CACI Products Co.,
CACI Products Company, 3333 North Torrey Pines Court La Jolla, California 92037,
https​:/​/ww​​w​.cos​​c​.bro​​cku​.c​​a ​/Off​​ering​​s​/4P9​​4​/sii​​​5​/zbu​​ildin​​.pdf
Russell, S, and Norvig, P. 2003. Artificial Intelligence: A Modern Approach, 2nd ed. Upper
Saddle River, NJ: Pearson Education.
Schriber, TJ, and Brunner, DT. 1997. Inside discrete-event simulation software: How it works
and why it matters. In Proceedings of the 1997 Winter Simulation Conference.
Siangsukone, T, Aswakul, C, and Wuttisittikulkij, L. 2003. Study of optimised bucket widths
in calendar queue for discrete event simulator. In Thailand’s Electrical Engineering
Conference (EECON-26).
Siow, R, Goh, M, and Thng, L-J. 2004. DSplay: An efficient dynamic priority queue structure
for discrete event simulation. In Simtect Simulation Conference.
Sivarajan, KN, McEliece, RJ, and Ketchum, JW. 1990. Dynamic channel assignment in cel-
lular radio. In IEEE 40th Vehicular Technology Conference, 631–637.
Smith, DC, Cypher, A, and Spohrer, J. 1994. KidSim: Programming agents without a pro-
gramming language. Communications of the ACM, 37(7), 55–67.
Splunter, S, Wijngaards, N, and Brazier, F. 2003. Structuring agents for adaptation. In E
Alonso et al. editors. Adaptive Agents and Multi-Agent Systems, LNAI, Vol. 2636, 174–
186, Heidelberg: Springer, Berlin.
Sprunt, B. 1990. Aperiodic Task Scheduling for Real-Time Systems. PhD dissertation.
Carnegie Mellon University.
Subramania Sharma, T, and Thazhuthaveetil, MJ. 2001. TWLinuX: Operating system sup-
port for optimistic parallel discrete event simulation. High Performance Computing–
HiPC 2001, 262.
Sycara, K. 1998. Multi-agent systems. Artificial Intelligence Magazine, 19(2), American
Association for Artificial Intelligence, 0738-4602-1998.
58 Artificial Intelligence (AI)

Takai, M, Bagrodia, R, Lee, A, and Gerla, M (1999, August). Impact of channel models on
simulation of large scale wireless networks. In Proceedings of the 2nd ACM interna-
tional workshop on Modeling, analysis and simulation of wireless and mobile systems
(pp. 7–14).
Tan, KL, and Thng, L-J. 2000. Snoopy calendar queue. In Proceedings of the 2000 Winter
Simulation Conference.
Tolk, A. 2012. Verification and validation. Engineering Principles of Combat Modeling and
Distributed Simulation, pp. 263–294, published by John Wiley and Sons, NJ, USA,
ISBN: 978-0-470-87429-5
Tropper, C. 2002. Parallel discrete-event simulation applications. Journal of Parallel and
Distributed Computing, Volume 62, Issue 3, March 2002, 327–335, Elsevier
West, J, and Mullarney, A. 1998. ModSim: A language for distributed simulation. In
Proceedings SCS Multi Conference on Distributed Simulation. San Diego, CA,
February, 155–159.
Whitaker, P. 2001. The Simulation of Synchronous Reactive Systems in Ptolemy II. Electronics
Research Laboratory, College of Engineering, University of California, https​:/​/pt​​olemy​​
.berk​​eley.​​edu​/p​​ublic​​ation​​s​/pap​​ers​/0​​​1​/sr/​​sr​.pd​f
Wooldridge, M. 2009. An Introduction to Multi-Agent Systems, 2nd edition. John Wiley &
Sons, UK, ISBN: 978-0-470-51946-2
Wooldridge, M, and Jennings, NR. 1995. Intelligent agents: theory and practice. The
Knowledge Engineering Review, 10(2), 115–152.
Xu, Z, and Mirchandani, PB. 1982. Virtually fixed channel assignment for cellular radio-
telephone systems: A model and evaluation. In IEEE International Conference on
Communications, ICC’92, Chicago, Vol. 2, 1037–1041.
Zeng, X, Bagrodia, R, and Gerla, M. 1998. GloMoSim: A library for parallel simulation of
large-scale wireless networks. In Proceedings of the 12th Workshop on Parallel and
Distributed Simulations.
Zhang, M, and Yum, TS. 1989. Comparisons of channel assignment strategies in cellular
mobile telephone systems. IEEE Transactions on Vehicular Technology, 38(4), 211–215.
Zhang, X, et al. 2001. A proposed approach to sophisticated negotiation. In AAAI Fall
Symposium on Negotiation Methods for Autonomous Cooperative Systems. November.
 2 Why Do Users React
Let’s Find Out

Differently to Applications
Infused with AI Algorithms?
Jayden Khakurel, Indu Manimaran, and Jari Porras

CONTENTS
2.1 Introduction..................................................................................................... 59
2.2 Related Work and Hypothesis Formulation..................................................... 61
2.2.1 Excitement........................................................................................... 62
2.2.2 Anger................................................................................................... 62
2.2.3 Desire................................................................................................... 62
2.2.4 Happiness............................................................................................. 63
2.2.5 Relax.................................................................................................... 63
2.3 Methodology.................................................................................................... 63
2.3.1 Participants..........................................................................................64
2.3.2 Procedure.............................................................................................64
2.4 Findings........................................................................................................... 65
2.4.1 Descriptive Statistics and Hypothesis Testing Outcomes.................... 65
2.4.2 Qualitative Feedback...........................................................................66
2.5 Discussions......................................................................................................66
2.6 Limitations and Future Work.......................................................................... 68
2.7 Conclusion....................................................................................................... 68
References................................................................................................................. 68

2.1 INTRODUCTION
In recent years, researchers and application developers have increasingly begun
to infuse applications with machine learning or natural language processing algo-
rithms that offer robust empirical performance to tackle real human problems and
enhance the quality of life in various domains (Khakurel et al. 2018; Inkpen et
al. 2019). Russell, Moskowitz, and Raglin (2017) point out, “Humans’ interaction
with information will only increase in the future, and this interaction will likely be
facilitated by artificial intelligent proxies” (p. 33). This supports the idea that appli-
cations infused with artificial intelligence (AI) algorithms will become ubiquitous
in our lifetimes, and humans will both interact and integrate with these programs.

59
60 Artificial Intelligence (AI)

Eventually, these programs will replace human–human with human–machine

communication (endowed with more or less artificial intelligence) (Hilbert and
Aravindakshan 2018). Evidence has emerged that there are similarities (e.g., user’s
alignment) and differences (e.g., less engagement) between human and artificial
intelligence interactions; however, what causes users to behave differently when
interacting with AI is still unknown (Rzepka and Berger 2018). Further, while
some studies (Yannakakis and Togelius 2018; Amershi et al. 2019) show that apps
with more or less AI are used to facilitate entire or partial automated communi-
cation, for example, via Chatbot, personal assistant, or, in some scenarios, video
games controlled by non-player characters, the AI may demonstrate unpredict-
able behaviours that can be disruptive, confusing, offensive, and even dangerous.
These unpredictable behaviours during human–AI interactions may backfire and
induce anger, anxiety, and/or feelings of discomfort among individuals, result-
ing in negative service evaluations, customer dissatisfaction, and increased switch
intentions (Feng et al. 2019). To examine this phenomenon, we explored the fol-
lowing research questions: (i) What causes users to behave differently toward an
app using AI? (ii) How can we minimize unpredictable behaviours and improve
user engagement with these apps and technologies?
To look at AI–human interactions, it is crucial to look at emotions in general.
Here, emerging insights into discrete state emotions (DSEs) indicate that each dis-
crete state emotion, such as excitement, happiness, anger, relaxation, and desire, has
a unique mechanism that causes a unique mental state with measurable outcomes
and has an influence on human decisions; indeed, DSEs are important in understand-
ing and predicting a user’s decisions and behaviours (Angie et al. 2011; Harris and
Isaacowitz 2015; Harmon-Jones et al. 2016). Therefore, in this study, we formulate
the main hypothesis that either the lack or presence of DSEs may cause differences
in interaction behaviours when users work with AI-infused apps. To test the main
hypothesis, five sub-hypotheses were formulated, which are explained in Section
2.2. The testing of these sub-hypotheses was carried out in two sessions using the
complete counterbalancing technique to control the order effect which is generally
recommended for conditions (K!) less than 4, K ≤ 4 (Allen Mike 2017). For instance,
in this study, there are two conditions (k! 2): condition A: apps game experience;
and condition B: dyadic game experience. Using all possible orders, two different
combinations – AB, BA – are generated. Two groups were formulated in which
group 1 (14 participants) interacted with conditions A and then B. During the app
game experience, participants held their phone with one hand and interacted with an
android game app based on the traditional game “rock–paper–scissors,” also known
as “roshambo,” but infused with machine learning algorithms to execute roshambo
moves. During dyadic game interactions, players who participated in the first session
played the roshambo game against other participants.
After each session, DSEs were collected using the Discrete Emotions Questionnaire
(DEQ) (Harmon-Jones et al. 2016). Having data from separate sessions enabled us
to identify the variation in DSEs between the two sessions. Further, we analyzed
the collected data using the statistical data analysis language R and the descriptive
statistical analysis functions available in R core (R Core Team 2017) and the psych
Let’s Find Out 61

library (Revelle 2017). The results expand the existing research on the influence of
DSEs on users’ interaction behaviours, providing informal guidelines that can be
used to minimize unpredictable behaviours and improve user engagement with apps
or technologies.
The remainder of this chapter is structured as follows. Section 2.2 provides related
work and hypothesis formulation. Section 2.3 then provides the methodology applied
in the study. We discuss the findings in Section 2.4, and finally, we present the dis-
cussion in Section 2.5 and promising avenues for further research in Section 2.6.

2.2 RELATED WORK AND HYPOTHESIS FORMULATION

We focus in particular on the interactions between users and AI, looking at users’
experiences, needs, and challenges. Rzepka and Berger (2018) conducted a litera-
ture review and aggregated the knowledge regarding human–AI interactions. Their
review revealed that (i) users’ interactions with AI systems trigger contradictory
behavioural responses; (ii) AI systems trigger perceptions of threat among users;
and (iii) users assign humanness and social characteristics to AI systems.
Väänänen et al. (2019) conducted a pilot survey and adopted a Geneva emotion
wheel to understand users’ positive and negative experiences and descriptions of
expected or potential experiences of AI in their everyday lives; they found that
the participants experienced some sort of negative experiences, such as a sense
of anger, disappointment, and irritation. Further, the participants expressed that
they did not want AI to feel too humanized. On the other hand, in terms of posi-
tive experiences, a sense of control and trust/reliability were the most reported
wanted experiences in addition to relief, feeling of safety, efficiency, satisfaction,
contentment, and pleasure. Subsequently, the authors suggested that understand-
ing, collaborating, and sensitive interactions could strengthen the human–AI
relationship.
Emotion, the key component that drives human behaviour, plays an essential role
in human cognitive processes, problem-solving competence, and decision-making
(de Freitas et al. 2005; Nass and Brave 2010; Khakurel 2018). Psychological studies
of emotions have been structured into two levels: the dimensional level that pos-
its emotions come in the form of dimensions such as arousal (high/low intensity),
valence (positive/negative), and motivation direction (approach/avoid) (Corener Paul
2013; Harmon-Jones et al. 2016, 2017), and the discrete level, which posits emotions
come in the form of discrete states such as anger, fear, joy, and so on. According to
Harmon-Jones et al. (2017), “each primary process or ‘basic’ discrete emotion is pos-
ited to evoke a specific response tendency that will address a specific evolutionarily
important need (e.g., protection from harm by fear, rejection of harmful substances
by disgust)” (p. 2).
Given the influence of a user’s discrete emotions, we derive the main hypothesis
that either the lack or presence of DSEs may cause differences in interaction behav-
iours when users work with AI-infused apps. To establish the main hypothesis, a set
of sub-hypotheses were formulated – H1, H2, H3, H4, and H5 – that focus on game-
based apps and that are based on the DSE states.
62 Artificial Intelligence (AI)

2.2.1 Excitement
Excitement refers to an individual’s response to a situation which increases the
adequacy; it supplements the routine modes of responses which appear inadequate
(Stratton 1928). We propose that when a player interacts with opponents, there
appear to be surprises (i.e., don’t know what to expect in the next move), which usu-
ally excite the other player more than when interacting with precise and pragmatic
games, which are more dry in nature unless the players are playing for the sake
of playing them, addicted to, or wanting to win the games. Such excitement may
cause other players to cry if the excitation is strong or can result in wild laughter.
Therefore, we hypothesize that playing with a human as opponent will lead to excite-
ment, whereas playing against AI-infused apps will lead to less excitement, affecting
a user’s interaction behaviour. H1: Playing with AI-infused apps will lead to less
excitement and thus generate different behavioural responses.

2.2.2 Anger
Williams (2017) and Mill et al. (2018) refer to the term “anger” as a multifaceted
construct that entails a negative activation and can be perceived as a threat to an
individual’s emotional well-being. In line with Charlton (2009) and Hadlington
and Scase (2018), we propose that a poor game app experience with an AI app,
including the app’s self-efficacy and related anxiety, can be a detrimental fac-
tor that may frustrate users, triggering an anger experience. We hypothesize that
playing against AI gaming apps triggers a stronger anger experience than when
playing against a human opponent, which reflects what Mill et al. (2018) refer to
as the “appraisal of an anger-provoking or frustrating situation triggers the anger
experience which, in turn, generates different behavioral responses” (p. 739). In
contrast, we propose that while playing with a human, individuals will use more
coping strategies, and anger will hence be less intense because both players could
solve the matter with a wide range of discussion. H2: Playing with AI-infused
apps will lead to frustration among users, triggering anger experiences and thus
generating different behavioural responses.

2.2.3 Desire
Zalta (Schroeder 2017) refers to desire as a state of mind that is commonly associ-
ated with a number of different effects. Furthermore, the author also points out that
a person with a desire tends to act in certain ways, feel in certain ways, and think in
certain ways. In the context of the current study, during a game experience, players
may have the desire to win the game, regardless of who they are playing against.
Previous studies have shown that when it comes to desire in game experiences, indi-
viduals take relative gains more vigorously than personal gains because they either
have the desire to defeat, or create a flow of experiences during the game, or do not
like receiving less than others (Messick and Thorngate 1967; Meyer-Parlapanis et al.
2017). However, we propose that when playing against AI-infused apps, the desire to
Let’s Find Out 63

win is heightened more because of the collision between two perceptual factors, i.e.,
time pressure and achieving the highest score. Therefore, we hypothesize that play-
ing with an AI game app increases the desire to win compared with playing against
a human opponent, leading the user to behave differently. H3: Increased feelings
of the desire “to win” while playing against AI-infused apps will evoke different
behavioural responses.

2.2.4 Happiness
Previous studies have conceptualized happiness as the positive emotions formed due
to general interaction between internal (endogenic) and external (exogenic) factors,
which are associated with and precede numerous successful outcomes (Lyubomirsky
et al. 2005; Dfarhud et al. 2014; Harmon-Jones et al. 2016). In the context of the
current study, we propose individuals are less happy during and after the game
experience with the AI because of (i) social disengagement (Baym et al. 2004), (ii)
presence of parameters related to game addiction (i.e., salience, mood modification,
tolerance, withdrawal, conflict, and relapse [Griffiths 2008; Hull et al. 2013]), and
(iii) frustrating experiences (i.e., error messages, dropped network connections, long
download times, and hard-to-find features [Ceaparu et al. 2004]). In line with Mehl
et al. (2010), in face-to-face interactions (i.e., conversations while having the game
experience), excitement may actually make individuals happier than the actual game
experience with the AI app. Therefore, we hypothesize that interpersonal communi-
cation and other DSE entities, namely excitement, may cause individuals to remain
happier while playing against a human opponent. H4: The interpersonal communica-
tion and other DSE entities resulting from the dyadic game experience will influence
happiness, generating more positive behavioural responses than the game experience
with an AI app.

2.2.5 Relax
Previous studies (Smith 2007) have organized relaxation states into four groups:
basic relaxation, core mindfulness, positive energy, and transcendence. These relax-
ation states are essential in creating a relaxation response (Benson et al. 1974). In the
context of the current study, we propose that a face-to-face game experience will let
players disclose personal items, build impressions, and compare values (Baym et al.
2004), which will eventually lead to basic relaxation. Therefore, we hypothesize that
playing a game against a human opponent makes a player more relaxed than play-
ing against an AI app, leading the user to behave differently. H5: Having a human
opponent during the game experience makes players more relaxed than having an
AI opponent.

2.3 METHODOLOGY
The following sections elaborate on the methodological approach used in the current
study.
64 Artificial Intelligence (AI)

2.3.1 Participants
A total of 26 participants (19 males, 7 females; mean age: 25.6 years; standard devia-
tion: 2.5; with adequate knowledge about technology; 12 participants were outgoing
and enjoyed being with a lot of people; 15 participants enjoyed more solitude) had
mobile devices using an Android operating system (OS) were voluntarily recruited
at the university by word-of-mouth. The sample size is sufficient based on the rec-
ommendation from Macefield (2009), where the author states that a group size of
3–20 participants is typically valid, with 5–10 participants demonstrating a sensible
baseline range in experimental studies. All of the participants were presented with
an informed consent form detailing participants’ right to confidentiality, risks, data
storage, the use of anonymized data, the voluntary nature of participation, and that
no health-related data would be collected. No incentives were provided in exchange
for participation.

2.3.2 Procedure
We adopted the complete counterbalancing technique to control the order effect,
which is generally recommended for conditions (K!) less than 4, K ≤ 4 (Allen Mike
2017). For instance, in this study, there are two conditions (k! 2), condition A: apps
game experience, and condition B: dyadic game experience. Using all possible
orders, two different combinations – AB, BA – were generated. Two groups (a total
of 26 participants) were formulated in which group 1 (14 participants) interacted
with conditions A and then B. Figure 2.1 shows the complete counterbalancing for
an experiment with two conditions adopted in this study.
During the first session, group 1 (14 participants) interacted with condition A,
where participants were asked to download an android-based game app based on the
traditional game “rock–paper–scissors”; this app used machine learning algorithms.
The participants were instructed to play exactly 30 rounds. While playing, players
made a hand gesture for rock, paper, or scissors before the timer ran out. The win-
ner of the round was decided based on the standard rules. The players then reported
the DSEs that they experienced while interacting with the game through the DEQ
adopted from Harmon-Jones et al. (2016). In a survey, the participants were asked to
rate on a 7-point Likert scale (1 = not at all to 7 = an extreme amount): “While playing
with a mobile game with integrated AI algorithms, to what extent did you experience

FIGURE 2.1  Complete counterbalancing for an experiment with two conditions.

Let’s Find Out 65

these emotions?” Furthermore, the participants were asked to provide qualitative

feedback of their experience.
Similarly, group 2 (12 participants) interacted with condition B (see Figure 2.1).
During the dyadic game experience, each pair was asked to play “rock–paper–scis-
sors” with another participant for a total number of 30 rounds. Hence, a player must
guess what their opponent will choose and pick the appropriate object to beat them
(University of Stirling 2017). Each player then reported the DSEs separately that
they experienced while interacting with the game through the DEQ adopted from
Harmon-Jones et al. (2016). In a survey, the participants were asked to rate on a
7-point Likert scale (1 = not at all to 7 = an extreme amount): “While with your part-
ners, to what extent did you experience these emotions?” Furthermore, the partici-
pants were asked to provide qualitative feedback on their experience.
During the second session, group 1 (14 participants) interacted with condition
B, whereas group 2 (12 participants) interacted with condition A to eliminate order
effects. The participants performed a similar task, filled the survey, and were pro-
vided feedback after the first session.
Quantitative data collected from the two sessions were analyzed using the sta-
tistical data analysis language R and the descriptive statistical analysis functions
available in R core (R Core Team 2017) and the psych library (Revelle 2017). We
first used the Mann–Whitney U test (Wohlin et al. 2012) to analyze the difference in
distributions between the data sets. A continuity correction was enabled to compen-
sate for non-continuous variables (Bergmann and Ludbrook 2000). The Bonferroni
correction was used to adjust the p-value to compensate for the family-wise error
rate in multiple comparisons (Abdi 2007). We calculated the effect size r using the
guidelines by Tofan et al. (2016) for the Mann–Whitney U test. We evaluated the
effect size as proposed by Cohen (1994): in r, a large effect is 0.5, a medium effect is
0.3, and a small effect is 0.1.

2.4 FINDINGS
2.4.1 Descriptive Statistics and Hypothesis Testing Outcomes
The results indicate that participants remained excited (M = 6.54, SD = 0.859),
had desire (M = 6.62, SD = 0.571), were happy (M = 5.77, SD = 0.704), and were
relaxed (M = 6.50, SD = 1.273) while playing against a human opponent compared
with the AI app. The results also indicate that individuals were angrier (M = 2.58,
SD = 1.79) while playing against the AI compared with a human opponent. To
establish the main hypotheses, the sub-hypotheses formulated in Section 2.3 were
tested, they are summarized in Appendix A. The results indicate that all the sub-
hypotheses are correct. The results show DSEs, namely excitement (p < 0.001;
r = 0.74), happiness (p < 0.001; r = 0.59), and relaxation (p < 0.001; r = 0.71), were
felt during the dyadic experience, whereas there were increased feelings of the
desire “to win” when playing against the AI-infused app. However, anger caused
because of frustration (p < 0.05; r = 0.38) was found to have medium effect on
the app’s game experience. The research findings support that either a lack or
66 Artificial Intelligence (AI)

presence of DSEs may cause differences in interaction behaviours when users

play with AI-infused apps.

2.4.2 Qualitative Feedback
We further analyzed our qualitative data that measured the overall experience after
playing the games to see what essential challenges affected the emotions of users
while interacting with the AI app. Many participants commented that the interac-
tion between humans was more fun, more competitive, and quicker than with the
AI app. For example, P6 commented, “Human is more fun definitely, but AI could
be an alternative when you want to play and no one is around.” The comment shows
that users may use the mobile app as an alternative method when lonely. Similarly,
P7 stated, “Human was more fun, AI takes more time to recognize the gesture, so I
feel that communication is easier and fun with humans.” Further, P12 commented,
“Human is more fun as the game is much quicker and competitive.” Consequently,
some participants also commented that AI game lack a certain level of intensity.
For example, P8 commented, “Playing with human is more fun. Rock, paper, scis-
sors should be more intense, and I didn’t feel the intensity. Good work on the app
though!” Here, further research should be conducted on why users like the highly
intense AI games and to what level the game should be intensified.
Further, out of 26 participants, 7 expressed that they had challenges regarding
gesture detection when using the app, which caused frustration. For example P4
commented, “The app didn’t recognize my hand gesture in a high percentage of
cases. I wasn’t sure if I should hold my hand vertically or horizontally.” P13 stated,
“The app recognizes rock all the time. It never recognizes scissors!” P3 noted, “Out
of 10 rounds only two rounds were recognized properly … makes me feel frustrated.”
Regarding enriching the user experience, the need for interactive visual feedback in
the form of either graphics or emojis was stated. For example, P12 said, “The app
should give some graphical feedback or emoji signs to make it more interactive.”

2.5 DISCUSSIONS
In light of concerns about why users behave differently with AI-infused apps/systems,
the current study performed two tests. From a theoretical point of view, the results
indicate that DSEs are important, and AI-based apps should have an ability to evoke
and include emotions such as happiness, excitements, and relaxation at the same level
as they have during dyadic interaction to develop mutual emotional attachments. The
finding is in line with the proposal presented by LaGrandeur (2015), who states that
inducing emotions is important to create safer and more attractive AIs, allowing both
humans and AI to develop mutual emotional attachments. Brave and Nass (2002) also
state that “any interface that ignores a user’s emotional state or fails to manifest the
appropriate emotion can dramatically impede performance and risks being perceived
as cold, socially inept, untrustworthy, and incompetent” (p. 82). Further, supporting
previous research (Thüring and Mahlke 2007), this study further reveals that poor
app experience triggers the anger experience among users that leads them to behave
Let’s Find Out 67

differently. Based on the current study’s findings, we suggest the below informal guide-
lines for application developers seeking to design better user experiences and mini-
mize unpredictable behaviours, such as anger experience. These guidelines may not
adequately address all AI-infused apps; however, they can be taken into consideration.
Improving the detection rate of gesture: although the model performed well dur-
ing development, the participants reported that they were frustrated while playing
because sometimes the app failed to detect their hand gestures. Thus, we recommend
developers to (i) use a public data set to train the model offline and update the appli-
cations more frequently to improve gesture detection n, (ii) use images that are taken
in different lighting conditions and with different backgrounds, (iii) to quantify the
detection performance, use Detection Error Trade-off (DET) curves and miss rate
versus False Positive Per Window (FPPW) as proposed by Lahiani and Neji (2018),
and (iv) apply the gesture recognition method based on a convolutional neural net-
work (CNN) and deep convolution generative adversarial networks (DCGAN) pro-
posed by Fang et al. (2019), which can train the model using fewer samples and
achieve better gesture classification and detection effects.
Considering interference time, response time, and progress indication with the
AI-infused app to improve the user interaction: some participants reported that they
were impatient because of the response time when interacting with the app. AI-infused
apps are built with (i) a set of powerful algorithms within the app itself that can be used
to collect the data sets, train the model, and display the results or (ii) can use different
algorithms on the data set on the server side to analyze patterns in data and make pre-
dictions, displaying the results on the app by synchronizing them between the app and
the server. In the latter scenario, there might be delays in the response because of mobile
computational power and server response time because of slow database queries, librar-
ies, and resource central processing unit (CPU) starvation, and slow application logic,
which may increase users’ cognitive load and affect their performance (Alnanih and
Ormandjieva 2016). Therefore, we recommend application developers (i) consider run-
ning the longer running operation as the background task; (ii) test the interference time;
and (iii) test the response time on apps utilizing different time frames, as mentioned by
Nielsen (1993). Furthermore, in order to reduce uncertainty among users, we recom-
mend developers to provide an accurate estimation of the waiting times using either a
determinate or an indeterminate progress indicator (Luo 2017).
Automatic recognition of emotional states and visualizing in the form of graphi-
cal feedback or digital pictograms/emojis to develop human–AI emotional attach-
ments: some participants were more interested in having graphics or having digital
pictograms such as emoji as a way to receive feedback. This may be because apps
that use emojis as a feedback method communicate a positive effect, specifically
joy (Riordan 2017), or the graphical feedback makes users less frustrated and more
engaged compared with textual feedback (Rieber et al. 1996). Stark and Crawford
(2015) point out, “Emojis can act as an emotional coping strategy and a novel form
of creative expression, even if, in both cases, working within real limits” (p. 1).
Therefore, the form of feedback should manifest the appropriate DSEs with a new
algorithm that describes the relationship between cognition and emotion better based
on behavioural science and neuroscience (Wang et al. 2016). Aiming at inducing
68 Artificial Intelligence (AI)

and evoking the emotions on apps through graphical feedback or digital pictograms/
emojis, developers could refer to the work of Gao et al. (2012) on how tactile behav-
iour on the screen can reflect users’ emotional state.

2.6 LIMITATIONS AND FUTURE WORK

We describe below several limitations. These limitations illustrate improve-
ments that can be considered in future research. First, we only investigated self-
reported DSEs with small samples and a case study on AI mobile games, and
the results may not be generalized and inapplicable to other AI-infused apps.
Therefore, the results from the current study can only be taken into consider-
ation as informal guidance. Furthermore, because AI could be used in various
app categories in different contexts, researchers may want to conduct further
research with larger samples to identify which DSEs appear in which context,
hence improving the AI–human relationship. Second, demographics such as
age, gender, and culture could be added to future work to identify if DSEs vary
during AI–human interactions.

2.7 CONCLUSION
In recent years, AI agency has been applied in forefront applications to simulate
the intelligent behaviour and critical thinking to perform a range of activities that
require human intelligence, such as decision-making, learning, sensing, and chal-
lenging humans through games. However, previous studies have shown that there is a
need to understand the constraints of a user’s interaction behaviours with AI-infused
apps. We formulated and tested a hypothesis using the perspective of DSEs. The con-
tributions of the current chapter are twofold. First, this research provides new knowl-
edge about the influence of DSEs on users’ interaction behaviours with AI-infused
apps. Second, it provides an informal guideline on how to minimize unpredictable
user interaction behaviours and improve their acceptance of these services or tech-
nologies in the future. In conclusion, the present study suggests DSEs should be an
indispensable part of AI-infused apps (i) for individuals to react the same way as
they act during human–human interactions and (ii) to minimize their adverse effects
to enrich human–AI relationships.
Appendix A: https://doi​.org​/10​.5281​/zenodo​.3361476
Acknowledgements. Thank you to Antti Knutas for your contribution during the
data analysis phase and all the reviewers for their valuable comments and consider-
able time and effort. The first author would like to thank INVEST Research Flagship
funded by the Academy of Finland Flagship Programme (decision number: 320162).

REFERENCES
Abdi H (2007) The Bonferonni and Šidák Corrections for Multiple Comparisons. In: Salkind
N (ed) Encyclopedia of Measurement and Statistics. SAGE Publications Inc., Thousand
Oaks, CA.
Let’s Find Out 69

Allen M (2017) The sage encyclopedia of communication research methods (Vols. 1–4).
SAGE Publications, Inc, Thousand Oaks, CA. doi: 10.4135/9781483381411
Alnanih R, Ormandjieva O (2016) Mapping HCI Principles to Design Quality of Mobile User
Interfaces in Healthcare Applications. Procedia Comput Sci 94:75–82. ISSN 1877-
0509, https​:/​/do​​i​.org​​/10​.1​​016​/j​​.proc​​s​.20​1​​6​.08.​​014
Amershi S, Weld D, Vorvoreanu M, et al (2019) Guidelines for Human-AI Interaction. In:
CHI Conference on Human Factors in Computing Systems Proceedings (CHI 2019).
Angie AD, Connelly S, Waples EP, Kligyte V (2011) The Influence of Discrete Emotions on
Judgement and Decision-Making: A Meta-Analytic Review. Cogn Emot 25:1393–1422.
https​:/​/do​​i​.org​​/10​.1​​080​/0​​26999​​31​.20​​​10​.55​​0751
Baym NK, Zhang YB, Lin MC (2004) Social Interactions Across Media: Interpersonal
Communication on the Internet, Telephone and Face-To-Face. New Media Soc. https://
doi​.org​/10​.1177​/1461444804041438
Benson H, Beary JF, Carol MP (1974) The Relaxation Response. Psychiatry. https​:/​/do​​i​.org​​
/10​.1​​080​/0​​03327​​47​.19​​7​4​.11​​02378​5
Bergmann R, Ludbrook J (2000) Different Outcomes of the Wilcoxon—Mann—Whitney
Test from Different Statistics Packages. Am Stat 54:72–77. https​:/​/do​​i​.org​​/10​.1​​080​/0​​
00313​​05​.20​​0​0​.10​​47451​3
Brave S, Nass C (2002) Emotion in Human–Computer Interaction. In: Julie A. J. L. (ed)
The Human-Computer Interaction Handbook. Erlbaum Associates Inc, Hillsdale, NJ,
81–96. doi: 10.1201/9781410615862.ch4
Ceaparu I, Lazar J, Bessiere K, et al (2004) Determining Causes and Severity of End-User
Frustration. Int J Hum Comput Interact 17:333–356. https​:/​/do​​i​.org​​/10​.1​​207​/s​​15327​​
590ij​​​hc170​​3_3
Charlton JP (2009) The Determinants and Expression of Computer-Related Anger. Comput
Human Behav 25:1213–1221. https​:/​/do​​i​.org​​/10​.1​​016​/j​​.chb.​​200​9.​​07​.00​1
Cohen S (1994) Perceived Stress Scale. Psychology 1–3. http:​/​/www​​.mind​​garde​​n​.com​​/prod​​
ucts/​​​pss​.h​​tm.
Corener Paul (2013) The Fine Points of Feelings : Why Discrete Emotions Matter and How
to Reveal Them. Quirks. https​:/​/em​​otive​​analy​​tics.​​com​/t​​he​-fi​​ne​-po​​ints-​​of​-fe​​eling​​-why-​​
discr​​ete​-e​​motio​​ns​-ma​​tter-​​and​​-h​​ow​-to​​-reve​​al​-th​​em/.
de Freitas JS, Gudwin R, Queiroz J (2005) Emotion in Artificial Intelligence and Artificial
Life Research: Facing Problems. In: Panayiotopoulos T, Gratch J, Aylett R, Ballin D,
Olivier P, Rist T (eds) Intelligent Virtual Agents. IVA 2005. Lecture Notes in Computer
Science, vol 3661. Heidelberg, Springer, Berlin. https://doi​.org​/10​.1007​/11550617​_ 52.
Dfarhud D, Malmir M, Khanahmadi M (2014) Happiness & Health: The Biological Factors-
Systematic Review Article. Iran J Public Health 43:1468–1477.
Fang W, Ding Y, Zhang F, Sheng J (2019) Gesture Recognition Based on CNN and DCGAN
for Calculation and Text Output. IEEE Access 7:28230–28237. https​:/​/do​​i​.org​​/10​.1​​109​/
A​​CCESS​​.2019​​​.2901​​930
Feng W, Tu R, Lu T, Zhou Z (2019) Understanding Forced Adoption of Self-Service
Technology: The Impacts of Users’ Psychological Reactance. Behav Inf Technol
38(8):820–832. doi: 10.1080/0144929X.2018.1557745
Gao Y, Bianchi-Berthouze N, Meng H (2012) What Does Touch Tell Us about Emotions
in Touchscreen-Based Gameplay? ACM Trans Comput Interact. https://doi​.org​/10​.1145​
/2395131​.2395138
Griffiths MD (2008) Diagnosis and Management of Video Game Addiction. New Dir Addict
Treat Prev 12:27–41.
Hadlington L, Scase MO (2018) End-user Frustrations and Failures in Digital Technology:
Exploring the Role of Fear of Missing Out, Internet Addiction and Personality. Heliyon.
https​:/​/do​​i​.org​​/10​.1​​016​/j​​.heli​​yon​.2​​​018​.e​​0 0872​
70 Artificial Intelligence (AI)

Harmon-Jones C, Bastian B, Harmon-Jones E (2016) The Discrete Emotions Questionnaire:

A New Tool for Measuring State Self-Reported Emotions. PLoS One 11:e0159915.
https​:/​/do​​i​.org​​/10​.1​​371​/j​​ourna​​l​.pon​​​e​.015​​9915
Harmon-Jones E, Harmon-Jones C, Summerell E (2017) On the Importance of Both
Dimensional and Discrete Models of Emotion. Behav Sci (Basel) 7:66. https://doi​.org​
/10​.3390​/ bs7040066
Harris JA, Isaacowitz D (2015) Emotion in Cognition. In: International Encyclopedia
of the Social & Behavioral Sciences. Elsevier, pp. 461–466. doi: 10.1016/
B978-0-08-097086-8.25003-4
Hilbert M, Aravindakshan A (2018) What Characterizes the Polymodal Media of the Mobile
Phone? The Multiple Media within the World’s Most Popular Medium. Multimodal
Technol Interact 2:37. https://doi​.org​/10​.3390​/mti2030037
Hull DC, Williams GA, Griffiths MD (2013) Video Game Characteristics, Happiness and
Flow as Predictors of Addiction Among Video Game Players: A Pilot Study. J Behav
Addict 2:145–152. https://doi​.org​/10​.1556​/JBA​.2​.2013​.005
Inkpen K, Choudhury M, Chancellor S, et al (2019) Where is the Human? Bridging the Gap
Between AI and HCI. https://michae​.lv​/ai​-hci​-workshop/. Accessed 2 Apr 2019
Khakurel J (2018) Enhancing the Adoption of Quantified Self-Tracking Devices. LUT
University. http://urn​.fi​/ URN​:ISBN​:978​-952​-335​-319-0
Khakurel J, Penzenstadler B, Porras J, et al (2018) The Rise of Artificial Intelligence under
the Lens of Sustainability. Technologies 6:100. https​:/​/do​​i​.org​​/10​.3​​390​/t​​echno​​logie​​​
s6040​​100
LaGrandeur K (2015) Emotion, Artificial Intelligence, and Ethics. In: Romportl J, Zackova
E, Kelemen J (eds) Beyond Artificial Intelligence. Topics in Intelligent Engineering and
Informatics, vol 9. Springer, Cham. https​:/​/do​​i​.org​​/10​.1​​0 07​/9​​78​-3-​​319​​- 0​​9668-​​1_7
Lahiani H, Neji M (2018) Hand Gesture Recognition Method Based on HOG-LBP Features
for Mobile Devices. Procedia Comput Sci 126:254–263. ISSN 1877-0509, https​:/​/do​​i​
.org​​/10​.1​​016​/j​​.proc​​s​.20​1​​8​.07.​​259
Luo G (2017) Toward a Progress Indicator for Machine Learning Model Building and Data
Mining Algorithm Execution. ACM SIGKDD Explor Newsl 19:13–24. https://doi​.org​
/10​.1145​/3166054​.3166057
Lyubomirsky S, King L, Diener E (2005) The Benefits of Frequent Positive Affect: Does
Happiness Lead to Success? Psychol Bull 131:803–855. https​:/​/do​​i​.org​​/10​.1​​037​/0​​033​-2​​
909​​.1​​31​.6.​​803
Macefield R (2009) How To Specify the Participant Group Size for Usability Studies: A
Practitioner ’s Guide. J Usability Stud 5:34–45.
Mehl MR, Vazire S, Holleran SE, Clark CS (2010) Eavesdropping on Happiness: Well-Being
Is Related to Having Less Small Talk and More Substantive Conversations. Psychol
Sci. https://doi​.org​/10​.1177​/0956797610362675
Messick DM, Thorngate WB (1967) Relative Gain Maximization in Experimental Games. J
Exp Soc Psychol. https​:/​/do​​i​.org​​/10​.1​​016​/0​​​022​-1​​031(6​7)900​39-X
Meyer-Parlapanis D, Siefert S, Weierstall R (2017) More Than the Win: The Relation between
Appetitive Competition Motivation, Socialization, and Gender Role Orientation in
Women’s Football. Front Psychol. https://doi​.org​/10​.3389​/fpsyg​.2017​.00547
Mill A, Kööts-Ausmees L, Allik J, Realo A (2018) The Role of Co-Occurring Emotions and
Personality Traits in Anger Expression. Front Psychol. https://doi​.org​/10​.3389​/fpsyg​
.2018​.00123
Nass C, Brave S (2010) Emotion in Human-Computer Interaction. doi: 10.1201/9781410615862.
ch4
Nielsen J (1993) Usability Engineering. Morgan Kaufmann Publishers Inc. doi:
10.1145/1508044.1508050
Let’s Find Out 71

R Core Team (2017) R: A Language and Environment for Statistical Computing. https://www​
.r​-project​.org/
Revelle W (2017) Psych: Procedures for Psychological, Psychometric, and Personality
Research. https://cran​.r​-project​.org​/package​=psych
Rieber LP, Smith M, Al-Ghafry S, et al (1996) The Role of Meaning in Interpreting Graphical
and Textual Feedback During a Computer-Based Simulation. Comput Educ 27:45–58.
https​:/​/do​​i​.org​​/10​.1​​016​/0​​​360 ​-1​​315(9​6)000​05-X
Riordan MA (2017) Emojis as Tools for Emotion Work: Communicating Affect in Text
Messages. J Lang Soc Psychol 36:549–567. https://doi​.org​/10​.1177​/0261927X17704238
Russell S, Moskowitz IS, Raglin A (2017) Human Information Interaction, Artificial
Intelligence, and Errors. In: Autonomy and Artificial Intelligence: A Threat or Savior?
Springer International Publishing, Cham, pp. 71–101. doi: 10.1007/978-3-319-59719-5_4
Rzepka C, Berger B (2018) User Interaction with AI-enabled Systems: A Systematic Review
of IS Research. ICIS Proc no. December:1–17.
Schroeder T (2017) Desire. In: Edward N Zalta (ed) The Stanford Encyclopedia of Philosophy,
Summer 201. Metaphysics Research Lab, Stanford University. https​:/​/pl​​ato​.s​​tanfo​​rd​.ed​​
u​/arc​​hives​​/sum2​​017​/e​​nt​r ie​​s​/des​​ire/
Smith JC (2007) The New Psychology of Relaxation and Renewal. Biofeedback.
Stark L, Crawford K (2015) The Conservatism of Emoji: Work, Affect, and Communication.
Soc Media Soc. https://doi​.org​/10​.1177​/2056305115604853
Stratton GM (1928) The Function of Emotion as Shown Particularly in Excitement. Psychol
Rev 35:351–366. https://doi​.org​/10​.1037​/ h0071406
Thüring M, Mahlke S (2007) Usability, Aesthetics and Emotions in Human-Technology
Interaction. Int J Psychol 42:253–264. https​:/​/do​​i​.org​​/10​.1​​080​/0​​02075​​907​01​​39667​4
Tofan D, Galster M, Lytra I, et al (2016) Empirical Evaluation of a Process to Increase
Consensus in Group Architectural Decision Making. Inf Softw Technol 72:31–47. https​
:/​/do​​i​.org​​/ http​​s:/​/d​​oi​.or​​g​/10.​​1016/​​j​.inf​​so​f​.2​​015​.1​​2​.002​
University of Stirling (2017) Can a Computer Learn Game Strategy as it Plays? Dep Comput
Sci Maths http:​/​/www​​.cs​.s​​tir​.a​​c​.uk/​​~kms/​​schoo​​ls​/rp​​s​​/ind​​ex​.ph​​p. Accessed 28 Nov.
2018.
Väänänen K, Pohjola H, Ahtinen A (2019) Exploring the User Experience of Artificial
Intelligence Applications : User Survey and Human-AI Relationship Model. In: CHI′19
Workshop on Where is the Human? Bridging the Gap Between AI and HCI. Glasgow,
p. 5.
Wang Z, Xie L, Lu T (2016) Research Progress of Artificial Psychology and Artificial
Emotion in China. CAAI Trans Intell Technol. https​:/​/do​​i​.org​​/10​.1​​016​/j​​.trit​​.201​6​​.11​.0​​03
Williams R (2017) Anger as a Basic Emotion and Its Role in Personality Building and
Pathological Growth: The Neuroscientific, Developmental and Clinical Perspectives.
Front Psychol 8. https://doi​.org​/10​.3389​/fpsyg​.2017​.01950
Wohlin C, Runeson P, Höst M, et al (2012) Experimentation in Software Engineering.
Springer Berlin Heidelberg, Berlin.
Yannakakis GN, Togelius J (2018) Artificial Intelligence and Games. Springer International
Publishing, Cham.
 3 AI vs. Machine Learning
vs. Deep Learning
R. Lalitha

CONTENTS
3.1 Introduction: Background and Driving Forces................................................ 73
3.2 Overview of Artificial Intelligence.................................................................. 74
3.3 Steps to Implement Artificial Intelligence Algorithms................................... 74
3.4 When/Where/How/Why to Use Artificial Intelligence?................................. 74
3.5 Examples for Artificial Intelligence Applications........................................... 74
3.6 Overview of Machine Learning...................................................................... 75
3.7 Steps to Implement Machine Learning Algorithms........................................ 75
3.8 When/Where/How/Why to Use Machine Learning?...................................... 75
3.9 Examples for Machine Learning Applications................................................ 76
3.10 Overview of Deep Learning............................................................................ 76
3.11 Steps to Implement Deep Learning Algorithms............................................. 76
3.12 When/Where/How/Why to Use Deep Learning?........................................... 76
3.13 Examples for Deep Learning Applications..................................................... 77
3.14 Comparisons of Artificial Intelligence, Deep Learning, and Machine
Learning.......................................................................................................... 77
3.15 Summary......................................................................................................... 79

3.1 INTRODUCTION: BACKGROUND AND DRIVING FORCES

While knowing about the futuristic issues of artificial intelligence (AI) and its appli-
cations, it is mandatory to understand about the correlation and differences between
artificial intelligence, machine learning (ML), and deep learning (DL) as they are
closely associated with one another. In a nutshell, machine learning forms the subset
of artificial intelligence, and deep learning forms the subset of machine learning.
Hence, any futuristic applications and ideas that emerge out of artificial intelligence
algorithms will certainly contain machine learning algorithms and deep learning
techniques. Therefore, it is essential to know where, when, how, and why machine
learning or deep learning algorithms must be applied for developing innovative
applications with them.
This chapter provides an overview of these domains and narrates their merits
and demerits. The reader will be able to understand the features of each of these
domains. The reader will also be able to know when and where these techniques are
suitable and not suitable for developing applications.

73
74 Artificial Intelligence (AI)

3.2 OVERVIEW OF ARTIFICIAL INTELLIGENCE

Artificial intelligence is a branch of computer science that deals with creating intelli-
gence artificially to a system. The system can be prepared to make decisions like human
brains. The system can plan, move from one place to another, and recognize objects
and sounds. The father of artificial intelligence, John McCarthy, says that “Artificial
Intelligence is the science and engineering of making intelligent machines especially
intelligent computer programs.” Artificial intelligence systems are created by studying
the process of thinking, learning, observing, and decision-making by the human brain.
The main objectives of creating AI systems are to create an expert system and
to implement human intelligence in machines. The best programming languages to
develop AI applications are Python, LISP, C++, Java, and Prolog. The humanoids,
Robots, are examples of physical devices that are upgraded with artificial intelli-
gence. They act as a substitute for human beings with efficiency in work, high speed
in performance, and enormous memory.

3.3 STEPS TO IMPLEMENT ARTIFICIAL INTELLIGENCE ALGORITHMS

1. Study and analyze the real intelligence in a human brain.
2. Understand the problem/scenario.
3. Apply heuristic techniques to solve the problem or else select the appropri-
ate AI technique.
4. Implement the technique to simulate the real intelligence.

3.4 WHEN/WHERE/HOW/WHY TO USE ARTIFICIAL INTELLIGENCE?

AI can be used in any domain where it is essential to exhibit intelligence in the sys-
tem where it is implemented. It can be used in any sector like transport, healthcare,
education, etc. When there is a need to simulate human intelligence and to make any
system to behave like a human, AI can be implemented.
Human intelligence is not permanent and may vanish as years pass by. Hence, to
retain expert knowledge for many decades, it is necessary to create expert systems by
implementing artificial intelligence. AI is also needed to enhance the speed in predic-
tion, forecasting, and decision-making by human efforts. The heuristic approach can
be used to implement artificial intelligence, which includes learning techniques and
experiences of the human expert. A large amount of data is processed and used with
intelligent algorithms, iterative procedures to make a system to work with intelligence.
As technology has become an important and integral part of daily life, artifi-
cial intelligence is needed to automate the processes, reduce the errors, and provide
accurate results at a greater speed.

3.5 EXAMPLES FOR ARTIFICIAL INTELLIGENCE APPLICATIONS

• Alexa is a well-known personal assistant with artificial intelligence that
can voice commands and can execute actions based on that. It is used to set
AI vs. Machine Learning vs. Deep Learning  75

reminders for the users to complete a task in time, to answer questions given
by the user, and to order items online.
• With the help of AI, smart home devices can be controlled and used eas-
ily. Smart voice assistant and smart assistant for old age people are a few
examples.
• AI can be used to diagnose disease in human bodies. For example, earlier
cancer detection with AI is possible; radiology assistant with AI is also
possible.
• Process automation in factories with AI.

3.6 OVERVIEW OF MACHINE LEARNING

Machine learning is the most promising and most relevant domain to apply artificial
intelligence in systems. It is the most common way to process big data. Machine
learning algorithms are designed in such a way that they are self-adaptive and are
able to get new patterns to itself through experience. It is a way of learning from data.
Tom Mitchell defines machine learning as follows: “A computer program is said to
learn from experience (E), with respect to some class of tasks (T), and performance
measures (P), if its performance at tasks Ti as measured by Pi improves with experi-
ence E.” A target is called label in machine learning. A variable is called feature.
The transformation which occurs for the variables is called feature creation. It com-
bines computer science, mathematics, and statistics. Computer science is needed for
implementing the algorithms. Mathematics is needed for developing machine learn-
ing models, and statistics is needed for generating inferences from the data. The best
programming languages for machine learning are Python, R Programming, LISP,
Prolog, and JavaScript.

3.7 STEPS TO IMPLEMENT MACHINE LEARNING ALGORITHMS

1. Identify the data set and prepare it for analysis.
2. Select the appropriate machine learning algorithm.
3. Develop an analytical model based on the selected algorithm.
4. Train the model with the test data.
5. Run and execute the model.

3.8 WHEN/WHERE/HOW/WHY TO USE MACHINE LEARNING?

Machine learning algorithms are used in situations where it is difficult to write code
for the rules and also in cases where huge volumes of data have to be processed for
prediction. It is used in places where process automation is needed, and time con-
sumption has to be minimized. Machine learning is used to execute certain repeti-
tive tasks, identify patterns in the input, and predict outcomes.
Machine learning is needed and important because as the data size grows
exponentially large, computation and prediction become complex. Machine
76 Artificial Intelligence (AI)

learning can be used to adapt to the changes in data by itself and can predict the
desired result easily. As the data size grows enormously in any domain, com-
putational algorithms are needed to derive meaningful insights from the data.
Hence, machine learning algorithms are needed and essential to provide precise
information to the user.

3.9 EXAMPLES FOR MACHINE LEARNING APPLICATIONS

• Speech recognition
• Weather forecasting and prediction
• Traffic prediction and sending alerts
• Filtering spam emails
• Product recommendation in online shopping
• Sentiment analysis
• Auto-driven cars

3.10 OVERVIEW OF DEEP LEARNING

Deep learning is a subset of machine learning. It has an artificial neural net-
work to carry out the tasks of machine learning. It enables the system to pro-
cess the data in a non-linear fashion. Deep learning can be defined as a class
of machine learning algorithms which are capable of extracting more features
from raw input data using multiple layers. It filters the input through many lay-
ers, and it will learn how to classify and predict the data. Deep means the many
number of layers that are used to transform data. To implement deep learning
techniques, many computational nodes will be created. Each node is trained to
analyze the given information and to make decisions like human brains. It is
exactly similar to how the human brain filters any information into deep layers
to understand in depth. The best programming languages for deep learning are
Python, R Programming, and LISP.

3.11 STEPS TO IMPLEMENT DEEP LEARNING ALGORITHMS

1. Provide the input to the system.
2. Classify the input.
3. Extract the features.
4. Generate the output.

3.12 WHEN/WHERE/HOW/WHY TO USE DEEP LEARNING?

Deep Learning is used when the data size is large and there is a lack of
domain understanding for feature extraction. It is widely used to solve com-
plex problems.
AI vs. Machine Learning vs. Deep Learning  77

3.13 EXAMPLES FOR DEEP LEARNING APPLICATIONS

• Automatic translation of text
• Instant visual translation
• Classifying objects in photos
• Identifying and detecting objects in images
• Automatic generation of captions for images
• Game play automation

3.14 COMPARISONS OF ARTIFICIAL INTELLIGENCE,

DEEP LEARNING, AND MACHINE LEARNING
Though artificial intelligence, machine learning, and deep learning are the subsets of
one another, there are few differences among them based on certain features. Tables
3.1– show the comparison and differences between them.
Table 3.1 shows the comparisons based on classification.
Table 3.2 shows the comparison between AI, ML, and DL based on characteristics.
Table 3.3 shows the differences between AI, ML, and DL based on performance
measures.

TABLE 3.1
AI vs. ML vs. DL: Based on Classification
S.
No. Artificial Intelligence Machine Learning Deep Learning
1 Meaning: Meaning: Meaning:
It makes the software It is a subset of artificial It is also called deep
think intelligently. It is intelligence. It relies on structured learning or
done by studying how patterns. The system uses hierarchical learning. It is
human brains think, statistical models to perform a based on artificial neural
learn, and decide to solve specific task without any networks. It is widely used
a problem. AI systems explicit instructions. It is for feature detection. It can
behave like humans widely used to make be referred as the procedure
without any fatigue, predictions and decisions to implement machine
emotion, and limitations learning
2 Types: Types: Types:
• Weak AI • Supervised learning • Unsupervised
• Strong AI • Unsupervised learning pre-trained networks
• Super-intelligence AI • Reinforcement • Convolutional neural
• Learning networks
• Recurrent neural
networks
• Recursive neural
networks
78 Artificial Intelligence (AI)

TABLE 3.2
AI vs. ML vs. DL: Based on Characteristics
S.
No. Artificial Intelligence Machine Learning Deep Learning
1 Key components: Key components: Key components:
The AI algorithms mainly The algorithms depend on The deep learning networks
depend on heuristics and structured data. Based on rely on the layers of artificial
the study of human the structured data, the neural networks
intelligence algorithms classify and
predict the required
information
2 Benefits: Benefits: Benefits:
• Error reduction is easy • Enormous volumes of • Possible to extract more
• Easy-to-handle data can be processed features in an incremental
repetitive tasks • Easy to identify manner
• Availability at all time patterns in data • Domain expertise is not
• Applicable to all fields • No human intervention needed
• Automation is easier • Easy to handle multiple • As the scalability of data
• Permanent memory is varieties of data increases, the performance
available will become better

TABLE 3.3
AI vs. ML vs. DL: Based on Performance Measures
S.
No. Artificial Intelligence Machine Learning Deep Learning
1 Performance: Performance: Performance:
The performance of AI algorithms is In machine learning, The performance
measured by comparing the observed performance measures measures are used to
values and inferred values in the data are used to evaluate the evaluate the deep
input. The intelligent agents are the learning algorithms. learning methods and
autonomous entities which direct their The performance models.
activities to achieve the goal. The metrics in machine The performance of
performance measure in AI is done learning are as follows: deep learning
through the intelligent agents 1. Classification algorithms can be
In AI, the different types of intelligent accuracy measured through the
agents are as follows: 2. Logarithmic loss following parameters:
1. Simple reflex agent 3. Confusion matrix 1. Programmability
2. Model-based reflex agent 4. Area under a curve 2. Latency
3. Goal-based agent 5. Mean absolute 3. Accuracy
4. Utility-based agent error 4. Size of the model
5. Learning agent 6. Mean squared error 5. Throughput
The performance measure is the 6. Efficiency
criterion that measures the success of 7. Rate of learning
an agent
AI vs. Machine Learning vs. Deep Learning  79

TABLE 3.4
AI vs. ML vs. DL: Based on Its Workflow
S.
No. Artificial Intelligence Machine Learning Deep Learning
1 Workflow: Workflow: Workflow:
1. Transform the real • Provide input data • Provide the training data
intelligence into the • Analyze the input • Identify the neural network
system • Identify the model
2. Process and train the patterns • Configure the model with the
system • Make the future learning process
3. Deploy and execute prediction • Train the model
the system • Generate feedback • Interpret the results

TABLE 3.5
AI vs. ML vs. DL: Based on Objectives and Limitations
1 Objective: Objective: Objective:
The main objective is to simulate The main objective is to Deep learning has been
human intelligence in a system and design the system for introduced as a new area of
to make the system to work smart learning by itself and machine learning to bring
and to solve complex problems for making predictions machine learning algorithms
closer to artificial intelligence
2 Limitations: Limitations: Limitations:
• Issues and challenges in • Flaws in input data • Need for a very large
integration may lead to amount of data
• Implementation is erroneous output • Training the system is
time-consuming • Bad input may complex and
• Challenges in interoperability bring down the time-consuming
with cross-platforms reliability of the • Requires more graphical
• Difficult to interpret the results output processing units and
machines

Table 3.4 shows the comparisons between AI, ML, and DL based on workflow.
Table 3.5 shows the objectives and limitations of AI, ML, and DL.
Table 3.6 shows the futuristic issues in AI, ML, and DL.
Table 3.7 shows the differences between AI, ML, and DL based on future scope
and tools.

3.15 SUMMARY
This chapter has provided an overview of artificial intelligence, machine learn-
ing, and deep learning and has provided solutions for when, where, how, and why
80 Artificial Intelligence (AI)

TABLE 3.6
AI vs. ML vs. DL: Based on Futuristic Issues
1 Futuristic issues: Futuristic issues: Futuristic issues:
• When it gives imperfect • Ethical issues • Availability of very
solutions, it leads to a huge loss • End-user satisfaction large data set for
in business environment • Variation in norms training
• Imperfect solutions lead to chaos and rules in different • Time consumption in
and will have a great social impact regions and countries deep learning networks
• More human intervention and • False correlation • Data overfitting occurs,
computation is needed to produce leads to when the number of
a perfect AI solution misinterpretation parameters exceeds the
• Availability of skilled resources • Time-consuming to number of observations
for developing solution with AI build correct • Information privacy
• End-user training cost and time structured data • Latency occurs when
• Data protection • Reusability and retraining is needed
• Compatibility with devices and integration for new data and
systems information
• Upgrading the algorithm • Leads to unstable
periodically and its implementation conditions when there
in AI-based system is needed to is variation in input data
produce reliable results

TABLE 3.7
AI vs. ML vs. DL: Based on Future Scope, Tools, and Areas of Application
1 Areas of application: Areas of application: Areas of application:
Examples: Examples: Examples:
• Medical diagnosis • Social media services • Speech recognition
• Aviation • Email spam and malware • Image recognition
• Robots filtering • Natural language processing
• Workspace communication • Virtual personal assistant • Self-driving cars
2 Future scope: Future scope: Future scope:
• It may spread across all • It will be widely used in • Wider scope for
spans of daily life digital marketing development of more deep
• It will be applied more for • It has more scope in the learning tools and standards
language translations field of education • Development of more
• It will be applied in all • It will have a greater simplified programming
branches of engineering impact on social media, frameworks
and for automation search engines, and • Deployment of transfer
• Development of expert predictions learning concept through
systems reusable components
3 Sample tools needed for Sample tools needed for Sample tools needed for
development: development: development:
• TensorFlow • Weka • Caffe
• Keras • PyTorch • Torch
• PyTorch • TensorFlow • DeepLearning4J
• Theano • KNIME • Cuda
AI vs. Machine Learning vs. Deep Learning  81

these techniques can be used. Artificial intelligence vs. machine learning vs. deep
learning is compared based on their meaning, types, examples, benefits, limita-
tions, futuristic issues, tools, future scope, areas of application, etc. Basic features
of artificial intelligence, machine learning, and deep learning are summarized in
this chapter.
 4 Ethical Reasoning and
AI and Big Data

Responsibility
Sweta Saraff

CONTENTS
4.1 Introduction..................................................................................................... 83
4.2 Ethics Reasoning in Artificial Intelligence.....................................................84
4.3 Ethical Responsibility in AI ........................................................................... 88
References.................................................................................................................90

4.1 INTRODUCTION
Do we have the capabilities to understand the infinite potentialities of Nature? This
question was contemplated years ago by Sri Aurobindo (1993).

“[O]ur science itself is a construction, a mass, of formulas and devices; masterful in

the knowledge of processes and in the creation of apt machinery, but ignorant of the
foundations of the being and of World-being, it cannot perfect our nature and therefore
cannot perfect our life.”

He further doubted the capabilities of science in reaching an ultimate truth or as a

matter of fact, any ultimate truth:

“One might ask whether science itself has arrived at any ultimate truth; on the con-
trary, ultimate truth even on the physical plane seems to recede as science advances.
Science started on the assumption that the ultimate truth must be physical and objec-
tive – and the objective ultimate (or even less than that) would explain all subjective
phenomena.” Sri Aurobindo (1972)

The overwhelming speed at which the age of objectivity is celebrated by science –

technology – data and artificial intelligence (AI) has elevated humanity at a material
plane. This current state of mindset to achieve and compete relentlessly for success
and the pleasures of life is trying to replace harmony with desire. Today everyone
is in an unnerving race of achieving material success in life. What is real success?
What we have achieved and we have lost? Can we still recover our losses? Is the
damage repairable?

83
84 Artificial Intelligence (AI)

With the advancement of technology, there is an unease regarding the increasing

rate at which the human race is snapping its ties with Nature, humanity, and their
conscience. In the path of seeking objective pleasure, there is decline in subjective
well-being. An attitude of liberated technology from the grasp of ethics is an insig-
nia of regressive society. The presumption of progress is rested on hollow pillars of
momentary achievement with no consideration of what is right or wrong unless it is
relatable to self.
We all measure our success or happiness in terms of material possessions which
are machines, clothes, hybrid food, luxury housing, and more dependability on oth-
ers. At the same time, we are forgoing pleasure from staying close to Nature, family,
and human values. This raises a question on measurement which was discussed by
David Bohm, the renowned physicist. He succinctly differentiates between “measur-
able” and “immeasurable”; or it may be understood as a difference between finite
and infinite. Bohm has postulated that the ultimate reality of Nature lies not in its
fragmentation but in its unification. Bohm (1993) opines that this immeasurable
universe encompasses all its parts and unites them in an irrefutable totality. The
fragmented objectivity is far from “ethical and moral basis of human peace and hap-
piness” (Chakraborty, 1998).
Professor S.K. Chakraborty expresses his concern that objective pathways may
lead to the compromise of “ethical-moral sentiments” which unites mankind with
each other and with our surroundings. Earlier philosophers like Gandhi and Tagore
were proponents of self-sustained living, which was simpler, closer to Nature, simi-
lar feelings for self and others, and an unwavering sense of values.

4.2 ETHICS REASONING IN ARTIFICIAL INTELLIGENCE

Understanding human cognition and reasoning involved in taking any decision
has always drawn attention. This field has been evolving with contributions from
different disciplines like social sciences, humanities, neurosciences, mathematics,
and artificial intelligence. Kay (2002) believed that it is imperative to decode the
inception and root cause of human intuition to comprehend the complexities of the
decision-making process. Cognition, action, and emotion are essential components
of decision-making phenomena. Belief, thought about an event or another person,
influences the behaviour and attitude, and these are compounded by psychological
affect and physiological reaction within a person.
People most often analyse the opportunity cost in taking any decision. The
decision-making process also incorporates one’s assumption about others’ expected
behaviour in a context, as per their past experiences. It operates on a prototypical
schema. There can be multiple outcomes and a chain of reactions following them.
Then arises the dilemma of selecting the best possible option or making a choice
and having confidence in the decision or analysis. A rational decision-maker always
analyses several possibilities from different perspectives to make a specific choice.
Some theorists proposed that humans used specific algorithms or normative reason-
ing (Goodwin & Wright, 1998), though Hoch et al. (2001) were of the opinion that
there may be no specific rules followed by humans and they generally use heuristics
AI and Big Data 85

to solve the problem if they are unique or novel. Rubinstein(1998) opined that people
choose the best alternative by integrating desirability with feasibility.
Economists believe that people try to make decisions based on their satisfaction
of wants or pursue a maximum satisfaction model. This model analyses the value
of choice for its efficacy in a particular situation and also in accordance with their
personal preferences, past experiences, etc. According to Oliveira (2007), one can
analyze the “Expected Utility Theory” either analytically or synthetically. When one
adopts the analytical view, they will observe the available choices and then decide
their outcomes, whereas in a synthetic method, they will ascertain the utility first
and then find out means to achieve it.
Based on these contemplations, the following pertinent questions arise – Is ratio-
nality overpowering ethics? Are we reasoning ethically? What strategies are we uti-
lizing in differentiating right from wrong? Does science have the capacity to reason
ethically? Can machines or robots reason like humans? Ethics, morality, and values
are both subjective and culture based. It does not follow the “one size fits all” rule.
Do these advanced algorithms follow policies of fairness above all? Are we careful
about the specificity and sensitivity measures taken by the predictive tools? It evokes
different opinions from utilitarian (consequential) as well as deontological (rule)
school of ethics. People are held accountable for their doings, and similar responsi-
bility needs to be fixed for decisions taken by automated machines. The engineers
and the software experts need to take the ethical responsibility for the outcomes.
It is quite expected that automated vehicles (AV) such as self-driven (SD) cars
would be facing such a dilemma in real-life situations where they must be pro-
grammed to take a quick decision. Which ethical rule will they follow? Would they
harm an innocent pedestrian to save the car speeding from the other side? To come
to an acceptable answer is quite difficult for AI engineers and scientists. We can also
assume that if we present this question to people from different geographical loca-
tions, ethnicity, culture, education, work types and levels, value orientation, age, and
gender, there will be disparities in responses.
Traditionalists may differ from pluralists in their outlook about ethical beliefs.
Their reasoning style is more consequential and dependent upon the probable out-
come. Some may consider saving more people over one due to higher order reasoning
and future orientation, whereas people with emotional attitude may be overwhelmed
by the thought of killing an innocent man. This dilemma still remains unanswered.
First we must understand that the philosophy of ethics is based on inductive reason-
ing. Here, the order and straightforward rules of logic do not apply. There is hardly
an agreement between nations regarding following the same laws and moral code of
conduct for delivering justice. What ethical principles or design will then be imbibed
by an automatic vehicle? Would one nation allow such self-driven cars which do not
fully adhere to their legal systems?
The requirement of ethical reasoning is not limited to self-driven cars. Artificial
intelligence today can predict stock prices, regulate capital markets, admissions, and
tax evasions, and prepare data on a range of requirements like weather forecasts,
cyclones, pollution levels, and epidemics. They are useful in maintaining applica-
tions in educational institutions, jobs, etc. AI has supported advancements and ease
86 Artificial Intelligence (AI)

of living at most levels to the extent of pushing us to dependency and thinking like
machines without emotions. Based on social media, AI can predict an individual’s
preferences, expected behaviour, or even reactions, intentions, beliefs, friends, etc.
Are these predictions on expected human opinion or behaviour reliable? Humans
by nature are dynamic, subjective thinkers who act according to situations. They
present different behaviours based on mood, situations, current wants, and expected
utility. The argument remains – Can a machine accurately predict the behavioural
patterns and emotional sensitivity of humans?
Another ethical challenge is taking informed consent before collection of data or
any information. Data collected from naïve internet users without their knowledge or
prior permission creates an environment of distrust and violation of personal space.
The predictive technology used by Google and Facebook provides us with segre-
gated information based on our search history, creating suggestiveness, which may
be misleading. They sell this information to different companies which are dealing
with products ranging from apparels, processed food, and travel agencies, making
people vulnerable to different frauds.
They can predict locations travelled, home, school, and office addresses, shopping
behaviour, suggestive friends list, chances of your partner’s infidelity, investments,
choice of candidates in an election, etc. All personal data are accessible easily over
the internet. Just like a detective, AI can draw up a conclusion about the future
behaviour, interests, or probable tendencies of an individual. This tool combines
various factors to produce a single score for a person, and this prediction is used
as a guide to take a future course of action. Does it follow the nuances of ethical
reasoning? Predictions can be a brilliant tool, but are they adaptive? Does this intel-
ligence conform to the principles of ethical sensitivity and awareness? Does it reason
ethically?
Descriptive ethics, normative ethics, and applied ethics are recognized as major
domains of ethical philosophy. Descriptive ethics take an empirical view of the
beliefs of the people. This domain defines the principles behind the moral reason-
ing for differentiating right actions from wrong actions. The theorists working on
descriptive ethics look into the values in which people repose their faith and the basic
features of virtuous conduct. The importance of descriptive ethics lies in the fact
that they are different from normative ethics. It believes that the ideals and values
of various cultures are dynamic in nature. They are always evolving with develop-
ment of societies and mankind. The theorists opine that ethical requirements of the
current generation may be different from their predecessors. So descriptive theorists
try to systematically inquire by collecting information and observations from dif-
ferent fields like basic sciences, applied sciences, and social sciences. Colby and
Kohlberg’s (2011) work on cognitive moral development is an example of descriptive
ethics. This field of ethics deals largely with describing and predicting human behav-
iour (Donaldson & Dunfee, 1994). A framework on ethics may include three steps
or major components (Figure 4.1): identifying the sensitivity or awareness, making a
moral reasoning, and engaging in moral conduct.
Normative ethics studies how people should act or what course of action can be
considered ethical (Hoffe, 1989). Normative ethical theories can be observed from
AI and Big Data 87

FIGURE 4.1  Components of ethical reasoning.

three perspectives: virtue ethics, deontology, and utilitarianism. Virtue ethics lays
emphasis on individual character, which includes how they are acquired, nurtured,
and applied in real-life scenarios. It is based on a set of stable dispositions where
decisions are not taken based on suitability in a particular context. Traits like hon-
esty, truthfulness, and care are attributes of the personality and not just part of a
habit. It includes choices, interests, values, attitudes, courage, and temperaments.
Virtue ethics differs from deontology as it focuses on the inherent disposition of an
individual rather than his or her adherence to an established set of rules. A virtuous
action is representative of an inner sense of morality and “a way of being” leading to
consistency in action and sensibilities.
Deontology ethics are well-defined or rule-based ethics (Waller, 2005). The main
tenet of the theory lies in engaging in a behaviour when the action is morally justifi-
able rather than considering its consequences. Deontology is different from conse-
quentialism as it believes in moral action rather than its effects (Flew, 1979). It deals
with personal and professional duties and legal rights. Such ethics may contrast with
personal or cultural values, but people have a moral obligation to follow a certain
code of rules. The premise of utilitarian philosophy of ethics is grounded in select-
ing an action based on its perceived consequences. The main thought is to maximize
88 Artificial Intelligence (AI)

utility in a conduct for general or optimal well-being of the humankind. In institu-

tions or people engaging in utilitarian or consequential ethical practices, the main
hindrance lies in limited perception or ability in understanding long-term conse-
quences or effects on a new theory, practice, or invention.
Even with the best motive, a person of high integrity and good will may still make
mistakes in justifying an action due to limitations of human perceptual ability and
foresightedness. Current advancements in AI have created a deep interest among
computer users, data scientists, psychologists, corporates, and the general public.
The idea of gaming has transitioned from single user to multi-users with chat-bots,
avatars, levels, badges, trophies, etc., creating a stimulating and challenging environ-
ment for players. This has shifted the focus of research in understanding the ethical-
ity of implicit and explicit stimulations used in such technologies. Do they improve
learning and development or they are creating blocks in the moral and cognitive
development of millennials?
The most important issue related to sustainability in ethical reasoning lies in
machine medical ethics: training ethics to machines as to how they should behave
adhering to values and rules in different situations with pragmatism and care.
Attempting to create algorithms to train an ethical AI is a very complex task, but still
efforts are being made in this respect (Churchland, 2011; Wallach & Allen, 2010). A
comprehensive and consistent moral theory which can balance social, cultural, and
demographic differences and guide the actions of an ethical, logical, and intelligent
robot is an urgent requirement in keeping pace with the innovations.
It must be programmed succinctly to deal with real-life situations. Since there are
no available solutions to manifold ethical dilemmas and mutually agreed-upon rules
and standards to follow, development of an acceptable theory of ethical machine
looks a herculean task.
Tatjana Kochetkova (2015) suggests that use of AI in health and medicine should
be limited to instances, where clear guidelines for ethical conduct already exist. A
hybrid approach (mixed top-down–bottom-up) is advised where advantages of both
top-down (use of moral principles in selection and control of action) and bottom-up
(supporting with theoretical background and experimental learning) can be utilized
(Allen et al., 2005). An AI system must be capable of understanding the legal, moral,
and cultural impact of their decisions.

4.3 ETHICAL RESPONSIBILITY IN AI
With power comes great responsibility, and systems dependent on machine learn-
ing and artificial intelligence are no exception (Dignum, 2017). According to
Dignum (2018),

Whatever their level of autonomy and social awareness and their ability to learn, AI
systems are artefacts, constructed by people to fulfil some goals. Theories, methods,
algorithms are needed to integrate societal, legal and moral values into technologi-
cal developments in AI, at all stages of development (analysis, design, construction,
deployment and evaluation).
AI and Big Data 89

Both ethical reasoning and responsibility form the edifice on which future moral robots
must be built. The primary goal is to train robots to behave according to the programmed
code of conduct. However, can they explain the reasons behind their conduct? So, this
responsibility of training and mentoring ethical robots lies on the software programmer.
The next question is whether the programmer is equipped with moral reasoning? Does
he understand the nuances of virtue, justice, and legality? How to solve the ambiguities
of an ethical dilemma and make the most fair and honest choice?
A framework of ethical responsibility envisaged on the teachings of Vedanta is
relevant for most of the cultures, societies, and overall development of humanity.

1. Delivering one’s duty without any doubt and lethargy is expected from all.
The world we live in will be better if each individual performs even the
smallest task assigned to them. There is no use talking about reforming the
world without caring for the pressing problems of the moment. The imme-
diate crisis, the task that lies ready to hand, must engage the individual’s
attention, and if he does it well, he will indirectly help in the betterment of
the world. Every man has a certain place to occupy and a certain function
to fulfil in the social economy.
2. Belief in respecting dignity of all mankind above self-interest and personal
desires is of paramount importance. Justice and benevolence are the pri-
mary attributes required to fulfil the ideals of social service. The laws made
for the benefit of society, if given more importance than the existence of
humanity, will lead to extremism by creating conflicts between duties. The
laws serve as the guiding path to deliver justice without any biasness. But
man has no call to act in blind obedience to them. The need is not confor-
mity to rules, but conformity to the law of reason. When the rules come
into conflict, we must fall back on the supreme commandment, and ask
ourselves which course is most conducive to the realization of reason in the
world. We must serve as moral rational beings, with acts expressive of the
pivotal purpose.
3. Radhakrishnan (1914) said: “The inner spirit is more important than out-
ward conformity to law. An action is good, not because of its external con-
sequences, but on account of its inner will. Virtue is a mode of being and
not of doing. It is not something to be found, but a function or an exercise of
the will.”
4. Reasoning credits us with accountability towards ideals of justice, human-
ity, and righteousness which must be delivered without any prejudices. If
each individual is given the power to play according to their conviction, the
universal goal of social balance and harmony would be destroyed. Logical
judgement with moral intent must not be sacrificed to fulfil the norms of
social convention. Myopic traditions and values are regressive and must not
be favoured at the cost of knowledge and growth. We must build institutions
which deliver harmony and progress with equal veracity.
5. No act should be done with a selfish interest or for the gratification of incli-
nations. An individual, if engaged in moral reasoning, serves humanity
90 Artificial Intelligence (AI)

with sensibilities and empathy. The highest ideal in life must be to serve
society by exercising logical judgement even in the moment of crisis.
6. Character, which is the habit of will, is not determined but self-created.
The rational self, if consciously agree to the immoral attributes or accept
them as its stable disposition, then such characteristics are selfish in nature.
Human beings cannot justify a mistake by blaming the genetic lineage. It is
his wilful desire to accept or reject. He must control the acquired behaviour
and reason with honesty and integrity.
7. Visionary idealism which lacks focus and intelligent direction is ineffec-
tive. If private interests are permitted, then there are chances of disrup-
tion in the society. A man cannot stop and debate within himself at every
point as to what his duty is. He must fulfil his duty with enthusiasm and
eagerness.
8. “Know the self to be sitting in the chariot, the body to be the chariot, the
intellect (buddhi) the charioteer, and the mind the reins” – Radhakrishnan
(1914). One must have control over senses for a steady and sharp intellect.
9. Knowledge should not be acquired for the purpose of completing a course
but must enable us to take a right perspective and a pragmatic frame of
reference about other things and our place in the world.
10. We must do the right whether the right is done or not. “Devoted each in his
own work, man attains perfection” Radhakrishnan (1914). Every man is
required to contribute to the national strength his quota of earnest work. It
is by the endorsement of this philosophy of work that the world as a whole
will flourish.
Bonnemains, Saurel, and Tessier (2018) in their paper on “Embedded
ethics: some technical and ethical challenges” propose development of
“formal tools” that are descriptive of different circumstances and “models”
of ethical philosophies which have the potential to reason not only auto-
matically but also ethically, giving proper justification. Today’s millennials
nurture the myth of a utopian world maybe, but such a world looks imper-
fect, even if it is the mightiest. An imperfect world, which is harmonious
and hopeful of future growth, is more blissful and happier.

REFERENCES
Allen, C, Smit, I, Wallach, W. Artificial morality: Top-down, bottom-up, and hybrid
approaches. Ethics and Information Technology, 7 (2005): 149–155.
Aurobindo Sri. The Life Divine. First Edn. Pondicherry: Sri Aurobindo Ashram; 1993. The
Divine Life; p. 1034. 1939-40, Tenth impression 1993.
Aurobindo Sri, Mother. On Science. Pondicherry: Sri Aurobindo Ashram; 1972. Drawbacks
and Limitations; p. 12.
Bohm, D. Science, spirituality, and the present world crisis. ReVision, 15(4) (1993): 147–152.
Bonnemains, V, Saurel, C, Tessier, C. Embedded ethics: Some technical and ethical chal-
lenges. Ethics and Information Technology, 20(1) (2018): 41–58.
Chakraborty, S.K. Values and Ethics for Organizations: Theory and Practice. New Delhi:
Oxford University Press, 1998, 153–171.
AI and Big Data 91

Churchland, P. Braintrust: What Neuroscience Tells Us about Morality? Princeton: Princeton

University Press, 2011, pp. 23–26.
Colby, A, Kohlberg, L. The Measurement of Moral Judgment, Vol. 1. Cambridge: Cambridge
University Press, 2011.
Dignum, V. Responsible autonomy. In Proceedings of the Twenty-Sixth International Joint
Conference on Artificial Intelligence (IJCAI’2017), 2017, pp. 4698–4704.
Dignum, Virginia. Ethics in artificial intelligence: Introduction to the special issue. Ethics
and Information Technology, 20(1) (2018): 1–3.
Donaldson, T, Dunfee, TW. Toward a unified conception of business ethics: Integrative social
contracts theory. Academy of Management Review, 19(2) (1994): 252–284.
Flew, A. Consequentialism. In A Dictionary of Philosophy, 2nd Ed. New York: St Martins,
1979, p. 73.
Goodwin, P, Wright, G. Decision Analysis for Management Judgment. Chichester: John
Wiley & Sons Ltd, 1998.
Hoch, SJ, Kunreuther, HC, Gunther, RE. Wharton on Making Decisions. New York: John
Wiley & Sons, Inc, 2001.
Höffe, O. Kant’s principle of justice as categorical imperative of law. In Kant’s Practical
Philosophy Reconsidered. Dordrecht: Springer, 1989, pp. 149–167.
Kay, J. Beware the pitfalls of over-reliance on rationality: Attempting to shoehorn complex
decisions into the framework of classical theory can be a mistake. The Financial Times,
2002, p. 9.
Kochetkova, T. An overview of machine medical ethics. In Machine Medical Ethics. Cham:
Springer, 2015, pp. 3–15.
Oliveira, A. A discussion of rational and psychological decision-making theories and mod-
els: The search for a cultural-ethical decision-making model. Electronic Journal of
Business Ethics and Organization Studies, 12(2) (2007): 12–13.
Radhakrishnan, S. The ethics of the Vedanta. The International Journal of Ethics, 24(2)
(1914): 168–183.
Rubinstein, A. Modeling Bounded Rationality. Cambridge, MA: MIT Press, (1998).
Wallach W, Allen C. Moral Machines: Teaching Robots Right from Wrong. Cambridge: The
MIT Press, 2010.
Waller, BN. Consider Ethics: Theory, Readings, and Contemporary Issues. New York:
Pearson Longman, 2005, p. 23.
 5 Online Liquid Level
Estimation in Dynamic
Environments Using
Artificial Neural Network
Thulasi M. Santhi and S. Sathiya

CONTENTS
5.1 Introduction..................................................................................................... 93
5.2 Liquid Level Measurement in Dynamic Environments..................................94
5.2.1 Influence of Temperature.....................................................................94
5.2.2 Influence of Inclination........................................................................ 95
5.2.3 Influence of Sloshes.............................................................................96
5.3 Sensor Design..................................................................................................97
5.3.1 Fibre Bragg Grating Sensor.................................................................97
5.3.2 Cantilever Beam.................................................................................. 98
5.3.3 Float Sensor.........................................................................................99
5.3.4 System and Working Principle.......................................................... 100
5.4 Introducing Neural Networks for Accurate Level Prediction....................... 101
5.4.1 Sampling of Sensor Output................................................................ 101
5.4.2 Artificial Neural Networks................................................................ 102
5.4.3 Activation Function........................................................................... 103
5.5 Wavelet Neural Network................................................................................ 103
5.5.1 Training of WNN.............................................................................. 104
5.6 Results............................................................................................................ 107
5.7 Conclusion..................................................................................................... 108
References............................................................................................................... 109

5.1 INTRODUCTION
A liquid level is an essential parameter to measure in almost every process system such
as the food and beverage industry, petrochemical plants, water reservoirs, and auto-
motive systems. Level measurements based on the principle of differential pressure,
vibrating wire, capacitive magnetic floating, and ultrasonic are conventionally used in
the industries. In automotive, mainly variable resistive, capacitive, float, and ultrasonic
sensors are used for measuring the level of the fuel tank. Resistive float-type sensors

93
94 Artificial Intelligence (AI)

are widely used as automotive fuel indicators more than other conventional sensors.
Many conventional level measurement techniques are accurate only when the liquid
contained in the fuel tank is stationary. With the presence of dynamic environments,
such as variation in temperature, an inclination of the fuel tank, and liquid slosh, most
of the methods fail to produce accurate level measurements which are very essential
[1]. The conventional sensors are incapable of compensating the effects of temperature
variations, inclination, or slosh while measuring the liquid level. The level measure-
ment in such environments requires the best effective system with high sensitivity,
which will result in highly accurate level measurement.
The proposed method of liquid measurement integrates the neural network
approach along with the measurement sensor for improving the system accuracy. As
the neural network is a computing system that predicts the output accurately by con-
sidering the past input data, the ability of a neural network is employed to improve
the accuracy of the proposed system in the dynamic environments [1, 2]. Another
level measurement technique with optical fibre has been tremendously increasing for
the past few years due to its advantages such as being unaffected from electromag-
netic interference, having low signal loss and distortion, and being lightweight. Thus,
designing a system with fibre Bragg grating (FBG), a type of optical fibre, will fur-
ther enhance the performance of the measurement system. The proposed level mea-
surement sensor based on FBG-embedded cantilever beam and a float combination
[3], integrated with the machine learning approach, specifically wavelet neural net-
work, contributes a highly accurate and sensitive system apt for level measurements
in dynamic environments [4]. The proposed design of the liquid level measurement
system is specifically formulated for the automobile fuel tank which is affected by
the dynamic environment when the vehicle undergoes different acceleration.

5.2 LIQUID LEVEL MEASUREMENT IN DYNAMIC ENVIRONMENTS

The liquid level measurements in static environments are easy to detect and accurate,
while it becomes very much erroneous in dynamic environments. The level measure-
ment of the fuel tank present in a moving automotive system is a common example of a
dynamic system. For the effective usage of fuel in an automobile, the measurement of
fuel quantity is essential which will help the drivers to determine the total time a vehi-
cle can drive without refuelling. The main disturbances which affect the measurement
accuracy of the fuel level during the movement of vehicles are temperature variations,
the inclination of the tank, and sloshes. All these disturbances and the variations must
be compensated to determine the exact level of the fuel left in the automobile fuel tank.

5.2.1 Influence of Temperature
The temperature has a great impact on the variation of liquid properties, notably den-
sity and viscosity. The movement of molecules in the liquid varies according to the
temperature variations. The volume of the liquid in the fuel tank seems to be increased
or decreased as a result of low or high densities, respectively (density is the ratio of mass
and volume). Viscosity is based on the cohesive forces between the molecules of the
Parameter Estimation in Dynamic Environments 95

TABLE 5.1
Dynamic Viscosity, Kinematic Viscosity, and Density Variations Due to
Temperature Change of Engine Oil SAE 15W-40 [5]
Temperature Dynamic Viscosity Kinematic Viscosity Density
S. No. (°C) (mPa·s) (mm2/s) (g/cm3)
1 0 1328.0 1489.4 0.8916
2 10 582.95 658.60 0.8851
3 20 287.23 326.87 0.8787
4 30 155.31 178.01 0.8725
5 40 91.057 105.10 0.8663
6 50 57.172 66.464 0.8602
7 60 38.071 44.585 0.8539
8 70 26.576 31.350 0.8477
9 80 19.358 23.006 0.8414
10 90 14.588 17.467 0.8352
11 100 11.316 13.648 0.8291

liquid. As temperature increases, the cohesive force between the molecules decreases
and results in a reduction of viscosity; in the same way, a decrease in temperature
causes an increase in cohesive force between molecules and viscosity increases. In
short, both density and viscosity have an inverse relation to temperature.
As the fuel level measurement is considered, the temperature dependency of the
engine oil SAE 15W-40 is given in Table 5.1 [5]. The variation in dynamic viscos-
ity, kinematic viscosity, and density over temperature from 0 to 100°C is mentioned.
It is clear from the table that these three parameters are decreasing concerning the
increase in temperature of the oil. In Figures 5.1 and 5.2, the three parameters are
plotted to analyze the effect of temperature on them.
The viscosity values have a remarkably high dependency on temperature as it
reduces rapidly to very low values up to 50°C, and for further increase in tempera-
ture, the rate of decrease in viscosity is found to be lesser. The density decreases
linearly as the temperature of the fuel increases. From the graphs, it is clear that
the consideration of temperature variation of liquid in the level measurement sys-
tem is significant. Due to the presence of environmental temperatures, the defects
in the coolant circulation system, and excessive loads, the engine temperature will
increase which results in the variation of liquid density and viscosity. As the proper-
ties of liquid vary because of the temperature, the liquid volume in the tank will also
vary accordingly. Thus, the fuel level measurement will give errors in level readings
unless the disturbances are compensated properly.

5.2.2 Influence of Inclination
The automobile is movable systems that pass through different geographical varia-
tions such as elevations, slopes, and irregularities in the road. In such cases, the
96 Artificial Intelligence (AI)

FIGURE 5.1  Dynamic and kinematic viscosity variations of engine oil SAE 15W-40 over
temperature.

FIGURE 5.2  Density variations of engine oil SAE 15W-40 over temperature.

baseline of the sensor will be in inclination with the normal horizontal line and the
liquid level will also get displaced in such a way that one side has the highest-level
reading while the other has the lowest. If in case the position of the sensor is at the
centre of the fuel tank, the level reading is at 45° inclination with the horizontal line
and 15° inclination with the horizontal line, thus considering the inclination is sig-
nificant in the level measurement system of a fuel tank.

5.2.3 Influence of Sloshes
Another main disturbance to be considered is the slosh of the liquid produced
inside the automobile fuel tank during the movement of the vehicle with variable
accelerations [6]. It is a highly unpredicted and undesired disturbance which seri-
ously affects the accuracy of the liquid level measurement system. The sloshes of
Parameter Estimation in Dynamic Environments 97

the liquid vary the liquid level dynamically with respect to the different accelera-
tion of the vehicle, which is difficult to compensate with the conventional liquid
measurement system. Therefore, the size and shape of the fuel tank, acceleration
of the vehicle, and inclination of the tank position are the parameters that vary the
slosh intensity.

5.3 SENSOR DESIGN
5.3.1  F ibre Bragg Grating Sensor
An optical fibre can be transformed into an FBG by exposing the core to intensive
laser light for introducing a periodic modulation in the refractive index. The broad-
band light signal passing through the FBG fibre gets refracted and reflected due
to the gratings. Because of the reflection and transmission of different wavelength
signals, interference occurs. The Bragg condition as given in Equation 5.1 must be
satisfied by the wavelength for the reflection of the optical signal:

lB = 2nL (5.1)

where λB is the Bragg wavelength, n is the refractive index of the optical fibre,
and Λ is the grating period. By changing the parameters n and Λ, the Bragg
wavelength can be varied. The grating length is also one factor that affects the
Bragg wavelength. Figure 5.3 shows the reflected signal from FBG looks like only
a single wavelength is reflecting, so it resembles peak signals. FBGs are passive,
robust, have less size, high sensitivity, and precision, which make the sensor suit-
able in many optical applications such as telecommunication fields and sensor
technologies, where wavelength selection is required [7]. The Bragg wavelength
change is described as

DlB
= C1 + C2 DT (5.2)
lB
where λB represents the Bragg wavelength under the unstrained condition of the FBG,
ΔλB is the Bragg wavelength shift due to the presence of the strain ε and temperature
variation ΔT, C1 = 0.78 × 10 −6 and C2 = 6.67 × 10 −6°C. For longer Bragg wavelengths,
the changes in wavelength occur due to the applied strain and temperature [3, 7, 8].
The axial strain and temperature are the physical parameters that the FBG sen-
sor can measure directly. To measure displacement by the FBG, the displacement
must be converted to the axial strain experienced on FBG by any transducer. The
basic block diagram presenting the working of the FBG sensor for measurement of
displacement is shown in Figure 5.4. The cantilever structure of the sensing element
gets deformed and induces a strain according to the external displacement. This
strain is experienced by the FBG, which converts the displacement into variations in
wavelength, bandwidth, light intensity, and finally to voltage using an optical spec-
trum analyser (OSA) connected to FBG circuitry. Calibrations can be done with this
information concerning the displacement for the measurement.
98 Artificial Intelligence (AI)

FIGURE 5.3  Example of reflection peaks from an FBG [7].

FIGURE 5.4  Schematic diagram of displacement measurement using FBG sensor.

5.3.2 Cantilever Beam
A cantilever beam is a beam that is fastened at one end and left free at the other end;
it means one end is fixed to support and connected, while the other end is left without
any support [9, 10]. When pressure or force is exerted on the free end of the beam,
the beam carries this load to the support, i.e., fixed end, from where the moment of
force and the shear stress can be managed. The tendency of an applied force to turn
or twist the beam is the moment of force, while the stress applied parallel to the
beam is the shear stress.
Due to the support on the cantilever beam’s fixed end, the free end can carry
a specific weight which effects in bending of the beam instead of breaking down
Parameter Estimation in Dynamic Environments 99

due to the shear stress. Without external bracing or support, cantilever construction
allows for overhanging structure, a stainless-steel cantilever beam is used in the
described methodology of design. The cantilever beam deflection for the applied
external force in the free end is illustrated in Figures 5.5 and 5.6.

5.3.3 Float Sensor
Float is a structure that is suspended or freely placed in the liquid surface in such a
way that it will float fully or partially over the liquid surface. To follow the variations
in liquid level, floats and float switches use the buoyancy principle which depends on

FIGURE 5.5  Typical cantilever beam.

FIGURE 5.6  Cantilever beam deflection under load at the fixed end.
100 Artificial Intelligence (AI)

the liquid level and its density. Float structure is made up of plastics or PVC mate-
rial having lesser density than the liquid whose level must be measured so that they
can float over the liquid surface. Most of the floats are hollow spherical or capsule-
shaped filled with air with lesser density than the fluid used in the applications.
For reed switch activation, commonly stainless-steel magnetic floats are used, and
they are tubed magnetic floats with a hollow tubed connection for suspending from
the top. In the measurement of the liquid level, with the consideration of strength,
corrosion resistance, and buoyancy, the magnetic floats are suitable in a wide range
of industries. The floats are made up of welding two symmetrical half-shells together
in an airtight manner and the perfection in the process decides the strength and
durability of the float. The welding should be done uniformly in such a way that it
produces a smoothly finished seam, which is like the surface of the rest of the float.

5.3.4 System and Working Principle

The proposed sensor system design for measuring the fuel level of an automobile
fuel tank is illustrated in Figure 5.7. This set-up is composed of an appropriate fuel-
filled horizontal cylindrical automobile fuel tank, a cantilever beam made up of
stainless steel, two FBGs mounted along with the optical fibre on the opposite sides
of a cantilever beam, a float suspended from the free end of the beam, an opti-
cal source, and an optical spectrum analyser (OSA). The float suspended from the
cantilever beam is submerged in the fuel. For increasing the buoyant force, a par-
tially immersed float having a large cross-sectional area is used, and this arrange-
ment results in the buoyancy force exerted by the liquid making vertical deflection
according to the changes in level. Thus, the FBGs attached to the deflected beam get
strained and the refractive index of the optical fibre core differs further shifting the
wavelength of reflected light [8].
A broadband source is used as the optical source of the proposed FBG-based sys-
tem which emits a wide range of wavelengths to the optical fibre. According to the
refractive index of the optical fibre, the FBG sensors reflect a particular wavelength.

FIGURE 5.7  Sensor architecture.

Parameter Estimation in Dynamic Environments 101

In normal cases, i.e., when FBGs are not strained or no change in the liquid level,
the refractive index does not experience any change. The increase or decrease in the
liquid level forces the cantilever beam to deflect up or down, respectively, resulting
in a change in the refractive index due to strain variations experienced by the optical
fibre. This forms a wavelength shift in the back-reflected light, and the wavelength
shift is detected by an optical spectrum analyser connected with a computer. The
OSA results are stored for further processing, analyzing, and calculations of the
liquid level [11]. This proposed design of the measurement system is used to improve
the sensitivity and accuracy of the liquid level measurement system. In the proposed
setup, the two FBGs are welded to both the surfaces of an elongated cantilever beam,
in a way to provide temperature compensation [12, 13]. The FBG sensors provide
high measurement accuracy and resolution and can measure strain experienced by
the cantilever beam at the point where it is embedded.

5.4 INTRODUCING NEURAL NETWORKS FOR

ACCURATE LEVEL PREDICTION
The accuracy improvement of any fluid level measurement system is mainly depen-
dent on signal processing and signal classification, particularly in dynamic environ-
ments. The signal flow of the designed sensor is shown in Figure 5.8. The acquired
signal from the sensor has to be processed by using suitable signal processing meth-
ods after sampling. Then the signal can be given to the neural network for the clas-
sification purpose.

5.4.1 Sampling of Sensor Output

The output voltage from any liquid level sensor continuously varies with time due to
dynamic variations in the fuel level. The stages of the measurement system are men-
tioned in Figure 5.9. Discretizing the continuous signal at some constant sampling
frequency fs by using sampling circuitry is necessary for digital signal processing

FIGURE 5.8  Signal flow of the designed level sensor system.

102 Artificial Intelligence (AI)

FIGURE 5.9  Overview of measurement system.

[14]. Ts is the sampling interval showing the time between two sample points of the
signal, which is equal to the reciprocal of sampling frequency:

Ts = 1/fs

5.4.2 Artificial Neural Networks

The artificial neural network (ANN) model is a data processing and classifying tech-
nique developed by the inspiration from the way biological neurons process informa-
tion. The neural networks have several neurons, interconnected in a manner to solve
complex problems. The ANN learns by examples similar to human beings. For appli-
cations such as pattern recognition or data classification, the properly configured ANN
using a suitable learning process can be used. The ANN learns by adjustments to the
weights of interconnections that exist between the neurons which is similar to the bio-
logical systems [15]. ANN has a great ability to find relations from complex or impre-
cise data. Too complex relationships and patterns can be extracted, and it can detect
the trend of the complex system by using neural networks. A trained neural network
categorizes the information given to it for analyzing as an expert [16]. By introducing
enough number of hidden neurons to the network, ANN can be trained effectively to
form any multivariable function with very high precision level [17].
A typical neural network configuration is given in Figure 5.10. The neural networks
are trained in such a way to obtain the specific target output from particular inputs. The

FIGURE 5.10  The typical configuration of an ANN.

Parameter Estimation in Dynamic Environments 103

network weights are adjusted and updated by comparing the output with the target to pro-
duce errors, till the error minimizes or network output becomes equivalent to the target.
For high accuracy, more input and target pairs are required for training the network [18].

5.4.3 Activation Function
The activation function used by the neuron decides the output produced by the neu-
ral network. To deliver the output signal, the activation function considers both the
inputs and the adjusted values of the weights [19]. This transfer function may come
under one of the categories mentioned below:

• Linear (ramp): The output signal is proportional to all inputs and corre-
sponding weights of the neuron.
• Threshold: The output will always be any one of the mentioned levels, com-
paring whether the total input given to a neuron is greater or less than the
specified threshold value.
• Sigmoid: The output has no linear relationship between the inputs but var-
ies continuously. This function resembles real biological neurons compared
to linear and threshold transfer functions.
• Wavelet: The output has no linear relationship between the inputs but var-
ies continuously. Different wavelets can be used as the activation function.

5.5 WAVELET NEURAL NETWORK

A trending technique used for signal classification is the machine learning approach. By
combining the concept of artificial neural network and wavelet, a derived neural net-
work is formed and named as wavelet neural network (WNN). WNN is a simple feed-
forward neural network that uses different wavelets as activation functions instead of the
binary, bipolar, or sigmoid functions as in conventional networks. For introducing non-
linear function approximation, accuracy, and precision in outputs and fault tolerance,
a relationship between neural network weights and wavelet transforms is defined [20].
The WNN considered is a three-layer network and the hidden layer activation
function utilizes the Mexican hat wavelet, and the structure is shown in Figure 5.11.
The Mexican hat wavelet function is defined as follows:

æ - x2 ö

( )
ç 2 ÷ø
f ( t ) = 1 - x2 eè (5.3)

In WNN, the signal is passed in a forward direction (forward signal propagation),

while the error is transmitted in the backward direction (backward error propa-
gation). The left to right travel of the signal implies the forward propagation, i.e.,
through the input layer, then through the hidden layer, and finally through the out-
put layer to the output end. Some intermediate results are produced by each layer
which will be given as the input to the succeeding layer, and the final output is com-
puted and given by the output layer. The backward propagation ensures the update of
104 Artificial Intelligence (AI)

FIGURE 5.11  Structure of WNN.

weights and bias values. The non-linear and complex relationships can be modelled
by the ANNs and it is applicable in many real-life practical systems. The response of
unseen data can be inferred by the ANN by learning from the initial input and output
relationships, i.e., an accurate prediction of output is possible from random inputs.
The neural networks can learn even the hidden complex relationship between input
and output by proper training to provide better results. All the mentioned features
introduced the neural network-based machine learning approach in industrial, tech-
nical, image processing, banking, finance, and many other fields as a very powerful
tool for the prediction of unknown, unexpected, and future outcomes.
Among the machine learning approaches, the WNN is faster and accurate, giving
better results compared to the back propagation network (BPN) and support vector
machine (SVM) techniques [4]. Training the programmed WNN algorithm with real-
time data provides true, faultless level readings. Thus, the WNN is useful in different
areas where exact true outputs are necessary, particularly for fuel level measurement in
the dynamic environments where the disturbance behaviour is unknown.

5.5.1 Training of WNN
Neural networks can undergo training to perform a particular task. Many engineering
tools are available for training neural networks. One of the powerful tools for train-
ing, analyzing, and simulating the neural network is the MATLAB software. The
training procedure modifies the weights and bias weights of a network by following a
training algorithm learning rule. Supervised learning and unsupervised learning are
Parameter Estimation in Dynamic Environments 105

two broad classifications of learning rules for data clustering. Supervised learning
always needs the support of a teacher for supervising and guiding the output accord-
ing to what desired response has to be given for each input signal. A training set is
provided with the learning rule for gaining proper network behaviour. While the
inputs are given to the neural network, comparisons are made between the network
outputs and the target values. To get the network outputs close to the target values,
the weights and biases of the network are updated using the learning rules [15].
Unsupervised learning gets trained by the available information only, and this
method never needs an external teacher. It is a self-organization technique because it
self-organizes the data given to the neural network and finds out their common proper-
ties. According to the network inputs, the weights and biases are adjusted and modi-
fied. In unsupervised learning, there is no need for target outputs and a majority of the
unsupervised learning algorithms are based on clustering operations. In the case of
clustering, the algorithm categorizes the input patterns into several finite classes. For
applications such as vector quantization, unsupervised learning is useful [15].
The Training Algorithm for WNN (Forward Calculations) [21]

X(n) External input vector

Wjk(n) Weight between input layer (k) and hidden layer (j)
Wij(n) Weight between output layer (i) and hidden layer (j)
aj(n) Dilation coefficient in the hidden layer at time n
bj(n) Translation coefficient in the hidden layer at time n

At time n, the net internal activity of neuron j is

m

Vj ( n ) = åWjk ( n ) * Xk ( n ) (5.4)
k =0

Vj(n) Sum of inputs to the jth hidden neuron

Xk(n) kth input at time n
The output of jth neuron is computed by passing Vj(n) through the wavelet ϕ(t)
where

f (Vj ( n ) - bj ( n ) )
f a, b (Vj ( n ) ) = (5.5)
aj ( n )
Sum of input neurons to output neuron is

V (n) = åWij ( n ) *f a, b (Vj ( n )) (5.6)

j =0

The output of the neuron is calculated by passing V(n) through the non-linear activa-
tion function, obtaining

Y (n) = s (V (n)) (5.7)

106 Artificial Intelligence (AI)

Learning Algorithm
At time n, the instantaneous sum of squared error is
1
E ( n ) = e2 ( n )
2

1
= éë y ( n ) - d ( n ) ùû (5.8)
2

2
d(n) The desired response at time n
η Learning rate
The method of steepest descent used to minimize the above cost function. Weight
between the hidden layer neuron j and input layer neuron k can be updated according to

¶E ( n )
DWjk ( n + 1) = -h * + m * DWjk ( n )
¶Wjk ( n )

Xk (n)
(
= h * e(n)* s (V (n)* Wij (n)* f a, b Vj ( n ) * ) ai(n)
+ m * DWjk (n) (5.9)

Weight between output layer neuron i and hidden layer neuron j can be updated
according to

¶E ( n )
DWij ( n + 1) = -h * + m * DWij ( n )
¶Wij ( n )

( )
= h * e ( n ) * d (V ( n ) * Wij ( n ) *f a, b Vj ( n ) + m * DWjk ( n ) (5.10)

Translation coefficient can be updated according to

¶E ( n )
Dbj ( n + 1) = -h * + m * Dbj ( n )
¶bj ( n )

1
(
= -h * e ( n ) * s (V ( n ) * Wij ( n ) * f a, b Vj ( n ) * ) aj ( n )
+ m * Dbj ( n ) (5.11)

Dilation coefficient can be updated according to

¶E ( n )
Daj ( n + 1) = -h * + m * Daj ( n )
¶aj ( n )

Vj ( n ) - bj ( n )
( )
= -h *e ( n ) * d (V ( n ) * Wij ( n ) * f a, b Vj ( n ) * + m * Daj ( n ) (5.12)
( aj ( n ) )
2
Parameter Estimation in Dynamic Environments 107

FIGURE 5.12  Flowchart of wavelet neural network training and validation.

The training and target data have been loaded to the network model once the training
is done. The neural network parameters such as learning rate, the maximum number of
epochs, and training function are fixed beforehand to call the training function. The trains
are nothing but the training function having training vectors, network objects, and target
vectors as parameters. The training steps involved are shown as a flowchart in Figure 5.12.

5.6 RESULTS
After calculating the samples from the mathematical model, the training data in dis-
turbances such as variations in temperature and inclination of the vehicles are saved
108 Artificial Intelligence (AI)

FIGURE 5.13  Final output: the red line shows the actual level and the blue line shows the
output of WNN.

as individual data. The saved data are loaded to the wavelet neural network and clas-
sified efficiently. After training and validation of the wavelet neural network model,
the proposed system is tested for its accuracy even in the dynamic environments.
The testing inputs are given randomly with the disturbances and dynamic variations
with respect to time. The neural network is tested to predict the level output and it is
compared with the known true values of fuel level of the tank.
The output of the neural network is given in Figure 5.13. The red line shows the
true level and the blue line shows the predicted level output, which are matching
with trained output a at highest accuracy. The MATLAB simulation and valida-
tion results carried out based on the given algorithm shows that the wavelet neu-
ral network predicted accurate results, which are exactly matching with the true
values of the level of the fuel tank at dynamic variations due to temperature and
inclinations.

5.7 CONCLUSION
WNN-based signal classification and signal processing method integrated with
FBG-embedded cantilever sensor was used for accurately predicting the level of
the fuel tank in the automotive system under dynamic conditions. Extensive train-
ing was done to ascertain an effective configuration for the WNN-based liquid level
measurement system. The WNN parameters selection and the signal pre-processing
configurations were all based on training with the values from the model. While
comparing with the existing level measurement methods, the results obtained from
the WNN-based FBG-embedded cantilever beam sensor system were having higher
sensitivity and higher accuracy in the dynamic environments.
Based on the accuracy, the WNN-based system is suitable to use in vehicles
used for racing competitions where vehicles are exposed to high manoeuvres. The
Parameter Estimation in Dynamic Environments 109

drivers of the automotive systems can monitor the dashboard for predicting the fuel
level accurately without any errors due to variations in accelerations, temperature,
and inclination. In conclusion, the proposed measurement system along with wave-
let neural network method for signal processing and classification is effective and
highly accurate in the determination of fuel level using FBG-embedded cantilever
sensor in dynamic environments.

REFERENCES
1. Edin Terzic, Romesh Nagarajah and Muhammad Alamgir, “A neural network approach
to fluid quantity measurement in dynamic environments”, Mechatronics, vol. 21, no. 1,
pp. 145–155, Feb. 2011.
2. Edin Terzic, Jenny Terzic, Romesh Nagarajah and Muhammad Alamgir, A Neural
Network Approach to Fluid Quantity Measurement in Dynamic Environments.
London: Springer-Verlag, 2012.
3. Dong-Sheng Xu, Hua-Bei Liu and Wei-Li Luo, “Development of a novel settlement
monitoring system using fibre-optic liquid-level transducers with automatic tempera-
ture compensation”, IEEE Transactions on Instrumentation and Measurement, vol. 67,
no. 9, pp. 2214–2222, Sept. 2018.
4. Bashir Shokouh Saljoughi and Ardesir Hezarkhani, “A comparative analysis of arti-
ficial neural network (ANN), wavelet neural network (WNN) and support vector
machine (SVM) data-driven models to mineral potential mapping for copper mineral-
izations in the Shahr-e-Babak region, Kerman, Iran”, Applied Geomatics, vol. 10, no. 3,
pp. 229–256, Sept. 2018.
5. https​:/​/wi​​k i​.an​​ton​-p​​aar​.c​​om​/in​​-en​/e​​​ngine​​-oil/​
6. Mengmeng Han, Jian Dai and Kok Keng Ang, “Hydrodynamic aspects of moving vehi-
cle with sloshing tanks”, MATEC Web of Conferences, 2018. https://doi​.org​/10​.1051/
matec conf/201821115002
7. Ilian Haggmark and Michael Fokine, “Fiber Bragg gratings in temperature and strain
sensors”, Laser Physics Group, May 2014.
8. Thulasi M Santhi and S Sathiya, “Performance enhanced liquid level sensing system
for dynamic environments”, IEEE 5th International Conference for Convergence in
Technology (I2CT), March 2019.
9. B Vasuki, S Sathiya and K Suresh, “A new piezoelectric laminated cantilever resonance
based hydraulic pump”, 2013 IEEE Sensors Applications Symposium Proceedings, vol.
978, no.1, pp. 4673–4637, April 2013.
10. S Sathiya and B Vasuki, “Higher resonant mode effect on the performance of piezo
actuated 2-DOF rectangular cantilever shaped resonators (2-DOF RCR) for liquid vis-
cosity and density sensing”, Microsystem Technologies, vol. 23, no. 7, pp. 2431–2445,
July 2017.
11. Hua-Fu Pei, Jian-Hua Yin, Hong-Hu Zhu, Cheng-Yu Hong, Wei Jin and Dong-Sheng
Xu, “Monitoring of lateral displacements of a slope using a series of special fibre Bragg
grating-based in-place inclinometers”, Measurement Science and Technology, vol. 23,
Issue 2, article id. 025007, 8pp, 2012.
12. Wenlong Liu, Yongxing Guo, Li Xiong and Yi Kuang, “Fiber Bragg grating based
displacement sensors: state of the art and trends”, Emerald Insight, vol. 39, no. 1, pp.
87–98, 2019.
13. T Guo et al. “Temperature-insensitive fiber Bragg grating liquid-level sensor based
on bending cantilever beam,” IEEE Photonics Technology Letters, vol. 17, no. 11, pp.
2400–2402, Nov. 2005.
110 Artificial Intelligence (AI)

14. M Pharr and G Humphreys. Sampling and Reconstruction. Physically Based Rendering:
From Theory to Implementation. Amsterdam: Elsevier/Morgan Kaufmann, 2004, pp.
279–367.
15. Neural Networks. http:​/​/www​​.doc.​​ic​.ac​​​.uk/^​nd/su​r pris​e_96/​journ​al /vol4/cs11/report​
.ht​m l
16. BD Ripley, “Statistical aspects of neural networks”, In OE Barndorff-Nielsen, JL
Jensen and WS Kendall (Eds.), Networks and Chaos—Statistical and Probabilistic
Aspects. London: Chapman and Hall, 1993, pp. 40–123.
17. H Demuth, M Beale and M Hagan, “Neural network toolbox 5 users guide”, MathWorks,
2007.
18. R Rojas, Neural Networks—A Systematic Introduction. New York: Springer, 1996.
19. U von Luxburg and G Rätsch (Eds.), Advanced Lectures on Machine Learning: ML
Summer Schools, Canberra, Australia, February 2–14, 2003. Tübingen, August 4–16,
2003, (Rev. lectures/Olivier Bousquet). Berlin, New York: Springer.
20. Ping Zhou, Chenyu Wang, Mingjie Li, Hong Wang, Yongjian Wu and Tianyou Chai,
“Modeling error PDF optimization based wavelet neural network modeling of dynamic
system and its application in blast furnace iron making”, Neurocomputing, vol. 285, pp.
167–175, April 2018.
21. MR Mosavi, “Wavelet neural network for corrections prediction in single-frequency
GPS users”, Neural Processing Letters, vol. 33, no. 2, pp. 137–150, April 2011.
 6 Computer Vision
Concepts and
Applications
Bettina O’Brien and V. Uma

CONTENTS
6.1 Introduction................................................................................................... 112
6.1.1 Evolution of Computer Vision........................................................... 113
6.2 Feature Extraction......................................................................................... 113
6.2.1 Types of Features............................................................................... 114
6.2.2 Feature Extraction Methods.............................................................. 114
6.2.2.1 I. Low-Level Features......................................................... 114
6.2.2.2 Texture Estimator................................................................ 115
6.2.2.3 Colour Histogram............................................................... 115
6.2.2.4 Colour Descriptor............................................................... 116
6.3 Object Detection............................................................................................ 120
6.3.1 Image Classification........................................................................... 120
6.3.1.1 Classification and Localization........................................... 120
6.3.2 Image Segmentation.......................................................................... 120
6.3.2.1 Semantic Segmentation....................................................... 120
6.3.2.2 Demerits of Sliding Window.............................................. 121
6.3.2.3 Instance Segmentation........................................................ 121
6.3.3 Region-based Methods...................................................................... 121
6.3.3.1 Region Proposal.................................................................. 121
6.3.3.2 Region-based Convolutional Neural Network(R-CNN)..... 121
6.3.3.3 Fast Region-based Convolutional Neural Network............. 122
6.3.3.4 Faster Region-based Convolutional Neural Network......... 122
6.3.4 Alternative Methods.......................................................................... 122
6.3.4.1 HOG Features..................................................................... 122
6.3.4.2 You Only Look Once (YOLO)........................................... 122
6.3.4.3 Demerits of YOLO.............................................................. 122
6.4 Computer Vision Hardware, Software, and Services.................................... 123
6.4.1 Computer Vision Hardware............................................................... 124
6.4.2 Software Libraries and Tools............................................................. 124
6.4.3 Computer Vision Services................................................................. 124
6.5 Applications of Computer Vision.................................................................. 125
6.5.1 Healthcare.......................................................................................... 125

111
112 Artificial Intelligence (AI)

6.5.2 Augmented Reality............................................................................ 125

6.5.3 Vision-based Self-Driving Cars......................................................... 126
6.5.4 Automatic Target Recognition and Detection................................... 127
6.5.4.1 Case Study: Robotic Path Planning Using Visual Percepts.... 127
6.6 Conclusion and Future Directions................................................................. 128
Bibliography........................................................................................................... 128

6.1 INTRODUCTION
Artificial intelligence (AI), which is the source of subfields like machine learning (ML),
deep learning, computer vision, etc., is a technology that defines intelligence as knowl-
edge acquisition and representation, where knowledge is acquired and applied to a prob-
lem. The objective of any AI problem [1] would be to elevate the success rate and not
the accuracy, whereas an ML problem would aim at utilizing the model to increase its
accuracy but does not concentrate on its success. In general, artificial intelligence is the
ability to engineer intelligent machines and programs, while machine learning is the
ability to learn without being explicitly programmed. AI pertains to decision-making
where the goal is to simulate natural intelligence in machines. But machine learning
learns from data to maximize the performance of the machine based on a specific task.
Deep learning [2, 3], a subset of machine learning, makes the computation of mul-
tilayered neural networks feasible, working with large amounts of data. Algorithms
with single-layered neurons worked just fine, but the weights remained the same
during the training phase, which didn’t have a very good impact on the accuracy.
So, algorithms were designed such that adjustment of weights made changes which
reduced the error. This, in turn, concentrates on delivering high accuracy for task-
specific applications such as object detection [1, 4], image segmentation [5, 6],
speech recognition [7], feature extraction [8, 9], language translation, and other
recent upgrades in technology. These image-related applications significantly co-
relate with another domain of AI: computer vision.
Computer vision [10, 11] is a simple phenomenon that aims at making computers
view and sense its surroundings. It has been the source for almost all computations
which implicate visual content, especially in the form of digital images. Machines
simply illuminate images as a series of pixels with their own set of colours [4] and
values. Hence, it helps to replicate how a human brain reads an image. This is the
basic way of how intelligence will be modelled in a machine. It ultimately aims at
extracting information from pixels. This field of computer science, which initially
worked based on statistics alone, is now switched to deep neural networks [7]. Here,
both the objective and accuracy of achieving the goal are taken care of.
Image processing [2] is another term that is often thought to resemble computer
vision, but it is not. The commonality is that both subfields of AI require images to
work with, but the method in which these images are used is varied. Image process-
ing, as the name implies, processes images by adding in functionalities like smooth-
ing, sharpening, adding contrast, and stretching, or performs other transformations
in images and the output would be an image. But computer vision aims to identify
images and interpret information to provide an output in terms of the image size,
Computer Vision Concepts and Applications 113

colour intensity, etc. It also tries to classify images, identify instances from an image,
etc. Trying to classify images or associate the information between images to differ-
entiate between individuals are applications that involve deep learning techniques.
Image processing uses probabilistic methods like the hidden Markov model, editing,
restoration techniques, and filtering. These methods use neural networks to incorpo-
rate more learning and to model intelligence.
Computer vision has a non-complete list of applications that tends to grow over
the years. It has a dramatic impact on almost all sectors like retail, healthcare, finan-
cial services, and so on. In general, computer vision [10, 11] is applied in automatic
inspection, identification systems, detecting events [12], controlling processes, mod-
elling or recreating objects or environments, navigation, etc. Section 6.2 gives a
detailed explanation of feature extraction and the techniques used to select the neces-
sary information from the image. In Section 6.3, methods used to identify and locate
objects and instances in images are dealt with. In Section 6.4, some of the important
hardware, software, and services in computer vision are discussed. In Section 6.5,
the applications in the field of healthcare, self-driving cars, and automatic target
detection are dealt with. A case study of how computer vision and deep learning
methods influence autonomous robotic path planning around a dynamic environ-
ment with obstacles is explained.

6.1.1 Evolution of Computer Vision

Computer vision, before deep learning came into existence, had a simple procedure
in processing images. Initially, the captured images were stored in a database. For
each image, data points regarding the properties and measurements of the image
were defined. This procedure is called data annotation. The same annotation proce-
dure was to be done for other new sets of images. After these processes, the objective
of application-specific problems was carried out. This traditional approach requires
much manual work to identify features during the pre-processing stage. Such a time-
consuming process can now be handled by the concept of feature extraction.

6.2 FEATURE EXTRACTION
Feature extraction is one of the indispensable preconditions for processing images.
In the field of machine learning, feature extraction [8] learns and identifies patterns
from images. It can be interpreted as the transformation from a set of raw data into
derived values called features. A feature [9] represents a function of characteristics
or measurements specifying the assessable property of an image. These features
must be both instructive and non-repetitive. Instead of using the entire data as a
whole, relevant information from features is extracted to perform the expected task.
The extracted features that are rich in the needful information are further used for
feature selection and classification tasks. Alternatively, feature extraction seems to
be a significant form of dimensionality reduction. In the problem of image classi-
fication, the extracted features should contain the relevant information required to
differentiate between classes.
114 Artificial Intelligence (AI)

6.2.1 Types of Features
The outcome of feature extraction would be a set of features referred to as the feature
vector, which is a representation of images. A feature is of good quality when it con-
tains discriminative information that can differentiate between objects. The types of
features can be classified as follows:

1. Local features are features derived from the results of the edge detection or
image segmentation process. They are individual results or features gained
from processing different partitions of an image.
2. Global features are features calculated from the overall image or a well-
ordered sub-section of an image.
3. Pixel-level features are calculated at each pixel in terms of colour and
texture. Colour holds features such as prominence, colour spread, and co-
occurrence, whereas texture holds features like contrast, sharpness, and
change in intensity.
4. Object-level features are calculated in terms of shape, size, texture, and
spatial distribution. Shape and size hold features like perimeter, bending
energy, Fourier shape descriptor, model dimensions, area, and solidity. The
texture holds the same features as pixel-level features. Spatial distribution
holds features like edge connectedness, length of edge, compactness, and
distance among peer neighbours.
5. Domain-specific features are conceptual features that are application
dependent; e.g., character recognition pertains to the formation of letters
alone, whereas face identification deals with facial features alone.
6. Semantic-level features are based on a subset of low-level features that
obtains a high-level concept. The inference or relation is identified between
two or more low-level features.

6.2.2 Feature Extraction Methods

In general, feature extraction is classified into low-level and high-level features.
Low-level features are the basic features of an image like shape, size, and texture.
Such features can be directly identified and extracted. They do not rely on individual
object description in the image. High-level features are those that concentrate on the
object’s shape, edge, or corners of the image.
6.2.2.1 I. Low-Level Features
6.2.2.1.1 a. Local Features
Local features generally give a representation of the image patches, which tends to
identify unique patterns or structures from an image. Local feature detection is a
preliminary procedure in a computer vision algorithm. It tries to find distinct fea-
tures that differentiate its surrounding pixels in terms of edges, corners, texture,
colour, or intensity. Such features extracted locally from an image are classified into
texture, shape, and colour features. These local features contribute to influential fac-
tors for images. Despite image occlusion or any presence of clutter, local features
Computer Vision Concepts and Applications 115

identify image compatibility. So, it helps in differentiating between patches in an

image, which leads to the application of image classification.

i. Texture

The texture of an image is a feature that identifies the region of inter-

est (RoI) by performing image segmentation. Later, image classification is
performed to classify between regions. Texture gains information from the
spatial representation of colours in an image. Since the texture is identified
in terms of regions, as a repeating pattern of local features, it cannot be
defined for a point in an image. An image’s texture can be seen in terms of
fine, coarse, grainy, smooth, sharp, etc.

These features are described in the tone and structure of a texture. Tone defines the
properties of texel’s pixel intensity, and structure defines the spatial representation or
relation between texels. Structurally, the texture is defined as a set of elementary texels in
a regular pattern, and statistically, it can be defined as a measure of intensity arrangement
in any particular region of an image. These measurement sets define the feature vector.

6.2.2.2 Texture Estimator
• The range is the simplest operator which calculates the difference between
maximum and minimum intensity values of neighbours in an image. Range
estimates texture by converting the original image into a brighter image.
The texture is read as the brightness value.
• Variance calculates the sum of squares of the differences between the
intensities of the central pixel and its neighbours, i.e. it simply calculates
the variance between neighbour regions.
ii. Colour
Colour is the primary image perception in understanding how
human brains perceive an image. It is a robust descriptor which sim-
plifies object detection and image segmentation problems. The colours
[4] that are perceived from an image are made up of light of different
wavelengths. A colour space depicts the colour representing its inten-
sity values. Colour space models can either be hardware-oriented based
on a three-colour stimulus like RGB or user-oriented.

6.2.2.3 Colour Histogram
A colour can be described in various shades (e.g., green colour being in shades of
dark or light green). But for the machine to interpret variations between shades of
the same colour, the probability of pixel intensities needs to be defined. Colour his-
tograms define a set of bins, where each bin holds the probability of pixels of a par-
ticular colour. A vector representation of a histogram is Equation 6.1:

{ }
H H éë0 ùû, H éë 1ùû ,¼, H éëi ùû ,¼, H éë N ùû (6.1)
116 Artificial Intelligence (AI)

Here, each H[i] in Equation 6.1 represents the number of pixels in the colour i in an
image and N represents the total number of bins. Histograms need to be normalized
if two different sized images are to be compared. The normalized vector representa-
tion is given in Equation 6.2 and the reduced form is given in Equation 6.3:

{ }
H ¢ H ¢ ëé0 ùû , H ¢ éë1ùû ,¼, H ¢ éëi ùû ,¼, H ¢ éë N ùû (6.2)

H ¢ éëi ùû = H éëi ùû / P (6.3)

P tends to the number of pixels in the image. The range of intensity values for a
greyscale image can be given in the range I(u,v), denoting the pixel intensities which
belong to [0, k − 1], where k represents the number of colours. For example, for an
8-bit image, k = 28 = 256 colours. So, in general, colour histograms deal with prob-
lems like over- and underexposure, brightness, contrast, and dynamic range.

6.2.2.4 Colour Descriptor
Descriptors provide a way to relate pixels in an image. A colour descriptor is catego-
rized under the general information descriptors. Since colour is the most significant
feature of any visual content, it has to be described exclusively. For any image, the
colour description can be performed using the following methods:

• Dominant colour descriptor (DCD)

• Scalable colour descriptor (SCD)
• Colour structure descriptor (CSD)

To trace the relation between colours among a group of images, the following meth-
ods are used:

• Colour layout descriptor (CLD)

• Group of frames (GOF) or group-of-pictures (GOP)
a. Global Features
Global features represent images in terms of a single vector. Mostly
the texture and shape descriptors are considered while addressing the
global features of an image.
i. Principal Component Analysis (PCA)
PCA is a dimensionality reduction tool that reduces a large set of
feature variables into a smaller set but still retains the needed infor-
mation from the larger set. PCA emphasizes variations in features
and identifies strong patterns from data sets. The set of uncorrelated
variables derived as the identified patterns from an image is known
as principal components and these components are simply eigen-
vectors of a covariance matrix and are orthogonal. For example, if
two dimensions, say, the height and weight in a data set are consid-
ered, PCA builds the coordinates (x,y) value for every point. Here,
Computer Vision Concepts and Applications 117

the combination of height and weight is called principal compo-

nents. The variance of a linear composite of variables i and j are
given by Equation 6.4:
P

åa x = å å a a s
i =1
i i i j ij (6.4)

where σ denotes the covariance between the ith and jth variables.
The variable which contributes less information along the axes
would be dropped off. In terms of matrix representation, a¢Ca,
where C represents the covariance matrix and a represents the vec-
tor of variable weights, PCA finds the weight vector a that mini-
mizes a¢Ca with a constraint defined in Equation 6.5:

åa 2
i = a¢a = 1 (6.5)

Hence, PCA projects data in the directions of maximum variance

and finds the most precise representation of data in a lower dimen-
sional space.
ii. Linear Discriminant Analysis (LDA)
LDA is another way that strives to reduce dimensionality while still
maintaining the class-biased information. So, LDA best applies to
classification problems. The basic objective of Fisher’s LDA is to
find a projection to a line such that the samples from various classes
are segregated. Let x be a p-dimensional predictor with k class,
where k represents linear discriminant functions, then Fisher’s
approach projects x to (k − 1) dimensional spaces to extract new
features in which the variance between classes is maximized and
the variance within classes are minimized.
The objective is to acquire a scalar y by projecting the samples x
onto a line y = wT x . To define a good projection vector, a measure
of separation between classes needs to be defined. So, the mean
vector of each class wi in x-space and y-space is given by

åx (6.6)
1
mi =
Ni xÎwi

åy (6.7)
1
mi¢ =
Ni yÎwi

Since y tends to w x , y = w x . Substituting y in Equation 6.7,

T T

å w x (6.8)
1
mi¢ = T
Ni yÎwi
118 Artificial Intelligence (AI)

Substituting Equation 6.6 in Equation 6.8,

mi¢ = wT mi (6.9)

The objective now can be framed to find the distance between

the projected means, which in turn yields better class separabil-
ity. Figure 6.1 shows a graph of randomly chosen means between
classes and how their separation measure is calculated from the
distance between means.

I. High-Level Features
High-level representations of an image are well-built features inclined to be
seemingly hidden. But minimizing the bridge between low-level and high-
level features is significant in emphasizing object detection problems. The
low-level local features after detection can be combined to classify images.
So, to learn in deep about these hidden features, one requires time and
computational power. This is where the most important segment of deep
learning is used. Deep neural networks use several layers to learn high-level
representations of an image.
a. Template Matching
Template matching is a high-level method that tries to identify areas
of an input image using a given image pattern or any patch. So given
an image, this method works by sliding the template over the whole
image to detect edges of instances. The matching approach completely

FIGURE 6.1  Distance between means and class separability of x and y spaces.
Computer Vision Concepts and Applications 119

FIGURE 6.2  Correlation analysis.

depends on the type of image and objective chosen. Template matching

can be performed based on features, area, template, or motion.
The matching process in Figure 6.2 is done pixel-by-pixel, where the
template image moves overall viable positions in the given input image
to compute an integer index which interprets how well a match is found.
For a greyscale image, the exactness cannot be precise, so a difference
in levels between the source image and template has to be known. Thus,
the distance dist is calculated by Equation 6.10:

n m

dist ( I , t , a, b ) = åå ( I ( a + i, b + j ) - t (i, j ))
2
(6.10)
i =1 j =1

where
(a,b) denotes the top corner coordinates of the template t
I     denotes the greyscale image with a grey-value template t of size n*m

Correlation is the measure of degree where two variables correlate with each
other in terms of general behaviour if not directly through values. Correlation is
given by Equation 6.11:

å (u - u¢) ( v - v¢)
N
i i
C= i =1
(6.11)
å (u - u¢) å ( v - v¢)
N 2 N 2
i i
i =1 i =1

where
u and v denote the template grey-level image and source image, respectively
u′ and v′ denote the average grey-image level in the template image and source
image, respectively
N denotes the image template size in terms of pixels and is given by
N = columns * row
C takes values that range between ±1, where greater values represent stronger
relationships between images
b. Hough Transform
Hough transform is a method for calculating global feature repre-
sentation, where it gradually builds classes by working with the local
features with the number of classes initially being unknown. Hough
transform is, in particular, used to detect shapes such as lines, circles,
120 Artificial Intelligence (AI)

or curves. It is a generalized localization problem aiming to find the

location of shapes in images.

6.3 OBJECT DETECTION
Object detection [9] is the principal challenge in computer vision. It is known to rap-
idly grow to give improved results in terms of accuracy and testing time when used
with machine learning [1, 13] and deep learning [14] approaches. One such latest
software system that implements object detection algorithms using deep learning
concepts is the Detectron, developed by Facebook. Nowadays, there are more deep
learning pre-trained models [3, 7] used for object detection like Convolutional
Neural Network (CNN), Region-based Convolutional Neural Network (RCNN),
You Only Look Once (YOLO), Fast R-CNN, etc. The basic structure for computer
vision is CNN implemented using deep learning. CNN combines image classifica-
tion, object classification, and localization. For the past few years, CNN has proved
to be a reliable method for object detection and classification due to its speed and
accuracy.

6.3.1 Image Classification
To classify images, the image given as input goes through a deep convolutional net-
work, and this network gives a feature vector to the fully connected layers that pro-
vide different class scores. Finally, the output delivers the content of the entire image
as a whole. This is the basic possible task in computer vision. There are various types
of other computer vision tasks that detect spatial pixels inside the image. These are
achieved through a convolutional neural network.

6.3.1.1 Classification and Localization

To know where and what the object detected is, object classification and localiza-
tion [15] algorithms are used. In addition to predicting a category-labelled object, a
boundary box is drawn around the region of the object in the image to show where
the object is located in that image. The distinction between classification with local-
ization and object detection is that in the localization scenario the number of objects
is known ahead of time, while in the latter bounding boxes are generated exactly
proportional to the identified quantity of objects found in the image.

6.3.2 Image Segmentation
6.3.2.1 Semantic Segmentation
In semantic segmentation [6], the output image is the decision of the category of
every pixel in that image. It is similar to image classification. Rather than assigning
a single category labelled to the entire image, the output of this segmentation [5]
produces a category label for each pixel of the input image. Semantic segmentation
does not differentiate instances, it only labels the category of that pixel. When two
similar objects are close together, the semantic model does not distinguish between
Computer Vision Concepts and Applications 121

the objects. Instead, it labels the whole mass of pixels. This is a small drawback for
semantic segmentation that can be fixed using instance segmentation. One potential
approach to attack semantic segmentation is through classification. The sliding win-
dow method can be approached to segment an image semantically.
For a sliding window, the input image is cropped into many small images and
each cropped image is treated as a classification problem. This will help to find
the category of the central pixel of the cropped image. The sliding window is com-
putationally expensive. To label a pixel, the full image needs to be cropped into
smaller images and this would be expensive to run. A model that would be even
more efficient than semantic segmentation is a fully convolutional network. Rather
than extracting single patches from the image and classifying it separately, it can be
imagined as a whole stack of convolutional layers.

6.3.2.2 Demerits of Sliding Window

• Expensive
Using CNN by cropping multiple images for detecting objects from the
image will be expensive. The method can be altered to pass the entire image
instead of passing the cropped image to detect objects.
• Inaccurate boxes
The algorithm can find and localize objects in an image, but the shape of
the box is not accurate, sometimes it has a rectangular object and at some
point, the object does not fit in the box at all.

6.3.2.3 Instance Segmentation
Given an input image, the objective is to identify and detect the location and identi-
ties of objects in the image, which is, perhaps, similar to detecting objects [16] but,
instead of predicting a bounding box for each object, the whole segmentation mask
has to be predicted so that a particular instance from the image would be captured.

6.3.3 Region-based Methods
6.3.3.1  Region Proposal
This method does not use deep learning as such, but it is mostly based on a tradi-
tional approach to computer vision. This region proposal network uses the basic
signal processing technique to generate several proposals. For a given input image,
the region proposal network gives an output of thousand boxes as to where the object
might be present. These are relatively fast to run.

6.3.3.2 Region-based Convolutional Neural Network(R-CNN)

Given an input image, the model executes the region proposal network to get RoI.
One of the problems in R-CNN is that the regions in the input image can be in dif-
ferent sizes. To run them all through the convolutional network for classification,
the input image needs to be of the same size. Each of the region proposals is taken
and wrapped to a fixed square size that is expected as input. This will run through a
convolutional network, which in turn makes classification decisions for each of those
122 Artificial Intelligence (AI)

images to predict categories. Besides, R-CNN also predicts a regression for each of
the input region proposals. The main disadvantage of R-CNN is that it is computa-
tionally expensive, and the runtime is slow.
6.3.3.3  Fast Region-based Convolutional Neural Network
Problems faced with R-CNN can be fixed by Fast R-CNN. Both methods are almost
the same but rather than each RoI being processed separately, the entire image is sent
through the convolutional network all at once. It gives a feature map that effects high
resolution that corresponds to the entire image. If there are fully connected layers down-
stream, those layers will be expecting fixed-size input. So, reshaping the cropped images
from the convolutional feature map is done using the RoI pooling layer, which is similar
to max pooling. Once the images are wrapped from a feature map, classification scores
can be predicted by running those images through fully connected layers. In terms of
speed, Fast R-CNN is ten times faster than the SPP net (which is in between R-CNN and
Fast R-CNN) and R-CNN. In terms of test time, Fast R-CNN is superfast since the test-
time is dominated by computing region proposals. So, Fast R-CNN ends up being bottle-
necked by computing these region proposals. This problem is solved by faster R-CNN.
6.3.3.4  Faster Region-based Convolutional Neural Network
In Faster R-CNN, self-generated region proposals are created. It runs the entire input
image through convolutional layers to get a feature map representing the entire image
in high resolution. There is a separate region proposal network that works above those
convolutional features, which in turn predicts its region proposals inside the network.
It looks like Fast R-CNN when the region proposals are predicted. Now the image is
cropped from those region proposals and passed to the network for classification.

6.3.4 Alternative Methods
6.3.4.1 HOG Features
The HOG is an important technique that can be used for object detection, mainly in
self-driving cars. It is a feature descriptor, powerful when used with a support vector
machine (SVM) that works for object detection. To determine whether the object is
found or not, each computed HOG descriptor is fed to the SVM classifier. By chang-
ing the image scale, HOG solves the scaling issues. HOG represents a single vector
which describes a segment of an image.
6.3.4.2 You Only Look Once (YOLO)
YOLO algorithm is a better solution to create accurate boxes around the images and is
also faster than CNN. YOLO is just a modified version of CNN. In YOLO, the image is
divided into multiple grids and for each grid, classification and localization algorithm
is implemented. There are multiple versions for YOLO models. The latest version is
YOLO9000 which is faster, stronger, and better. It is trained on 9000 classes.
6.3.4.3 Demerits of YOLO
• Detecting one object multiple times can be confusing and time-consuming
to correct them. This can be solved by Non-Max suppression. This helps
Computer Vision Concepts and Applications 123

to remove boundary boxes that have a low probability that is close to high-
probability boxes.
• No possibility to detect multiple objects in one grid. With the use of Anchor
box, each object is assigned to each anchor box in a single grid to detect
multiple objects. This can also be solved by selecting a small grid, but when
objects are too close to each other this method can fail.

6.4 COMPUTER VISION HARDWARE, SOFTWARE, AND SERVICES

The advancement of tools and libraries in computer vision has evolved over the
years. The evolution of hardware and software has grown such that services are now
available and are offered by standard companies, as depicted in Figure 6.3. Various
hardware products and services are available in the market, out of which the most
significant ones are listed below.

FIGURE 6.3  Computer vision revenue by world markets.

124 Artificial Intelligence (AI)

6.4.1 Computer Vision Hardware

a. Commercial Products
Some of the most frequently used products are 3D Scanners, Bandit,
CycloVision, DALSA (Digital Imaging), Eltec, Genex Technologies,
IQinVision, NanoSystem, Optel Vision, Robocams, and Vista Imaging.
b. Research Systems
Some of the hardware products used in research systems are BiTEC,
Binocular Camera Platforms, Homemade 3D Laser Scanner, and VLSI
Vision Chips.

6.4.2 Software Libraries and Tools

Recent advancements in GPUs, software tools, libraries, and frameworks have made
it possible to build various image-processing applications. Out of the available ones,
the basic and most widely used tools are as follows.

a. OpenCV
OpenCV is the most fundamental multi-platform open-source library
that has in-built techniques and algorithms to implement image-related
problems. It works well for a smaller image data set. But for larger data
sets, since OpenCV does not support GPU, it uses CUDA.
b. MATLAB
Due to the ability of prototyping, MATLAB is the most used research
tool for building computer vision applications. The ease of debugging using
MATLAB is higher. But its demerit is that it is a pay-and-use platform.
c. TensorFlow
TensorFlow is an important tool that is gaining attention nowadays. It
provides a way to integrate deep learning with computer vision. It also
has APIs that can be used for other applications like classification, detec-
tion, and so on. One demerit of TensorFlow is that it consumes a lot of
resources.
d. CUDA
CUDA is a platform developed by NVIDIA to enable increased comput-
ing performance in GPUs. The toolkit includes a primitive library that uses
image and video functions. Memory distribution and power consumption is
one downside to this platform.

6.4.3 Computer Vision Services

a. Google Cloud Vision API
Cloud and Mobile Vision service developed by Google allows users to
call simple APIs to their application. The so-called APIs hold resourceful
machine learning models and their functionalities. It also allows barcode
scans and text detection using the OCR functionality.
Computer Vision Concepts and Applications 125

b. Amazon Rekognition
Rekognition is a deep learning-based service developed by Amazon to
analyse images and videos. The basic technique supported by this service
is object detection and segmentation. This service provides the best analysis
for facial recognition as well as sentimental analysis.
c. Microsoft Azure Computer Vision API
This service invented by Microsoft provides the same functionalities
with images and videos. Besides, it provides handwriting recognition.

6.5 APPLICATIONS OF COMPUTER VISION

Through the years, the applications of computer vision have been exponentially
expanding in almost all the fields. The phrase application here can be explained as
a high-level function which tends to handle an objective with a high level of com-
plexity. In general, computer vision is used in the fields of healthcare, manufactur-
ing, security systems, OCR recognition, tracking systems, agriculture, autonomous
vehicles, robotics, and much more. In this section, applications relating to healthcare
[17], autonomous vehicle, automatic target detection, and a case study on how com-
puter vision is used in robotic path planning are discussed.

6.5.1 Healthcare
Computer vision combined with machine learning has made adequate progress in
the field of healthcare [2]. The possible areas in healthcare where technology can
directly intervene are medical imaging analysis, monitoring of patients using gad-
gets, predictive analysis of diseases, and much more. Medical imaging is the process
of making visual representations of the inner body organs for analysis. Such analysis
requires feature extraction techniques [17], which is one of the key concerns. Other
disciplines like electroencephalography (EEG) and electrocardiography (ECG) pro-
vide graphical representations of the body organs, whereas on computer vision the
concept of image co-registration is used. Image co-registration is the transformation
of multiple images or photographs from different sources and time into one coordi-
nate system. Co-registration implies the integration of information from different
images into a single image for analysis. This concept has a significant impact in the
field of healthcare as well as in satellites.
A fundamental machine learning algorithm for visual perception is the polyhar-
monic extreme learning machine (PELM) algorithm. This algorithm visually learns
the image details after which it tries to classify images with a particular disease, based
on the learned parameters. This is where machine learning and deep learning are used
along with computer vision for disease prediction. One such important accomplishment
in the healthcare field is the prediction of breast cancer [18] using deep learning.

6.5.2 Augmented Reality
Augmented reality can be discerned as adding digital information to a real-
world environment. AR just creates additional graphics and sounds to an existing
126 Artificial Intelligence (AI)

FIGURE 6.4  General AR-based system.

environment, whereas virtual reality (VR) virtually creates a wholly artificial envi-
ronment of the real world. AR tends to change the perception of the reality observed
from a physical space, which completely involves computer vision techniques. There
are various types of AR [19], under which generic digital augmentation is one. This
type visualizes multiple 3D objects in its environment. Figure 6.4 gives a general
outline of how an AR-based system works and how the hardware, software, and
servers integrate to visualize virtual objects in a captured real-world environment.
To randomly include objects into a physical space, a combination of three tech-
niques is required.

• Simultaneous localization and mapping (SLAM) [15]: It involves a map

construction of an unknown environment and based on the virtually
included 3D object’s location, a regular update of the map is done.
• Depth tracking: It measures the distance between the added virtual objects
and the AR device.
• Image projection: The captured information from input sensors is pro-
cessed and then projected onto the augmented setup.

6.5.3 Vision-based Self-Driving Cars

Autonomous driving [13] has been a major application of computer vision along with
deep learning techniques, where the computer itself perceives and learns from a non-
deterministic environment. Object detection is the most significant strategy based on
which vision-based self-driving cars work. The general procedure involves stereo
vision, space estimation to drive, lane recognition, traffic light recognition, object
detection and tracking, semantic scene understanding, mapping and localization,
Computer Vision Concepts and Applications 127

temporal and spatial recognition of the vision, and vision fusion from various sen-
sors. Stereo vision [1, 7] involves the perception of the environment from sensors,
which can be categorized as direct and mediated perception. Perception [20] can be
taken directly from the working space regarding the texture, 3D shape, etc., without
any knowledge of object categories. The basic methodology of how object detection
and classification with localization are performed for a given image has been dis-
cussed in the previous section.

6.5.4 Automatic Target Recognition and Detection

Automatic target detection is the potential of any learning algorithm to identify,
sense, and detect the target from an environment based on the input sensors. From
a given environment, the spatial and temporal features are learned, which would be
useful in recognizing the target. An application of robotic navigation is discussed as
a case study in which the robot’s path is semantically generated, given a target label.

6.5.4.1 Case Study: Robotic Path Planning Using Visual Percepts

The objective [8] is to navigate a robot to a target object in an undeterministic envi-
ronment, i.e. through a previously unseen physical space. The environment is inter-
preted as representations that capture spatial and semantic information [12]. The
workflow of the paper is depicted in Figure 6.5. The representations are in terms
of detection (Det), segmentation (SSeg), depth of the image, and RGB. This prob-
lem is perceived using visual observations, handled by Markov’s decision process.
Semantic segmentation and detection are performed using state-of-the-art computer
vision algorithms, where the features are identified and classified using a convolu-
tional neural network (CNN) and a long short-term memory (LSTM) network to
remember previous actions of the robot in the environment towards the goal: the
target label. Thus, deep networks are used to learn navigation policies. During the
training phase, an optimal path planning algorithm is used to calculate the progress
towards the goal, which in turn helps in identifying what action to perform during
the testing phase. The representations Det and SSeg, in particular, with no require-
ments of domain adaptation, identify layouts of scenes, location of the given target,
and the obstacles along the path. For simulation, a renderer generates bounding boxes
to identify objects. Thus, given a target label in an environment, the environment is
interpreted into representations that identify and capture the objects and their loca-
tion. Based on these representations and the memory about previous actions from

FIGURE 6.5  Path planning using visual percepts.

128 Artificial Intelligence (AI)

LSTM, the next action towards the goal will be learned. Finally, the model generates
a path from the start to the target.

6.6 CONCLUSION AND FUTURE DIRECTIONS

The incorporation of deep learning and artificial neural networks in almost all
fields of computer vision has made it possible to replicate human vision. Computer
vision works basically with images and includes techniques such as image analysis,
scene analysis, and image understanding. Computer vision without artificial intel-
ligence works well but including deep learning methods achieves accurate results.
Researchers have identified that computer vision is more effective in identifying
image patterns than human cognitive systems. This chapter includes the most impor-
tant aspects of computer vision like feature extraction, object detection, and image
segmentation. The applications of computer vision have been growing exponentially,
especially in terms of control systems, detecting events, modelling environments,
target detection, automatic inspection, navigation, etc.
The future of computer vision has great scope because of its faster, reasonably
economic, and more accurate result-providing capabilities. Computational time is
one inconvenience that should be noted, because of the immense mathematical oper-
ations involved for each iteration. Further methods are being researched so that the
computationally time and power can be gradually reduced.

BIBLIOGRAPHY
1. K. Grauman and B. Leibe, Visual Object Recognition (Synthesis Lectures on Artificial
Intelligence and Machine Learning). Morgan and Claypool, 2011.
2. Muhammad Imran Razzak, Saeeda Naz, and Ahmad Zaib, “Deep learning for medical
image processing: Overview, challenges and the future,” in Classification in BioAapps,
pp. 323–350. Springer: Cham, 2018.
3. N. Buduma and N. Locascio, Fundamentals of Deep Learning: Designing Next-
Generation Machine Intelligence Algorithms. O’Reilly Media Inc., 2017.
4. F.S. Khan, R.M. Anwer, J. Van De Weijer, A.D. Bagdanov, M. Vanrell, and A.M.
Lopez, “Color attributes for object detection,” in 2012 IEEE Conference on Computer
Vision and Pattern Recognition, pp. 3306–3313. IEEE, 2012.
5. Y.-J. Zhang, Advances in Image and Video Segmentation. IGI Global, 2006.
6. M. Seyedhosseini and T. Tasdizen, “Semantic image segmentation with contextual hier-
archical models,” IEEE Transactions on Pattern Analysis and Machine Intelligence
38(5): 951–964, 2015.
7. M. Alam, M.D. Samad, L. Vidyaratne, A. Glandon, and K.M. Iftekharuddin, “Survey on
deep neural networks in speech and vision systems,” arXiv Preprint ArXiv:1908.07656,
2019.
8. I. Guyon, S. Gunn, M. Nikravesh, and L.A. Zadeh, Feature Extraction: Foundations
and Applications, vol. 207. Springer, 2008.
9. Y. Zhang, Y. Chen, C. Huang, and M. Gao, “Object detection network based on feature
fusion and attention mechanism,” Future Internet 11(1): 9, 2019.
10. R. Szeliski, Computer Vision: Algorithms and Applications. Springer Science &
Business Media, 2010.
Computer Vision Concepts and Applications 129

11. S.J. Prince, Computer Vision: Models, Learning, and Inference. Cambridge University
Press, 2012.
12. A.T. Mousavian, M. FiŠer, J. KoŠecká, A. Wahid, and J. Davidson, “Visual representa-
tions for a semantic target driven navigation,” in 2019 International Conference on
Robotics and Automation (ICRA), pp. 8846–8852. IEEE, 2019.
13. J. Stilgoe, “Machine learning, social learning and the governance of self-driving cars,”
Social Studies of Science 48(1): 25–56, 2018.
14. C.C. Aggarwal, Neural Networks and Deep Learning. Springer International
Publishing, 2018.
15. T. Taketomi, H. Uchiyama, and S. Ikeda, “Visual SLAM algorithms: A survey from
2010 to 2016,” IPSJ Transactions on Computer Vision and Applications 9(1): 16, 2017.
16. L. Xu, Y. Li, Y. Sun, L. Song, and S. Jin, “Leaf instance segmentation and count-
ing based on deep object detection and segmentation networks,” in 2018 Joint 10th
International Conference on Soft Computing and Intelligent Systems (SCIS) and 19th
International Symposium on Advanced Intelligent Systems (ISIS), pp. 180–185. IEEE,
2018.
17. J. Gao, Y. Yang, P. Lin, and D.S. Park, “Computer vision in healthcare applications,”
Journal of Healthcare Engineering, 2018, p. 5157020.
18. D. Wang, A. Khosla, R. Gargeya, H. Irshad, and A.H. Beck, “Deep learning for identi-
fying metastatic breast cancer,” arXiv Preprint ArXiv:1606.05718, 2016.
19. R. Silva, J.C. Oliveira, and G.A. Giraldi, “Introduction to augmented reality,” National
Laboratory for Scientific Computation 11, 2003.
20. O.I. Abiodun, A. Jantan, A.E. Omolara, K.V. Dada, N.A. Mohamed, H. Arshad, K.V.
Dada, N.A. Mohamed, and H. Arshad, “State-of-the-art in artificial neural network
applications: A survey,” Heliyon 4(11): e00938, 2018.
 7 Generative Adversarial
Network
Concepts, Variants,
and Applications
K. Rakesh and V. Uma

CONTENTS
7.1 Introduction................................................................................................... 132
7.2 Overview........................................................................................................ 133
7.2.1 Deep Learning .................................................................................. 133
7.2.2 Deep Generative Models .................................................................. 133
7.2.3 Generative Adversarial Networks ..................................................... 134
7.3 GAN Architecture ........................................................................................ 135
7.3.1 General Structure............................................................................... 135
7.3.2 Adversarial Process........................................................................... 136
7.3.3 Background Mathematics.................................................................. 137
7.4 GAN Variations............................................................................................. 138
7.4.1 Overview............................................................................................ 138
7.4.2 Techniques......................................................................................... 138
7.4.2.1 Architecture-based Variant Class ...................................... 138
7.4.2.2 Formulation-based Variant Class ....................................... 139
7.5 Applications .................................................................................................. 141
7.5.1 Image Generation and Prediction...................................................... 141
7.5.2 Image Translation.............................................................................. 142
7.5.3 Image Editing.................................................................................... 143
7.5.4 3D Object Generation........................................................................ 143
7.5.5 Video Manipulation........................................................................... 143
7.5.6 Audio Generation and Translation..................................................... 144
7.5.7 Medical Image Processing................................................................. 145
7.6 Conclusion and Future Directions................................................................. 145
Bibliography........................................................................................................... 145

131
132 Artificial Intelligence (AI)

7.1 INTRODUCTION
Humans always prefer automation of tasks. The invention of computers and com-
puter programming has made it possible to automate some tasks by creating a series
of instructions, referred to as algorithm. As we approach more and more complex
problems, algorithms start to demand a certain degree of intelligence that is similar
to human cognitive functions. This kind of intelligence that is induced to a machine
preparing them to mimic our intelligence is termed as artificial intelligence (AI). A
subfield of AI that is responsible for most of the recent advances of many fields is
machine learning (ML). ML algorithms are statistical models built by finding pat-
terns and inferences from the data that is provided. One class of ML algorithms that
learns a significant amount of cognitive tasks is deep learning (DL) [1].
Deep learning comprises multiple layers of artificial neural network (ANN), a
computing system that learns low-level to high-level features of the input data. DL is
the technology behind many of the current awesome products and services like voice
assistants, recommendation systems, self-driving cars, medical image analysis, com-
puter vision systems, and many more [2]. The discussion on the basics of deep learning
and deep neural network (DNN) is made in Section 7.2.1. Though we made signifi-
cant progress in recognition and prediction tasks, the creative applications of AI are
demanding more attention and in recent years more research momentum is noticed
in this application. The research works on generative tasks have many applications in
areas like content writing, film industry, game development, advertisement, etc.
The class of deep learning networks that is capable of generating data from the
input is known as deep generative networks [3]. Over the years, many generative
models have been proposed. All these different approaches for generative modelling
have been extensively discussed in Section 7.2.2. Out of these models, generative
adversarial networks (GAN) is the most popular model used in generation-based
research works and applications. It provides state-of-the-art results and there pre-
vails a debate on social and cultural aspects of its applications.
GAN can mimic the distribution of a data set, such that new data of the same
distribution can be generated, which will be similar to the input data [4]. The GAN
and its architecture are detailed in Section 7.3. Their architecture can be changed
accordingly to our needs or use cases. This flexibility has provided a lot of variations
while dealing with the distribution. Even the distribution of two different data sets
has been utilized to produce some interesting results. The overview of the variations
in GAN and the techniques used in these variations are discussed in Section 7.4.
The various research directions on GAN show several types of GAN that can
be used for different architecture and functions. These diverse types of GAN have
made it possible to use them in a variety of applications. Some of the interesting
applications include generating photo-realistic images, converting an object in an
image to another object, changing features of a person, generating missing areas in
an image, image editing, converting a text to image, future scene prediction, creating
3D objects, medical image generation, and many more. Many of these applications
and their impact on the industry have been discussed in Section 7.5. The future
directions of GAN are discussed in Section 7.6.
Generative Adversarial Network 133

7.2 OVERVIEW
7.2.1 Deep Learning
Machine learning algorithms have been sufficient for basic classification tasks like
spam detection. But for solving complex tasks like image classification, a huge num-
ber of parameters have to be understood by the algorithm. This is not possible in
classical machine learning algorithms. Then, there evolved a particular type of ML
algorithm that was modelled after the human brain known as the artificial neural
network. ANN comprises many connection units called neurons. Weight and bias
values are set as parameters to connections that prevail between neurons in one layer
to every other neuron of the adjacent layer. These parameter values are automatically
figured out using algorithms like backpropagation as the network progresses along
with the data samples. ANN shows significant results than classical ML algorithms.
More and more complex problems have been solved by adding more layers and pro-
gressing deep into ANN. These deep structures are termed as deep learning.
Deep learning is capable of learning even from unsupervised data which makes
it learn features automatically without much human effort. Deep learning has pro-
gressed over the years with the rise of graphical processing units (GPUs) and a huge
increase in data and research over different ANN architectures and techniques.
Many standard deep learning architectures have been proposed for various appli-
cations. DL architectures that are mostly applied in industries include deep neural
networks, convolutional neural networks (CNN), recurrent neural networks (RNN),
deep belief networks (DBN), AutoEncoders, and GAN.
Each network performs better in different applications. DNN is used for basic
prediction and classification tasks of structured data like medical records. These data
sets have rows of data and many parameters that are tuned to provide the required
results. CNN performs better at image processing, object recognition, and image
segmentation. RNN processes sequential data like time-series data and paragraphs
of text for prediction of the upcoming data. Autoencoders are good at capturing the
essential data from the input, and it is used in data compression and image process-
ing. DBN is used in image clustering, recognition, and generation. At present, GAN
is used for the generation of images and videos with high accuracy.

7.2.2 Deep Generative Models

Deep generative models work mainly by learning the distribution of the input data
and exploring through the latent space. The model describes a probability density
function (PDF) that represents the distribution of the input data. Mostly the genera-
tive model performs maximum likelihood from the PDF [2]. Each generative model
differs from each other in the implementation of maximum likelihood. The repre-
sentation of the density function can be explicitly defined or implicitly defined in a
generative model with each one having its advantages and disadvantages.
In explicitly defined functions, a tractable PDF is crafted to provide the data dis-
tribution or approximates an intractable PDF. Some explicit PDF generative models
that have tractable PDF include PixelRNN, NADE, MADE, and non-linear IOC.
134 Artificial Intelligence (AI)

Explicit PDF generative models that use approximation are Variational AutoEncoder
(VAE) and Boltzmann machine.
In implicitly defined functions, the density function is not mentioned and the
model will internally figure out a way to draw samples from the same distribution
provided. Some implicit PDF generative models are generative stochastic networks
(GSN) and GAN.
The explicitly defined model requires defining the density function which is hard
to be established. Because the density function should be able to model the whole
distribution, defining such a function is not possible. In the explicitly defined models
that have been mentioned, a tractable density function is modelled or approximation
of a density function is utilized. The tractable density function is a density function
designed in such a way that it can model most of the distribution. But, for a com-
plex distribution like distributions in Nature, the distribution of speech, and natural
image, designing a parametric function that can capture the distribution perfectly
is difficult. The problem is that it cannot handle the vast distribution of some data
sets and this limits the usability of these models. The models that approximate the
density function learn the distribution using two approaches. One by approximating
an intractable density function by placing a lower bound in log-likelihood and maxi-
mizing the bound. Another approach uses the Markov chain to use an estimation of
the density function or its gradient. Both of these approaches include approximation
which does not guarantee to capture the complex distribution. These complications
led to the design of implicit density function which is mostly automatically learned
by the model itself. The GAN is the only implicit density function that shows promis-
ing results.
A lot of research has been done to track and improve these promising models.
Most recent researches show results of GAN to be state-of-the-art in generative mod-
elling. GAN has various applications when compared to other types of generative
models.

7.2.3 Generative Adversarial Networks

GAN was proposed by Ian Goodfellow in 2014 [5]. GAN contains two networks
that compete against each other. One network helps the other to figure out the dis-
tribution of the data. There is no need for defining the PDF for the network. The
architecture automatically figures it out internally which removes one huge prob-
lem. More than 12,000 citations have been made at present on the original research
paper of GAN by Ian Goodfellow. It gained popularity for its robust architecture
and ability to add many different kinds of models replacing the generator or dis-
criminator in the architecture of GAN. Over the years, many different variations of
GAN architecture made up of a combination of other neural network models have
been proposed and used. The main concept in GAN that makes the whole archi-
tecture effective is adversarial training. Facebook’s AI Research Director LeCun
referred to adversarial training as “The most interesting idea in the last 10 years of
ML.” Apart from images and videos, GAN is also used for the generation of text,
music, speech, and 3D objects.
Generative Adversarial Network 135

7.3 GAN ARCHITECTURE
 eneral Structure
7.3.1  G
GAN is a combination of two neural networks: the generator model and the discrimi-
nator model. The architecture of vanilla GAN is represented in Figure 7.1.
The generator model tries to generate data similar to the distribution of training
data set and the discriminator network checks whether the generated data is agree-
ably realistic.
In the vanilla GAN which was first introduced by Ian Goodfellow in 2014 [5], a
high-dimensional input noise is generated randomly and given as input to the gen-
erator. Any distribution can be used to generate these random noise values. The
generator network has randomly assigned weights and biases as parameters. Both
the weight and bias are numerical values that are subjected to mathematical opera-
tions along the process. The generator network processes the input noise with its
parameters and produces output with the required dimension of the output image.
These processes depend on the functions that are used in the neural network. The
generated output will be mostly noisy and does not have any similarity with the
distribution of our training data set. The data set can be an image, video, text, audio,
or even 3D objects. Then, the generated output and the real data are passed into the
discriminator. The discriminator classifies the real and fake images after processing
it with its parameters, randomly assigned weights, and biases. From the output of
the discriminator, the generator loss and discriminator loss are calculated and then
backpropagation is performed. It helps the generator and discriminator to update
their randomly assigned weights.
The generation, discrimination, and backpropagation are repeated many times
with new samples of input data; this is referred to as training of GAN. Throughout

FIGURE 7.1  GAN architecture.

136 Artificial Intelligence (AI)

the training, the generator model gets better at generating more realistic data and the
discriminator improves in classifying generated and real data [6, 7]. Both the genera-
tor model and the discriminator model converge at a point, after which both do not
improve. Now, the discriminator model is discarded and the generator alone is used
for generating the samples. The discriminator model was used as a tool for improv-
ing the generator and not used again in the generation of samples. However, the dis-
criminator is a good classifier and sometimes used for validation of other models [4].

7.3.2 Adversarial Process
The generator and discriminator networks compete against each other to improve
each other in their respective tasks. This adversarial technique was inspired by game
theory, a mathematical modelling study of the interaction between two rational deci-
sion-makers. The interaction is performed based on the zero-sum game, in which the
participant either wins or loses. Here, when the discriminator successfully identifies
the real and fake data, the generator is penalized and vice versa when it fails to iden-
tify. GAN is mainly based on the Nash equilibrium to balance the generator and the
discriminator. Nash equilibrium is a solution state when many players are playing
a game and each knows the other player’s strategies and makes decisions without
changing his or her strategy. The generator and discriminator know the strategies
of each other but make the best decisions by not changing their strategy, making
the models to compete with each other and achieve the Nash equilibrium state and
converge at a point.
Both the generator and discriminator models are modelled as differentiable func-
tions. A value function is defined using the generator and discriminator functions in
the form of a two-player minimax game and then optimized throughout the training.
If the discriminator is completely optimized before starting to optimize the gen-
erator, then the generator will not get better. So, k-steps of discriminator optimiza-
tions are performed before one step of optimizing the generator. Then, the generator
model tries to increase the chances of the discriminator making mistakes.

  Algorithm
   Step 1: Iterate through the number of iterations.
  Step 2: Iterate through the discriminator for a fixed
number of times.
    Step 2a: Sample noise and minibatch of n examples from
given examples.
   Step 2b: Calculate the error and gradients.
   Step 2c: Update the discriminator weights.
   Step 3: Sample minibatch of n noise samples.
  Step 4: The generator is provided with the noise sample
and output is generated.
  Step 5: Output is sent to the discriminator and loss value
is calculated.
  Step 5: The weights of the Generator are updated using the
loss value from Discriminator.
Generative Adversarial Network 137

FIGURE 7.2  Divergence of distribution.

7.3.3 Background Mathematics
Every data set has a distribution of data. Every model tries to understand the distri-
bution of input data and captures relevant information from the distribution using
probability estimation. The divergence of distribution between two data sets is pro-
vided in Figure 7.2. The generative models capture joint probability and discrimina-
tive models capture conditional probability. The generator and discriminator models
are mathematically differentiable functions G and D, respectively. A value function
is defined using these differentiable functions [5].

min G max D V ( D, G )

( ( ))
V ( D, G ) = E x ~ pdata ( x ) éë log D ( x ) ùû + Ez ~ pz ( z ) é log 1 - D G ( z ) ù
ë û
The whole training process is designed in such a way that D tries to maximize the
probability of labelling the generated and input data correctly. Log D(x) is the loga-
rithmic probability of discriminator predicting the real data label. The generator
( ( ))
tries to minimize log 1- D G ( z ) , the log probability of discriminator predicting
generated images label correctly. E denotes the expectation of a random variable.
In Equation 7.1, there are two parts. The first part handles the real data sample
and the latter part after the plus symbol deals with the generated samples. But, prac-
tically, the equation may not fare well as the discriminator can reject generated sam-
ples easily with high confidence. To change this in the equation, the generator
function is changed. Rather than decreasing the log probability of discriminator
making correct predictions, the generator increases the log-probability of discrimi-
( )
nator making mistakes é log D G ( z ) ù . The value function is optimized by using
ë û
backpropagation. This eliminates the use of a Markov chain which adds a huge
advantage over other generative models.
The value function is based on Jensen–Shannon divergence (JSD), which mea-
sures the similarity between two distributions. JSD is based on Kullback–Leibler
138 Artificial Intelligence (AI)

divergence (KLD). KLD measures the difference of one distribution to another one.
The difference in the distance between the two distributions can be estimated by
KLD. Minimizing KLD can minimize the distance between the distributions. Log-
likelihood is the estimator that helps to minimize KLD.
The GAN internally finds the density function (PDF) of the input data distribu-
tion. A sample from the density function provides a relative likelihood. Likelihood
expresses different plausible parametric values of the sample. The generator of GAN
tries to model data that have a similar distribution of input data using the density
function. At first, the generated distribution will be different from the real input dis-
tribution. This difference is then reduced with minimizing KLD by maximizing the
log-likelihood of the distributions.

7.4 GAN VARIATIONS
7.4.1 Overview
The vanilla GAN invented in 2014 by Ian Goodfellow was popular mainly because
of the adversarial process. Many different variations of GAN which improve its
functionality, accuracy, and applications are proposed over the years. The varia-
tions may possess changes in architecture or formulations. Every variation provides
a variety of applications for GAN. The number of GAN variant papers proposed
has exponentially increased each year. The architectural variation includes changes
by including concepts and components of other deep learning architectures. These
changes can be an increase in the number of generators or discriminators, the way
they are connected, and the introduction of other deep learning components in the
discriminator or generator. The variations are mainly classified into architecture-
based variants and formulation-based variants [8, 9].

7.4.2 Techniques
7.4.2.1 Architecture-based Variant Class
The architectural variant class has two subclasses: condition-based variants and
structure-based variants.

I. Condition-based Variants
Condition-based variants involve conditional input that changes the
output according to it. This class of variants started with the condi-
tional GAN (CGAN) proposed in 2014 [10]. It is based on vanilla GAN
with additional conditional input to generator and discriminator. This
extracts features based on the modelling of the conditional input. The
InfoGAN proposed in 2016 provides the condition as latent information
to the generator. Auxiliary Classifier GAN (ACGAN) is an extension
of CGAN that adds conditional input only to the generator. All these
GANs include modifying the objective function according to the condi-
tion structure.
Generative Adversarial Network 139

II. Structure-based Variants

The structure-based variants include changes in the architecture of GAN.
The architectural change includes an increase in generators or discrimina-
tors, the addition of convolution layer (network), or encoder–decoder.
a. Increasing Components
The generators are increased in multiple generator GAN (MGAN)
and discriminators are increased in multiple discriminator GAN
(MD-GAN).
b. Convolutional Components
The addition of the convolutional network to the GAN architec-
ture was first proposed in deep convolutional GAN (DCGAN) in 2015
[11]. Based on DCGAN, the Laplacian Pyramid GAN (LAPGAN) was
introduced which has a series of DCGAN that processes images from
coarse to fine. The 3D convolution was utilized in 3D-GAN for 3D
object generation.
c. Encoder–Decoder Components
The encoder–decoder-based GAN variant includes an encoder,
a decoder, or an auto-encoder network in the GAN architecture.
Adversarial AutoEncoders (AAE) in 2015 was the first paper to com-
bine encoders with GAN. In the same year, variational AutoEncoder
was also combined with GAN. Then, Adversarial Learning and
Inference (ALI) combined a generator and inference model, which
is monitored by a discriminatory model. The energy-based GAN
(EBGAN) which uses a different energy-based optimization function
also utilizes the encoder–decoder network in the discriminator. Later,
many architectures, including Adversarial Generator-Encode (AGE),
Boundary Equilibrium GAN (BEGAN), Bidirectional GAN (BIGAN),
and Decoder-Encoder GAN (DEGAN) were designed by combining
encoder–decoder networks in different strategies [8].
d. U-NET Components
A convolutional network called U-NET which is in the form of
encoder–decoder style with connections between the encoder and
decoder layers (skip connections) has been used extensively in recent
works. The GAN architectures like progressive GAN (ProgGAN) and
Pixel-to-Pixel GAN (Pix2Pix) [12] provided profound results by utiliz-
ing U-NET in their models. ProgGAN uses U-NET in the generator and
decoder network in discriminator. ProgGAN gradually increases the
decoder in the generator model and discriminator model to improve the
image quality of the generated image. Pix2Pix uses U-NET as generator
and PatchGAN, a series of deconvolutional layers, as the discriminator.

7.4.2.2 Formulation-based Variant Class

The formulation-based variant class includes GAN architectures that have modified
objective function, normalization method, regularization method, and loss function.
The variant typically tries to solve a problem that GAN faces by mathematically
140 Artificial Intelligence (AI)

modifying the formulation. The original GAN was unstable and hard to train. It had
many issues in training and convergence. Many proposed variants use these alterna-
tive objective functions to solve these issues.

I. Divergence-based Modification
Original GAN used JSD for the objective function. f-GAN and
Wasserstein GAN (WGAN) improved stability of GAN training by chang-
ing JSD to other divergence formulations. f-GAN utilizes f-divergence for
the objective function. WGAN formulated the objective function using
Wasserstein distance. WGAN claims to be more stable than GAN, uses
JSD, and also performs weight clipping to avoid mode collapse, a condition
where data generated by the generator are not diverse or even produces the
same sample for different inputs.
II. Loss Function-based Modifications
In some GANs, loss function is tuned to achieve desired results. Least-
square GAN (LSGAN) uses the least-square loss function and f-GAN uses
gradient-based loss function. Discover cross-domain relationship GAN
(DiscoGAN) [13] attempts to automatically discover relationships and fea-
tures between different domains. This is achieved by modifying the objec-
tive function with reconstruction loss. Cycle consistency GAN (CycleGAN)
[14] introduces an alternative loss function, named cycle-consistency loss.
It helps CycleGAN to minimize the distribution of data between two
domains. StarGAN aims at the multi-domain image-to-image translation.
It uses domain classification loss along with the reconstruction loss.
III. Technique-based Modifications
Another technique to optimize the objective function was performed
in energy-based GAN (EBGAN), where discriminator is formulated as an
energy function. The mode collapse problem was also handled in unrolled
GAN (UGAN) by using unrolled optimization in the objective function.
IV. Regularization
GAN has to converge properly to achieve good results. The generator
and discriminator model training has to be stable to achieve convergence.
The unstable training process could lead to mode collapse. The regulariza-
tion technique is a trick that could lead to the stabilization of the training
process. Many different kinds of regularization are used to solve this prob-
lem. Regularization techniques like mode regularization, Lipschitz regu-
larization, Jacobian regularization, and orthogonal regularization are used
in variants like MRGAN, LSGAN, JRGAN, and BigGAN, respectively
[15]. WGAN-LP uses a penalty function that uses Lipschitz constraint.
V. Normalization
Normalization is another trick performed to scale the values in GAN. It
also aids in the stabilization of GAN training and convergence. The nor-
malization techniques like gradient penalty and spectral normalization are
widely used [15]. The gradient penalty is used in WGAN-GP, DRAGAN,
and BWGAN. WGAN-GP uses the gradient penalty in WGAN. Along with
Generative Adversarial Network 141

the gradient penalty, DRAGAN uses deep regret analytics and Banach
Wasserstein GAN (BWGAN) uses dual normalization in Banach space.
Spectral Normalization is used in Spectral normalization GAN (SNGAN)
and Self-attention GAN (SAGAN) [16]. The attention method is also used
along with spectral normalization in SAGAN.

7.5 APPLICATIONS
7.5.1 Image Generation and Prediction
The creation of visual content remains challenging for humans. The generation of
new images can be useful in many areas where more visual data is needed. More
data can provide more possibilities and opportunities. In the case of DL, huge data
sets are required to train models. These data can be hard to collect and pre-pro-
cess. Data augmentation is a process done to increase the data set by performing
transformations like cropping, scaling, padding, and flipping to the existing images.
The augmented images are extremely useful in creating machine learning models
for medical and geological image classification, where collecting medical images
is hard, and augmented images serve the purpose of training these models. These
problems can be tackled by GAN. They have the potential to create many images
similar to the input image.
All the GAN variants are capable of data augmentation, depending on the type
of data they process. DCGAN [11] and StyleGAN [17] are famous for the genera-
tion of human faces from scratch. Figure 7.3 provides an example of the image
generation. BigGAN [18] is capable of generating photo-realistic images. Images
of objects are also generated which can be used in websites, clothing lines, media,
and many more areas.

FIGURE 7.3  Image generation.

142 Artificial Intelligence (AI)

7.5.2 Image Translation
Image translation is the process of converting an image into another modified image
or other forms of data like text, audio, etc. Image-to-image translation leads to many
fascinating applications. Image translation includes the conversion of objects from
one domain to objects of another domain. For example, images of a horse can be con-
verted into images of zebra and vice versa. Many neural networks like CycleGAN
[14], Pix2Pix [12], and BigGAN [18] have demonstrated this using various data sets.
The conversion between day and night images, seasons, colourization of black and
white images, and many more have been demonstrated using Pix2Pix. Even alter-
ing human features like hair colour and facial expression has been translated using
StyleGAN [17] and DiscoGAN [13]. An example of converting label image from
CMP facades data set [19] to real-world buildings using Pix2Pix [12] is provided in
Figure 7.4. Even the conversion of faces into cartoon or anime characters has been
demonstrated. Image super-resolution is another application where GAN shines.
Image super-resolution is the problem where fewer dimension images are con-
verted into images with higher resolution. Variants like Super Resolution GAN
(SRGAN) [20] and Enhanced Super Resolution GAN (ESRGAN) [21] have achieved
remarkable results in image super-resolution tasks. Style transfer is another problem
where the style or feature of one image is transferred to the input image. The image
which has to be referred for the features or style is processed as conditional input
and the main input is the image where style has to be applied. Many variants under
conditional GAN class are capable of solving this problem with good accuracy.
Image translation into textual data has been achieved in DCGAN, StackGAN
[22], AttnGAN, GAN-INT-CLS, and OP-GAN. Image inpainting can also be per-
formed in GAN. In some images, certain parts of the pictures are missing or hid-
den by overlapping objects. Image inpainting is a process where these images are
translated into images that are filled with accurate details in missing regions. Image
inpainting can also be used to remove watermarks in images.

FIGURE 7.4  Image translation using Pix2Pix.

Generative Adversarial Network 143

FIGURE 7.5  Image editing using GAN.

7.5.3 Image Editing
Image editing is the technique of manually providing input to the model that alters
the image accordingly. This technique enables us to paint objects and scenes like
trees, architectural structures, sea, and sky into the image. The variants like IGAN,
GP-GAN, and GauGAN are capable of image editing. This is under development
as the technique has a long way to follow, even though the results seem stunning
and natural in some cases. Unlike other models, GauGAN [23] has achieved some
remarkable quality in images. Figure 7.5 describes a made-up example of image
editing using GAN.

7.5.4 3D Object Generation

The virtual environment has been extensively created and experienced by humans.
These environments are used in gaming industries, virtual reality, and many more
areas. Virtual environments are filled with 3D objects that are hard to create, involv-
ing a lot of man-hours as well. Altering and recreating 3D objects will still delay
such projects on a large scale.
Automating the creation of 3D objects and environments can increase the devel-
opment process on a large scale. The automated 3D object creation can also pave the
way for real-time object creation and simulations. 3D GAN [24] and 3D Scene-GAN
[25] are two prominent 3D object generation architectures. 3D GAN can generate
high-quality 3D objects with sample images and objects. 3D Scene-GAN tries to
create the whole environment with the sample images. These variants laid the foun-
dation for the whole application in these areas. There are still more fascinating appli-
cations in these areas that are yet to be discovered.

7.5.5 Video Manipulation
Video generation has been an interesting area of work for humans. Creative video
content creation is an important profession and many industries have been built upon
this platform. Video content has to be created by using video capture devices or
graphically creating each scene through software. In each case, the creator needs
144 Artificial Intelligence (AI)

plenty of time and resources to complete the task. The generation and manipulation
of video content using GAN can make it easy for the creators.
Video is nothing but a sequence of pictures, the same as images. Video can be
created, manipulated, and predicted. Generation of video can be done using the
GAN models used for image generation and style transfer. Video manipulation has
created a great discussion in society, as there occurs a possibility of changing faces
and expressions in video content. Face swapping GAN (FSGAN) [26] is a recent
addition to the class of GANs that can manipulate videos. Few research papers have
proposed works on video prediction. Video prediction guesses the next frame and it
can evolve machines into a better content creator.

7.5.6 Audio Generation and Translation

GANs are capable of generating audio with ease like other forms of data. The gen-
eration of audio initiates other operations with audio content. The adversarial models
like MidiNet, WaveGAN [27], and MuseGAN [28] have received a good response
from the research community. A representation of using GAN to generate music is
provided in Figure 7.6. These models have guided audio manipulation and transla-
tion models. FusionGAN [29] can fuse different genres of music and create quality
music. GANs can also perform voice impersonation of one input over the other.
The translation of audio to text and vice versa has been performed in GAN-TTS.

FIGURE 7.6  Audio generation.

Generative Adversarial Network 145

The conversion of audio content to the image has been demonstrated in Wav2Pix
architecture.

7.5.7 Medical Image Processing

Image prediction is used in medical fields. The anomalies in medical reports are
hard to detect and need medical expertise. The introduction of GAN can greatly help
professionals automate prediction tasks. AnoGAN [30] is one such GAN that can
predict tumours in medical images. Generating medical images also have been dem-
onstrated in MedGAN [31]. The PET scans have been translated into CT scans using
MedGAN. Medical reports like X-rays, CT scans, and MRI scans have been trans-
lated into clear, understandable images. Motion correction and positron-emission
tomography (PET) have also been performed in MedGAN.

7.6 CONCLUSION AND FUTURE DIRECTIONS

Generative adversarial network has great potential and a lot of applications. They
can be used in nearly every possible field and production sector. The rapid research
developments had provided hope for developing a lot of creative applications using
GAN. Each variation of GAN provides new features. These variations are envel-
oping various kinds of neural networks and techniques that help gather complex
features and generate more relatable and realistic images as discussed. The vast col-
lection of GAN implies the variety of its applications. The impact of GAN had been
on every aspect of creative tasks. Every form of natural and synthetic data has been
manipulated using GAN with proper research work. These works on GAN take us
closer and closer to the possibility of creating automated machines that are creative.
There has been a vast production area eagerly waiting to utilize these ideas and
research works. GANs would be widely used in medical fields to predict diseases and
abnormalities. They can help generate augmented images, which in turn improves
the quality of image classification. They would also be used in creating content for
media and industries. The graphical rendering machines and game development
teams can utilize GANs to create and modify 2D images and 3D objects on the fly.
The 3D object generation quality of GAN is not up to mark at present. If the quality
increases down the line, these rendering techniques with GAN can rapidly minimize
the time taken for developing games. In the near future, many tasks that are consid-
ered impossible today can be demonstrated using GAN.

BIBLIOGRAPHY
1. Francois, Chollet, Deep Learning with Python Manning Publications, 2017.
2. Goodfellow, Ian, Yoshua Bengio, and Aaron Courville, Deep Learning. MIT Press,
2016.
3. Nielsen, Michael A., Neural Networks and Deep Learning, Vol. 25. Determination
Press, 2015.
4. Langr, Jakub, and Vladimir Bok, GANs in Action Manning Publications, 2018.
146 Artificial Intelligence (AI)

5. Goodfellow, Ian, et al., “Generative adversarial nets,” in Advances in Neural

Information Processing Systems, Curran Associates, 2672–2680, 2014.
6. Wang, Kunfeng, et al., “Generative adversarial networks: Introduction and outlook,”
IEEE/CAA Journal of Automatica Sinica, IEEE, 4(4): 588–598, 2017.
7. Creswell, Antonia, et al., “Generative adversarial networks: An overview,” IEEE Signal
Processing Magazine, IEEE, 35(1): 53–65, 2018.
8. Pan, Zhaoqing, et al., “Recent progress on generative adversarial networks (GANs): A
survey,” IEEE Access, IEEE, 7: 36322–36333, 2019.
9. Wang, Zhengwei, Qi She, and Tomas E. Ward, “Generative adversarial networks: A
survey and taxonomy,” arXiv Preprint ArXiv:1906.01529, 2019.
10. Mirza, Mehdi, and Simon Osindero, “Conditional generative adversarial nets,” arXiv
Preprint ArXiv:1411.1784, 2014.
11. Radford, Alec, Luke Metz, and Soumith Chintala, “Unsupervised representation
learning with deep convolutional generative adversarial networks,” arXiv Preprint
ArXiv:1511.06434, 2015.
12. Isola, Phillip, et al., “Image-to-image translation with conditional adversarial net-
works,” in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, IEEE, 1125–1134, 2017.
13. Kim, Taeksoo, et al., “Learning to discover cross-domain relations with genera-
tive adversarial networks,” in Proceedings of the 34th International Conference on
Machine Learning, Volume 70, JMLR, 1857–1865, 2017.
14. Zhu, Jun-Yan, et al., “Unpaired image-to-image translation using cycle-consistent
adversarial networks,” in Proceedings of the IEEE International Conference on
Computer Vision, IEEE 2223–2232, 2017.
15. Kurach, Karol, et al., “A large-scale study on regularization and normalization in
GANs,” in Proceedings of the 36th International Conference on Machine Learning,
JMLR, Volume 97, 3581–3590, 2019.
16. Zhang, Han, et al., “Self-attention generative adversarial networks,” arXiv Preprint
ArXiv:1805.08318, 2018.
17. Karras, Tero, Samuli Laine, and Timo Aila, “A style-based generator architecture for
generative adversarial networks,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, IEEE, 4401–4410,2019.
18. Brock, Andrew, Jeff Donahue, and Karen Simonyan, “Large scale gan training for high
fidelity natural image synthesis,” arXiv Preprint ArXiv:1809.11096, 2018.
19. Tylecek, Radim, The cmp facade database. Tech. rep., CTU–CMP–2012–24, Czech
Technical University, 8 p, 2012.
20. Ledig, Christian, et al., “Photo-realistic single image super-resolution using a genera-
tive adversarial network,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, Elsevier, 4681–4690, 2017.
21. Wang, Xintao, et al., “Esrgan: Enhanced super-resolution generative adversarial net-
works,” in Proceedings of the European Conference on Computer Vision (ECCV),
Springer, Workshops, 0–0, 2018.
22. Zhang, Han, et al., “Stackgan: Text to photo-realistic image synthesis with stacked gen-
erative adversarial networks,” in Proceedings of the IEEE International Conference on
Computer Vision, IEEE, 5907–5915, 2017.
23. Park, Taesung, et al., “Semantic image synthesis with spatially-adaptive normalization,”
in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
IEEE, 2337–2346, 2019.
24. Wu, Jiajun, et al., “Learning a probabilistic latent space of object shapes via 3d gener-
ative-adversarial modeling,” in Advances in Neural Information Processing Systems,
Curran Associates, 82–90, 2016.
Generative Adversarial Network 147

25. Yu, Chong, and Young Wang, “3D-scene-GAN: Three-dimensional scene reconstruc-
tion with generative adversarial networks,” in ICLR Workshop, ICLR, 2018.
26. Nirkin, Yuval, Yosi Keller, and Tal Hassner, “Fsgan: Subject agnostic face swapping
and reenactment,” in Proceedings of the IEEE International Conference on Computer
Vision, IEEE, 7184–7193, 2019.
27. Donahue, Chris, Julian McAuley, and Miller Puckette, “Adversarial audio synthesis,”
arXiv Preprint ArXiv:1802.04208, 2018.
28. Dong, Hao-Wen, et al., “Musegan: Multi-track sequential generative adversarial
networks for symbolic music generation and accompaniment,” in arXiv preprint
arXiv:1709.06298, 2018.
29. Ma, Jiayi, et al., “FusionGAN: A generative adversarial network for infrared and visible
image fusion,” Information Fusion, Elsevier, 48: 11–26, 2019.
30. Schlegl, Thomas, et al., “Unsupervised anomaly detection with generative adversarial
networks to guide marker discovery,” in International Conference on Information
Processing in Medical Imaging. Springer, 146–157, 2017.
31. Armanious, Karim, et al., “Med GAN: Medical image translation using GANs,”
Clinical Orthopaedics and Related Research (CoRR), Elsevier, 79, p.101684, 2018.
 8 Detection and
Classification of
Power Quality
Disturbances in Smart
Grids Using Artificial
Intelligence Methods
Gökay Bayrak and Alper Yılmaz

CONTENTS
8.1 Introduction................................................................................................... 149
8.1.1 Signal Processing (SP)-based PQD Detection Methods.................... 151
8.1.2 Artificial Intelligent (AI) Methods for PQD Detection..................... 151
8.2 Wavelet Transform (WT)-based PQD Detection Methods........................... 153
8.2.1 Wavelet Transform (WT)................................................................... 154
8.2.2 Proposed DWT-based PQD Detection Method................................. 155
8.3 AI-based PQD Classification Methods.......................................................... 155
8.3.1 Deep Learning Structures................................................................. 158
8.3.1.1 SAE-based Methods .......................................................... 158
8.3.1.2 DNN-based Methods ......................................................... 160
8.3.1.3 DBN (Deep belief network)-based Methods...................... 160
8.3.1.4 CNN-based Methods.......................................................... 160
8.3.2 Proposed Deep Learning and WT-based Hybrid PQD
Classification Method........................................................................ 161
8.4 Results............................................................................................................ 165
8.5 Conclusion..................................................................................................... 167
References............................................................................................................... 168

8.1 INTRODUCTION
Integration of DG units to the grid improves the distribution system performance,
reliability, stability, security, power quality (PQ), and voltage profile. Also, DG-based
microgrid should have fewer losses, high-degree voltage support, and a reliable
power quality disturbance (PQD) detection system [1]. The interconnection of DG

149
150 Artificial Intelligence (AI)

with a grid may reduce the cost, losses, complexity, a load of lines, interdependence,
and inefficiencies associated with low-carbon emission in microgrids [2].
The DG-based power plants must adapt the defined electrical standards to provide
reliable grid operation and grid stability in cases of any fault or disturbances at the
point of common coupling (PCC). The PQDs [3], islanding detection and protection,
and electromagnetic interference are critical issues to adapt DG systems to the grid.
According to Ref. [4], this fault and disturbance conditions costs in the European
Union (EU) are predicted to be high.
DGs inclusion in the system has a lot of positive effects, but it creates technical
difficulties facing the integration of DG. These difficulties are voltage regulation,
increased distribution losses, harmonic control and harmonic injection problem,
islanding condition control, and the sensitivity of existing protection schemes.
The conventional power system is designed to have a unidirectional flow of power
from the high-voltage to the low-voltage side. Today, power distribution systems
with bidirectional power flows have been created due to grid-integrated DG systems,
which have increased in numbers rapidly. As a result, classical analysis, operation,
and design methods cannot manage this complex grid structure. DG integration into
the main grid is an essential issue if considering this aspect. The DG systems are
expected not to operate out of the defined threshold values of the related standards’
voltage and frequency. In a microgrid, unintentional islanding conditions cause sev-
eral damages for both the system operators and the DG unit side. An incorrect detec-
tion in islanding conditions with conventional methods also leads to switching of
DGs and economic losses. The voltage, frequency, and current balance of DG sys-
tems operating with the grid are also essential parameters for microgrids.
Figure 8.1 shows the general structure of a DG system. There are some technical
electrical criteria to switch DG systems from the grid, and these conditions must be
considered to provide the sustainable operation, reliability, and high-power quality
of the grid. IEEE 929-2000 defines some technical electrical rules for microgrids to
provide a reliable power system [5].

FIGURE 8.1  A general schematic for DG systems.

Power Quality Disturbances in Smart Grids 151

PQ parameters are limit values that allow an electrical device to operate as

intended without any significant loss of performance and life expectancy. Any power
system problem that leads to voltage, current, and frequency deviations can cause
malfunction of hardware on the consumer side. Voltage sag/swell and interruption,
current/voltage harmonics, transients, notches, spikes, flicker, and voltage imbal-
ances are among the main factors affecting PQ in systems. Also, abnormal operating
conditions of the grid affect the DG system, so the DG system has to be out of island-
ing condition, which is a security problem in a DG system. Several international
standards have been established to regulate PQ. IEEE Std. 1547 [6] and IEEE Std.
519 [7] determine harmonic voltage and current harmonic limits. IEEE Std. 1159 [8]
also offers some suggestions for monitoring PQDs in DG systems.

8.1.1 Signal Processing (SP)-based PQD Detection Methods

Many methods have been proposed to detect and classify PQDs in microgrids in the
last decade [9]. Today, PQ analysers are used to detect short-term and long-term voltage
disturbances using effective value conversion (true RMS method). These analysers per-
formed the harmonic analysis of a signal using the fast Fourier transform (FFT) method.
These methods are easily applicable but have many drawbacks. These disadvantages are
difficulties in the selection of threshold values, noisy measurements, incorrect detection
tendencies, etc. Thus, intelligent methods are more reliable to detect PQDs [10]. Signal
processing-based methods use both time and frequency components of a signal. Thus,
this characteristic is a useful property for detecting disturbances compared to conven-
tional methods. SP-based methods used in fault detection can also be used to extract the
features to be given to the classifier input [10]. Fourier transform (FT), wavelet transform
(WT), Hilbert–Huang transform (HHT), s-transform, curvelet transform, Kalman filter,
Gabor transform, etc. constitute SP-based methods for detecting PQDs. Figure 8.2 shows
a classification of SP-based PQD detection methods used in smart grids.
PQD detection is proposed in Ref. [11], and this method used Dbn wavelets to
detect events accurately. The discrete wavelet transform (DWT)-based method is
investigated for wind turbines in Ref. [12], and UWT decomposition coefficients are
used to obtain detailed coefficients. The WT method was proposed to detect fault
conditions for the grid-connected PV-based microgrid investigated in Ref. [13]. WT
method is also used to identify PQDs in a hydrogen energy-based microgrid in Ref.
[14]. Threshold values were used in Ref. [15] for a DG system by determining har-
monics, entropy, and energy by DWT decomposition. In another study, three-level
decomposition is performed in a DG system using WT to the classification of PQDs
[16]. The un-decimated WT (UWT)-based disturbance detection method was sug-
gested for microgrids in Ref. [17] for the first time. This method was tested for real-
time applications, and it overcame the limitations of WT-based methods.

8.1.2 Artificial Intelligent (AI) Methods for PQD Detection

Various artificial intelligence (AI)-based algorithms have been used in the literature
to classify any faults and PQDs in microgrid applications. Some of the AI-based
152 Artificial Intelligence (AI)

FIGURE 8.2  SP-based methods using in smart grids.

methods for detection of PQDs are fuzzy expert system-based classifiers, neural net-
work (NN)-based classifiers, decision tree (DT) and random forest (RF) classifiers,
support vector machine (SVM)-based classifiers, and deep learning (DL)-based clas-
sifiers. Figure 8.3 shows a classification of AI-based PQD detection methods used
in smart grids. The classifier is an algorithm that takes raw data or extracted futures
data as an input and makes a decision about the normal operating state of the system.
Artificial neural network (ANN) has been used for fault detection in PV systems and
wind turbines in Refs. [18, 19]. A SVM is a classification method that is used suc-
cessfully in many pattern recognition problems. It is also originally a proposed and
defined classifier to distinguish two classes from each other [20]. A multiclass SVM
method that performs multiple PQDs is proposed in Ref. [21]. A DT-based algorithm
developed for the classification of complex PQDs is proposed in Ref. [22]. Adaptive
probabilistic neural networks (APNN) and probabilistic neural networks used the
classification of PQDs [23, 24].
Complex learning networks gradually replace the few layers used in regression
and classification problems with developing information technologies. DL algorithms
can be considered as multilayers of ANNs. In the study where the DL algorithm was
used for the first time in islanding detection is Ref. [25]. In another study [26], a new
approach based on single spectrum analysis, curvelet transform, and deep convolu-
tional neural network (DCNN) has been performed to realize the classification of
PQDs. A novel DL method proposed using raw data for the classification of PQD
Power Quality Disturbances in Smart Grids 153

FIGURE 8.3  AI-based classification methods in smart grids [9].

[27]. This method recognizes automatic feature extraction and selection using any
filter or signal processing method. The proposed algorithm is compared to other DL
network structures – long short-term memory (LSTM) network, ResNet50, stacked
auto-encoder (SAE), and gated recurrent unit network – in this study. Results show
that the proposed DCNN-based method has higher accuracy and less training cost.
In Refs. [28, 29], 1D time series PQD data converted 2D image data and then 2D
convolutional neural network (CNN) performed the classification of PQDs. The data
type for this method is not suitable because PQDs are 1D time series signals. While
there is a significant correlation between the pixels on the horizontal and vertical
axes in the image data, there is only one direction correlation in the PQD signals.
Results are compared to other AI-based networks but noise independence, training
time, and model size of the proposed network are unspecified.

8.2 WAVELET TRANSFORM (WT)-BASED

PQD DETECTION METHODS
Fourier transform (FT) methods are the most commonly used SP-based methods
founded by J. Fourier. PQDs generally have stationary and non-stationary signal char-
acteristics [30]. These methods do not allow the local analysis of frequency components,
so they cannot use non-stationary signals. This problem is solved by the windowing
154 Artificial Intelligence (AI)

process using a short-time Fourier transform (STFT) [31]. STFT has a fixed window
size, and it does not contain decomposed low- and high-frequency components simul-
taneously [32, 33]. WT-based methods have flexible time-frequency representation,
locally decompose discontinuities, and sudden changes in signal in high-grade deriva-
tives where other signal processing methods fail to detect PQDs.

8.2.1 Wavelet Transform (WT)

WT, which is often used in pattern recognition applications, can decompose sig-
nals at different scales and resolutions. WT-based methods used in DG systems can
be examined under four headings as continuous WT (CWT), discrete WT (DWT),
wavelet packet transform (WPT), and undecimated WT (UWT). WT uses the mother
wavelet (ψ(x)) function as a scalable window concept instead of the windowing con-
cept used in the STFT. The CWT is using the parameter b which is the shifting factor,
and the scaling factor is s. The CW function (y s,b ) is expressed as in Equation 8.1.
CWT is determined as in Equation 8.2 by using the wavelet function [34]:

1 æ t -b ö
y s ,b = yç ÷ (8.1)
s è s ø

¥
1 æ t -b ö
CWT ( s, b ) =
s ò f ( t )y çè
-¥
s ÷ø
dt (8.2)

CWT method is used for limited real-time processes because of intensive computa-
tion. DWT is designed by using filter banks to remove the disadvantages caused by
the computational load in CWT. DWT is performed to signal as in Equation 8.4,
using g j,k(n) main wavelet function (Equation 8.3):

DWT ( j, k ) = ååS ( n ) g
nÎZ kÎZ
j, k (n), g j , k Î Z , j Î N , k Î Z (8.3)

-j
(
g j ,k ( n ) = a0 2 g a0 j n - kb0 (8.4) )
The relation between main wavelet function and S(n) can be expressed as in Equation 8.5:

S (n) = ååd
nÎZ kÎZ
j ,k g j ,k ( n ) (8.5)

The DWT applied signal is observed approximation coefficients (cAn) and detail
coefficients (cDn). Equation 8.6 gives the formula of the wavelet coefficients:

cA éë n ùû = åS ( k ) g éë2n - k ùû ;
k
cD éë n ùû = åS ( k ) h éë2n - k ùû (8.6)
k
Power Quality Disturbances in Smart Grids 155

The UWT method uses filter banks like DWT, but the coefficient lengths are the
same as the original signal [17]. Detail and approximation coefficients are upsam-
pled using multi-resolution analysis (MRA) by the “á trous” algorithm in the UWT
method [35]. UWT can reduce oscillations and noise, and it is useful for selecting
threshold values rather than DWT [36]. UWT detail and approach coefficients can
be calculated by using Equation 8.7.

( ) åc
cUj éë k ùû = h0( ) * cUj -1 éë k ùû =
j U
j -1 é k + 2 j n]h0 ( k ) ù
ë û
n

( ) åc
wUj éë k ùû = h0( ) * cUj -1 éë k ùû =
j U
j -1 é k + 2 j n]h1 ( k ) ù (8.7)
ë û
n

8.2.2 Proposed DWT-based PQD Detection Method

In this part of chapter, a DWT-based method using threshold values of WTs to detect
PQDs is proposed for smart grids. The developed MATLAB/Simulink model of
a developed electrical power distribution system is shown in Figure 8.4 [30]. This
system consists of a grid-connected PV plant, transformers, inductive and resistive
loads, an induction motor, a capacitor bank, and three-phase non-linear loads. Table
8.1 shows the generated PQD signals with a developed model following IEEE stan-
dards [5,6]. Single-phase information is used for PQD events in the proposed model
indicated in Figure 8.4.
The flowchart of the proposed DWT-based PQD detection method is shown in
Figure 8.5. Firstly, the voltage signal of a developed grid-connected PV system is
acquired at PCC. The two-level decomposition process is performed using the high-
pass and low-pass filters with daubachies4 (db4) mother wavelet, which reduces the
number of samples at each filter output. The selected db4 type contains a minimum
sample and is a short wavelet type. Thus, it reduces the computational load. In the
last stage, the algorithm decides the nominal condition or PQDs according to the
determined threshold value. Different experimental studies have been conducted.
The most appropriate parameters have been identified by selecting the type of wave-
let, sampling frequency, wavelet level, and threshold value to be used in the proposed
DWT-based method for PQD detection. PQD is detected when the specified coef-
ficients of the d2 threshold value are exceeded. After the PQD detection, the event
duration time and the overvoltage/undervoltage condition amplitude are determined
WT coefficients [17]. The PQD conditions are detected by the proposed method cor-
rectly and in a short time. The voltage signal, detail coefficients, and approximation
coefficients for PQD signals are shown in Figure 8.6.

8.3 AI-BASED PQD CLASSIFICATION METHODS

Fuzzy expert system-based classifiers, neural network (NN)-based classifiers,
decision tree (DT) and random forest (RF) classifiers, SVM-based classifiers, and
DL-based classifiers are AI-based methods for detecting PQDs.
 156

FIGURE 8.4  MATLAB/Simulink model of developed electrical power distribution system [30].
Artificial Intelligence (AI)
Power Quality Disturbances in Smart Grids 157

TABLE 8.1
PQD Events Obtained from Developed Model
PQD Event Causes of Event
Voltage sag Short-circuit faults, switching on a large load
Voltage swell Short-circuit faults, switching of large capacitors
Voltage interruption Short-circuit faults, component failures, CB tripping
Harmonics Non-linear power electronic-based drivers, arc welders
Oscillatory transient Switching of large capacitor bank and inductive loads
Notch Non-linear power electronic-based drivers

In the training phase with ANN, PQD signals or extracted futures are given as
input to the system, and the weight coefficients of each node in the model are deter-
mined. In the test phase, the unknown signal is provided as an input to the trained
ANN system, and the class of test signal is decided according to the value calculated
in the output node. As a result of each output layer, the probability of input signal
belonging to that class is obtained.
SVM classifier, which is a classification method used successfully in many pat-
tern recognition problems, was originally a proposed and defined classifier to distin-
guish the two classes. Multiple classifications are performed with the help of kernel
function. The purpose of kernel functions is to move the features that cannot be
separated linearly in the existing space into a higher dimensional space and to make
the features linearly separable in this new high dimensional space. Decision tree
(DT) is a supervised learning method, and it has tree-shaped architecture. DT estab-
lishes the relationship between the class and the attributes. Random forest (RF) is
a further supervised learning method, and it has multiple tree-shaped architectures.
RF classifier improves the generalization performance of DT.

FIGURE 8.5  Proposed DWT-based PQD detection method.

158 Artificial Intelligence (AI)

FIGURE 8.6  PQD detection using the proposed DWT-based method. (a) Voltage sag. (b)
Voltage swell. (c) Voltage interruption. (d) Oscillatory transient.

8.3.1 Deep Learning Structures

Complex learning networks gradually replace the few layers used in regression and
classification problems with developing information technologies. DL algorithms
can be considered as multi-layers of ANNs.
A two-step procedure can be followed to train the deep neural network (DNN):

1. Initiation of unsupervised learning algorithms such as stacked auto-encoder

(SAE) for determining weight values.
2. Fine-tuning of initiation weights using a supervised learning algorithm to
provide better classification.

8.3.1.1 SAE-based Methods
AE structures provide better results than other AI methods in extracting features
from data, removing unnecessary information, and compressing useful information.
AE consists of an encoder and decoder network. Represented features are defined as
Power Quality Disturbances in Smart Grids 159

FIGURE 8.6  Continued.

Equation 8.8 using the SAE structure where σ is the activation function and θ is the
training parameters:

( )
hi = f0 ( xi ) = s f wT × xi + b (8.8)

The reconstructed sample x̂i can be obtained by Equation 8.9:

( )
x̂i = gq ¢ (hi ) = s g w¢T × hi + b¢ (8.9)

A two-step procedure can be followed to train the DNN and SAE:

1. Initiation of unsupervised learning algorithms such as SAE for determining

weight values.
2. Fine-tuning of initiation weights using a supervised learning algorithm to
provide better classification.
160 Artificial Intelligence (AI)

FIGURE 8.6  Continued.

8.3.1.2 DNN-based Methods
DNN algorithms can be considered as multilayers of ANNs.

8.3.1.3 DBN (Deep belief network)-based Methods

Restricted Boltzmann machine (RBM) is a type of generative stochastic neural net-
work (GSNN) which has been widely used as greedy-layer wise pre-training strategy
to train the DBN [37]. DBN-based methods could automatically learn features from
the input data different from the SAE. DBN structures solve the vanishing prob-
lem using the backpropagation (BP) algorithm for fine-tuning DBNs. DBN maps the
learned features into the label space by adding the classification layer to classify the
PQDs or faults in microgrids.

8.3.1.4 CNN-based Methods
CNN is capable of capturing the shift-variant properties of input data [38]. CNN-
based PQD classification methods are learned features from the raw data without
using any signal processing methods compared to other DL structures. Different
Power Quality Disturbances in Smart Grids 161

FIGURE 8.6  Continued.

from traditional methods, CNN methods were performed for detection and classifi-
cation of PQDs without using an additional method of pre-processing, feature extrac-
tion, and classification. Besides, the number of training parameters in classification
algorithms is reduced using CNN.
The signal processing and classification-based hybrid method and DCNN-based
method are shown in Figure 8.7. DCNN has stacked units to extract features. Stacked
units consist of pooling layers, convolutional layers, and batch-normalization layers.

8.3.2 Proposed Deep Learning and WT-based

Hybrid PQD Classification Method
In this part of the chapter, a hybrid method based on DWT and feed-forward deep
neural network (FDNN) is proposed for the classification of PQDs occurring in grid-
connected PV systems. In the developed process, the nominal and PQD signals were
subjected to five-level DWT decomposition, and then the energy values of the detail
coefficients were calculated. The calculated energy values constitute the inputs of
162 Artificial Intelligence (AI)

FIGURE 8.7  (a) Signal processing and classification-based hybrid method. (b) DCNN-
based PQD classification method [27].

the multilayer FDNN. At the last stage, PQDs were classified using the softmax
activation function. Figure 8.8 shows the flow diagram of the proposed method. The
classified power quality events by proposed FDNN-based method are as follows:

• Class 1 – Nominal signal

• Class 2 – Voltage sag
• Class 3 – Voltage swell
• Class 4 – Voltage interruption
• Class 5 –Harmonics
• Class 6 – Flicker
• Class 7 – Oscillatory transient
• Class 8 – Impulsive transient
• Class 9 – Notch

Several parameters for PQD signals:

• The fundamental frequency value is 50 Hz, and the sampling frequency is

10 kHz.
• The number of cycles of fundamental frequency is ten cycles (0.2 s–2000
points).
• The number of samples for each class is 2000.
• The amplitude of the PQD signals is per unit (p.u.).

Most of the available literature has gone into using a mathematical model or simula-
tion to build data sets due to the difficulty of obtaining real signals related to PQD. In
this study, the integral-based mathematical model was used to generate PQD signals
Power Quality Disturbances in Smart Grids 163

FIGURE 8.8  FDNN-based model flowchart.

[39, 40]. Train, validation, or test data sets can be generated randomly by the soft-
ware created based on the equations specified in PQD parameters, possible values in ​​
Table 8.2. Mathematical definitions of five of the most common PQDs mentioned in
IEEE standards are given in Table 8.2. Other PQDs were generated with the help of
an integral-based mathematical model.
A five-level decomposition process was performed using the high-pass and low-
pass filters with daubachies4 mother wavelet, and then the energy values of the detail
 164

TABLE 8.2
Mathematical Model of PQDs [39]
Class PQD Signals Equations Threshold Parameters
C1 Nominal 49.8 £ f £ 50.2, -p £ j £ p
x ( t ) = sin ( 2p ft - j )
C2 Sag T £ t2 - t1 £ 9T
x ( t ) = é1 - a u ( t - t1 ) - u ( t - t2 ) ù sin ( 2p ft - j )
( )
ë û
0.1 £ a < 0.9
C3 Swell T £ t2 - t1 £ 9T
x ( t ) = é1 + b u ( t - t1 ) - u ( t - t2 ) ù sin ( 2p ft - j )
( )
ë û
0.1 £ b < 0.8
C5 Transient 300 Hz £ fn £ 900 Hz;
x ( t ) = sin ( 2p ft - j ) + b exp - ( t - t1 ) / t sin 2p fn ( t - t1 ) - J ((u ( t - t2 ) - (u ( t - t I ) )
( ( ) )
8 ms £ t £ 40 ms; - p £ J £ p x

0.5T £ t2 - t1 £ 3T
C7 Harmonics 7
n = {3, 5, 7} ; 0.05 £ a n £ 0.15
x ( t ) = sin ( 2p ft - j ) + n
åa sin ( np ft - J ) n
n =3
-p £ Jn ,Jn¢ ,Jn¢¢ £ p
n¢ = {3, 5, 7} ; 0.05 £ a n¢ £ 0.15
n¢¢ = {1, 3, 5}
a n¢¢ = 1 n¢¢ = 1; 0.05 £ a n¢¢ £ 0.15 n¢¢ = {3, 5}
Artificial Intelligence (AI)
Power Quality Disturbances in Smart Grids 165

FIGURE 8.9  FDNN structure.

coefficients were calculated. FDNN structure is shown in Figure 8.9. Energy values
of five detail coefficients are given as input to the network. At the output of the sys-
tem, there are labelled 8 class outputs determined.
The energy feature extraction technique is applied to DWT coefficients of the
single-phase PQD signal. The energy feature extraction technique is used to DWT
coefficients of the single-phase PQD signal. The energy of the DWT coefficients
is found in Equation 8.10 to represent the level of j: decomposition, d: detail coef-
ficients, N: a number of detail coefficients.

Ej = åD
n =1
jn j = 1,¼, l (8.10)

8.4 RESULTS
The performance scores of FDNN are given in Table 8.3. FDNN classifier has the
best score with a training accuracy (Train Acc) of 100% and with a test accuracy

TABLE 8.3
Detailed Performance Report for FDNN Classifier Test Set
Classes Precision Recall F1 Score Support
C1 95.17% 99.11% 97.10% 338
C2 100% 97.20% 98.76% 321
C3 99.64% 97.89% 98.55% 285
C4 99.00% 100% 99.50% 298
C5 100% 100% 100% 301
C6 100% 100% 100% 298
C7 99.66% 99.66% 99.66% 292
C8 99.06% 98.13% 98.59% 321
C9 99.64% 99.64% 99.64% 290
Accuracy 99.09% 2744
Macro avg 99.17% 99.11% 99.13% 2744
166 Artificial Intelligence (AI)

FIGURE 8.10  Learning curves.

Power Quality Disturbances in Smart Grids 167

TABLE 8.4
Performance Comparison of Existing Methods
Method No. of Futures Number of PQDs Accuracy
Reference [23] 5 11 98.2%
Reference [41] 5 14 97.28%
Reference [42] 19 9 99.66%
Reference [43] 10 5 98.51%
Reference [44] 45 22 99.09%
Reference [45] 2 10 94.2%
Reference [21] 6 9 97.22%
Proposed method 5 9 99.13%

(Test Acc) of 99.13%. Training and test set were randomly distributed, and the test
set was determined to be approximately 30%.
The training lasted 200 Epoch – about 608 sec. Learning curves are shown in
Figure 8.10.
As shown in Table 8.4, the PQDs classification accuracy corresponding to the WT
and FDNN-based algorithm is higher than all the other investigated methods.

8.5 CONCLUSION
This chapter presents SP-based and AI-based PQD detection and classification meth-
ods used in smart grids. In Section 8.1, SP-based PQD detection methods used in
smart grids are classified, and then wavelet transform (WT)-based detection meth-
ods are presented. Current WT-based methods were introduced for the detection
of PQDs. Some PQD events are investigated in MATLAB Simulink by using the
developed discrete WT method.
The classification of PQDs with AI-based methods are also discussed in
the last section. Deep neural networks are presented in the chapter to show the
effectiveness of AI-based methods. A hybrid method based on DWT and feed-
forward deep neural network (FDNN) is proposed for the classification of PQDs
occurring in grid-connected PV systems. In the developed process, the nominal
and PQD signals were subjected to five-level DWT decomposition, and then the
energy values of the detail coefficients were calculated. The calculated energy
values constitute the inputs of the multilayer FDNN. At the last stage, PQDs
were classified using the softmax activation function. The results show that the
proposed hybrid WT and FDNN methods have 99.13% accuracy in classifying
PQDs.
The proposed detection and classification methods for detecting PQDs are suit-
able for smart grids, and the obtained results in this chapter will contribute to select-
ing the proper detection or classification methods of PQDs using AI-based methods
in smart grids.
168 Artificial Intelligence (AI)

REFERENCES
1. Biswas, S., Goswami, S. K., & Chatterjee, A. (2012). Optimum distributed generation
placement with voltage sag effect minimization. Energy Conversion and Management,
53(1), 163–174.
2. Hung, D. Q., Mithulananthan, N., & Bansal, R. C. (2010). Analytical expressions for DG
allocation in primary distribution networks. IEEE Transactions on Energy Conversion,
25(3), 814–820.
3. Bollen, Math H. J. (2000). Understanding power quality problems. In Voltage Sags and
Interruptions. IEEE Press.
4. Chapman, David. (2001). The cost of poor power quality. Power Quality Application
Guide 1–4.
5. IEEE. Std. 929 (2000). IEEE Recommended Practice for Utility Interface of Photovoltaic
(PV) Systems. Institute of Electrical and Electronics Engineers, Inc., New York.
6. F II I (1993). IEEE Recommended Practices and Requirements for Harmonic Control
in Electrical Power Systems. New York.
7. IEEE (2014). Std. IEEE Recommended Practices and Requirements for Harmonic
Control in Electrical Power Systems, 519.
8. Chen, C. I., & Chen, Y. C. (2014). Intelligent identification of voltage variation
events based on IEEE std 1159–2009 for SCADA of distributed energy system. IEEE
Transactions on Industrial Electronics, 62(4), 2604–2611.
9. Mahela, O. P., Shaik, A. G., & Gupta, N. (2015). A critical review of detection and
classification of power quality events. Renewable and Sustainable Energy Reviews, 41,
495–505.
10. Khokhar, S., Zin, A. A. B. M., Mokhtar, A. S. B., & Pesaran, M. (2015). A comprehen-
sive overview on signal processing and artificial intelligence techniques applications
in classification of power quality disturbances. Renewable and Sustainable Energy
Reviews, 51, 1650–1663.
11. Latran, M. B., & Teke, A. (2015). A novel wavelet transform based voltage sag/swell detec-
tion algorithm. International Journal of Electrical Power and Energy Systems, 71, 131–139.
12. Karegar, H. K., & Sobhani, B. (2012). Wavelet transform method for islanding detec-
tion of wind turbines. Renewable Energy, 38(1), 94–106.
13. Mohammadnian, Y., Amraee, T., & Soroudi, A. (2019). Fault detection in distribution
networks in presence of distributed generations using a data mining–driven wavelet
transform. IET Smart Grid, 2(2), 163–171.
14. Bayrak, G. (2018). Wavelet transform-based fault detection method for hydrogen
energy-based distributed generators. International Journal of Hydrogen Energy,
43(44), 20293–20308.
15. Deokar, S. A., & Waghmare, L. M. (2014). Integrated DWT–FFT approach for detection
and classification of power quality disturbances. International Journal of Electrical
Power and Energy Systems, 61, 594–605.
16. Lopez-Ramirez, M., Cabal-Yepez, E., Ledesma-Carrillo, L. M., Miranda-Vidales,
H., Rodriguez-Donate, C., & Lizarraga-Morales, R. A. (2018). FPGA-based online
PQD detection and classification through DWT, mathematical morphology, and SVD.
Energies, 11(4), 769.
17. Yılmaz, A., & Bayrak, G. (2019). A real-time UWT-based intelligent fault detection
method for PV-based microgrids. Electric Power Systems Research, 177. http:​/​/www​​
.ncbi​​.nlm.​​nih​.g​​ov​/pu​​bmed/​​​10598​​4.
18. Dhimish, M., Holmes, V., Mehrdadi, B., & Dales, M. (2018). Comparing Mamdani
Sugeno fuzzy logic and RBF ANN network for PV fault detection. Renewable Energy,
117, 257–274.
Power Quality Disturbances in Smart Grids 169

19. Bangalore, P., & Patriksson, M. (2018). Analysis of SCADA data for early fault detec-
tion, with application to the maintenance management of wind turbines. Renewable
Energy, 115, 521–532.
20. Zhang, L., Zhou, W., & Jiao, L. (2004). Wavelet support vector machine. IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 34(1), 34–39.
21. Thirumala, K., Pal, S., Jain, T., & Umarikar, A. C. (2019). A classification method for
multiple power quality disturbances using EWT based adaptive filtering and multiclass
SVM. Neurocomputing, 334, 265–274.
22. Puliyadi Kubendran, A. K., & Loganathan, A. K. (2017). Detection and classification of
complex power quality disturbances using S‐transform amplitude matrix–based deci-
sion tree for different noise levels. International Transactions on Electrical Energy
Systems, 27(4), e2286.
23. Lee, C. Y., & Shen, Y. X. (2011). Optimal feature selection for power-quality distur-
bances classification. IEEE Transactions on Power Delivery, 26(4), 2342–2351.
24. Khokhar, S., Zin, A. A. M., Memon, A. P., & Mokhtar, A. S. (2017). A new optimal fea-
ture selection algorithm for classification of power quality disturbances using discrete
wavelet transform and probabilistic neural network. Measurement, 95, 246–259.
25. Kong, X., Xu, X., Yan, Z., Chen, S., Yang, H., & Han, D. (2018). Deep learning
hybrid method for islanding detection in distributed generation. Applied Energy, 210,
776–785.
26. Liu, H., Hussain, F., Shen, Y., Arif, S., Nazir, A., & Abubakar, M. (2018). Complex
power quality disturbances classification via curvelet transform and deep learning.
Electric Power Systems Research, 163, 1–9.
27. Wang, S., & Chen, H. (2019). A novel deep learning method for the classification of
power quality disturbances using deep convolutional neural network. Applied Energy,
235, 1126–1140.
28. Balouji, E., & Salor, O. (2017). Classification of power quality events using deep learn-
ing on event images. In 2017 3rd International Conference on Pattern Recognition and
Image Analysis (IPRIA), 216–221. IEEE.
29. Mohan, N., Soman, K. P., & Vinayakumar, R. (2017). Deep power: Deep learning
architectures for power quality disturbances classification. In 2017 International
Conference on Technological Advancements in Power and Energy (TAP Energy), 1–6.
IEEE.
30. Bayrak, G., & Yilmaz, A. (2019). Assessment of power quality disturbances for grid
integration of PV power plants. Sakarya University Journal of Science, 23(1), 1–1.
31. Heydt, G. T., Fjeld, P. S., Liu, C. C., Pierce, D., Tu, L., & Hensley, G. (1999). Applications
of the windowed FFT to electric power quality assessment. IEEE Transactions on
Power Delivery, 14(4), 1411–1416.
32. Cohen, L. (1995). Time-Frequency Analysis. Prentice-Hall.
33. Samantaray, S. R., Samui, A., & Babu, B. C. (2011). Time-frequency transform-based
islanding detection in distributed generation. IET Renewable Power Generation, 5(6),
431–438.
34. Strang, G., & Nguyen, T. (1996). Wavelets and Filter Banks. Wellesley-Cambridge
Press.
35. Shensa, M. J. (1992). The discrete wavelet transform: Wedding the a trous and Mallat
algorithms. IEEE Transactions on Signal Processing, 40(10), 2464–2482.
36. Zafar, T., & Morsi, W. G. (2013). Power quality and the un-decimated wavelet trans-
form: An analytic approach for time-varying disturbances. Electric Power Systems
Research, 96, 201–210.
37. Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with
neural networks. Science, 313(5786), 504–507.
170 Artificial Intelligence (AI)

38. Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., & Nandi, A. K. (2020). Applications of
machine learning to machine fault diagnosis: A review and roadmap. Mechanical
Systems and Signal Processing, 138. http:​/​/www​​.ncbi​​.nlm.​​nih​.g​​ov​/pu​​bmed/​​​10658​​7.
39. Igual, R., Medrano, C., Arcega, F. J., & Mantescu, G. (2018). Integral mathemati-
cal model of power quality disturbances. In 2018 18th International Conference on
Harmonics and Quality of Power (ICHQP), 1–6. IEEE.
40. Igual, R., Medrano, C., Arcega, F. J., & Mantescu, G. (2017). Mathematical Model of
Power Quality Disturbances. Mendeley Data, 1.
41. Thirumala, K., Prasad, M. S., Jain, T., & Umarikar, A. C. (2016). Tunable-Q wavelet
transform and dual multiclass SVM for online automatic detection of power quality
disturbances. IEEE Transactions on Smart Grid, 9(4), 3018–3028.
42. Abdoos, A. A., Mianaei, P. K., & Ghadikolaei, M. R. (2016). Combined VMD-SVM
based feature selection method for classification of power quality events. Applied Soft
Computing, 38, 637–646.
43. Erişti, H., & Demir, Y. (2012). Automatic classification of power quality events and
disturbances using wavelet transform and support vector machines. IET Generation,
Transmission and Distribution, 6(10), 968–976.
44. Dalai, S., Dey, D., Chatterjee, B., Chakravorti, S., & Bhattacharya, K. (2014). Cross-
spectrum analysis-based scheme for multiple power quality disturbance sensing device.
IEEE Sensors Journal, 15(7), 3989–3997.
45. Borrás, M. D., Bravo, J. C., & Montaño, J. C. (2016). Disturbance ratio for optimal
multi-event classification in power distribution networks. IEEE Transactions on
Industrial Electronics, 63(5), 3117–3124.
 9 Robust Design of
Artificial Neural Network
Methodology to Solve
the Inverse Kinematics of
a Manipulator of 6 DOF
Ma. del Rosario Martínez-Blanco,
Teodoro Ibarra-Pérez, Fernando Olivera-Domingo,
and José Manuel Ortiz-Rodríguez

CONTENTS
9.1 Introduction................................................................................................... 171
9.1.1 Kinematics of Robotic Manipulators................................................. 173
9.1.2 Artificial Neural Networks................................................................ 174
9.1.3 Inverse Kinematics Solution with Artificial Neural Networks......... 178
9.1.4 Robust Design of Artificial Neural Networks................................... 179
9.2 Robust Design of Artificial Neural Networks Methodology......................... 180
9.3 Kinematics Analysis of Robotic Manipulator Called Ketzal........................ 184
9.3.1 Data Set Description.......................................................................... 189
9.3.2 Description of Reduction Data Filter Algorithm............................... 191
9.3.3 Data Set Analysis of Training and Test............................................. 194
9.3.4 Planning and Experimentation Stage................................................ 195
9.3.5 Analysis and Confirmation Stage...................................................... 196
9.4 Conclusions and Discussions......................................................................... 203
Future Scope...........................................................................................................205
Acknowledgements................................................................................................. 205
References............................................................................................................... 205

9.1 INTRODUCTION
With recent advances in electronics, mechanics, computer science, and robotics,
robotic manipulators have become the main focus of interest for the development
of industrial applications [1]. Among the multiple articular configurations available,
generally, the manipulators that have more degrees of freedom (DOF) offer greater

171
172 Artificial Intelligence (AI)

control flexibility for complicated tasks; so the scientific community has taken an
interest in this type of robotic manipulators during the last decades [2–4].
During the last years, several investigations have been carried out in the field of
robotics due to its high impact in various areas such as space exploration [5], indus-
trial [6], military [7], and medical applications [8], among others. The use of these
manipulators has contributed considerably to the development and application of
new methods and technologies applied to the field of robotics, mainly in the control
of movement in real time [9].
In the field of robotics manipulators, the main challenge of motion control is to
find a precise and reliable solution for inverse kinematics. The calculation necessary
for the resolution of inverse kinematics requires expensive processing and is compu-
tationally complex [10].
In general, the solutions of the inverse kinematics of a robotic manipulator
are based on three methods: the geometric, iterative, and analytical or algebraic
methods. The singularities and uncertainties, or configurations where manipulator
mobility is reduced, are the main complications in the aforementioned methods for
the control of movement in robotic manipulators. For example, for the algebraic
method, closed-form solutions are not guaranteed and such solutions must exist
for the first three joints if the geometric method is used. In the case of an itera-
tive method, it converges to a single solution that depends on the initial point; so
it requires high-performance hardware without guaranteeing the accuracy of the
calculations [3, 11].
Therefore, the traditional methods used in inverse kinematics problems with geo-
metric, iterative, and algebraic approaches are sometimes complex and generally
unsuitable for multiple joint configurations that can present robotic manipulators
nowadays. As a consequence, various approaches based on artificial neural networks
(ANNs) have been proposed due to the great advantages they present in their parallel
distribution, non-linear mapping, the ability to learn through examples, and the high
performance in the ability to generalization, among others [4, 12].
Generally speaking, the application of ANNs has reduced the error in con-
trolling movement in real time, allowing increasing accuracy [8]. A well-trained
neural network can work with very fast response times, so it is ideal for real-time
applications, compared to other conventional methods where the response time
is longer [13].
Most of the research in ANNs has focused on specific applications of model
development and training algorithms to improve the convergence and accuracy of
the results obtained; however, finding the parameters to achieve adequate learning in
the training of neural networks remains a difficult and complex task [9, 10, 14–17].
In general, the quality in the development and training of ANNs is highly reliable;
however, structural parameters, just like the number of hidden layers and neurons
per hidden layer, represent an important role in the accuracy of the expected results
and these are usually proposed due to the previous experience of the researcher in
trial and error procedures, consuming time and resources without guaranteeing that
the optimal configuration of the parameters to achieve a better performance in the
neural network is obtained [3, 10, 18–21].
Robust Design of Artificial Neural Networks 173

In this work, the use of a robust design methodology based on the fractional fac-
torial design of experiments is proposed to get the optimal parameters of an ANN
architecture for the calculation of the robotic manipulator of six DOF.
The chapter is organized as follows: First, the introduction is addressed, where
the bases of the kinematics of robotic manipulators are presented, as well as those
of the artificial neural networks, to approach the solution of the inverse kinemat-
ics solution with artificial neural networks. The robust design of artificial neural
networks applied in this chapter is also addressed in the introduction. In the second
section, the proposed methodology and its different phases are presented. The third
section describes Ketzal – the robotic manipulator used, performing a kinematic
analysis of it, which allows obtaining a data set to be used in training. The data set
description and the reduction data filter algorithm used are also presented, followed
by data set analysis of training and test. Finally, the different phases of the proposed
methodology are addressed in Planning and Experimentation Stage and Analysis
and Confirmation Stage subsections. In the fourth section, the conclusions are ana-
lyzed and finally, in the fifth section, the future work is presented.

9.1.1 Kinematics of Robotic Manipulators

The morphology of robotic manipulators refers to the description of the components,
parts, and mechanical structure [22]. A robotic manipulator is usually made using
consecutive rigid mechanisms called links and connected by joints to form an open
kinematic chain that allows the robot to move [23].
A robot can be modelled as an articulated open chain of several rigid links con-
nected in series by prismatic or rotating joints actioned by actuators. The intercon-
nection between two links consecutively represents an articulation [4, 21].
From the mechanical point of view, it is an open kinematic chain as long as each
of the links is connected through joints to the next and previous links, except for the
first and last, where the first is usually always fixed to support and the latter is usually
free, that is to say, that a closed loop is not formed, otherwise it would be a closed
kinematic chain [22].
The last link is usually a terminal element as a clamping tool or final actuator.
Each of the joints forms a DOF. The joints can produce rotational and linear move-
ment, where the latter ones are called linear prismatic joints [22].
The kinematics of a manipulator can be represented analytically as the descrip-
tion of the movement of the robot arm as a function of time according to a fixed
reference system without considering the forces/torques that cause them, that is, the
dynamics of the system [24, 25].
The kinematics of the manipulator describes the relationship between the joint
angle, the position, and the orientation of the robot’s end. The robot kinematics anal-
ysis includes the solution of two aspects: direct kinematics and inverse kinematics
[3], as shown in Figure 9.1.
Direct kinematics consists of obtaining the position and orientation of the robot’s
end given the value of the articular coordinates. The inverse kinematics consists
in determining the configuration that the robot must adapt for a known position
174 Artificial Intelligence (AI)

FIGURE 9.1  Direct and inverse kinematics.

and orientation of its end, so the complexity in the solution to solve this problem is
mainly due to the manipulator geometry, being able to obtain multiple solutions and
singularities that do not always correspond to physical solutions [12, 24].
In this work, the direct kinematics model applied to a six DOF manipulator was
developed, obtaining a set of data on position, orientation, and joint values cor-
responding to the robotic manipulator’s workspace. A filter was used to decrease
the volume of the data and the methodology of Robust Design of Artificial Neural
Networks (RDANN) was applied to obtain the optimal parameters of an ANN archi-
tecture to solve the problem of inverse kinematics for the proposed manipulator.

9.1.2 Artificial Neural Networks

An artificial neuron represents the mathematical model of the functioning of a bio-
logical neuron. The ANNs are inspired in the functioning of the human brain that
processes the information very efficiently, particularly these represent a computa-
tional structure inspired by the biological neurons in our brains that make complex
systems and can learn by experience [26].
A standard neural network consists of the interconnection of several simple pro-
cessors called neurons, each producing an activation. Input neurons receive infor-
mation from the environment and also get activated by weighted connections from
previously active neurons [27].
Achieving learning in a neural network consists of finding the synaptic weights
between the interconnections to minimize the cost function and ensure that the neural
network shows the desired behaviour. The cost or loss function is regularly optimized
with the gradient descent method. Although the method is very effective, a globally
optimized solution is not always guaranteed because it can fall to a local minimum [20].
An artificial neuron can be conceived as a mathematical model of the function-
ing of a biological neuron, as shown in Figure 9.2, where the inputs xi represent the

FIGURE 9.2  Artificial neuron model.

Robust Design of Artificial Neural Networks 175

signals captured by the dendrite that come from another neuron, where the weights
Wi represent the intensity of the synapse that connects two neurons that may be posi-
tive (excitatory) or negative (inhibitory). θ is the transfer function or threshold that
the neuron must overcome to enter the activation state. Finally, the accumulated sum
in the node from all the input signals and multiplied by the synaptic weights go to the
output signal through the transfer function or activation function [28].
In essence, an artificial neuron mimics the behaviour of a biological neuron, that
is, it receives varying signals from other neighbouring neurons and can easily pro-
cess the information from the other neurons to let it pass to another neuron with a
certain level of energy [29].
The signals that enter the neuron are weighted by the multiplication of weights,
which represents a process similar to the strength of the synaptic connection in a
biological neuron; the weighting is added by a neural node and the output is calcu-
lated as the sum of the weighted inputs plus a value b called bias. The output of the
neuronal node is used as input to a transfer function that responds to the artificial
neuron [30]. The output signal of a neuron net can be represented as described below.
Figure 9.3 shows the simplification of an artificial neuron model, where it can be
observed that the inputs to a neuron, xj (j = 1, 2, …, n), are represented by the input
array component, the synaptic weights are represented by the matrix Wi,j (j = 1, 2, …,
n) and b represents the bias. The response of the artificial neuron to the input signals
can be described mathematically as follows:

æ n
ö
yi = f (net ) = f ç
ç
è
å
i =1
x j wij + b ÷ (9.1)
÷
ø
where yi is the value obtained at the end of the artificial neuron and f is the transfer
function used. Usually, a neuron can have one or more inputs.
A model of neurons with multiple inputs is shown in Figure 9.4. In this diagram,
the inputs x1, x2, …, xi are shown, which are multiplied by the corresponding weights
w1,1, w1,2, …, wi,j belonging to the synaptic weight matrix. The subscripts of the
matrix represent the neurons involved in the link, where the first subscript represents
the target neuron and the second represents the source of the input signal to the target
neuron [14]. For example, w2,5 indicates that this weight is the connection from the
fifth input and the second neuron.

FIGURE 9.3  Artificial neuron model with transfer function.

176 Artificial Intelligence (AI)

FIGURE 9.4  Neuron model with multiple inputs.

FIGURE 9.5  Architecture with two hidden layers.

An ANN is a set of interconnected neurons distributed in parallel that acts as an

information processing system. Networks achieve learning due to an external source
or set of training data by making changes in their structure to predict linear or non-
linear trends [18].
Generally, an ANN is formed by the interconnection of several neurons. The form
of connection between neurons varies according to the type of network, but they are
usually grouped by layers that can be classified according to the location of the layer
in the network as the input layer, hidden layer, and output layer. The architecture of
the neural network refers to the arrangement of neurons within layers and the pat-
terns of connection between them [28].
A simplified descriptive model of an ANN with three layers is shown in Figure
9.5, where it is noticeable that the network has R1 inputs, i1 neurons in the first layer,
i2 neurons in the second layer, and i3 neurons in the third layer. An input constant is
known as bias and a value of 1 is added to each neuron. Each intermediate output
connects to the input of the next layer. In this way, a layered approach can be taken
to analyze the entire network [29].
According to the existence of feedback connections or not in the neural network,
two architectures can be described: forward propagation (known as perceptron) and
backpropagation. The first architecture has no feedback and does not maintain a
record of the previous values of its output and the activation state of the neurons. The
Robust Design of Artificial Neural Networks 177

second architecture has a closed-loop connection and therefore it maintains a record

of its preceding states, so the next state depends on the preceding states in addition
to the input signals [28].
The Artificial Neural Propagation Network (ANNBP) was proposed by Rumelhart
and McClelland in 1986 and it is a widely used algorithm. The fundamental archi-
tecture of the network consists of three layers: the input layer, the hidden layer, and
the output layer. For such networks, there are no restrictions on the number of hidden
layers and has a solid mathematical foundation based on the algorithm of the gradi-
ent descent. The backpropagation refers to the method in which the synaptic weights
between the connections change after being processed by the neuron layer through
an activation function, producing outputs for the next layer [20, 28, 31].
The problem with this type of network is mainly related to the parameters that
must be initially established before performing any training. The user must choose,
according to their experience, the network architecture to be used and determine
several of the parameters of the chosen network. These parameters determine suc-
cess or failure in network training. There is currently no procedure that fully guar-
antees the optimal configuration of these parameters [14].
Currently, most of the selection of these parameters is based on the practice of trial
and error, so a large amount of ANN models are developed and compared to choose
the best model, as described in Figure 9.6. If the value of a parameter is modified and
it does not have a significant effect on the network performance, another parameter
is changed to improve network performance until the parameters that guarantee the
desired performance are found for a specific problem.
The main disadvantage of this procedure is that while one parameter is evaluated,
the others remain at a single level. Consequently, the best parameter evaluated in a
particular design variable may not necessarily be the best parameter for the final
model. Regularly, this practice does not guarantee the interaction and effects of the
variables between their different levels, because it only combines one parameter at
the same time [32].
This problem has motivated researchers in the search for design solutions and
strategies to optimize design parameters in ANNs. One of the alternatives for this

FIGURE 9.6  The traditional approach based on trial and error.

178 Artificial Intelligence (AI)

problem could be to evaluate all possible combinations of levels in the variables

involved, that is, a complete factorial design. This practice could be complex and
computationally expensive due to the multiple combinations that could be generated
even for a small number of parameters and levels [33].
However, the number of experiments could be substantially reduced with the frac-
tional factorial method, which is a statistical method based on the robust design of
the Taguchi philosophy. This method makes it possible to find the optimal adjust-
ment parameters, involving design and noise variables, allowing the noise factors
involved in the design to be insensitive during the process. This technique, based on
the philosophy of robust design, is a very powerful optimization method that differs
from traditional practices [14, 34].

9.1.3 Inverse Kinematics Solution with Artificial Neural Networks

In recent years, the ANNs have contributed significantly to the development of vari-
ous fields of engineering and science [10, 35–38]. The main interest in the use of
ANNs is due to its characteristics related to non-linearity, high robustness, parallel-
ism, fault tolerance, and its great capacity for learning and generalization, based on
the learning process through complex, non-linear, and multiple examples of input
and output relations [13].
The multilayer perceptron (MLP) algorithm trained with backpropagation (BP) is
one of the most used techniques in modelling, classification, and optimization appli-
cations. The cost or loss function is regularly optimized with the gradient descent
method; although the algorithm is very effective, it tends to converge very slowly and
its main disadvantage is that an optimal and global solution is not always guaran-
teed, because the gradient descent method runs the risk of getting caught in a local
minimum [14, 16, 39].
The quality in the data sets is another of the most important factors to consider
in the training of ANNs, because a low quality generally leads to low quality in the
knowledge extracted, so the efficiency in the algorithms of extraction of the knowl-
edge largely depends on the quality of the data sets [40].
In general, the quality in the development and training of artificial neural
networks is highly reliable; however, structural parameters, just like the number
of hidden layers and neurons per hidden layer, represent an important role in the
accuracy of the expected results and generally these are usually proposed based
on the previous experience of the investigator in trial and error procedures, con-
suming time and resources without guaranteeing the optimal configuration of
the parameters to achieve a better performance in the neural network [2, 3, 18,
19, 21].
The process optimization, Taguchi methods, and experiment design constitute
some of the most appropriate methods to achieve this objective since they can be
applied to artificial neural networks design and training. Robust design is a statisti-
cal technique widely used to analyze the relationship between the variability factors
that influence the results of a process and that can be used to systematically find the
optimal configuration to obtain the desired result [14, 32, 41].
Robust Design of Artificial Neural Networks 179

This chapter proposes a systematic and experimental strategy according to the

requirements of the problem. The RDANN methodology was used to find the opti-
mal parameters in an artificial neural network architecture of inverse propagation
applied to the solution of the inverse kinematics of a robotic manipulator of six DOF.

9.1.4 Robust Design of Artificial Neural Networks

In the 1920s, the design of experiments by multifactor intervention by Sir Ronald
A. Fisher was used for the first time to determine the results in agricultural prac-
tices [42].
This technique is known as factorial design of experiments, where a full factorial
design allows identifying all possible combinations, implying a significant amount
of experiments concerning the number of factors or variables involved and where
the variables involved determining the performance or functionality of a system or
product that can be controlled [43].
The robust parameter design method was proposed by Genichi Taguchi. This
engineering method applied to the design of products or processes is focused on the
decrease or sensitivity to noise. The method has proven to be a powerful and efficient
procedure in the design of products or processes allowing optimal and consistent
operation under certain conditions [44].
The main objective of the method is to find the selection of the factors involved
that minimize the variability of the response to different inputs of the system
by properly selecting the levels in the controllable design variables. The robust
design approach allows studying all the parameter space with a small num-
ber of experiments using one orthogonal array (OA) and fractionated factorial
design [45].
An OA represents a lesser part of the conventional complete factorial design
and allows a well-adjusted comparison of the levels in the variables involved in the
design of experiments. The design parameters involved represent the columns of the
OA and the rows are the individual experiments by combining levels in the design
parameters [46].
To determine the combination of levels, Taguchi proposed a two-phase proce-
dure. First, the most significant control factors are determined to reduce variability.
The factors that are significant to affect the sensitivity of the system are chosen
below. The second phase aims to adjust the response to the desired values [47].
The Taguchi method can be described in four stages:

Quality characteristics and parameter design. Generate a brainstorm

1.
about the most important design parameters and quality characteristics in
the process. At this stage, there are variables that the user can control and
others that cannot. These are called design and noise factors, respectively.
Design of experiments. Choose the most appropriate OA by selecting the
2.
levels of the design factors that reduce the effects produced by noise, that is,
that the product response maintains the minimum variation and its average
approximates the desired objective.
180 Artificial Intelligence (AI)

An OA can evaluate several factors involved in a small number of exper-

iments, allowing to reduce the time in conducting experiments and obtain-
ing more information with fewer experiments. The mean and variance of
the response in each configuration proposed by the OA are combined into a
single performance measure known as the signal-to-noise ratio (S/N).
3. Analysis of results for the determination of the optimal conditions. An
excellent quality indicator is the S/N ratio, in which the effect produced on
the performance of the process or product can be evaluated by changing a
parameter. Taguchi used the S/N ratio to assess the variation in system per-
formance. For the static characteristic, Taguchi classified them into three
types of relationship:
(a) Smaller-the-better (STB)
(b) Larger-the-better (LTB)
(c) Nominal-the-best (NTB)

To obtain the optimal factors, for the NTB case, Taguchi developed a
two-phase optimization procedure. For both other cases, Taguchi recom-
mends the direct minimization of the expected loss.
In dynamic characteristics the relationship

SN i = 10 × log10 ( bi / MSEi ) (9.2)

is used to evaluate the S/N ratio, and MSE represents the average of the
squared errors, that is, the difference between the measured response and
what is estimated.
Confirmation test in optimal conditions. The main objectives in the
4.
design of parameters are to reduce the design variation of the processes or
products to different environments or conditions. At this stage, a confirma-
tion experiment is performed using the optimal design conditions.
Currently, the ANNs can be trained to solve complex and difficult prob-
lems for conventional computers and humans, including various fields such
as pattern recognition and speech, vision, or control systems. In the field of
robotics, ANNs have been successfully applied in recent decades [13, 48],
mainly in the solution to the problem of the inverse kinematics in robotic
manipulators with a high number of joints [1–3, 49].

9.2 ROBUST DESIGN OF ARTIFICIAL NEURAL

NETWORKS METHODOLOGY
Determining the solution of inverse kinematics in manipulators with a high DOF
number is a complex task [2]. To solve this problem, three methods are generally
used: the geometric method, the algebraic method, and the iterative method [3].
Generally, most solutions using traditional methods with geometric, iterative, and
algebraic approaches sometimes turn out to be complex and generically inadequate
Robust Design of Artificial Neural Networks 181

for the multiple articular configurations that robotic manipulators may present due to
the high degree of precision required in the manipulator positioning [50].
In geometric and algebraic methods, a closed solution can only be guaranteed as
long as the manipulator meets some special conditions, such as the spherical wrist
[51]. For their part, iterative methods tend to be too slow and do not meet the require-
ment for implementation in real-time control systems [49].
This problem has been addressed by researchers during the last decades, who
have proposed various methods of solution through the use of ANNs, so they con-
stitute one of the most appropriate methods in solving this problem. An important
problem in the calculation of the inverse kinematics using ANNs is the determina-
tion of the parameters of the network. The choice of basic ANN parameters deter-
mines success in ANN training [9, 15, 52].
However, there are no rules to identify the optimal selection of these parameters.
Generally, the traditional trial and error technique produces a low capacity for gen-
eralization and low performance, consuming a great deal of time and computational
resources [3].
To overcome this disadvantage easily and efficiently, the RDANN methodology is
used. This methodology, proposed in Ref. [14] and shown in Figure 9.7, describes a
systematic process to find the optimal parameters of an ANNBP architecture applied
to the inverse kinematics solution of a six DOF robotic manipulator.
According to Figure 9.8, the steps to obtain the optimal ANN parameters are
described below:

Planning stage
1.
In this stage, the objective function and the design and noise variables
are identified.

FIGURE 9.7  RDANN methodology.

182 Artificial Intelligence (AI)

FIGURE 9.8  Ketzal robot manipulator.

(a) The objective function. The requirements of the problem allow defin-
ing the objective function. The objective function in this research is the
prediction or classification of errors between the expected values and
the output values of the ANNBP in the testing stage. The performance
or value of the mean square quadratic error (MSE) of the ANNBP is
used to achieve this goal, which is expressed mathematically as follows:
N

å (q )
1 2
MSE = i
predicted
- qiDeduced (9.3)
N i =1

where N is the number of attempts, qiPredicted is the array of values

predicted by the neuronal network, and qiDeduced is the array of the values
originally inferred or calculated by the forward kinematics [14, 18].
(b) Design and noise variables. Users can select some of the design vari-
ables that may vary during the optimization process or iteration pro-
cess according to the requirements of the problem. Among the different
parameters that affect the performance of an ANNBP, four design vari-
ables were chosen as they are the variables that can be controlled by the
user, as shown in Table 9.1.

TABLE 9.1
Design Variables and Their Levels
Design Variables Level 1 Level 2 Level 3
A L1 L2 L3
B L1 L2 L3
C L1 L2 L3
D L1 L2 L3
Robust Design of Artificial Neural Networks 183

TABLE 9.2
Noise Variables and Their Levels
Noise Variables Level 1 Level 2
U Set 1 Set 2
V 9:1 8:2
W Tr-1/Ts-1 Tr-2/Tst-2

Here A is the number of neurons in the first hidden layer, B is the

number of neurons in the second hidden layer, C is the momentum, and
D is the learning rate.
The noise variables are shown in Table 9.2. Under various conditions
of noise or distortion in the process, three variables were determined.
These variables in most cases are not controlled by the user as is the
case with the initialization of the set of weights U, which is normally
selected randomly. In training and test data sets V, the designer must
decide the size of the training set versus the test size. Once V is deter-
mined, the designer makes the random selection to decide the data to be
included in the training and test set W.
U represents the initial set of random weights, V is the size of the train-
ing set versus the size of the test set, i.e. V = 90%/10%, 80%/20%, and
W is the random selection of the set of training and set of tests, that is,
W = Training1/Test1, Training2/Test2. These types of variables cannot be
controlled by the designer because they are randomly determined.
Experimentation Stage
2.
Choosing the right elements in an OA is critical to success at this stage. An
OA allows evaluating the main effects of interaction between factors involved
through a small number of experiments. Taguchi suggests the use of two
cross-configurations of OAs with an L9(34) and L4(32), as shown in Table 9.3.

TABLE 9.3
OA with L9(34) and L4(32) Configuration
Trial No. A B C D S1 S2 S3 S4 Mean S/N
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1
184 Artificial Intelligence (AI)

Analysis Stage
3.
The S/N ratio considers the mean and the variation in the measured
responses allowing a quantitative evaluation of the design parameters. The
unit of measure is a decibel and the formula used is

S / N = 10 × log10 ( MSD ) (9.4)

where MSD is the measure of the mean square deviation in the perfor-
mance of the neural network.
Since it is desirable to have a higher signal and less noise, the highest S/N
ratio indicates the best design parameters. A statistical program (JMP) was
used at this stage to determine the optimal parameters of the network.
Confirmation Stage
4.
The robust measurement value is predicted by the optimal design condi-
tion. A confirmation experiment is performed by calculating the perfor-
mance robustness measurement and verifying if the value of the robust
measurement is close to the value deduced.

Once the aspects involved in each of the stages of the RDANN methodology were
identified, the training data set regarding the position, orientation, and articular val-
ues of the robotic manipulator was generated, based on the analysis of the direct
kinematics of the six DOF robotic manipulator, to subsequently perform the 36
workouts described in the proposed OA.

9.3 KINEMATICS ANALYSIS OF ROBOTIC

MANIPULATOR CALLED KETZAL
The solution to the direct kinematic problem refers to finding a homogeneous trans-
formation matrix T that relates the position and orientation of the end of the robot
concerning a fixed reference system located at the base of the robot since the move-
ment of the robotic manipulator describes the relationship between the angles of the
joints, the position, and orientation of the final effector [53].
Figure 9.8 shows the structure and reference coordinates of the Ketzal robot used
in this work, which is a six DOF robotic manipulator taken from an open-source, 3D
printable, and low-cost project [54].
The direct kinematics problem is to find the position and orientation of the final
effector with respect to a reference system, given the array of joint angles, ϴ = (ϴ1,
ϴ2, ϴ3, ϴ4, ϴ5, ϴ6) of the robotic manipulator.
The objective of inverse kinematics is to calculate the array of joint angles ϴ
given the position and orientation of the final effector with respect to a reference
coordinate system. Direct kinematics is represented as a 4 × 4 homogeneous trans-
formation matrix that represents the position and orientation of the final effector of
the robotic manipulator with respect to a reference coordinate system and which is
given by the following equation [24].
Robust Design of Artificial Neural Networks 185

é nx ox ax px ù
ê ú
éR 0
6 P ù ê ny
6
0
oy ay py ú
T60 = ê ú=ê (9.5)
ë0 1û nz oz az pz ú
ê ú
êë 0 0 0 1 úû
where R60 is the 3 × 3 rotation matrix which contains the orientation array [n o a] of
the final effector, and P60 is the position array [p] of the final effector in the reference
coordinate system.
The Denavit–Hartenberg (DH) method was used to analyze the direct kinematics
of the Ketzal robot by implementing four basic transformations that depend exclu-
sively on the geometric characteristics of the links [55]. The DH parameters of the
Ketzal robot are shown in Table 9.4.
These transformations consist of a succession of rotations and translations that
allow the reference system of element i to be related to the system of element i − 1,
where the homogeneous transformation matrix for each joint is given by the follow-
ing equation [10].

i -1
Ai = Rot z,q iTrans x ,diTrans x ,a i Rot z,a i (9.6)

Thus,

éci -si ca i si sa i ai ci ù
ê ú
si ci ca i -ci sa i ai si ú
i -1
Ai = ê (9.7)
ê0 sa i ca i di ú
ê ú
ë0 0 0 1 û

TABLE 9.4
DH Parameters of the Ketzal Robot
Link Offset (cm) Joint Angle (rad) Link Length (cm) Twist Angle (rad)

d1 = 20.2 q1 = q1 a1 = 0 p
a1 =
2

d2 = 0 q 2 = q2 a2 = 16 a2 = 0
d3 = 0 p a3 = 0 p
q3 = q3 + a3 =
2 2

d4 = 19.5 q 4 = q4 a4 = 0 p
a4 = -
2
d5 = 0 q 5 = q5 a5 = 0 p
a5 =
2

d6 = 6.715 q6 = q6 a6 = 0 a6 = 0
186 Artificial Intelligence (AI)

where i is the number of the link, αi is the joint rotation, ai is the length of the link, di
is the displacement of the link, ci = cos(ϴi), and si = sin(ϴi); therefore, ϴi is the rota-
tion angle of the joint.
The matrix that describes the location of the final system with respect to a refer-
ence system located at the base of the robot, also known as the homogeneous trans-
formation matrix, is obtained by making the product of the six matrices obtained
from (9.5) [10].

T60 = 0 A1 × 1 A2 × 2 A3 × 3 A4 × 4 A5 × 5 A6 (9.8)

Therefore, each of the transformation matrices of the links can be calculated as

shown below:

éc1 0 s1 0 ù
ê ú
s1 0 -c1 0 ú
0
A1 = ê (9.9)
ê0 1 0 a0 + a1 ú
ê ú
ë0 0 0 1 û
éc2 -s2 0 a2 * c2 ù
ê ú
s2 c2 0 a2 * s2 ú
1
A2 = ê (9.10)
ê0 0 1 0 ú
ê ú
ë0 0 0 1 û
é -s3 0 c3 0ù
ê ú
c3 0 s3 0ú
2
A3 = ê (9.11)
ê 0 1 0 0ú
ê ú
ë 0 0 0 1û
é c4 0 - s4 0 ù
ê ú
s4 0 c4 0 ú
3
A4 = ê (9.12)
ê0 -1 0 a3 + a4 ú
ê ú
ë0 0 0 1 û
éc5 0 s5 0ù
ê ú
s5 0 -c5 0ú
4
A5 = ê (9.13)
ê0 1 0 0ú
ê ú
ë0 0 0 1û
éc6 -s1 0 0 ù
ê ú
s6 c6 0 0 ú
5
A6 = ê (9.14)
ê0 0 1 a5 + a6 ú
ê ú
ë0 0 0 1 û
Robust Design of Artificial Neural Networks 187

é nx ox ax px ù
ê ú
ny oy ay py ú
T06 = ê (9.15)
ê nz oz az pz ú
ê ú
ë0 0 0 1û
where

nx = -(c1 × (c2 × (c3 × c6 × s5 + (c4 × c5 × c6 - s4 × s6 ) × s3 ) + (c3 × (c4 × c5 × c6 - s4 × s6 )

(9.16)
-c6 × s3 × s5 ) × s2 ) - (c4 × s6 + c5 × s4 ) × s1 )

ny = -(c1 × (c4 × s6 + c5 × c6 × s4 ) + (c2 × (c3 × c6 × s5 + (c4 × c5 × c6 - s4 × s6 ) × s3 )

(9.17)
+ (c3 × (c4 × c5 × c6 - s4 × s6 ) - c6 × s3 × s5 ) × s2 ) × s1 )

nz = c2 × (c3 × (c4 × c5 × c6 - s4 × s6 ) - c6 × s3 × s5 )
(9.18)
- (c3 × c6 × s5 + (c4 × c5 × c6 - s4 × s6 ) × s3 ) × s2

ox = c1 × (c2 × (c3 × s5 × s6 + (c4 × c5 × s6 + c6 × s4 ) × s3 )

(9.19)
+ (c3 × (c4 × c5 × s6 + c6 × s4 ) - s3 × s5 × s6 ) × s2 ) + (c4 × c6 - c5 × s4 × s6 ) × s1

oy = -(c1 × (c4 × c6 - c5 × s6 × s4 ) - (c2 × (c3 × s6 × s5 + (c4 × c5 × s6 + s4 × c6 ) × s3 )

(9.20)
+ (c3 × (c4 × c5 × s6 + s4 × c6 ) - s3 × s5 × s6 ) × s2 ) × s1 )

oz = -(c2 × (c3 × (c4 × c5 × s6 + c6 × s4 ) - s3 × s5 × s6 )

(9.21)
- (c3 × s5 × s6 + (c4 × c5 × s6 + c6 × s4 ) × s3 ) × s2 )

a x = c1 × (c2 × (c3 × c5 - c4 × s3 × s5 ) - (c3 × c4 × s5 + c5 × s3 ) × s2 ) + s1 × s4 × s5 (9.22)

a y = -(c1 × s4 × s5 - (c2 × (c3 × c5 - c4 × s3 × s5 ) - (c3 × c4 × s5 + c5 × s3 ) × s2 ) × s1 (9.23)

az = c2 × (c3 × c4 × s5 + c5 × s3 ) + (c3 × c5 - c4 × s3 × s5 ) × s2 (9.24)

px = a2 × c1 × c2 + a3 × c1 × (c2 × c3 - s2 × s3 ) + a4 × c1 × (c2 × c3 - s2 × s3 ) + (a5 + a6 )

(9.25)
× (c1 × (c2 × (c3 × c5 - c4 × s3 × s5 ) - (c3 × c4 × s5 + c5 × s3 ) × s2 ) + s1 × s4 × s5 )

py = a2 × c2 × s1 + a3 × (c2 × c3 - s2 × s3 ) × s1 + a4 × (c2 × c3 - s2 × s3 ) × s1 - (a5 + a6 )

(9.26)
× (c1 × s4 × s5 - (c2 × (c3 × c5 - c4 × s3 × s5 ) - (c5 × c4 × s5 + c5 × s3 ) × s2 ) × s1 )
188 Artificial Intelligence (AI)

pz = a0 + a1 + a2 × s2 + a3 × (c2 × s3 + c3 × s2 ) + a4 × (c2 × s3 + c3 × s2 ) + (a5 + a6 )

(9.27)
× (cc2 × (c3 × c4 × s5 + c5 × s3 ) + (c3 × c5 - c4 × s3 × s5 ) × s2
It can be seen that Equations 9.16–9.24 represent the orientation array [n o a] of
the end of the robot based on joint coordinates (q1, q2 , q3, q4, q5, q6) and Equations
9.25–9.27 represent the position array [p] according to the joint coordinates and the
lengths of the links (a0, a1, a2, a3, a4, a5, a6).
When solving (9.8), the position and initial orientation of the final effector is
obtained with respect to the reference system located at the base of the robot as
shown below:

é0 0 1 42.215ù
ê ú
0 -1 0 0 ú
T06 = ê (9.28)
ê1 0 0 20.2 ú
ê ú
ë0 0 0 1 û
Where the initial position of the final effector of the Ketzal robot was px = 42.215
cm, py = 0 cm, and pz = 20.2 cm and the orientation obtained was nx = 0, ox = 0, ax = 1,
ny = 0, oy = −1, ay = 0, nz = 1, oz = 0, and az = 0. This confirms that the data obtained
coincide with the initial position and orientation of the robot arm with the calculation
of direct kinematics.
The simulation of the graphic representation of the Ketzal robot was carried out
with the Robotics Toolbox for MATLAB software [56]. Figure 9.9 shows the initial
position and orientation of the final effector of the robotic manipulator.

FIGURE 9.9  Graphic representation of the robotic manipulator.

Robust Design of Artificial Neural Networks 189

9.3.1 Data Set Description

Based on the geometry and dimensions of the manipulator, the workspace of the
robotic manipulator can be represented by the position array, orientation array, and
joint values. This representation can be defined as an infinite set of coordinates of
position, orientation, and articular values.
Therefore, it is very important to define the size of the training data set for the
neural network, since the volume of the data is an important factor that has a decisive
influence on the available processing capabilities [57].
In this study, the possibility of obtaining representative samples in the entire
workspace was analyzed, so that the selected data guaranteed a representative sam-
ple of the entire original set, minimizing the risk that some or a large part of the
population was not represented [58, 59].
The data set was generated from Equations 9.16–9.27 and according to the geom-
etry of the robot, the articular values of (ϴ1ϴ6) were established in the range
shown in Table 9.5.
Generally, the amount of data depends on the spatial resolution used in the range
of joint values. For example, if a jump of Δϴ1 = π/2 is used to generate the range of
values in ϴ1, only five values [0, π/2, π, 3π/2, 2π] will be considered; on the contrary,
if a jump of Δϴ1 = π/5 is used, 11 values are considered, allowing to have a better
spatial resolution in the data set.
In this study, four data sets were generated by applying the direct kinematics
model described above. The data sets generated can be described by a matrix of rows
× columns represented by the following equation:

é a11 a12 ... a1c ù

ê ú
ê a21 a22 ... a2c ú
Data _ set = ê ú (9.29)
ê ... ... ... .... ú
êa ar 2 ... arc ú
ë r1 û
where the subscript r represents the number of data generated and the subscript c
represents the number of variables used in this work, where the elements ar1, ar2, and
ar3 correspond to the data of the position array [p] = {px, py, pz}, the elements ar4ar12
correspond to the data of the three-orientation array [n o a] = {nx, ny, nz, ox, oy, oz, ax,

TABLE 9.5
Angular Ranges in the Joints of the
Ketzal Robot
rad θ1 θ2 θ3 θ4 θ5 θ6
Minimum 0 0 2π 0 2π 0
Maximum 2π π p 2π p 2π
2 2
190 Artificial Intelligence (AI)

ay, az}, and finally the elements ar13ar18 correspond to the array of articular values
[ϴ] = {ϴ1, ϴ2, ϴ3, ϴ4, ϴ5, ϴ6}.
The spatial resolution used in this work was 17 × 17 × 17 × 17 × 17 × 17 × 17 with
a total of 24 137 569 data with 18 variables. The jump values where set to Δϴ1 = π/8,
Δϴ2 = π/16, Δϴ3 = π/16, Δϴ4 = π/8, Δϴ5 = π/16, and Δϴ6 = π/8. The 18 variables con-
sidered in the data set generated correspond to the position array [p], orientation
array [n o a], and joint value array [ϴ].
To guarantee better performance in the knowledge extraction algorithms, the data
pre-processing stage is fundamental, since the efficiency in the results of perfor-
mance and knowledge extraction depends largely on the quality of the data, and low
quality regularly leads to low quality in the knowledge extracted [57].
The techniques generally used in data pre-processing are divided into two areas:
data preparation and data reduction. Data preparation is generally an obligatory
technique and consists of a series of algorithms whose purpose is to prepare the data
so that the knowledge extraction algorithms can be executed without difficulties dur-
ing processing. This area includes data transformation, normalization, cleaning, and
probably the recovery of lost values [40].
On the other hand, data reduction techniques are not always mandatory, but
sometimes they help to obtain better results, where the main objective in this type
of techniques is to obtain a reduced amount of the original data, keeping as much
as possible the integrity of the information in the data, highlighting the methods of
feature selection, instance selection, grouping, compaction, and sampling [60, 61].
Theoretically, obtaining a complete data set will allow for a broader understand-
ing of the problem; however, if the population is too large, the processing time
required will be much more expensive, so it would be unfeasible.
In practice, sampling is one of the most appropriate processes to solve this prob-
lem due to the advantages obtained in the performance of time and cost of process-
ing, even better results can be obtained from the knowledge extracted compared to
the processing of data sets without sampling [57].
Therefore, data sampling can be widely applied in various engineering fields such
as statistics, machine learning, and data mining [59].
Systematic sampling, either alone or in combination with any other technique, is
one of the most used methods because of its simplicity and ease of use; in its simplest
form, it is also known as linear systematic sampling (LSS) and was introduced in
1944 in Ref. [62].
The LSS method can be seen as a sampling technique by cluster in its simple
form, where the procedure is to divide the population N into k groups of n elements
each. Taking only one group of the sample accordingly, each of the k groups has a
probability of 1/k to be selected; in this case, the application of systematic sampling
is feasible because N = k∙n [63].
LSS allows all units of the data set to have the same opportunities to be selected,
for example, if you have an initial set of size N from which you want to select a new
set of size n, a random number is selected for sampling between 1 and k to select
the instance k of the data set where k represents the sampling interval. From the
selected instance, the kth item is selected forward up to the n instances proposals.
Robust Design of Artificial Neural Networks 191

This procedure ensures that each unit has the same possibility of being included in
the sample [64].
The sampling interval constant is usually taken as the integer closest to N/n,
known as the inverse of the sampling fraction. This method has two advantages
mainly. Firstly, the selection of the first unit is determined by a sampling interval or
sampling period chosen, guaranteeing a random sample among a set of samples of
a given interval. Secondly, systematic samples are distributed in a good way among
the population, guaranteeing a representative sample of the entire original data set,
that is to say, there is less risk that some or a large part of the population will not be
represented, maintaining a constant and uniform distribution between the data [57].
Although it is one of the most common and simple methods, it has two main
drawbacks. Firstly, the sampling variance cannot be taken impartially based on the
only systematic sample taken. Secondly, when population size N cannot be divided
equally by the desired sample size n, systematic sampling cannot be performed, that
is, when N is not an integer multiple of the desired sample size n where N ¹ n∙k.
In this case, k is not an integer value and an inefficient sampling effect can prob-
ably be produced, that is, if at any time the characteristics of the population were
periodic and they coincide with a selected sampling interval, the representativeness
of the sample could be biased [63].
In this work, the data set proposed N1 is not an integer multiple of the desired size
n due to the spatial resolution used in generating the data set. It is clear that in this
case the criteria for applying the method LSS is not met, where N = n∙k.
To solve this problem, the reduction data filter (RDF) algorithm was designed
based on the LSS method in such a way that it allows obtaining a representative dis-
tribution of the original data set, minimizing the risk that some or part of the popula-
tion is not represented and that allows obtaining a constant and uniform distribution
for training and test data sets.

9.3.2 Description of Reduction Data Filter Algorithm

Working with a volume of data greater than 24 million × 18 variables could be a
difficult task for a common computer. An RDF algorithm was designed based on the
LSS method for solving the problem above, where the task processing is much easier
for a conventional processor.
Considering the data set A with a population of N1 = 24 137 569 where the sample
size desired is n = 1000, it is observed that N1 is not an integer multiple of n, due to
k = N1/n = 24 137 569, where k is not an integer that meets the criteria for applying
the LSS method.
To solve the problem, the RDF algorithm, shown in Figure 9.10, helps to fulfil
this purpose.
In general, the steps to follow for the implementation of the filter are described
below:

1. The values of the original population size N and the desired sample size n
are set.
192 Artificial Intelligence (AI)

FIGURE 9.10  RDF algorithm based on linear systematic sampling.

2. The sampling interval constant k = N/n is calculated, where k must not be an

integer value.
3. A random number between 1 and k is generated.
4. The random value is rounded to an integer r1.
5. Element r1 of the original data set is selected as the first filtered data.
6. The cumulative sum between the random number and the constant k is
made. In this step, the cumulative sum is a positive non-integer value.
7. The cumulative value of the sum is rounded to an integer r2.
8. Element r2 of the original data set is selected as the second filtered data and
steps 6, 7, and 8 are repeated until reaching the desired n samples.

According to the flowchart shown in Figure 9.11, in the first section the constant
n = 1000, which represents the amount of desired samples, is initialized. The vari-
able N represents the number of samples for the data set previously generated by
Equations 9.16–9.27.
The variable k represents the sampling interval, also known as the inverse of the
sampling fraction with a value of k = 24 137 569. The start variable represents the
random value between 1 and k. The index variable is the decimal value of the index
in the instance, while the index_r variable is the rounded-up integer value of the
index in the instance.
In the loop section of the iterative control structure for, the instance selection is
made from the kth element forward until reaching the 1000 instances.
In this process, the index variable accumulates the values of the indexes with
decimals and the value of the sampling interval k. Finally, the p variable calculates
the percentage of reduction applied by the proposed algorithm.
Robust Design of Artificial Neural Networks 193

FIGURE 9.11  Dispersion matrix of the position data set D: (a) Before filtering (b) After
filtering.
194 Artificial Intelligence (AI)

9.3.3 Data Set Analysis of Training and Test

Once the RDF was applied to data set A, a reduction to 1000 data with 18 vari-
ables was obtained. According to the pre-established range of motion for each of the
joints, the distribution of the position data was analysed before and after applying
the filter by means of scatter matrix diagrams shown in Figure 9.11, where it was
observed that the position data maintained a constant and uniform distribution in
their workspace with respect to the originally generated data set, as shown in Figure
9.11b, allowing to obtain a representative distribution of the original population.
Figure 9.12 shows the analysis of the distribution of the position and orientation
data set, corresponding to the input data for the training of the ANN. As shown in
Figure 9.12b, after applying the RDF algorithm, it was observed that the position and
orientation data maintained a constant and uniform distribution in their workspace
with respect to the data set A, allowing to obtain a representative distribution of the
original data set.
The distribution of the set of input data used for training and testing in the ANN
is shown in Figure 9.13. As described previously in Section 9.2, one of the noise
variables that are not controlled by the user is the random selection of training and
test sets, in this case, two sets with proportions of 80:20 and 90:10 were selected,
respectively.
Figure 9.13a shows the distribution of the input data of the position array [p];
at the top, three graphs of the distribution of the position data without a filter are
shown, of which the first corresponds to the position data, the second to the training
data, and the third to the test data with an 80:20 proportion, respectively.
In the lower part of the same figure, three graphs of data distribution with filters
are shown, of which the first corresponds to the position data, the second to the train-
ing data, and the third to the test data with a proportion of 90:10, respectively.
Figure 9.13b shows the distribution of the input data of the orientation array [n o
a]; at the top, three graphs of the distribution of the orientation data without a filter
are shown, of which the first corresponds to the orientation data, the second to the
training data, and the third to the test data with an 80:20 proportion, respectively.
In the lower part of the same figure, three graphs of data distribution with filters
are shown, of which the first corresponds to the orientation data, the second to the
training data, and the third to the test data with a proportion of 90:10, respectively.
Figure 9.14 shows the distribution of the output data set corresponding to the joint
array [ϴ]. The first three graphs in the upper part of the figure correspond to the
distribution of joint data without a filter in training and test data with proportions
of 80:20, respectively, and the next three graphs in the lower part correspond to the
distribution of the data of the joints with the filter applied in training and tests data
with proportions of 90:10, respectively.
In this graph it can be seen that the filtered data maintains a constant and uniform
distribution in the data regardless of the proportion used, maintaining a representa-
tive distribution with respect to the original population of the data.
The filtered data set was used to perform the experimentation stage and to deter-
mine the optimal parameters of the neural network as described in the next section.
Robust Design of Artificial Neural Networks 195

FIGURE 9.12  Distribution analysis of position and orientation data set A: (a) Before filter-
ing (b) After filtering.

9.3.4 Planning and Experimentation Stage

In the planning stage, the design and noise variables with their respective levels were
chosen, as shown in Table 9.6.
Table 9.7 shows the noise variables and their respective levels, where U is the ini-
tial set of random weights, V is the size of the training set versus the size of the test
set, that is, V = 90%/10%, 80%/20%, and W is the random selection of the training set
and test set, that is, W = Training1/Test1, Training2/Test2.
196 Artificial Intelligence (AI)

FIGURE 9.13  Distribution analysis of input data set with training and test: (a) Position (b)
Orientation.

OA cross-configuration was used in the experimental stage with L9(34) and L4(32),
where 36 different ANN architectures were trained and tested, as shown in Table 9.8.
The analysis of variance (ANOVA) was used with the JMP statistical program to
obtain the signal-to-noise ratio.

9.3.5 Analysis and Confirmation Stage

Once the results of the MSE were obtained in the 36 trainings of the proposed OA,
a statistical analysis in JMP was performed to identify the possible optimal values
corresponding to the possible best network topologies, as shown in Figure 9.15.
 Robust Design of Artificial Neural Networks

FIGURE 9.14  Distribution analysis of output data set with training and test.
197
198 Artificial Intelligence (AI)

TABLE 9.6
Design Variables and Their Levels
Design Variables Level 1 Level 2 Level 3
A 6 9 12
B 0 3 0
C 0.1 0.2 0.3
D 0.01 0.1 0.2

TABLE 9.7
Noise Variables and Their Levels
Noise Variables Level 1 Level 2
U Set 1 Set 2
V 9:1 8:2
W Tr-1/Ts-1 Tr-2/Tst-2

TABLE 9.8
OA with L9(34) and L4(32) Configuration
Trial
No. S1 S2 S3 S4 Mean S/N
1 0.34028409 0.33660998 0.33523831 0.33120539 0.3358344 −39.042614
2 0.34294367 0.34775807 0.34162374 0.34110013 0.3433564 −41.071318
3 0.33780647 0.33966576 0.34055964 0.33274657 0.3376946 −39.708489
4 0.31766289 0.32279780 0.32111454 0.31413515 0.3189276 −38.378724
5 0.33082154 0.32892629 0.32892339 0.31509698 0.3259421 −33.015696
6 0.32950029 0.31278398 0.31908650 0.31464315 0.3190035 −32.598799
7 0.30203761 0.29094929 0.30443676 0.29524616 0.2981675 −33.657940
8 0.32171345 0.31260463 0.32582416 0.30687483 0.3167543 −31.330813
9 0.30484374 0.30907752 0.31075605 0.29823255 0.3057275 −34.659736

The optimal ANN architecture is determined using the signal-to-noise ratio.

After carrying out the analysis of the signal-to-noise ratio, the values closest to the
red line on the X-axis are selected, whose values are highlighted in Table 9.9.
Figure 9.16 shows a cube diagram showing the effect on the interaction of the
factors involved and the response in the obtained design. The two cubes show all the
combinations of the configuration variables of the proposed factors and the adjusted
mean of each combination. The cube on the left shows the measurements of the
responses when the learning rate is equal to 0.1 and the cube on the right shows the
responses when the learning rate is 0.03.
The average value obtained from the MSE and the S/N ratio are two of the factors
that are considered to determine the appropriate levels in the variables involved in
Robust Design of Artificial Neural Networks 199

FIGURE 9.15  S/N analysis for the determination of optimal network parameters: (a) Best
topology with normal profile (b) Best topology with desirability profile (c) Best topology with
maximizing desirability profile.
200 Artificial Intelligence (AI)

TABLE 9.9
Best Values to Design the ANN
with Normal, Desirability and
Maximized Profiles
Trial No. A B C D
1 12 0 0.01 0.1
2 12 0 0.01 0.1
3 9 3 0.01 0.1

FIGURE 9.16  Variable response and interaction analysis.

the design of the network. The cube on the left describes that the best average value
of the MSE obtained is given by 12 neurons in the first layer, a moment equal to 0.01,
and a learning rate equal to 0.1. However, an increase in neurons in the second layer
can improve the S/N ratio but tends to decrease the yield by approximately 3% above
the MSE average.
On the other hand, in the cube on the right can be seen that 12 neurons in the first
layer, a moment equal to 0.01, and a learning rate equal to 0.3 are sufficient to obtain
the best performance. However, there is a tendency that coincides with the cube on
the left in including a second hidden layer to improve the S/N ratio, but the network
performance tends to decrease by 3% above the MSE average.
From the results obtained in the JMP statistical analysis, three trainings were
carried out with the parameters obtained in each of the profiles, and tests were per-
formed to validate the data obtained with the designed ANN. In the final validation,
the statistical correlation and chi-square tests were performed for the best and worst
prediction, as shown in Figures 9.17 and 9.18, respectively.
The cross-validation method was used with three data sets A, B, and C with data
sizes of 600, 1000, and 1500 elements with 18 variables each, respectively. Each
data set was divided into four subsets of the same size, as shown in Figure 9.19. The
validation set in each training was used to estimate the generalization error, that is,
Robust Design of Artificial Neural Networks 201

FIGURE 9.17  Best joints predicted and correlation test: (a) Best joints predicted (b)
Correlation test.
202 Artificial Intelligence (AI)

FIGURE 9.18  Worst joints predicted and correlation test: (a) Worst joints predicted (b)
Correlation test.
Robust Design of Artificial Neural Networks 203

FIGURE 9.19  Cross-validation for data set n.

TABLE 9.10
Results of the Fourfold Cross-Validation
MSE
Fold
Number A B C
1 0.30089341 0.31660459 0.33148271
2 0.32114314 0.32018527 0.33704913
3 0.30725039 0.32869244 0.34501222
4 0.31404578 0.32941529 0.34597482
Average 0.31083318 0.32372439 0.33987972
Std Deviation 0.00872260 0.00633206 0.00688028

the incorrect classification rate of the model with data different from those used in
the training process.
Table 9.10 shows the average MSE value and the standard deviation obtained dur-
ing the workouts performed for each of the sets proposed during the validation of the
neural network using the cross-validation method.
The average value of MSE obtained, and highlighted in Table 9.10 for data set B,
is approximately equal to the value obtained in the confirmation step. However, as it
is shown in Table 9.10, the average MSE value obtained during training with data set
A is approximately 4% less than that obtained with set B and the average MSE value
is approximately 5% higher than that obtained with the set B.
The use of the RDANN methodology allowed solving problems in the design
of the ANN critical parameters. The density of the data used was 80% for training
data and 20% for the test. The best architecture obtained in this work was 12:12:6, a
momentum = 0.01, a learning rate = 0.1, trainrp training algorithm and a MSE = 1E-4.

9.4 CONCLUSIONS AND DISCUSSIONS

In this research work, an RDF filter based on the LSS method was designed with the
aim of reducing the data set, and the RDANN methodology was applied to solve the
inverse kinematics problem in a 6 DOF robotic manipulator. During the optimization
204 Artificial Intelligence (AI)

process, an OA was selected in L9 (34) and L4 (32) configuration considering four

design variables and three noise variables.
Proper preparation of the data in the pre-processing stage was essential during
ANNs training. Consequently, the effect produced by applying the data filter allowed
obtaining a wider range and better predictive resolution in the workspace.
The RDANN methodology was successfully applied to solve the problem of
inverse kinematics in a six DOF robotic manipulator, where the most significant fac-
tors in this study were the number of hidden neurons in the first layer, the learning
rate, and the momentum constant.
The best topology obtained in this study was 12:12:6 with a pulse constant = 0.01,
a learning rate = 0.1, MSE = 1E-4, and the trainrp learning algorithm. The hyper-
bolic sigmoid tangent transfer function was used in the input and output layers. The
parameters obtained for the optimized network architecture showed high perfor-
mance and generalization capacity.
The integration of ANNs and optimization techniques is a powerful tool in the
design of structural parameters and improvement in the performance of ANNs. In
this work, the RDANN methodology allowed determining the optimal parameters
in the ANN design.
To validate the quality in the measurement of the results in the prediction of the
model, the cross-validation method confirmed that the proposed data set, the struc-
tural and learning parameters in the ANN architecture obtained by the RDANN
methodology, is statistically reliable, because the average MSE value obtained and
highlighted in Table 9.10 with data set B is approximately equal to the value obtained
in the confirmation stage.
According to the results obtained by the method of cross-validation, the per-
formance of the neural network could be improved to the use a smaller data set;
however, the corresponding statistical tests must be performed to confirm the gener-
alization capacity, since, as observed in the cross-validation method for this case, the
network performance is independent of the size and partition between the training
and test data.
The systematic and experimental strategy used in this work is an alternative that
considers the concept of robustness in the ANNs design process, due to the simulta-
neous incorporation of noise and design variables.
The RDANN methodology significantly decreases training time and effort in the
modelling phase compared to the traditional trial and error method that is usually
proposed by the previous experience of the researcher, where he can spend sev-
eral days to months testing different architectures and consuming a lot of time and
resources without guaranteeing the optimal configuration and the network perfor-
mance. Consequently, the RDANN methodology may take a few hours to determine
the optimal parameters of the network, allowing more time to solve the problem in
question.
It is advisable to use the RDANN methodology to obtain an optimal architecture
and minimize the training time instead of using a more complex architecture that
uses a long training time and possibly the effect of over-adjustment in the data and
reduced capacity in the generalization of the network; so it is not advisable to use
Robust Design of Artificial Neural Networks 205

a great architecture and stop the training of the network in a certain time to avoid
overfitting the data.
The initialization of the weights significantly affects the performance of the
ANN because a large amount of noise is introduced into the training data and
has an impact on the effects produced at the start, specifically on the joints ϴ1,
ϴ2 , ϴ4, and ϴ6. In this case, the negative value does not make physical sense
for the joints, because they cannot make movements outside the mechanically
realizable range.

FUTURE SCOPE
The robust design methodology can be used to find a better artificial neural net-
work configuration with high performance and generalization capability, reducing
the time spent determining the optimal architecture of the artificial neural network
compared to the trial and error approach. In the field of robotics, it is possible to
direct future work towards the use of the robust design of artificial neural networks
methodology presented here for the solution of the inverse kinematics of robots that
do not have an algebraic solution or even to find the direct kinematics of parallel
robots, which is very complex, unlike in serial manipulators.

ACKNOWLEDGEMENTS
This work was partially supported by CONACYT – Becas Nacionales de Posgrado
con la Industria under contract 431101/640582 and paid salary license program
for research IPN-COTEBAL under assignments CPE/COTEBAL/14/2018, CPE/
COTEBAL/33/2019, and CPE/COTEBAL/23/2020. This work was supported
by OMADS S.A. of C.V., an enterprise dedicated to innovation and technological
development.

REFERENCES
1. S. Li, Y. Zhang, and L. Jin, “Kinematic Control of Redundant Manipulators Using
Neural Networks,” IEEE Transactions on Neural Networks and Learning Systems, vol.
28, no. 10, pp. 2243–2254, Oct. 2017, doi:10.1109/TNNLS.2016.2574363.
2. Z. Zhou, H. Guo, Y. Wang, Z. Zhu, J. Wu, and X. Liu, “Inverse Kinematics Solution for
Robotic Manipulator Based on Extreme Learning Machine and Sequential Mutation
Genetic Algorithm,” International Journal of Advanced Robotic Systems, vol. 15, no.
4, p. 1729881418792992, Jul. 2018, doi:10.1177/1729881418792992.
3. G. Jiang, M. Luo, K. Bai, and S. Chen, “A Precise Positioning Method for a Puncture
Robot Based on a PSO-Optimized BP Neural Network Algorithm,” Applied Sciences,
vol. 7, no. 10, p. 969, 2017, doi:10.3390/app7100969.
4. R. Köker, “Reliability-Based Approach to the Inverse Kinematics Solution of Robots
Using Elman’s Networks,” Engineering Applications of Artificial Intelligence, vol. 18,
no. 6, pp. 685–693, Sep. 2005, doi:10.1016/j.engappai.2005.01.004.
5. M. Ono, M. Pavone, Y. Kuwata, and J. Balaram, “Chance-Constrained Dynamic
Programming with Application to Risk-Aware Robotic Space Exploration,” Autonomous
Robots, vol. 39, no. 4, pp. 555–571, Dec. 2015, doi:10.1007/s10514-015-9467-7.
206 Artificial Intelligence (AI)

6. W. Ji, and L. Wang, “Industrial Robotic Machining: A Review,” The International

Journal of Advanced Manufacturing Technology, vol. 103, no. 1, pp. 1239–1255, Jul.
2019, doi:10.1007/s00170-019-03403-z.
7. G. R. Lucas, “Industrial Challenges of Military Robotics,” Journal of Military Ethics,
vol. 10, no. 4, pp. 274–295, Dec. 2011, doi:10.1080/15027570.2011.639164.
8. B. Long, J. Yang, X. Chen, Y. Sun, and X. Li, “Medical Robotics in Bone Fracture
Reduction Surgery: A Review,” Sensors, vol. 19, no. 16, Feb. 2019, doi:10.3390/
s19163593.
9. A.-M. Zou, Z.-G. Hou, S.-Y. Fu, and M. Tan, “Neural Networks for Mobile Robot
Navigation: A Survey,” in Advances in Neural Networks - ISNN 2006, 2006, pp.
1218–1226.
10. A. R. J. Almusawi, L. C. Dülger, and S. Kapucu, “A New Artificial Neural
Network Approach in Solving Inverse Kinematics of Robotic Arm (Denso
VP6242),” Computational Intelligence and Neuroscience: CIN, vol. 2016, 2016,
doi:10.1155/2016/5720163.
11. D. Martins, and R. Guenther, “Hierarchical Kinematic Analysis of Robots,”
Mechanism and Machine Theory, vol. 38, no. 6, pp. 497–518, Jun. 2003, doi:10.1016/
S0094-114X(03)00005-3.
12. B. Karlik, and S. Aydin, “An Improved Approach to the Solution of Inverse Kinematics
Problems for Robot Manipulators,” Engineering Applications of Artificial Intelligence,
vol. 13, no. 2, pp. 159–164, Apr. 2000, doi:10.1016/S0952-1976(99)00050-0.
13. L. Jin, S. Li, J. Yu, and J. He, “Robot Manipulator Control Using Neural Networks:
A Survey,” Neurocomputing, vol. 285, pp. 23–34, Apr. 2018, doi:10.1016/j.
neucom.2018.01.002.
14. J. M. Ortiz, M. del R. Martínez, J. M. C. Viramontes, and H. R. Vega, “Robust Design
of Artificial Neural Networks Methodology in Neutron Spectrometry,” Artificial
Neural Networks - Architectures and Applications, Jan. 2013, doi:10.5772/51274.
15. X. Wu, and Z. Xie, “Forward Kinematics Analysis of a Novel 3-DOF Parallel
Manipulator,” Scientia Iranica. Transaction B, Mechanical Engineering, vol. 26, no. 1,
pp. 346–357, 2019, doi:10.24200/sci.2018.20740.
16. A. Hasan, Ali T. Hasan, and H. M. A. A. Al-Assadi, “Performance Prediction Network
for Serial Manipulators Inverse Kinematics Solution Passing Through Singular
Configurations,” International Journal of Advanced Robotic Systems, vol. 7, no. 4, pp.
11–24, Jan. 2011.
17. M. I. Petra, and L. C. da Silva, “Inverse Kinematic Solutions Using Artificial Neural
Networks,” Applied Mechanics and Materials, vol. 534, pp. 137–143, Feb. 2014,
doi:10.4028/www.scientific.net/AMM.534.137.
18. L. Aggarwal, K. Aggarwal, and R. J. Urbanic, “Use of Artificial Neural Networks for the
Development of an Inverse Kinematic Solution and Visual Identification of Singularity
Zone(s),” Procedia CIRP, 2014, vol. 17, pp. 812–817, doi:10.1016/j.procir.2014.01.107.
19. R. R. Kumar, and P. Chand, “Inverse Kinematics Solution for Trajectory Tracking
Using Artificial Neural Networks for SCORBOT ER-4u,” in 2015 6th International
Conference on Automation, Robotics and Applications (ICARA), Queenstown, New
Zealand, Feb. 2015, pp. 364–369, doi:10.1109/ICARA.2015.7081175.
20. Y. Zhang, D. Guo, and Z. Li, “Common Nature of Learning Between Back-Propagation
and Hopfield-Type Neural Networks for Generalized Matrix Inversion with Simplified
Models,” IEEE Transactions on Neural Networks and Learning Systems, vol. 24, no. 4,
pp. 579–592, Apr. 2013, doi:10.1109/TNNLS.2013.2238555.
21. R. Köker, T. Çakar, and Y. Sari, “A Neural-Network Committee Machine Approach to
the Inverse Kinematics Problem Solution of Robotic Manipulators,” Engineering with
Computers, vol. 30, no. 4, pp. 641–649, Oct. 2014, doi:10.1007/s00366-013-0313-2.
Robust Design of Artificial Neural Networks 207

22. F. Reyes, Robótica - Control de Robots Manipuladores. Alfaomega Grupo Editor,

2011.
23. A. O. Baturone, Robótica: Manipuladores y Robots Móviles. Marcombo, 2005.
24. Lee, “Robot Arm Kinematics, Dynamics, and Control,” Computer, vol. 15, no. 12, pp.
62–80, Dec. 1982, doi:10.1109/MC.1982.1653917.
25. J. J. Craig, Robótica. Pearson Educación, 2006.
26. X. He, S. Xu, and SpringerLink (Online Service), Process Neural Networks: Theory
and Applications. Springer Berlin Heidelberg, 2010.
27. S. A. Kalogirou, “Artificial Neural Networks and Genetic Algorithms in Energy
Applications in Buildings,” Advances in Building Energy Research, vol. 3, no. 1, pp.
83–119, Jan. 2009, doi:10.3763/aber.2009.0304.
28. A. K. Jain, Jianchang Mao, and K. M. Mohiuddin, “Artificial Neural Networks: A
Tutorial,” Computer, vol. 29, no. 3, pp. 31–44, Mar. 1996, doi:10.1109/2.485891.
29. J. Zupan, “Introduction to Artificial Neural Network (ANN) Methods: What They Are
and How to Use Them,” Acta Chimica Slovenica, vol. 41. 1994.
30. S. A. Ziaee, E. Sadrossadat, A. H. Alavi, and D. Mohammadzadeh Shadmehri,
“Explicit Formulation of Bearing Capacity of Shallow Foundations on Rock
Masses Using Artificial Neural Networks: Application and Supplementary Studies,”
Environmental Earth Sciences, vol. 73, no. 7, pp. 3417–3431, Apr. 2015, doi:10.1007/
s12665-014-3630-x.
31. T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, and Y. Uchikawa, “Trajectory Control
of Robotic Manipulators Using Neural Networks,” IEEE Transactions on Industrial
Electronics, vol. 38, no. 3, pp. 195–202, Jun. 1991, doi:10.1109/41.87587.
32. J. M. Ortiz, Ma. del R. Martinez, and H. R. Vega, “Robust Design of Artificial Neural
Networks Applying the Taguchi Methodology and DoE,” in Electronics, Robotics
and Automotive Mechanics Conference (CERMA’06), Sep. 2006, vol. 2, pp. 131–136,
doi:10.1109/CERMA.2006.83.
33. Ş. Karabulut, “Optimization of Surface Roughness and Cutting Force During AA7039/
Al2O3 Metal Matrix Composites Milling Using Neural Networks and Taguchi Method,”
Measurement, vol. 66, pp. 139–149, Apr. 2015, doi:10.1016/j.measurement.2015.01.027.
34. S. Sholahudin, and H. Han, “Simplified Dynamic Neural Network Model to Predict
Heating Load of a Building Using Taguchi Method,” Energy, vol. 115, pp. 1672–1678,
Nov. 2016, doi:10.1016/j.energy.2016.03.057.
35. T. P. Teixeira, C. M. Salgado, R. S. de F. Dam, and W. L. Salgado, “Inorganic Scale
Thickness Prediction in Oil Pipelines by Gamma-Ray Attenuation and Artificial
Neural Network,” Applied Radiation and Isotopes, vol. 141, pp. 44–50, Nov. 2018,
doi:10.1016/j.apradiso.2018.08.008.
36. V. Uraikul, C. W. Chan, and P. Tontiwachwuthikul, “Artificial Intelligence for
Monitoring and Supervisory Control of Process Systems,” Engineering Applications
of Artificial Intelligence, vol. 20, no. 2, pp. 115–131, Mar. 2007, doi:10.1016/j.
engappai.2006.07.002.
37. K. López-Linares et al., “Fully Automatic Detection and Segmentation of Abdominal
Aortic Thrombus in Post-Operative CTA Images Using Deep Convolutional Neural
Networks,” Medical Image Analysis, vol. 46, pp. 202–214, May 2018, doi:10.1016/j.
media.2018.03.010.
38. R. A. Teixeira, A. de P. Braga, and B. R. de Menezes, “Control of a Robotic Manipulator
Using Artificial Neural Networks with On-line Adaptation,” Neural Processing Letters,
vol. 12, no. 1, pp. 19–31, Aug. 2000, doi:10.1023/A:1009694129740.
39. B. K. Bose, “Neural Network Applications in Power Electronics and Motor Drives—
An Introduction and Perspective,” IEEE Transactions on Industrial Electronics, vol.
54, no. 1, pp. 14–33, Feb. 2007, doi:10.1109/TIE.2006.888683.
208 Artificial Intelligence (AI)

40. S. García, J. Luengo, and F. Herrera, Data Preprocessing in Data Mining, vol. 72.
Springer International Publishing, 2015.
41. S. S. Mahapatra, and A. Patnaik, “Parametric Optimization of Wire Electrical Discharge
Machining (WEDM) Process Using Taguchi Method,” Parametric Optimization of
Wire Electrical Discharge Machining (WEDM) Process Using Taguchi Method, no.
Journal, Electronic, 2006.
42. R. K. Roy, A Primer on the Taguchi Method, Second Edition. Society of Manufacturing
Engineers, 2010.
43. M. Ibrahim, N. Zulikha, Z. Abidin, N. R. Roshidi, N. A. Rejab, and Mohd Faizal
Johari, “Design of an Artificial Neural Network Pattern Recognition Scheme Using
Full Factorial Experiment,” Applied Mechanics and Materials, 465–466, pp. 1149–
1154, Dec. 2013, doi:10.40​28/ww​​w​.sci​​entif​​i c​.ne​​t ​/AMM​​.465-​​46​6​.1​​149.
44. J. Limon-Romero, D. Tlapa, Y. Baez-Lopez, A. Maldonado-Macias, and L. Rivera-
Cadavid, “Application of the Taguchi Method to Improve a Medical Device Cutting
Process,” The International Journal of Advanced Manufacturing Technology, vol. 87,
no. 9–12, pp. 3569–3577, Dec. 2016, doi:10.1007/s00170-016-8623-3.
45. T. Y. Lin, and C. H. Tseng, “Optimum Design for Artificial Neural Networks: An
Example in a Bicycle Derailleur System,” Engineering Applications of Artificial
Intelligence, vol. 13, no. 1, pp. 3–14, Feb. 2000, doi:10.1016/S0952-1976(99)00045-7.
46. M. S. Packianather, and P. R. Drake, “Modelling Neural Network Performance through
Response Surface Methodology for Classifying Wood Veneer Defects,” Proceedings of
the Institution of Mechanical Engineers, vol. 218, no. 4, pp. 459–466, Apr. 2004.
47. T. Y. Lin, and C. H. Tseng, “Optimum Design for Artificial Neural Networks: An
Example in a Bicycle Derailleur System,” Engineering Applications of Artificial
Intelligence, vol. 13, no. 1, pp. 3–14, 2000, https​:/​/do​​i​.org​​/10​.1​​016​/S​​0​952-​​1976(​99)00​
045-7​.
48. K. Vahab, “An Optimized Artificial Neural Network for Human-Force Estimation:
Consequences for Rehabilitation Robotics,” Industrial Robot: An International
Journal, vol. 45, no. 3, pp. 416–423, Jan. 2018, doi:10.1108/IR-10-2017-0190.
49. S. Kucuk, and Z. Bingul, “Inverse Kinematics Solutions for Industrial Robot
Manipulators with Offset Wrists,” Applied Mathematical Modelling, vol. 38, no. 7, pp.
1983–1999, 2014, doi:https​://do​i.org​/10.1​016/j​.apm.​2013.​10.01​4.
50. L. Zhang, and N. Xiao, “A Novel Artificial Bee Colony Algorithm for Inverse
Kinematics Calculation of 7-DOF Serial Manipulators,” Soft Computing, vol. 23, no.
10, pp. 3269–3277, May 2019, doi:10.1007/s00500-017-2975-y.
51. D. L. Pieper, “The Kinematics of Manipulators under Computer Control,” Physiol.D.,
Stanford University, Ann Arbor, 1969.
52. M. Tarokh, and M. Kim, “Inverse Kinematics of 7-DOF Robots and Limbs by
Decomposition and Approximation,” IEEE Transactions on Robotics, vol. 23, no. 3,
pp. 595–600, Jun. 2007, doi:10.1109/TRO.2007.898983.
53. S. Tejomurtula, and S. Kak, “Inverse Kinematics in Robotics Using Neural
Networks,” Information Sciences, vol. 116, no. 2, pp. 147–164, Jan. 1999, doi:10.1016/
S0020-0255(98)10098-1.
54. A. Larrañaga, “3D Printable Robotic Arm,” GitHub, 2018. https://github​.com​/AngelLM
(accessed Sep. 18, 2019).
55. J. Denavit, and R. S. Hartenberg, “A Kinematic Notation for Lower-Pair Mechanisms
Based on Matrices.,” Transactions of the ASME Journal of Applied Mechanics, vol. 22,
pp. 215–221, 1955.
56. P. Corke, Robotics, Vision and Control: Fundamental Algorithms in MATLAB, vol. 73.
Springer, 2011.
Robust Design of Artificial Neural Networks 209

57. H. Liu, and H. Motoda, Instance Selection and Construction for Data Mining. Springer
US, 2001.
58. P. R. Krishnaiah and C. R. Rao, Eds., Sampling. Elsevier Science Publisher, 1988.
59. B. Gu, F. Hu, and H. Liu, “Sampling and Its Application in Data Mining,” Technical
Report TRA6/00, Department of Computer Science, National University of Singapur,
2000.
60. H. Liu, and H. Motoda, “On Issues of Instance Selection,” Data Mining and Knowledge
Discovery, vol. 6, no. 2, pp. 115–130, Apr. 2002, doi:10.1023/A:1014056429969.
61. H. Brighton, and C. Mellish, “Advances in Instance Selection for Instance-Based
Learning Algorithms,” Data Mining and Knowledge Discovery, vol. 6, no. 2, pp. 153–
172, Apr. 2002, doi:10.1023/A:1014043630878.
62. W. G. Madow, and L. H. Madow, “On the Theory of Systematic Sampling, I,” The
Annals of Mathematical Statistics, vol. 15, no. 1, pp. 1–24, 1944.
63. S. A. Mostafa and I. A. Ahmad, “Remainder Linear Systematic Sampling with Multiple
Random Starts,” Journal of Statistical Theory and Practice, vol. 10, no. 4, pp. 824–
851, Oct. 2016, doi:10.1080/15598608.2016.1231094.
64. L. H. Madow, “Systematic Sampling and its Relation to Other Sampling Designs,”
Journal of the American Statistical Association, vol. 41, no. 234, pp. 204–217, Jun.
1946, doi:10.1080/01621459.1946.10501864.
 10 Generative Adversarial
Network and Its
Applications
A. Vijayalakshmi and Deepa V. Jose

CONTENTS
10.1 Introduction................................................................................................... 211
10.2 Discriminative Learning vs. Generative Learning........................................ 212
10.3 Deep Generative Model................................................................................. 213
10.4 Variational Auto Encoders............................................................................ 214
10.5 Generative Adversarial Network................................................................... 215
10.6 Architecture of Generative Adversarial Network.......................................... 216
10.7 Variations of GAN Architectures.................................................................. 217
10.7.1 Fully Connected GAN (FCGAN)..................................................... 217
10.7.2 Laplacian Pyramid of Adversarial Networks (LAPGAN)................ 217
10.7.3 Deep Convolutional GAN (DCGAN)............................................... 217
10.7.4 Conditional GAN............................................................................... 217
10.7.5 Least-Square GAN............................................................................ 218
10.7.6 Auxiliary Classifier GAN.................................................................. 219
10.7.7 InfoGAN............................................................................................ 220
10.8 Applications of GAN..................................................................................... 220
10.8.1 Image generation............................................................................... 220
10.8.2 Image Translation.............................................................................. 221
10.8.3 Anomaly Detection............................................................................ 221
10.9 Conclusion..................................................................................................... 221
References............................................................................................................... 222

10.1 INTRODUCTION
Generative adversarial networks (GAN) is a deep learning unsupervised machine
learning technique that was introduced in 2014 by Ian J. Goodfellow and co-authors
[1]. The objective of GAN is generating data from scratch wherein it consists of two
neural networks that generate new data points from random uniform distribution
to get an accurate prediction. GANs are modelled as a methodology to generative
modelling that uses deep learning techniques like convolutional neural networks. In
order to generate new samples, the model automatically learns the patterns in the
input image data. GANs consist of two models as represented in Figure 10.1: genera-
tor and discriminator.
211
212 Artificial Intelligence (AI)

FIGURE 10.1  Generator and discriminator model.

The generator is a neural network that generates new data instances whose objec-
tive is to fool the discriminator. The discriminator evaluates the instances generated
by the generator for authenticity of the instances so as not to be fooled by the genera-
tor. The generator feeds as the input a random vector at each training step, thereby
preventing the model from producing the same sample every single time. The dis-
criminator receives the synthetic sample produced by the generator and outputs the
probability that the input image is a real training sample or a fake sample that is
generated by the generator predicting the authenticity of the input image.
Table 10.1 describes the steps taken by a generative adversarial network based on
the diagram depicted in Figure 10.1.
GAN is a deep learning method where two neural networks compete with each
other. Deep learning is a subset of machine learning based on artificial neural net-
works wherein there is a learning involved by improving its own computer algorithms
[2]. Neural network consists of layers made up of different numbers of nodes which
are connected to nodes of adjacent layers. Each node within the layer is assigned
a weight and the node with heavier weight will exert more effect on the next layer
of nodes. Deep learning is used in discovering hierarchical models that epitomize
probability distribution on data that is involved in artificial intelligence applications.

10.2 DISCRIMINATIVE LEARNING VS. GENERATIVE LEARNING

The two different approaches to classification techniques in machine learning are
generative and discriminative models. Discriminative model builds the output based

TABLE 10.1
Steps taken by GAN
• A generator takes random numbers as input and returns an image.
• The image so generated is fed to the discriminator together with the actual images from the data set.
• The real and fake images are taken by the discriminator and returns a probability
• The discriminator is in a feedback loop with the training set of images.
• The generator is in a feedback loop with the discriminator.
Generative Adversarial Network 213

FIGURE 10.2  Process of discriminative model.

FIGURE 10.3  Process of generative modelling.

on the observed data and quality of data rather than depending on the probabilistic
distribution. When provided with a good amount of data, discriminative models tend
to give better results. Figure 10.2 describes the process of a discriminative model.
Generative model understands the distribution of data and performs the task of
classification. Generative model creates similar data as the training data after learn-
ing the distribution of data. Figure 10.3 shows the process of generative modelling.
The training data consists of observations with many features for a problem of
image generation. Generative models can generate new sets of features that look
similar to the one that is created using the original data. In order to accomplish this,
the model should include a random element that builds a probabilistic distribution
model which affects each of the individual samples generated by the model.

10.3 DEEP GENERATIVE MODEL

In a deep learning model, the major challenge is the scarcity of training data.
Generative models are profoundly used in generating new data. For a healthcare
industry, where there is scarcity of data, generative models can be helpful in gener-
ating data for training the model and hence will play a major role in increasing the
size of data set. Translating satellite images to Google maps, converting a black and
white image to colour etc. are some of the applications of a generative model.
The generative model generates new data points by learning the data distribu-
tion from the training data set. To make it learn the exact distribution of data, a
neural network is used [3]. Deep generative model has difficulty in approximating
many inflexible probabilistic calculations in the generative context [1]. Deep genera-
tive models build generative models and also algorithms that are used for learning
214 Artificial Intelligence (AI)

FIGURE 10.4  Types of generative models.

the model using ideas from deep learning. Different types of generative models are
shown in Figure 10.4.
Explicit probabilistic model provides explicit parametric specification of the dis-
tribution of data that have controllable functions. Figure 10.5 shows a simple explicit
probabilistic model.
The explicit probabilistic model for Figure 10.4 is

( ) ( ) ( )
p x, z a = p x z p z a

Implicit probabilistic models define a stochastic process wherein after training of

the model, the samples are generated based on the underlying distribution of data.
Variational auto encoders (VAEs) and generative adversarial networks are examples
of generative model approaches.

10.4 VARIATIONAL AUTO ENCODERS

An artificial neural network that is engaged in training models that produce a recon-
structed input as the output of a network is termed as auto encoders. Auto encoders
consist of two connected networks, encoder and decoder, wherein an encoder takes
an input and produces a feature, as demonstrated in Figure 10.6.
Incorporating a decoder network and loss function in order to learn the features
and train the model, the auto encoder network is presented in Figure 10.7.
Generative Adversarial Network 215

FIGURE 10.5  Explicit probabilistic model.

FIGURE 10.6  Block diagram of auto encoder

FIGURE 10.7  Auto encoder network.

The neural network uses dimensionality reduction techniques in order to restruc-

ture the inputs so that input values will be better for a training model. A variational
auto encoder (VAE) is an architecture that is developed consisting of an encoder,
a decoder, and a loss function which in turn is trained in order to minimize the
error between the encoded–decoded data and the initial data that is generated while
reconstruction. The purpose of the decoder is to reconstruct the data and encoder
represents a neural network [4].
The objective of both encoders and variational auto encoders is to generate an
input data of lower dimension. Considering the input, x ∈ Rn, the objective of an
auto encoder is to produce h ∈ Rd where d < n. The output is designed such that it
contains the very vital features of the input, x, in order to reconstruct it. The differ-
ence between an auto encoder and a VAE lies in the fact that VAE is probabilistic
and a generative model. Hence, by sampling from the distribution, samples x ∈ Rn
can be obtained [5].

10.5 GENERATIVE ADVERSARIAL NETWORK

The objective of GAN is to generate images (mostly) from scratch [1, 6–8]. GAN
belongs to the category of generative model and it has the capability to handle sharp
density functions and generate preferred samples. GAN achieves this with the help
216 Artificial Intelligence (AI)

FIGURE 10.8  Summary of GAN training algorithm.

of a pair of networks with backpropagation signals. The samples generated by GAN

can be used in applications like image synthesis, image editing, image super-reso-
lution and classification, etc. [8]. The algorithm for generative adversarial networks
taken from original paper by Goodfellow, 2014, is depicted in Figure 10.8. The algo-
rithm consists of the process of training both the generator model and the discrimi-
nator model in parallel.

10.6 ARCHITECTURE OF GENERATIVE ADVERSARIAL NETWORK

Generative adversarial network belongs to the class of generative models that
can generate new content from the available data [1]. GAN is well accepted by
the deep learning community and its application spreads across in the domains
of computer vision, natural language processing, time series synthesis, semantic
segmentation, etc. [8].
Figure 10.9 represents the architecture of GAN that involves two major modules:
a discriminator (D) that helps in differentiating between a real image and generated
image and a generator (G) that produces images to deceive the discriminator in the
model. For a given distribution z∼pz, G defines a probability distribution pg as the
distribution of the samples G(z). The objective of GAN is to learn the generator’s
distribution pg that estimates the real data distribution pr. The joint loss function for
the discriminator D and generator G leads to the optimization of GAN.

ë ( )
min G max D V ( D, G ) = E x ~ pdata éë log D( x ) ùû + Ez ~ pz ( z ) é log 1 - D G ( Z ) ù
û
Generative Adversarial Network 217

FIGURE 10.9  Architecture of GAN.

10.7 VARIATIONS OF GAN ARCHITECTURES

10.7.1 Fully Connected GAN (FCGAN)
The first architecture for GAN depicted a fully connected neural network for both
the components, namely generator and discriminator [7], and this architecture was
implemented on simple image data sets like MNSIT handwritten digit data set.

10.7.2 Laplacian Pyramid of Adversarial Networks (LAPGAN)

CNN is considered to be best for image data [7]. The Laplacian pyramid of adversar-
ial networks (LAPGAN) offers a solution to the problem of difficulty in training the
generator and discriminator using convolutional neural networks. LAPGAN decom-
poses the process of generation using multiple scales. The image is decomposed
into a Laplacian pyramid. Further, a conditional, convolutional GAN is trained to
produce each layer [7, 8].

10.7.3 Deep Convolutional GAN (DCGAN)

Deep convolutional GAN is one of the effective network models of GAN. This net-
work works with continuous training of a pair of deep convolutional discriminator
and generator. DCGAN applied deconvolutional neural network architecture for the
generator and this model has demonstrated a good performance for CNNs visualiza-
tion [8]. This model consists of convolutional layers without max pooling and the
convolution in this model allows spatial down-sampling up-sampling operators that
would be learned during training the sample data [7]. Figure 10.10 illustrates the
architecture of the deep convolutional network.

10.7.4 Conditional GAN
GAN framework is extended to conditional setting [9] in which both generator and
discriminator class is conditioned with extra auxiliary information like class labels
218 Artificial Intelligence (AI)

FIGURE 10.10  Components of GAN architecture.

by feeding this information as an extra layer. This model provides better representa-
tion for multimodal data generation. Figure 10.11 illustrates the structure of a condi-
tional adversarial network.

10.7.5 Least-Square GAN
GAN works with the basic concept of simultaneously training the discriminator and
generator where the function of the generator lies in generating fake image samples
that look real. There are innumerable works showing the application of GAN in gen-
eration of new image samples but the quality of such generated images is limited.
GAN uses sigmoid cross-entropy loss function in the discriminator. LSGAN is pro-
posed based on the assumption that the sigmoid loss function will lead to vanishing
gradient when updating the generator [10]. Least-square GAN (LSGAN) transforms
the GAN through L2 loss function as a replacement for log loss function and hence
we could say that LSGAN uses the least square loss function for the discriminator.
In this method labels for fake and real data, a and b, are used as a–b coding scheme
is used in the discriminator. The objective function can be written as

1 1
( ( ) )
E x ~ pdata ( x ) éê( D( x ) - b ) ùú + Ez ~ pz ( z ) éê D G ( Z ) - a ùú
2 2
min D VLSGAN ( D) =
2 ë û 2 ë û
Generative Adversarial Network 219

FIGURE 10.11  Architecture of GAN.

1
( ( ) )
Ez ~ pz ( z ) éê D G ( Z ) - c ùú
2
min G VLSGAN (G ) =
2 ë û
In the above example, c represents the value of fake data that the generator and dis-
criminator want to believe. The advantage of LSGAN is that unlike regular GAN,
LSGAN produce superior quality images and they are more stable in the learning
procedure.

10.7.6 Auxiliary Classifier GAN

The auxiliary GAN (AC-GAN) aimed at improving the training of generative adver-
sarial networks for synthesizing images [11]. In this proposed method, a class label
in addition to the noise is available for each of the generated samples. The objective
functions defined by LS and LC are described as follows:

ë ( û ) ë (
LS = E é log P S = real Xreal ù + E é log P S = fake X fake ù
û )

ë ( û ) ë (
LC = E é log P C = c Xreal ù + E é log P C = c X fake ù
û )
220 Artificial Intelligence (AI)

The discriminator is trained so that LS + LC are maximized, and the generator is
trained so that the value of LC-LS is maximized. This method provided a new metric
for image discriminability. The result of the work proves to be more discriminable
than that of a model that generates images with low resolution.

10.7.7 InfoGAN
InfoGAN incorporates the information theory for the transformation of noise in an
image to latent codes that have predictable consequences in the output [12]. In this
method, the generator is split into two parts: traditional noise vector and latent code
vector. Further, the mutual information between the codes is maximized. This algo-
rithm is implemented by including a regularization term − λ I(c;G(z,c) − to the objec-
tive function of GAN.

( )
min G max D VI ( D, G ) = V ( D, G ) - l I c; G ( z, c )

Lambda is the regularization term, I(c;G(z,c) is the mutual information between the
latent code c and G(z,c), which is the generator output. InfoGAN is unsupervised and
it learns interpretable and untangled information on images. The advantage of this
method is that it adds negligible computation cost on top of GAN and hence makes
the whole system very easy to train the images.

10.8 APPLICATIONS OF GAN
The ability of GAN to generate immense new samples on demand led to its use
in a variety of real-time applications. The currently existing GAN applications can
be segregated into three broad categories: image generation, image translation, and
anomaly detection. This section gives an overview of the various applications under
these three main categories.

10.8.1 Image generation
The major applications related to this category belong to the generation of image data
sets of handwritten digits, photographs of small objects, and human faces based on the
MNIST, CIFAR-10, and Toronto Face Database [1]. Generation of quite realistic human
faces is described in Ref. [13], which was later adopted to generate images of other
objects. Realistic photographs [14] is another commonly used and widely accepted
application. It seemed to be so perfect and was unable to differentiate between the syn-
thetically generated and the real-time ones. Creation of impressive cartoon characters
[15], especially characters from the Japanese comic books, was extensively used.
Similarly, another fancy application is the photos to emojis [16] where the human
faces or signboards, etc. will be converted to various emojis. Face frontal view gen-
eration [17] found immense use in human face identification and verification applica-
tions. Generating new human poses [18] seems to be useful in textile and modelling
industries.
Generative Adversarial Network 221

10.8.2 Image Translation
Another quite useful application of GAN is the image translation and editing. In
many scenarios, we need to convert images into different image formats or perform
various operations on it to extract information. Image-to-Image translation [19] is
used in many areas like creation of Google maps from satellite images, converting
sketches or black and white images to colour, etc.
Image-to-image translation includes producing another engineered rendition of a
given picture with a particular adjustment, for example, making an interpretation of
a late spring scene to winter. Training a model for picture to-picture interpretation
commonly requires an enormous data set of combined models. These data sets can
be troublesome and costly to make, and at times infeasible. The CycleGAN [20] is
a technique that does the automatic training of the image to-image interpretation
models without combined models. The models are prepared in an unsupervised way,
utilizing an assortment of pictures from the source and target domain that should not
be connected in any way. This procedure is an effective method for accomplishing
outwardly amazing outcomes on a scope of a variety of use areas.
Similarly, text-to-image translation [21] and semantic image-to-photo translation [22]
are quite useful in many scientific and learning applications. Photograph editing is another
vibrant application for reconstructing photographs [23]. This also helps in enhancing spe-
cific features and adding or removing features. To visualize the faces based on age, face
ageing [17] application is useful. Blending the features from different images is possible
with photo blending [24]. Besides, all the best quality view of images can be assured
through the high resolution using super resolution [25] methods. If we want to fill some
areas of images which were removed for some reason, it is possible with photo inpaint-
ing [26]. This will retain the originality of the image. Clothing translation [27] is now a
popular application installed in cloth marts where the customer can view how the attire
looks on them. In fact, video prediction [28], especially in sports and entertainment, is
another application related to this area. We can also see a variety of 3D object generation
[29] currently used in design applications to augment the user experience.

10.8.3 Anomaly Detection
Enhancing the security for comfortable living is always the need of the hour, espe-
cially in this era of smart manufacturing. To ensure security and prevent intrusion
from any sort through predictions is currently in practice in many scenarios wherein
we are completely connected to the internet. Security analysis [30] can be achieved
through GANs. Similarly, intrusion detection [31–33] is another area where GANs
are of much help where models can be devised to predict the chances of intrusions by
monitoring the continuous data from the actuators and the sensors.

10.9 CONCLUSION
The enormous attraction received by the generative adversarial network is not
only because of the ability of the model to learn deep mappings but also because
222 Artificial Intelligence (AI)

of its capability to make use of unlabelled image data. This chapter gives a brief
introduction on GAN and its variations with possible applications. It is very much
clear that the influence of deep neural networks offers enormous openings for novel
­applications.

REFERENCES
1. I. J. Goodfellow et al., “Generative Adversarial Networks,” ArXiv14062661 Cs Stat,
Jun. 2014, Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1406​
.2661.
2. X. Dong, J. Wu, and L. Zhou, “How deep learning works –The geometry of deep learn-
ing,” ArXiv171010784 Cs Stat, Oct. 2017, Accessed: Dec. 16, 2020. [Online]. Available:
http://arxiv​.org​/abs​/1710​.10784.
3. R. Salakhutdinov, “Learning Deep Generative Models,” Annu. Rev. Stat. Its Appl., vol.
2, no. 1, pp. 361–385, Apr. 2015, doi: 10.11​46/an​nurev​-stat​istic​s-010​814-0​20120​.
4. C. Doersch, “Tutorial on Variational Autoencoders,” ArXiv160605908 Cs Stat, Aug.
2016, Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1606​.05908.
5. R. Wei, C. Garcia, A. El-Sayed, V. Peterson, and A. Mahmood, “Variations in
Variational Autoencoders - A Comparative Evaluation,” IEEE Access, vol. 8, pp.
153651–153670, 2020, doi: 10.1109/ACCESS.2020.3018151.
6. K. Wang, C. Gou, Y. Duan, Y. Lin, X. Zheng, and F.-Y. Wang, “Generative Adversarial
Networks: Introduction and Outlook,” IEEECAA J. Autom. Sin., vol. 4, no. 4, pp. 588–
598, 2017, doi: 10.1109/JAS.2017.7510583.
7. A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath,
“Generative Adversarial Networks: An Overview,” IEEE Signal Process. Mag., vol. 35,
no. 1, pp. 53–65, Jan. 2018, doi: 10.1109/MSP.2017.2765202.
8. Z. Wang, Q. She, and T. E. Ward, “Generative Adversarial Networks in Computer
Vision: A Survey and Taxonomy,” ArXiv190601529 Cs, Jun. 2020, Accessed: Dec. 16,
2020. [Online]. Available: http://arxiv​.org​/abs​/1906​.01529.
9. M. Mirza and S. Osindero, “Conditional Generative Adversarial Nets,” ArXiv14111784
Cs Stat, Nov. 2014, Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​
/1411​.1784.
10. X. Mao, Q. Li, H. Xie, R. Y. K. Lau, Z. Wang, and S. P. Smolley, “Least Squares
Generative Adversarial Networks,” p. 9.
11. A. Odena, C. Olah, and J. Shlens, “Conditional Image Synthesis With Auxiliary
Classifier GANs,” ArXiv161009585 Cs Stat, Jul. 2017, Accessed: Dec. 16, 2020.
[Online]. Available: http://arxiv​.org​/abs​/1610​.09585.
12. X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel, “InfoGAN:
Interpretable Representation Learning by Information Maximizing Generative
Adversarial Nets,” ArXiv160603657 Cs Stat, Jun. 2016, Accessed: Dec. 16, 2020.
[Online]. Available: http://arxiv​.org​/abs​/1606​.03657.
13. T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive Growing of GANs for
Improved Quality, Stability, and Variation,” ArXiv171010196 Cs Stat, Feb. 2018,
Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1710​.10196.
14. A. Brock, J. Donahue, and K. Simonyan, “Large Scale GAN Training for High Fidelity
Natural Image Synthesis,” ArXiv180911096 Cs Stat, Feb. 2019, Accessed: Dec. 16,
2020. [Online]. Available: http://arxiv​.org​/abs​/1809​.11096.
15. Y. Jin, J. Zhang, M. Li, Y. Tian, H. Zhu, and Z. Fang, “Towards the Automatic Anime
Characters Creation with Generative Adversarial Networks,” ArXiv170805509 Cs, Aug.
2017, Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1708​.05509.
Generative Adversarial Network 223

16. Y. Taigman, A. Polyak, and L. Wolf, “Unsupervised Cross-Domain Image Generation,”

ArXiv161102200 Cs, Nov. 2016, Accessed: Dec. 16, 2020. [Online]. Available: http://
arxiv​.org​/abs​/1611​.02200.
17. G. Antipov, M. Baccouche, and J.-L. Dugelay, “Face Aging With Conditional
Generative Adversarial Networks,” ArXiv170201983 Cs, May 2017, Accessed: Dec. 16,
2020. [Online]. Available: http://arxiv​.org​/abs​/1702​.01983.
18. L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and L. V. Gool, “Pose Guided Person
Image Generation,” p. 11.
19. P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-Image Translation with
Conditional Adversarial Networks,” in 2017 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), Honolulu, HI, Jul. 2017, pp. 5967–5976, doi: 10.1109/
CVPR.2017.632.
20. J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired Image-to-Image Translation
using Cycle-Consistent Adversarial Networks,” ArXiv170310593 Cs, Aug. 2020,
Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1703​.10593.
21. H. Zhang et al., “StackGAN: Text to Photo-realistic Image Synthesis with Stacked
Generative Adversarial Networks,” ArXiv161203242 Cs Stat, Aug. 2017, Accessed:
Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1612​.03242.
22. T.-C. Wang, M.-Y. Liu, J.-Y. Zhu, A. Tao, J. Kautz, and B. Catanzaro, “High-Resolution
Image Synthesis and Semantic Manipulation with Conditional GANs,” ArXiv171111585 Cs,
Aug. 2018, Accessed: Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1711​.11585.
23. G. Perarnau, J. van de Weijer, B. Raducanu, and J. M. Álvarez, “Invertible Conditional
GANs for image editing,” ArXiv161106355 Cs, Nov. 2016, Accessed: Dec. 16, 2020.
[Online]. Available: http://arxiv​.org​/abs​/1611​.06355.
24. H. Wu, S. Zheng, J. Zhang, and K. Huang, “GP-GAN: Towards Realistic High-
Resolution Image Blending,” ArXiv170307195 Cs, Aug. 2019, Accessed: Dec. 16, 2020.
[Online]. Available: http://arxiv​.org​/abs​/1703​.07195.
25. C. Ledig et al., “Photo-Realistic Single Image Super-Resolution Using a Generative
Adversarial Network,” in 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Honolulu, HI, Jul. 2017, pp. 105–114, doi: 10.1109/CVPR.2017.19.
26. D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros, “Context Encoders:
Feature Learning by Inpainting,” in 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Las Vegas, NV, USA, Jun. 2016, pp. 2536–2544, doi:
10.1109/CVPR.2016.278.
27. D. Yoo, N. Kim, S. Park, A. S. Paek, and I. S. Kweon, “Pixel-Level Domain Transfer,”
ArXiv160307442 Cs, Nov. 2016, Accessed: Dec. 16, 2020. [Online]. Available: http://
arxiv​.org​/abs​/1603​.07442.
28. C. Vondrick, H. Pirsiavash, and A. Torralba, “Generating Videos with Scene Dynamics,”
ArXiv160902612 Cs, Oct. 2016, Accessed: Dec. 16, 2020. [Online]. Available: http://
arxiv​.org​/abs​/1609​.02612.
29. J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum, “Learning a
Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling,”
ArXiv161007584 Cs, Jan. 2017, Accessed: Dec. 16, 2020. [Online]. Available: http://
arxiv​.org​/abs​/1610​.07584.
30. S. R. Chhetri, A. B. Lopez, J. Wan, and M. A. Al Faruque, “GAN-Sec: Generative
Adversarial Network Modeling for the Security Analysis of Cyber-Physical Production
Systems,” in 2019 Design, Automation & Test in Europe Conference & Exhibition
(DATE), Florence, Italy, Mar. 2019, pp. 770–775, doi: 10.23919/DATE.2019.8715283.
31. H. Zenati, C. S. Foo, B. Lecouat, G. Manek, and V. R. Chandrasekhar, “Efficient GAN-
Based Anomaly Detection,” ArXiv180206222 Cs Stat, May 2019, Accessed: Dec. 16,
2020. [Online]. Available: http://arxiv​.org​/abs​/1802​.06222.
224 Artificial Intelligence (AI)

32. H. Zenati, M. Romain, C. S. Foo, B. Lecouat, and V. R. Chandrasekhar, “Adversarially

Learned Anomaly Detection,” ArXiv181202288 Cs Stat, Dec. 2018, Accessed: Dec. 16,
2020. [Online]. Available: http://arxiv​.org​/abs​/1812​.02288.
33. D. Li, D. Chen, J. Goh, and S. Ng, “Anomaly Detection with Generative Adversarial
Networks for Multivariate Time Series,” ArXiv180904758 Cs Stat, Jan. 2019, Accessed:
Dec. 16, 2020. [Online]. Available: http://arxiv​.org​/abs​/1809​.04758.
 11 Applications of
Artificial Intelligence in
Environmental Science
Praveen Kumar Gupta, Apoorva Saxena,
Brahmanand Dattaprakash, Ryna Shireen Sheriff,
Surabhi Hitendra Chaudhari, Varun Ullanat,
V. Chayapathy

CONTENTS
11.1 Introduction................................................................................................... 225
11.1.1 Artificial Intelligence......................................................................... 225
11.1.2 Artificial Intelligence and Its Roles in the Environment................... 226
11.2 Technological Solutions................................................................................. 226
11.2.1 Autonomous and Connected Electric Vehicles................................. 226
11.2.2 Conservation Biology........................................................................ 227
11.2.3 Next-Generation Weather and Climate Prediction............................ 229
11.2.4 Smart Earth........................................................................................ 231
11.3 AI in the Monitoring Environment................................................................ 232
11.3.1 Monitoring Soil.................................................................................. 232
11.3.2 Monitoring Water.............................................................................. 233
11.3.3 Monitoring Air................................................................................... 235
11.4 Risks of Artificial Intelligence...................................................................... 235
11.4.1 Bias.................................................................................................... 235
11.4.2 Liability............................................................................................. 235
11.4.3 ASI (Artificial Superintelligence)...................................................... 236
11.5 Conclusion and Future .................................................................................. 236
Acknowledgement.................................................................................................. 237
References............................................................................................................... 237

11.1 INTRODUCTION
11.1.1 Artificial Intelligence
In 1956, artificial intelligence (AI) was established as an educational field of study
[1]. In the field of computer applications and science, artificial intelligence can be
described as the intelligence that is indicated by machines, as compared to the

225
226 Artificial Intelligence (AI)

natural intelligence manifested in humans [2]. The idea was based on the assump-
tion that human intelligence “can be described so accurately that a machine can be
trained to simulate it” [3]. The human mind is capable of conjuring logical reasons
and explanations for events that occur beyond the measure of their control. Through
logic and emotions, some problems may be straightforward for humans to solve, but
rather difficult to computationally reciprocate. For this reason, there are two classes
of models: functionalist and structuralist. Where the functionalist data correlates
data to the equivalent obtained computationally, a structural model casually imper-
sonates the cardinal functioning of the mind [4]. Artificial intelligence is broken into
three domains: mundane, expert, and formal tasks. Each of these three domains, in
turn, contains the fields artificial intelligence plays a role in. Formal tasks are known
for the role they play in the mathematical and gaming fields. Mathematics is not
exclusive to arithmetic but includes geometry, calculus, and logical math as well.
Artificial intelligence has been used to verify existing theorems as well as theorems
being worked on. Expert tasks include artificial intelligence applications in diagnosis
in the medical field, analysis of scientific data, analysis of financial data as well as
in engineering. In engineering, artificial intelligence is applied to error probing and
analysis, manufacturing processes, efficiency testing, and monitoring all manufac-
tured items and processes. Mundane tasks describe the role of artificial intelligence
in language processing, perception, logic, and reasoning. Perception includes speech
and voice applications as well as computer vision [5].

11.1.2 Artificial Intelligence and Its Roles in the Environment

Artificial intelligence is now being used to improve environmental conditions, which is
in a dire state and requires improvement as soon as possible. From water quality testing
to air quality testing, the applications of artificial intelligence are limitless and we just
need to create suitable algorithms for it to be achieved. Currently, IBM is in the lead in
attempting to reach new heights in the field of research related to improving environ-
mental conditions. From their new radical recycling process, to obtaining real-time data
from oceans, to testing soil and water conditions, IBM methods of bettering the envi-
ronment using artificial intelligence show the capability of machine intelligence. The
further subtopics in this chapter will describe the various methods IBM and the rest of
the world have undertaken to ameliorate the environmental state [6].

11.2 TECHNOLOGICAL SOLUTIONS
11.2.1 Autonomous and Connected Electric Vehicles
The automotive industry is on the verge of the age of automated vehicles. Connected
and Autonomous Electric Vehicles (CAEV) works on three underlying technolo-
gies: Connected Vehicles (CV), Autonomous Vehicles (AV), and Electric Vehicles
(EV). The amalgamation of facilities offered by CAEVs plays a significant function.
CAEVs can provide services for individuals and products shipped, such as self-driv-
ing, advanced communications, and greatly enhanced mobility. Vehicle knowledge
AI in Environmental Science 227

and networking are part of an evolving response to the current problems of travel in
the world. These advance swiftly to provide fully controlled and partly autonomous
vehicles. By 2030, roads are going to be dominated by these vehicles, generating
significant advantages such as improvised equity, safer roads, and transport access.
Thus, a considerable reduction of greenhouse gases and congestion of traffic can be
expected.
The Society of Automotive Engineers states that there are six levels of driving
autonomy from level 0 (no automation) to level 5 (self-driving). Intermediate levels
(levels 1, 2, 3, and 4) are supposed to gain positive advantages, such as safer roads.
The automation level 5 is necessary to achieve the maximum economic benefit of
AV. That’s when cars don’t need drivers. An autonomous vehicle must be adaptive,
technical, and networking to its environment. Another significant consideration is
that energy is required to reduce GHG emissions as desired. An electricity source
should be incorporated in a vehicle to grid [V2G], a grid to the vehicle [G2V] to
coordinate its charging and unload operations and autonomously reduce its energy
usage. Many advantages can be seen as on-demand technologies such as mobility as
a service, dissipation of traffic congestion as a service, electricity storage as a service,
cellular charging as a service, sensing as a service, computation as a service, and
content distribution as a service. Such programs may be used by citizens, municipal
governments, and non-governmental organizations, and even by corporations. Such
services include a portal for accepting demands for facilities, prices, bill process-
ing, and vehicle assignment. Yet, in comparison to current systems for the software
administration of automotive clouds and V2G, this framework handles heterogeneous
vehicle service distribution. Therefore, the platform has to use appropriate strategic
planning and customized selection plans by integrating service characterizations and
accurate pricing specifications to achieve effective management [7].

11.2.2 Conservation Biology
The revolutionary advances in artificial intelligence have unlocked the ability to
quickly process a variety of signals, accurately identify risks, and provide real-time
alerts to the conservationists. Several AI models have been developed to take in
these signals, such as images, video, and audio, and use it to monitor the threat
faced by different wildlife populations. Machine learning (ML), a subset of AI, is
consistently being employed to develop algorithms for predicting the risk of extinc-
tion of several species [8], assess the overall footprint of fisheries [9], and identify
animals using sensor data obtained from biodiversity hotspots [10]. Initiatives by
tech giants such as Microsoft’s “AI for Earth” and Google’s “AI for Social Good”
are also contributing greatly to wildlife and biodiversity conservation. Deep learn-
ing (DL), yet another subset of AI, is highly effective in identifying patterns. Unlike
machine learning, deep learning can work on unlabelled data sets, such as video,
audio, and images, and hence is becoming a powerful tool in conservation biology.
For example, a fully trained deep learning model can be fed hours of signals col-
lected from the field to recognize unique species of organisms with remarkable accu-
racy. The data can also be collected from open-source repositories such as Google
228 Artificial Intelligence (AI)

Earth for satellite data [11] and Move Bank for animal tracking data [12] and fed
into ML or DL algorithms. In 2019, Google along with several organizations such
as Conservation International, developed “Wildlife Insights,” an AI-powered cloud-
based platform to store and process the millions of photos taken by biologists and
conservationists to streamline the conservation monitoring process. With the help
of AI, Wildlife Insights can label up to 3.6 million photos every hour, several times
faster than human experts [13].
Many projects are being conducted around the world using a combination of
AI and signals from the field to address specific problems. For example, Microsoft
AI for Earth and Gramener has developed a deep learning algorithm to assist the
Nisqually Foundation in the United States to identify salmon species from underwa-
ter camera trap videos. The algorithm was able to process several hours of video to
first identify the frames with salmon in it, and then subsequently identify the type of
salmon in those frames [14].
Automated browser-based tools for detecting and counting animals in images and
identifying individual animal species are also developed. One such tool is Wildbook, a
software framework produced by Wild Me in Portland, Oregon. Wildbook uses a com-
bination of neural networks and computer vision for animal detection and identification.
The team also uses AI to gather pictures from YouTube and Twitter of a particular ani-
mal. Researchers and scientists can start projects for a particular species by submitting
many manually annotated images to train the algorithm. Some examples of these proj-
ects include that for whale sharks (Rhincodon typus), manta rays (Manta birostris and
M. alfredi), Iberian lynxes (Lynx pardinus), and giraffes (Giraffa sp.) (Figure 11.1) [15].

FIGURE 11.1  Salmon species identification solution developed by Gramener and

Microsoft [17].
AI in Environmental Science 229

Another example is “iNaturalist” one of the world’s most powerful citizen sci-
ence platforms. It was recently improved by the collaboration between the California
Academy of Sciences (the Academy) and the National Geographic Society.
Crowdsourced images, which can be clicked even with the simplest of cameras,
can be submitted on to this platform to help scientists all over the world track the
response of various species to climate change and devise solutions for the same. The
iNaturalist AI has learned to recognize more than 24,000 types of plants, animals,
and fungi by analyzing more than 65 million images submitted by ordinary citi-
zens. The network also shares its findings with the Global Biodiversity Information
Facility (GBIF), enabling scientists’ access to verified data [16].
AI is especially helpful in providing quick solutions for high-risk situations such
as poaching and climate change. For example, in the African subcontinent, monitor-
ing elephant population can be cumbersome due to its very dispersed population
across a large surface area. Here, AI can be used along with aerial images generated
by organizations such as “Save the Elephants” to identify concentrations of elephants
most susceptible to poaching [17].
The new tools presented by AI along with the rapid influx of data generated have
presented new opportunities for conservation biologists all over the world. AI has
enabled us to process large amounts of data in very less amount of time with high
accuracy, sometimes even better than human experts. It saves time and effort for
these experts, enabling them to focus on devising solutions based on the information
generated by the AI algorithms.

11.2.3 Next-Generation Weather and Climate Prediction

Severe thunderstorms, tornadoes, and hurricanes are high-impact weather phenom-
ena that cause severe damages, loss of property, and life losses. High-impact activi-
ties may have a positive effect on society, such as the effect of renewable energy
on savings. Predicting these incidents has substantially improved with expanded
computational capacities, increasing computational power, and enhanced model
mechanics; however, there is a scope for improvement. By enhancing precision,
artificial intelligence and data science technology, especially machine learning and
data mining, fill the difference between the predictive system forecasts and feed-
back in real-time. Artificial intelligence strategies often derive potentially inac-
cessible knowledge from prediction models by merging model performance with
studies to get the forecasters and consumers with additional decision help. Applying
AI strategies perhaps with a physical knowledge of the atmosphere will signifi-
cantly improve the predictive ability for multiple high effect weather forms. The
AI method is also a key contributor to computational sustainability in the growing
field. AI techniques will analyze big data,” offer clarity into weather patterns of
high effect, and strengthen our awareness of high-impact weather [18]. It has helped
to achieve 90% accuracy in predicting cyclones, transition zones between masses
of air at Earth’s surface called weather fronts, and narrow corridors of moisture in
the atmosphere called atmospheric rivers which cause high precipitation and is not
possibly predicted by humans (Figure 11.2).
230 Artificial Intelligence (AI)

FIGURE 11.2  An atmospheric river over California [19].

The huge data sets needed and the Earth atmosphere’s intrinsic unpredict-
ability make forecasting upcoming events very difficult indeed. Existing com-
puter systems are needed to make decisions on several phenomena on a massive
scale.
These include issues such as how the Sun heats the Earth’s atmosphere, how pres-
sure changes affect wind patterns, and how temperature-changing processes (ice to
water to vapor) impact the flow of energy through the atmosphere.
We also need to remember the rotation of the Earth in space which helps to churn
the atmosphere all day long. Some tiny shifts in one aspect will affect the potential
events profoundly.
AI may be used to find correlations and create a valid hypothesis, generalizing the
data, utilizing computer-generated mathematical programs, and analytical problem-
solving techniques on broad results sets.
Given the inherent difficulty of weather forecasting, scientists are now using
AI for climate prediction to achieve improved and precise performance. By using
deep mathematical machine learning, AI may learn to predict the future from past
weather data.
The Numerical Weather Prediction (NWP) is an example. This model analyses
and examines large satellite data sets as well as other sensor data sets to produce
short-term weather predictions and long-term temperature projections.
Many companies currently also invest heavily in forecasting AI environment.
IBM, for instance, recently acquired the Weather Service, integrating its data with
its in-house Watson AI growth.
This has led to the development of IBM’s Deep Thunder which provides hyper-
local weather forecasts to customers within a resolution of 0.2–1.2 miles [20].
AI in Environmental Science 231

11.2.4 Smart Earth
While AI offers significant potential to solve the environmental challenges con-
fronting the Earth left undirected, it also can accelerate deterioration of the climate.
This research focuses on leveraging AI systems today and providing maximum
significant effect in pressing environmental challenges. This proposes strategies
by which AI can effectively reinvent mainstream industries and processes tackling
climate change, provide security for food and water, preserve ecosystems, and
foster human well-being. This problem seems closely related to the ongoing issue
of how to guarantee AI isn’t detrimental to humans. The main aim for creating
“safe” AI is proving that it is value-aligned – the vision of a positive scope based
on the ideals of humanity, offering secure implementation of the technology to
humanity. This implies that checks and balances built to ensure that emerging AI
processes stay “secure” will include the well-being of the natural environment as
a fundamental aspect [21].
In the 1940s, the first concrete steps were taken toward artificial intelligence. AI
is extensively used in our daily lives and caused a revolutionary vision consisting of
six factors:

• Big data: Machines give exposure where huge quantities of data, either
organized (in databases and spreadsheets) or unstructured (such as text,
audio, video, and images), can be accessed. All this evidence records our
experiences and increases the understanding of the world for humans. Big
data will only get larger, as trillions of sensors are installed in devices,
products, shoes, autonomous vehicles, and elsewhere. AI-assisted analy-
sis of this knowledge helps us to use this evidence to uncover historical
patterns, to forecast more accurately, to render decisions more efficient,
and more.
• Power consumption: Emerging developments such as cloud computing
and graphics processing units have become simpler and easier for compli-
cated AI-enabled applications to handle large volumes of data by parallel
processing.
• Globally linked: Social networking sites have radically transformed the
way people communicate. Such expanded openness has facilitated informa-
tion exchange and enabled knowledge sharing, leading to the development
of “collective intelligence,” also comprising open platforms that involve
devising AI software and app sharing.
• Open-source applications and information: Open-source applications and
information facilitate the modernization and implementation of AI, as seen
in the success of open-source machine learning frameworks and platforms
such as TensorFlow, Caffe2, PyTorch, and Parl​.a​i.
• Enhanced algorithms: Scientists also achieved progress in various areas of
AI, especially in “deep learning,” involving layers of neural networks, con-
structed in a way that is inspired by the approach of the human brain to pro-
cessing information. Another recent research field is “extreme learning,” in
232 Artificial Intelligence (AI)

which the AI programmer learns by trials and errors guided by a reward

feature, with little or no initial input data.
• Accelerating returns: Market pressures drive the growth of AI since com-
panies use enhanced analytics and open-source tools to increase their mar-
ket benefits and raise their yields by, for instance, rising consumer product
personalization or using smart automation to boost productivity.

The combination of these variables helped to push AI from in vitro (in research labo-
ratories) to in vivo (in daily lives). Known businesses, as well as start-ups, will now
be exploring developments and innovations in AI. Many people have begun exploit-
ing AI technologies to traverse towns, shop online, find tips for movies, screen out
unwelcome emails, or share a journey to work, whether they know it or not.
Therefore, AI is being considered for its extensive potential. Fifty-four percent
acknowledged making substantial investments in AI in a 2017 PwC study of global
executives, while a shortage of technical expertise remains a major concern. The
realized prices are expected to take off as companies continue to invest in software,
technology management, assets, and AI-enabled innovations: rising from $1.4 bil-
lion in annual sales from AI-enabled applications in 2016 to $59.8 billion by 2025,
according to one research study [22].

11.3 AI IN THE MONITORING ENVIRONMENT

Artificial intelligence is making its way through many domains like healthcare, auto-
motive, finance and economics, cybersecurity, military, hospitality, and now video
games, too. But the most important and positive impact it has is on its application in
environmental monitoring, which is also the need of the hour. Real-time monitoring
of soil, air quality, oceans, and lakes’ health are some of the basic applications of
AI. For example, IBM Research has been extensively exploring the application of
AI in monitoring the environment and has come up with various prototypes for the
same [6]. The best example for cleaning up of oceans using AI and Data Science is
the Ocean Cleanup Project undertaken for clearing the debris in the Great Pacific
Garbage Patch [23].

11.3.1 Monitoring Soil
The agriculture industry consumes more than 70% of the world’s water usage, annu-
ally. Small farms produce 80% of the food source for the world population, so ensur-
ing the quality and safety of water supplied is essentially critical. IBM’s AgroPad is
a real-time enabled prototype that is being used for chemical analysis of soil or water
samples using AI (Figure 11.3) [6]. It is a business card-sized paper device with a
microfluidic chip that performs the analysis of the sample of water or soil and results
are obtained in seconds [6]. Machine learning and vision algorithms are used to
translate the measured colour of the composition and intensity into concentrations of
chemical constituents present in the sample [6]. This is comparatively more reliable
than human vision-based tests alone. The parameters that can be measured are pH,
AI in Environmental Science 233

FIGURE 11.3  IBM’s AgroPad to analyze soil and water samples [24].

chlorine, magnesium, nitrogen dioxide, and aluminium [6]. This technology could
revolutionize digital agriculture.
AI-enabled agriculture bots are helping farmers in protecting their crops from
weeds and also combat the labour challenge. Monitoring of weed and timely spray-
ing can be possible by incorporating computer vision.
Two research students have devised a smart farming system that works on the
principles of AI and IoT (Internet of Things). The device called E-parirakshak has
been designed to monitor agricultural fields based on factors like fertility, soil tem-
perature, and moisture, controlling water levels, blowers, and sprinklers from a
remote location. The data is pushed to the cloud from where it can be retrieved and
analyzed later. It also helps in the detection of diseases in crops and updates the
farmers in real time.

11.3.2 Monitoring Water
Plankton is considered to be biological and natural sensors of aquatic health.
These are a collection of organisms that serve as the basis of the oceanic food
chain and the primary source of protein for billions of people. These organisms
include bacteria, protozoa, archaea, algae, and other floating organisms. These
organisms can alter their behaviour even after slight changes in the quality of the
water are detected. Therefore, IBM researchers have developed small, autonomous
microscopes in water for plankton monitoring, determining different species, and
tracking their movement. [6]. These findings can be further used for their response
to changes from temperature to oil spills to run-offs. The microscope consists of
an imager chip to capture the shadow of plankton as it swims over, creating a digi-
tal image of its health [6].
234 Artificial Intelligence (AI)

FIGURE 11.4  The Interceptor [25].

The Ocean Cleanup Project based in the Netherlands is an initiative to extract plas-
tic pollution from oceans, mainly the Great Pacific Garbage Patch. The Interceptor
is a technology developed to avoid new plastic reaching the water, not only to clean
up the plastic (Figure 11.4). It is solar-powered, autonomous, and remote-sensing
enabled with scalability and cost-effectivity.
The very first Interceptor which is a prototype is currently deployed in the
Cengkareng Drain in Jakarta, Indonesia (Figure 11.4). The local government of
Malaysia has incorporated the Interceptor as an addition to their clean-up plans to
stop the flow of plastic into rivers using barriers. It has been deployed in the Klang
River which runs through Kuala Lumpur (Figure 11.5, left and right view).

FIGURE 11.5  The Interceptor deployed in Jakarta (left) and Klang River, Malaysia (right) [25].
AI in Environmental Science 235

11.3.3 Monitoring Air
Air pollution is of great concern in many cities around the world. It fluctuates greatly
based on certain parameters like weather, humidity, wind, and temperature [6].
Certain pollutants interact with water molecules to create smog and while winds
can disperse pollutants, they can also carry forward other pollutants with them. IBM
has launched an initiative called the Green Horizon that utilizes the concept of the
Internet of Things (IoT) and AI to predict air quality and bring down pollutant lev-
els [6]. These systems can make sense out of the huge amount of data and make
forecasts far more effectively. The Green Horizon initiative is making an impact in
Beijing and will move to other cities like Johannesburg and New Delhi [6].
Blue Sky Analytics is a big data and AI start-up with a mission to provide action-
able intelligence starting with air pollution [26]. In October 2019, it launched BreeZo,
an app that provides real-time and contextual air quality data and helps people mini-
mize exposure to air pollution. The company is also building an AI-enabled app
called Zuri, for mapping farm and forest fires, and Zorro, for monitoring industrial
emissions.

11.4 RISKS OF ARTIFICIAL INTELLIGENCE

Artificial intelligence, while being a field of vast successes and further scope for
improvement, seems to slowly prove itself to be highly disruptive. It should be con-
sidered that since artificial intelligence uses algorithms, it cannot mimic the behav-
iour of humans from an emotional and realistic sense, thus creating discrepancies
in the overall life of individuals. Algorithms cannot consider the various kinds of
reactions a person could portray in response to a single stimulus, and this is where
the issue lies, and branches out to several other problems [26].

11.4.1 Bias
Biases are a huge problem in an everyday human’s life, and for a person with a bias
to feed biased data, and train a system to follow these biased data sets, we are essen-
tially curating a system that in turn logically produces results to match our sense of
bias. This could mean bias towards, for instance, ethnic minorities, religious minori-
ties, selective employment conditions, etc. [26].

11.4.2 Liability
Legal action has been a subject where artificial intelligence continually fails in,
because as humans, for an error in the actions of a machine with artificial intel-
ligence, who would we pass the blame to? Would it be the machine in its entirety or
the person who made it? The fact to be considered is, whether we are ready to accept
that in fields of high precision, medicine for instance, would humans be able to trust
a machine to replace himself and consider in retrospect the small margin of error it
could result in?
236 Artificial Intelligence (AI)

11.4.3 ASI (Artificial Superintelligence)

This would be a potential condition in which the artificial intelligence expressed by a
machine eventually surpasses the general intelligence of humans as well. To prevent
the possibility of this, ethical reforms and morals need to be transmitted. But tech-
nology is a little young for this, and in turn, this is not yet possible [26].
While AI allows everyone to effectively operate the effects of climate change
and preserve the environment and corporation, banking, healthcare, medicine, law,
education, etc., it’s not without risks. Several notable people like the late physicist
Stephen Hawking and Tesla CEO Elon Musk have cautioned of the potential risks of
AI unchecked [26].
The six categories of risk associated with AI as per the World Economic Forum
report are as follows [27]:

• Performance: The assumptions of the AI black box may not be available

to people, which makes it difficult to determine whether they are right or
beneficial. Deep learning can be dangerous when applications like early
warning systems are required for natural disasters.
• Security: Potentially AI could be “taught,” enabling hackers to introduce
their data sets or manipulate existing systems for malicious reasons.
• Control risks: When AI systems communicate with each other autono-
mously, they can deliver unpredictable results. Two species, for example,
came up with their language which humans could not comprehend.
• Economic risks: Companies that adopt AI at a slower level are economi-
cally impacted as their AI rivalry progresses. We can already see the clos-
ing of brick and mortar stores as the market is shifting to digitization.
• Social risks: As AI leads to automation, it is destroying the jobs in almost
all the industries. Autonomous weapons systems could also speed up global
conflicts and worsen them.
• Ethical risks: AI makes generalized conclusions in decision-making, this
often leads to increased bias decisions. Data processing also questions the
safety aspects.

To combat such risks, the WEF states “the stability, the accountability and the integ-
rity of the AI system must be ensured for government and business.” More collab-
oration between public and private entities, technologists, policymakers, and also
philosophers and further innovation in science is required to avoid and consider the
possible danger of artificial intelligence.

11.5 CONCLUSION AND FUTURE

Artificial intelligence is advancing at lightning speed, with discoveries and applica-
tions each day. While we cannot ignore the risk factor it has in our lives, integrat-
ing it with water and electricity systems to make a sustainable city is a future goal
that many countries wish to approach quickly. In place of climate change, artificial
AI in Environmental Science 237

intelligence technology-based products and discoveries could pave a path to at least

lessen the effects of climate change, if not eradicate it as a whole. The current applica-
tions of artificial intelligence in environmental science have been discussed, whether
it is in real-time testing of air pollution levels or to test the salinity of the ocean and
seawater for the thriving of marine life, and while we can use these techniques and
discoveries to make our lives easier, we should not focus on a complete reliance
on artificial intelligence. Some artificial intelligence algorithms are rather opaque,
and this, in turn, can affect us in several ways. For instance, we may not be able to
identify how much of the information being provided is in turn being used. Making
any form of assumption or prediction beyond the training that has been given can
result in unpredictable results, even though the character of generalization that most
machine learning algorithms possess entails it to give us a result.
In short, we can claim that artificial intelligence is helping us save the environ-
ment and thus improve our sustainability. Admittedly, it has adverse effects as well,
as demonstrated above. The key is to find a balance between using and exploiting
artificial intelligence. The techniques are not cheap either, so a lot of financial and
computational power is required for this purpose. Having said that, it can be argued
that the recently exploited idea of applying AI in environmental science has been a
success, and the future seems quite prospective in this particular field.

ACKNOWLEDGEMENT
The authors listed in this chapter wish to express their appreciation to the RSST
trust Bangalore for their continuous support and encouragement. As a corresponding
author, I also express my sincere thanks to all other authors whose valuable contribu-
tion and important comments made this chapter to this form.
Conflict of Interest The authors listed in this chapter have no conflict of inter-
est as known best from our side. There was also no problem related to funding.
All authors have contributed equally with their valuable comments which made the
chapter to this form.

REFERENCES
1. Russell, S.J. and Norvig, P. 2010. Artificial Intelligence-A Modern Approach, Third
International Edition.
2. Poole, D.I., Goebel, R.G. and Mackworth, A.K. 1998. Computational intelligence (p.
142). New York: Oxford University Press.
3. McCarthy, J., Minsky, M.L., Rochester, N. and Shannon, C.E. 2006. A proposal for
the Dartmouth summer research project on artificial intelligence, august 31, 1955. AI
Magazine, 27(4), pp. 12–12.
4. Lieto, A., Lebiere, C. and Oltramari, A. 2018. The knowledge level in cognitive archi-
tectures: Current limitations and possible developments. Cognitive Systems Research,
48, pp. 39–55.
5. Tutorialspoint​.co​m. 2020. Artificial intelligence - research areas - tutorialspoint. https​
:/​/ww​​w​.tut​​orial​​spoin​​t​.com​​/arti​​ficia​​l​_ int​​ellig​​ence/​​a rtif​​i cial​​_ inte​​llige​​nce​_ r​​​esear​​ch ​_ ar​​
eas​.h​​tm [Accessed 16 Jan. 2020].
238 Artificial Intelligence (AI)

6. IBM Research - Energy and Environment. 2020. IBM research - energy and envi-
ronment. https​:/​/ww​​w​.res​​earch​​.ibm.​​com​/e​​nergy​​-and-​​env​ir​​onmen​​t / [Accessed 15 Jan.
2020].
7. Autonomous & Connected Vehicles. https​:/​/au​​tomot​​ive​.r​​icard​​o​.com​​/ hybr​​id​-el​​ectro​​nic​
-s​​ystem​​s​/aut​​onomo​​us​-co​​​nnect​​ed​-ve​​hicle​s [Accessed 21 Jan. 2020].
8. Darrah, S.E., Bland, L.M., Bachman, S.P., Clubbe, C.P. and Trias‐Blasi, A. 2017. Using
coarse‐scale species distribution data to predict extinction risk in plants. Diversity and
Distributions, 23(4), pp. 435–447.
9. Kroodsma, D.A., Mayorga, J., Hochberg, T., Miller, N.A., Boerder, K., Ferretti, F.,
Wilson, A., Bergman, B., White, T.D., Block, B.A. and Woods, P. 2018. Tracking the
global footprint of fisheries. Science, 359(6378), pp. 904–908.
10. Mac Aodha, O., Gibb, R., Barlow, K.E., Browning, E., Firman, M., Freeman, R., Harder,
B., Kinsey, L., Mead, G.R., Newson, S.E. and Pandourski, I. 2018. Bat detective—Deep
learning tools for bat acoustic signal detection. PLoS Computational Biology, 14(3).
https​:/​/do​​i​.org​​/10​.1​​371​/j​​ourna​​l​.pcb​​​i​.100​​5995
11. Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D. and Moore, R. 2017.
Google earth engine: Planetary-scale geospatial analysis for everyone. Remote Sensing
of Environment, 202, pp. 18–27.
12. Kranstauber, B., Cameron, A., Weinzerl, R., Fountain, T., Tilak, S., Wikelski, M.
and Kays, R. 2011. The movebank data model for animal tracking. Environmental
Modelling & Software, 26(6), pp. 834–835.
13. Google. 2020. Using AI to find where the wild things are. https​:/​/ww​​w​.blo​​g​.goo​​gle​/p​​
roduc​​ts​/ea​​r th​/a​​i​-fin​​ds​-wh​​ere​-t​​he​​-wi​​ld​-th​​ings-​​a re/ [Accessed 21 Jan. 2020].
14. partner​.microsoft​.c​om. 2020. Gramener case study. https​:/​/pa​​r tner​​.micr​​osoft​​.com/​​en​
-us​​/case​​-stud​​ies​​/g​​ramen​​er/ [Accessed 21 Jan. 2020].
15. Nature​.co​m. 2020. AI empowers conservation biology. https​:/​/ww​​w​.nat​​ure​.c​​om​/ar​​ticle​​
s​/d41​​586​- ​0​​19​- 00​​746-1​ [Accessed 21 Jan. 2020].
16. National Geographic Society Newsroom. 2020. The California academy of sciences
and national geographic society join forces to enhance global wildlife observation
network iNaturalist: A growing community of iNaturalist users—and the artificial
intelligence they help power—help observe and monitor more than 165,000 species
around the world. https​:/​/ bl​​og​.na​​tiona​​lgeog​​raphi​​c​.org​​/2018​​/06​/2​​6​/the​​-cali​​forni​​a​-aca​​
demy-​​of​-sc​​ience​​s​-and​​-nati​​onal-​​geogr​​aphic​​-soci​​ety​-j​​oin​-f​​orces​​-to​- e​​n hanc​​e​-glo​​bal​-w​​
ildli​​​fe​-ob​​serva​​tion-​​netwo​​rk​-in​​atura​​list/​ [Accessed 21 Jan. 2020].
17. Medium. 2020. How AI can save earth’s biodiversity. https://medium​.com/​@kesari​/
how-​​ai​-ca​​n​-sav​​e​-ear​​ths​-b​​iodiv​​ersit​​y​-945​​55d57​​dd28 [Accessed 21 Jan. 2020].
18. McGovern, A., Elmore, K.L., Gagne, D.J., Haupt, S.E., Karstens, C.D., Lagerquist, R.,
Smith, T. and Williams, J.K. 2017. Using artificial intelligence to improve real-time
decision-making for high-impact weather. Bulletin of the American Meteorological
Society, 98(10), pp. 2073–2090.
19. Cho, R., 2018. Artificial intelligence–a game-changer for climate change and the envi-
ronment. https://phys​.org​/news​/2018–06-artificial-intelligence game​-changer​-climate​.​
html [Accessed 09 March 2020].
20. McFadden, C., Young, C., Lang, F., English, T. and Lang, F. 2020. AI might be the
future for weather forecasting. Int​eres​ting​engi​neering​​.com. https​:/​/in​​teres​​tinge​​ngine​​
ering​​.com/​​ai​-mi​​ght​-b​​e​-the​​-futu​​re​-fo​​r​-wea​​​ther-​​forec​​astin​g [Accessed 21 Jan. 2020].
21. Bakker, K. and Ritts, M. 2018. Smart earth: A meta-review and implications for envi-
ronmental governance. Global Environmental Change, 52, pp. 201–211. [Accessed 09
March 2020].
22. Pwc​.co​m. 2020. https​:/​/ww​​w​.pwc​​.com/​​gx​/en​​/news​​-room​​/docs​​/ai​-f​​or​-th​​​e​-ear​​th​.pd​f
[Accessed 21 Jan. 2020].
AI in Environmental Science 239

23. https://theoceancleanup​.com/. [Accessed 09 March 2020].

24. Chris Albrecht. September 6, 2018. IBM AgroPad combines paper, AI, and the cloud
to analyse soil and water. https​:/​/th​​espoo​​n​.tec​​h ​/ibm​​-agro​​pad​-c​​ombin​​es​-pa​​per​-a​​i​-and​​
-clou​​d​-to-​​analy​​ze​​-so​​il​-an​​d​-wat​​er/. [Accessed 09 March 2020].
25. https://theoceancleanup​.com​/rivers/. [Accessed 09 March 2020].
26. Ambika Choudhury. March 1, 2020. How blue sky analytics is utilizing ML & Geo-
spatial data to minimize air pollution. https​:/​/an​​alyti​​csind​​iamag​​.com/​​this-​​gurga​​on​-ba​​
sed​-s​​tartu​​p​-is-​​utili​​sing-​​m l​-ge​​ospat​​ial​-d​​atase​​t​-to-​​m inim​​ise​-e​​xpos​u​​re​-to​​-air-​​pollu​​tion/​.
[Accessed 09 March 2020].
27. How to manage AI’s Risks and Rewards. January 11, 2018. World economic forum.
https​:/​/ww​​w​.wef​​orum.​​org ​/a​​genda​​/2018​​/01​/ h​​ow​-to​​-mana​​ge ​-ai​​s​-ris​​k s​​-an​​d​-ben​​efits​/.
[Accessed 21 Jan. 2020].
 12 A Genetic Algorithm-
based Artificial
Intelligence Solution for
Optimizing E-Commerce
Logistics Vehicle Routing
Suresh Nanda Kumar and
Ramasamy Panneerselvam

CONTENTS
12.1 Introduction................................................................................................. 241
12.2 Transport Costs and Goods Characteristics in Logistics............................ 242
12.3 Logistics Challenges on Festive Days......................................................... 242
12.4 AI in Logistics............................................................................................. 243
12.5 Genetic Algorithms.....................................................................................244
12.6 Vehicle Routing Problem............................................................................. 245
12.7 Performance Measures................................................................................248
12.8 The Time-Dependent Vehicle Routing Problem with Time Windows.......248
12.9 Application of Artificial Intelligence in Route Planning and
Optimization............................................................................................... 249
12.10 Development of GA for Time-Dependent Vehicle Routing Problem
with Time Windows.................................................................................... 249
12.10.1 The GA-based Crossover for the Time-Dependent VRPTW....... 251
12.10.2 Evaluation of Fitness Function...................................................... 253
12.10.3 Mutation........................................................................................ 253
12.11 Comparison of Algorithms in Terms of Vehicles Utilized......................... 257
12.12 Conclusion...................................................................................................260
References............................................................................................................... 261

12.1 INTRODUCTION
Every manufacturing or service organization as well as e-commerce organization
makes use of warehouse storage and distribution to other stores or to customers using
transportation services of vehicles and carriers such as delivery trucks. Physical
distribution is the movement of goods by means of the outbound logistics process,

241
242 Artificial Intelligence (AI)

outward from the end of the assembly lines/shops of factories to the customers.
Supply-chain management comprises logistics as one of its fields and links logistics
with the firm’s other functions like engineering and production and links closely
with the communications network. It includes the procurement of raw materials and
components, upstream, transforming them into finished products, storing them, and
then transporting them to the customers efficiently and effectively. Logistics is the
process of transporting, storage, handling, and packaging of goods and materials,
from the beginning to the end of the manufacturing, sales, and finally the disposal
in order to satisfy customer requirements and also to add competitiveness to the
business.

12.2 TRANSPORT COSTS AND GOODS

CHARACTERISTICS IN LOGISTICS
Transportation can be considered to be one of the key activities of economic impor-
tance to the organization, among the different activities of the constitute logistics
for business applications. It is common knowledge that around 30–65% of the
companies’ logistics expenses are transportation costs (Chang, 1998). The costs of
transportation consist of components such as the mode of transportation, transport
infrastructure, packaging materials such as containers and pallets, transport termi-
nals like railway stations, ports, airports, etc., labour, and time. It also depends on
the property of the goods transported, i.e. the value, volume, weight, and stowability.
Hence, logistics and supply chain managers must understand the various aspects
of transportation deeply. The transportation achieves the efficient movement of
goods and provides delivery of products to the customers on time. Transport impacts
logistics activities and also has a great impact on the production and sale of products.
Hence, transportation cost is a necessary cost and ways must be found to minimize
the transportation costs. Value of transportation costs varies with the type of indus-
try. Transportation costs constitute a very big part of the selling price. Hence, it
affects profits more, and therefore it needs to be minimized.

12.3 LOGISTICS CHALLENGES ON FESTIVE DAYS

With a sudden surge in demand during festive days, the logistics for the e-commerce
players feel the pressure like never before. With home deliveries being the norm now,
supply chain managers need to ensure that they adhere to their SLAs and avoid miss-
routes to provide on-time deliveries to their customers.
With rising customer expectations, simple efficiency in operation no longer cuts
the bill for e-commerce players. Companies now need to ensure that the customers
are serviced in the time-window of their choice. Time-definite delivery is the new
benchmark that companies need to follow.
Nearly 50% of the costs incurred by logistics companies come from executing
the first- and last-mile delivery services. Logistics firms are trying to bring down
costs by increasing efficiency, reliability, and speed of the transportation activity.
With the increasing adaptation of digital technologies by the logistics industry, many
AI Solution for Optimizing E-Commerce 243

companies these days are applying artificial intelligence (AI) to their logistics and
supply chain activities with the hope of maximizing their resources by minimizing
the time and cost of planning and executing the routes, the number of vehicles to be
used, and the time of delivery of the goods and services to the customers.
Artificial intelligence is the simulation of human intelligence processes by
machines, particularly computer systems. Artificial intelligence is extensively
applied for improving logistics experience by increasing reliability, reducing the
cost of transportation, faster processing, and deciding optimal routes for last-mile
operations. Given the intense competition in the industry, customer satisfaction
has become the main factor of competitive advantage and for business dominance;
e-commerce firms like Amazon and Flipkart are investing hundreds of millions of
dollars with a view to improve their delivery processes in terms of speed and effi-
ciency and to achieve next-day or even same-day delivery. Artificial intelligence is
the main technology used to realize these goals by finding the optimal route, calcu-
lating exact delivery times, and deciding the correct product mix in the warehouse in
order to minimize logistics expenses.

12.4 AI IN LOGISTICS
The efficiencies gained using AI have achieved the greatest visibility in the fields
of vehicle routing, network/route planning and optimization, and forecasting
demand. Innovations that have been achieved using machine learning in these
fields have led to the emergence of more agile businesses and also becoming more
dynamic. As a result, food delivery aggregators such as Foodpanda, Zomato,
Swiggy, and Uber are able to provide an excellent customer experience. AI can
be applied in order to achieve improved planning of capacity, route optimization,
dynamic pricing, optimizing resource allocation, and vehicles with agility and
nimbleness to the areas where demand is high in order to decrease the waiting
time of the customers.
The package carrier United Parcel Service (UPS) makes use of an application
called ORION (On-Road Integrated Optimisation and Navigation), which is a GPS
tool that used expansive fleet telematics and advanced algorithms to analyze vast
quantities of data that enables UPS’ drivers make deliveries on-time and at the same
time being cost-effective. UPS has also implemented projects to improve efficiencies
for its employees and to help reduce their environmental impact.
Dynamic route planning and optimization can be achieved based on traffic condi-
tions and other factors. A reduction of even 1 mile for a driver each day in one year
will lead to a savings of nearly $50 m for UPS. The deployment ORION will save
for UPS nearly 100 million miles per year, which amounts to consumption of nearly
10 million gallons of fuel less, annually. It leads eventually to a decrease in carbon
dioxide emissions by about 100,000 metric tonnes.
AI will eventually set a new standard of efficiency across supply-chain and logis-
tics processes in the months to come. The scenario is changing quickly, creating a
“new normal” in how global logistics companies manage data, run operations, and
serve customers, in a manner that’s automated, intelligent, and more efficient.
244 Artificial Intelligence (AI)

Irrespective of how these developments are viewed, AI and related technologies

are going to radically transform the international supply-chain management and
logistics industries.

12.5 GENETIC ALGORITHMS
Meta-heuristics combine different mechanisms to find solutions of good quality.
In contrast to classical heuristics, they accept non-improving solutions during the
search to overcome local optima. Normally, meta-heuristics produce better results
than classical heuristics even though there is increased computation times (Gendreau
et al. 2002).
Genetic algorithm is a technique in AI and was first developed by J. Holland at the
University of Michigan in 1975. Solutions to a combinatorial problem are encoded
as chromosomes. The chromosomes are evaluated for their fitness by an evaluation
function and good properties of a generation of solutions are propagated to the fol-
lowing generations.
A genetic algorithm starts with a collection of chromosomes known as the ini-
tial population. Each of the chromosomes represents a solution to the given prob-
lem. There could be a random generation of the initial population, in which case it
will take more time for the algorithm to converge to the solution or some heuristic
method can be used for the generation of the initial population generated and since
the population is now closer to the solution, it would take less time to converge.
The GA’s next step is known as selection. This mechanism involves selecting
potential chromosomes as the parent chromosomes, which in turn depends on the
fitness values of the parent chromosomes and computed by using the fitness evalu-
ation function. The parent chromosomes thus selected will experience recombina-
tion by means of the crossover operations to generate the offspring chromosomes.
Mutation occurs in a small number of newly produced offspring chromosomes.
Mutation is carried out in order that a degree of randomness is introduced so that
the GA is preserved and prevented from getting converged to a local optimum.
The gene sequence undergoes a random swap in a mutation operation or randomly
undergoes a bit negation in the event of the offspring being encoded in bits. The
formation of new population is now made possible by substituting the parent chro-
mosomes with the offspring chromosomes.
The genetic algorithm will continue through this process until a stopping crite-
rion is met, which can be one of the following:

• Predefined number of generations has been produced.

• There was no improvement in the population, which would mean that the
GA has found an optimal solution.
• A predefined level of fitness has been reached.

Genetic algorithms have been used extensively to obtain solution for the vehicle
routing problem with time windows (VRPTW). The genetic algorithm is based on
natural reproduction, selection, and evolution based on Darwin’s theory.
AI Solution for Optimizing E-Commerce 245

12.6 VEHICLE ROUTING PROBLEM

Vehicle routing problem (VRP) involves the optimal transportation of goods by
means of movement and delivery of goods. The VRP determines the least cost routes
from a central depot to a set of customers. The VRP has been widely studied ever
since its introduction in 1959. Many variants to the VRP have been researched and
discussed by many researchers.
The principle of the VRP is described below.
The route of a vehicle to deliver goods must satisfy the following constraints:

• Every customer on the route must be visited exactly one time.

• All the routes begin at the depot and conclude at the depot. The vehicles
must return to the depot, once all the customers on the route have been
serviced.
• The sum of all customer demands on a route must not exceed the vehicle
capacity servicing the route.

The vehicle routing problem is pictorially depicted in Figure 12.1. From this figure,
one can note that each vehicle originates at the central depot, visits a set of customer
or supplier nodes once, fulfil their demands or perform the pickups at the nodes,
and then returns to the depot once again. The nodes can be either supplier sites in
the case of e-commerce company supplier site pickups or customer sites where the
e-commerce company delivers to the customer sites. It can also be different manu-
facturers to whom the suppliers deliver raw materials, components, parts, and other
supplies of items. In this process, the number of customers served by all the vehicles
is equal to the total number of customers under consideration/in the network. The
various colours depicted in the figure represent the various routes of the vehicles to
fulfil customer demand. This output was generated in the Heuristic Lab software
environment, using a plugin developed in this research. If the number of vehicles
under consideration is only one, then that vehicle has to start its journey from the
depot and visit each customer node exactly once and return to the depot. In this

FIGURE 12.1  An example solution to a vehicle routing problem.

246 Artificial Intelligence (AI)

process, the total cost/total time of transportation is minimized. This is called as

simple travelling salesman problem (TSP).
The vehicle routing problem has been widely studied in the operations research
and optimization literature. The earliest works were done in seminal papers of the
researchers Dantzig and Ramser (1959) and then later the VRP was discussed by
another famous pair of researchers, Clarke and Wright (1964). Now, VRP can be
solved by using different heuristic and meta-heuristics methods. The VRP is so
widely studied because of its extensive application in a wide range of problems and
its ability in identifying efficient methods for reducing distribution networks opera-
tional costs.
The total distance between any two locations in the network is known and the
demand for the product is specified for each service location (node) in the route.
In the time-dependent vehicle routing problem (TDVRP), the time to serve the
customers vary along by considering traffic conditions like congestion, accidents,
etc. in the route. In order to collect the items from various suppliers, the third
party logistics service providers (3PL) or logistics transportation service provider
must visit all scheduled suppliers during different hours required by the suppli-
ers. Some suppliers may further request visits within certain time windows. But
it is not easy to meet the time window constraints since the delivery processes
are frequently affected by real-time traffic flow conditions. The traffic congestion
occurring during the peak hours might cause severe delays of pickups and deliver-
ies. Therefore, in order to pick up items from the suppliers more efficiently, it is
necessary to address several challenging problems among customers and suppli-
ers, in a 3PL setting, whereby the e-commerce fulfilment centre’s vehicle visits the
suppliers’ sites to pick up the items ordered: (1) to satisfy suppliers’ specific time
windows; (2) to reduce the travel time and also the distance travelled, which in
turn is affected by peak-hour urban traffic congestion, and also the travel speeds
are not uniform.
An improved ACO algorithm was developed by Yu et al. (2009) for the VRP.
The algorithm’s effectiveness was tested by the data of 14 data points. Pisinger
and Ropke (2007) proposed a general heuristic for the VRPTW, CVRP, MDVRP,
site-dependent VRP, and open VRP. Schyns (2015) employed ACO algorithm to
solve the VRP, including mixed batch arrival and dynamic capacity with time
windows. Bell and McMullen (2004) compared the ant colony algorithm with
the TS algorithm and genetic algorithm and found that the ant colony algorithm
required less computation time. Meng et al. (2019) have developed a GA-based
solution for what they call the customer-oriented vehicle routing problem consid-
ering energy conservation and emission reduction (C-VRPESER). The objective
is to minimize the total cost based on the customers’ group which can lead to an
en-group distribution vehicle routing optimization. Planning the correct schedul-
ing vehicles, which are of different kinds for each customer group, will lead to
an improved utilization rate of the vehicles. The constraint of vehicle capacity
can also be met. The optimization objective will include minimizing the vehicle
fixed costs (minimizing the number of vehicles used), time penalty cost, fuel
cost, and carbon emissions.
AI Solution for Optimizing E-Commerce 247

Most vehicle routing models assume constant travel times throughout the day.
In reality, however, travel times fluctuate because of predictable events such as
congestion during rush hours or from unpredictable events such as road accidents/
incidents, unexpected weather conditions etc. (Ichoua et al. 2003) and also uncer-
tainties like volatile demand for service. Many VRPs are based on the assump-
tion that there is a constant and deterministic travel times between the various
customer nodes and the central depot (Kok et al., 2012) or it is the same as the
distance travelled between customer nodes. But in an actual scenario, the vehicles
do not travel at a constant speed (due to factors like congestion, weather condi-
tions, etc.), which impacts the transportation cost because of increased fuel con-
sumption (Kuo 2010).
The TDTSP and the TDVRP were studied by Ehmke et al. (2012, 2016). The
time-dependent travel times were used on the basis of Floating Car Data (FCD)
from Stuttgart, Germany. The travel times so obtained were then converted into
planning data sets using data-mining techniques. The authors applied two differ-
ent kinds of travel-time planning sets. The time-dependent vehicle routing problem
with time windows (TDVRPTW) makes the basic and important assumption that
the travel times are a function of the present time. Traffic congestion affects travel
speeds and hence it affects the total time to complete the route and the number of
vehicles utilized and the total transportation cost can also be found. Travel time
between customers (nodes) and depot has been identified as one of the major fac-
tors that increases the adverse impacts of congestion. Traffic congestion also has a
direct effect on carriers’ cost structure and the relative weight of wages and overtime
expenses (Figliozzi, 2009).
TDVRPs satisfy a property which is known as the “non-passing property,”
and in simpler terms, the First-In First-Out (FIFO) property (Ichoua et al., 2003).
According to Doerfler (2017), supply-chain management professionals face a lot of
delivery challenges, especially in handling routing of the vehicles. She says that
according to Hani Mahmassani, the William A. Patterson Distinguished Chair in
Transportation and Director of the Northwestern University Transportation Center
in Evanston, Illinois, routing is a form of roulette because traffic congestion, road
construction, and weather conditions can adversely affect the time it takes to deliver
packages to their destination. A civil and environmental engineering professor, He
says “if you plan that a certain route should take two hours with travel time and it
takes twice as long, you lose your schedule synchronization quickly”. He also adds
that “if you plan for the worst case or even the 95th percentile of travel time, the extra
time may be unnecessary, and you’ve wasted the use of the vehicle. Thus, there is a
loss of productivity.”
The TDVRP takes the first aspect into account by assuming that travel times
depend on the time of the day. All of these uncertainties if not addressed properly
will result in an overall increase in logistics costs. Hence, vehicle routing is a com-
plex logistics management problem and represents a key class of problems to solve
in order to lower costs and optimize logistics resources. The main purpose of VRP is
to find optimal routes for multiple vehicles visiting a set of locations by minimizing
both time and cost.
248 Artificial Intelligence (AI)

12.7 PERFORMANCE MEASURES
The quality of solution obtained for the vehicle routing problem is judged in terms
of the number of vehicles utilized and also the total distance travelled. If a single
vehicle is used to satisfy the demands of all the customers, the cost of the number
of vehicles as well as the cost of transportation will be the minimum. If the total
demand exceeds the capacity of the vehicle, then the use of single vehicle becomes
infeasible, which necessitates the use of multiple vehicles. This then gives rise to the
multiple Traveling Salesman Problem (mTSP), which in other words is called the
vehicle routing problem. In this case, the objective of the problem is to minimize
the number of vehicles used and also the total distance travelled by the vehicle and
thereby the total costs of transportation.
When multiple vehicles are used to satisfy the demands of the customers, the
objectives are as follows:

• Minimize the total distance of travel of all the vehicles put together and
hence the total cost of transportation of satisfying the demands of all the
customers.
• Minimize the number of vehicles required to meet the demands of the
customers.

12.8 THE TIME-DEPENDENT VEHICLE ROUTING

PROBLEM WITH TIME WINDOWS
The time-dependent vehicle routing problem is an extension of the basic VRPTW.
The TDVRPTW is being researched increasingly because of its practical impor-
tance to transportation problems. This is due to the fact that the traffic conditions are
different and vary during different times of the day and as a result manufacturers,
suppliers, and e-commerce retailers need to schedule their pickups and deliveries at
the appropriate times of the day, considering the time windows of the customers and
suppliers to make their order fulfilment needs efficient and faster. The methodology
adopted by Kumar and Panneerselvam (2015) study was based on genetic algorithm.
In a discussion with a vehicle driver about delivery of prawn eggs for breeding in
Andhra Pradesh from Tamil Nadu, he described how he had to change routes near
Chennai East Coast Road to avoid traffic congestion during peak hours from morn-
ing 6 a.m. to 10 a.m. and then from evening 5 p.m. to 11 p.m. He had to deliver the
prawn eggs to the customer in 10 hours’ time with a maximum relaxation of 2 hours.
An important real-life property found in transportation problems in the retail indus-
try is time-dependent travel times, which is also known as dynamic travel times,
where travel time depends on the time of departure and traffic conditions such as
rush hours. Esmat and Mirmohammadi (2015) develop a tabu-search-based meta-
heuristic to solve the TDVRPTW. Here they present an Integer Linear Programming
(ILP) and the tabu search algorithm to handle the problem in large-scale instances.
Duygu et al. (2014) propose a tabu search and adaptive large neighbourhood search-
based solutions to solve the time-dependent vehicle routing problem. The conducted
AI Solution for Optimizing E-Commerce 249

experiments confirm that the proposed procedure is effective to obtain very good
solutions to be performed in real-life environment.

12.9 APPLICATION OF ARTIFICIAL INTELLIGENCE IN

ROUTE PLANNING AND OPTIMIZATION
According to a report by RedSeer (2019), sales during the festive season in September
to October, 2019, in India by the major e-commerce retailers such as Amazon and
Walmart-owned Flipkart generated a sales by the e-tailers, a whopping $3 billion of
Gross Merchandise Value (GMV) with a large chunk of customers from Tier-2 cities
or Bharat. RedSeer had forecasted a total sale of $3.7 billion in the first phase of the
festive season. The actual sales of $3 billion are nearly 80% of RedSeer’s forecasted
sales. E-commerce companies such as Flipkart and Amazon could generate up to
US$6 billion or Rs 39,000 crore in sales this festive season, according to consulting
firm RedSeer (as compared to GMV of $30 billion for Singles Day in China, about
$8 billion for Cyber Monday, and about $6 billion for Black Friday in the United
States in 2019). However, these events don’t come without its challenges.

12.10 DEVELOPMENT OF GA FOR TIME-DEPENDENT VEHICLE

ROUTING PROBLEM WITH TIME WINDOWS
The genetic algorithm – Random Sequence Insertion-based Crossover (RSIX) (a
detailed description on the development of the genetic algorithm-based solution for
TDVRPTW) – is given. Traffic congestion is a common occurrence in many urban
cities worldwide. This phenomenon of traffic congestion results in a significant varia-
tion in vehicle travel speeds during peak hours of traffic, especially during the morn-
ing and evening times. This situation is unlike other vehicle routing models, where
a constant travel speed is assumed irrespective of congestion. Urban vehicle route
designs, which do not take into consideration such significant travel speed varia-
tions, result in solutions that are inefficient and suboptimal. Routes that are designed
without taking traffic situations into consideration will lead freight and passenger
vehicles into congested arterial roads and streets and thereby not only increase sup-
ply chain and logistics costs but also contribute to environmental ill-effects that are
associated with freight and passenger traffic in urban locations such as the emission
of greenhouse gases, noise, and air pollution.
Travel time between customers (nodes) and depot has been identified as one of
the major factors that increases the adverse impacts of congestion. Traffic congestion
also has a direct effect on carriers’ cost structure and the relative weight of wages
and overtime expenses (Figliozzi, 2009). Dynamic vehicle routing problem is based
on the time-dependent variations in travel times. A genetic algorithm is developed
for solving the TDVRPTW. The VRP with pickup and delivery is used for picking
up or to deliver to the various customers or from the various supplier sites. These
sites specify a time window and the multiple vehicles with having various capacities
and with real-time variations in the times of travel from one node to the other are
considered.
250 Artificial Intelligence (AI)

To load and start the time-dependent VRP variant in HeuristicLab, a valid file
to load the time-dependent VRP data is required. So, the data from the Solomon
(CVRPTW) instances is used and the travel time matrix is added. The dimension of
the travel time matrix must match the number of cities. For the travel times, random
values between 0.5 and 1.0 were generated using random function in Excel.
So, the objectives of the research (Nanda Kumar and Panneerselvam, 2017) were
to minimize the total distance travelled and to minimize the total number of vehicles
used to cover all the customer nodes. The minimization of the total distance travelled
is treated as the primary measure and the corresponding number of vehicles utilized
to serve all the customer nodes is treated as the secondary measure.
In the genetic algorithm, a chromosome is a chain of integers (genes) and each of the
integers represents a customer node. This chromosome represents all the customers. The
customers on it are separated to several routes, each of them representing a sequence of
deliveries that must be covered by a vehicle. The chromosome consists of information on
the number of vehicles used to service the customer nodes, separately so as to minimize
the number of vehicles used to service the customer nodes. This information is called
as the vehicle-part. In the vehicle information, the number of vehicles is equal to the
number of routes in the customer chromosome. The number on each of the vehicle genes
represents the number of customers to service. The sum of these numbers in vehicle
information should be equal to the total number of customers. The application of the
RSIX crossover method to a pair of chromosomes yields a pair of offspring. The off-
spring chromosome is chosen based on the fitness value. Lower fitness value indicates a
desirable solution, since it will lead to less distance travelled. For example, in Figure 12.2,
there is an illustration of a possible solution of 3 vehicles for 8 customer nodes. There are
3 genes in the vehicle-part of chromosome, which means that the 8 customers are divided
into 3 routes. The 3 on the first gene in vehicle-part represents route 1 that services 3 cus-
tomer nodes: 1, 5, and 4. The 2 on the second gene means that route 2 services customers
2 and 8. The 3 on the third gene indicates that route 3 services customers 7, 6, and 3.
Vehicle travel times in the cities and urban areas vary due to various reasons and
factors, like congestions in traffic, accidents, road repairs, VIP movement, and delays
due to bad weather. If these travel time variations are ignored, when forming plans
for the routes needed for pickup and delivery of goods by vehicles from customers
and suppliers, one may end up in creating route plans that may make the vehicles end
up in urban heavy traffic congestions. As a result of this, vehicles may be spending

FIGURE 12.2  An example of the customer and vehicle chromosome.

AI Solution for Optimizing E-Commerce 251

unnecessary time in traffic jams and get delayed, resulting in long waiting times
for the customers/suppliers and not having exact information regarding the times at
which the vehicles arrive actually. Due to this, the time windows during which the
demand or supply nodes have to be visited become difficult to be satisfied.
Consideration of time-dependent travel times and also the demand information
arising during real-time to solve the VRP can help in reducing the costs of travel that
occurs if the changes in traffic scenarios are ignored.
The total time taken for travel between two locations depends on the specific
time of departure. Hence, to take these external influences into consideration, the
VRPTW is extended and is studied as TDVRPTW. In this particular scenario, the
driving time that changes with the time of the day is suitably represented by a time-
dependent function.

12.10.1 The GA-based Crossover for the Time-Dependent VRPTW

The skeleton of a generalized genetic algorithm is presented below:

1. Generate an initial population, which consists of a set of chromosomes to

represent the characteristic of all the individuals of the initially selected
population.
2. Evaluate each of the chromosomes in the initial population using the fitness
function.
3. Select a subpopulation for crossover operation.
4. Perform crossover operation between different pairs of chromosomes in the
subpopulation to produce two offspring.
5. Perform mutation for each of offspring based on a mutation probability.
6. Replace the offspring in the current population along with their fitness
function values.
7. Sort the chromosomes of the current population either in the descending
order of their fitness function values in the case of maximization problem or
in the ascending order of their fitness function value in the case of minimiz-
ing problem.
8. Repeat step 3 to step 7 for the specified number of generations.
9. Find the chromosome with best fitness function value for implementation.

The steps of the proposed genetic algorithm (SNRPGA) for the time-dependent
vehicle routing problem with time windows are presented below. Kumar and
Panneerselvam (2015) developed a crossover technique for the genetic algorithm,
named Random Sequence Insertion-based Crossover (RSIX).
A schematic view of the crossover operation (RSIX) used in this algorithm is
shown in Figure 12.3. A crossover operation is a major process which is applied on
two chromosomes from the current subpopulation to produce two offspring. This
operation modifies the arrangement of genes of the chromosomes considered and
the resultant chromosomes are called offspring. There are many methods for cross-
over operation according to different problems. “Random Sequence Insertion-Based
252 Artificial Intelligence (AI)

FIGURE 12.3  Crossover operation.

Crossover” method is proposed to perform the crossover operation between any two
chromosomes, which is explained in detail below.
The steps of the Random Sequence Insertion-based Crossover (RSIX) method for
the TDVRPTW are presented below.
Consider two parent chromosomes (P1 and P2) with seven genes in each of them
as shown in Figure 12.3a. The gene elements in each of the parent chromosomes are
from 1 to 7, but in some random order.

Step 1: The route chromosomes of the initial population are generated ran-
domly by using a random seed. The random seed is used to initialize the
new pseudo-random number generator for the time-dependent VRP prob-
lem instances. For the travel times, random values between 0.5 and 1.0 have
been generated.
Step 2: Two chromosomes from the chromosome pool are randomly chosen
as parents.
Step 3: Generate two crossover points, which will lead to three chromosome
segments in each chromosome as shown in Figure 12.3a.
Step 4: Copy the middle segment of the chromosome P1 in Figure 12.3a imme-
diately after the middle segment of the chromosome P2 in Figure 12.3a and
AI Solution for Optimizing E-Commerce 253

copy the middle segment of the chromosome P2 in Figure 12.3a imme-

diately after the middle segment of the chromosome P1 in Figure 12.3a,
as shown in Figure 12.3b, i.e. swapping operation of the middle crossover
genes segment takes place, as in Figure 12.3b.
Step 5: In each chromosome in Figure 12.3b, check whether each of the
genes in the second segment is present in the first segment, third segment,
and fourth segment in it. If so, shade that gene as shown in Figure 12.3c.
Similarly, check whether each of the genes in the third segment is present
in the first segment, second segment, and fourth segment in it. If so, shade
that gene as shown in Figure 12.3c.
Step 6: Write the genes of each of the chromosomes in Figure 12.3c from left
to right after dropping the shaded genes to give the results as shown in
Figure 12.3d. This results in obtaining two new offspring having gene seg-
ments that underwent crossover and they are added into the next generation
as shown in Figure 12.3d.
Step 7: The final two offspring of the crossover operation are as shown in
Figure 12.3e.
• The new offspring are tested for fitness values.
• The smaller the “fitness-value,” the stronger the road chromosome
obtained.

12.10.2 Evaluation of Fitness Function

Every solution has a fitness value assigned to it, which measures its quality. A route
for a vehicle with respect to a chromosome/offspring is obtained by assigning its
genes (customer nodes) serially from left to right and then copying its first gene
(customer node) at the end of the route for the vehicle routing problem. A vehicle is
assigned to each route to deliver the items to the customer nodes in the route and/or
pickup items from the customer nodes in the route. Every road or route chromosome/
offspring has its own fitness value, defined through a fitness function. The fitness
function value of a route chromosome is the sum of the distances of the arcs (pairs
of adjacent customer nodes) in the route of the route chromosome. In this case, the
smaller the fitness function value, the stronger the road chromosome.
The fitness function is formulated for this VRP, and the quality of a solution s
depends on the total cost or the distance travelled for all the vehicles:

fs = å distance
r
s ,r

where distances,r denotes the cost or distance travelled in route r in solution s.

12.10.3 Mutation
After performing the crossover operation, mutation operation is performed to further
perturb the arrangement of the genes in each of the offspring. This operation selects
254 Artificial Intelligence (AI)

two gene positions randomly and interchanges them for a given probability. This
means that a random number is selected for each offspring. If this random number
has a lesser value, or equal in value to specified probability of mutation, chromosome
undergoes mutation. Else, mutation will not be carried out on that offspring.
Mutation helps to prevent the genetic algorithm from converging to a local opti-
mum. Other genetic algorithms parameters can also influence the GA efficiency.
Crossover probability which is specified in the GA determines the rate at which the
crossover occurs.
In this section, a genetic algorithm with the modified random sequence insertion-
based crossover (RSIX) method for the vehicle routing problem with time windows
(TDVRPTW) is proposed.
This crossover technique (RSIX) is applied to develop a new genetic algorithm
(SNRPGA) (Nanda Kumar and Panneerselvam, 2017)
The steps of the proposed genetic algorithm (SNRPGA) for the time-dependent
vehicle routing problem with time windows are presented below.

Step 1: Input the following:

• Number of customer nodes (n)
• Number of vehicles (k)
• Capacity of the vehicles (a)
• Set Generation Count (GC) = 1.
• Maximum number of generations to be carried out (MNG) = 1000
Step 2: Generate a random initial population (L) of 100 (N) chromosomes (suit-
able solutions routes for the problem).
Step 3: Evaluate the fitness function f(x) of each chromosome in the popula-
tion L.
Step 4: Selection – sort the population L by the objective function (fitness func-
tion) value in the ascending order, since the objective of the study is mini-
mization of the total distance travelled. Copy a top 30% of the population to
form a subpopulation S rounded to the whole number. Smaller fitness value
is preferred here.
Step 5: Randomly select any two unselected parent chromosomes from the
subpopulation S. Let them be c1 and c2 using tournament selection.
Step 5.1: Perform two-point random crossover using the random sequence
insertion-based crossover (RSIX) described in the earlier section
among the chromosomes c1 and c2 to obtain their offspring d1 and d2
by assuming a crossover probability of 0.7.
Step 5.2: Perform mutation on each of the offspring using a mutation prob-
ability of 0.3.
Step 5.3: Evaluate the fitness function with respect to the total distance
travelled and number of vehicles utilized value for each of the offspring
d1 and d2.
Step 5.4: Replace the parent chromosomes c1 and c2 in the population with
the offspring d1 and d2, respectively, if the fitness function of the off-
spring is less than that of the parent chromosomes.
AI Solution for Optimizing E-Commerce 255

Step 6: Increment the generation count (GC) by 1, i.e. GC = GC + 1.

Step 7: If GC ≤ MNG, then go to step 4; else, go to step 8.
Step 8: The topmost chromosome in the last population serves as the solution
for implementation. Print the tour along with the total distance travelled
and the number of vehicles used.
Step 9: Stop.

The algorithm developed by Nanda Kumar and Panneerselvam (2017) was compared
with some existing algorithms to find the superiority of their new algorithm. So, the
SNRPGA (Nanda Kumar and Panneerselvam 2017) was applied to the TDVRPTW
and was compared with another existing algorithm using a complete factorial experi-
ment with two factors: Problem Size (Factor A) and Algorithm (Factor B) with four
replications in each experimental combination.
In the first comparison of the new algorithm, SNRPGA with the existing algo-
rithm developed by Demir (2012) for the TDVRPTW in terms of the total distance
travelled, it was found by the authors that there was significant difference among
them. Further, through Duncan’s multiple rage test, they proved that the proposed
SNRPGA is superior to the existing algorithm developed by Demir (2012) in terms
of the total distance travelled.
The results of the factorial experiment in terms of the total distance travelled are
shown in Table 12.1. The application of ANOVA to the data given in Table 12.1 gives
the results as shown in Table 12.2.
From the ANOVA results shown in Table 12.2, one can infer that the factors
“Algorithm” and “Problem Size” have significant effects on the total distance trav-
elled. Since, there is a significant difference among the two algorithms compared in
terms of the total distance travelled, Duncan’s multiple range test is next conducted
to identify the best algorithm by arranging the algorithms in the descending order of
their mean total distance travelled from to right.
The standard error used in this test is computed as shown below using the mean
sum of squares of the interaction terms (Problem Size × Algorithm) and the number
of replications under each of the algorithms (24). One can notice the fact that the
mean sum of squares of the interaction term AB is used in estimating the standard
error (SE), because the F ratio for the factor “Algorithm” is obtained by dividing its
mean sum of squares by the mean sum of squares of the interaction term ABij (Nanda
Kumar and Panneerselvam, 2012):

SE = ( MSS AB ¸ n ) = ( 21194.060 ¸ 24 )
0.5 0.5
= 29.72

The least significant ranges (LSR) are calculated from the significant ranges of
Duncan’s multiple range tests table for α = 0.05 and 5 degrees of freedom as shown
in Table 12.3.
The treatment means for the Factor B (Algorithm) in terms of the total dis-
tance travelled are arranged in the descending order from left to right, as shown
in Figure 12.4.
256 Artificial Intelligence (AI)

TABLE 12.1
Results of the Factorial Experiment in Terms of the Total Distance
Travelled for TDVRPTW
Problem Class (Factor A) Algorithm (Factor B)
Class Replication of Class ALNS (Demir) TDVRPTW (SNRPGA2)
Random 1 1 R1 971 831
2 R1 932 675
3 R1 948 717
4 R1 1048 664
Clustered 1 1 C1 822 593
2 C1 826 567
3 C1 827 585
4 C1 827 580
Random Clustered 1 1 RC1 1207 867
2 RC1 1114 771
3 RC1 1258 847
4 RC1 1457 811
Random 2 1 R2 740 525
2 R2 701 688
3 R2 731 590
4 R2 794 506
Clustered 2 1 C2 585 422
2 C2 585 430
3 C2 586 432
4 C2 586 430
Random Clustered 2 1 RC2 777 762
2 RC2 783 573
3 RC2 923 732
4 RC2 962 606

TABLE 12.2
Analysis of Variance for Total Distance Travelled
Degrees F Ratio
Source of Sum of of Mean Sum Calculated F (α =
Variation Squares Freedom of Squares Ratio 0.05) Remarks
Algorithm (B) 697352.302 1 697352.302 146.138 4.12 Significant
Problem Size (A) 1313751.984 5 262750.397 55.062 2.47 Significant
Problem Size × 105970.302 5 21194.060 4.441 2.47 Significant
Algorithm (A × B)
Error 171787.094 36 4771.864
Total 2288861.682 47
AI Solution for Optimizing E-Commerce 257

TABLE 12.3
Duncan’s Multiple Range Tests
Significant Standard LSR = Significant Range ×
No. of treatments – 1 (j) Range Error Standard Error
2 2.872 29.72 85,356

FIGURE 12.4  Results of Duncan’s multiple range test for distance travelled.

From the Duncan’s multiple range test performed as shown in Figure 12.4, it
is also clear that the TDVRPTW is superior in performance when compared to
the existing algorithm used in this study for comparison, in terms of total distance
travelled.

12.11 COMPARISON OF ALGORITHMS IN
TERMS OF VEHICLES UTILIZED
In this section, a comparison is made among the proposed algorithm SNRPGA and
an existing algorithm for the TDVRPTW in terms of number of vehicles utilized
using a complete factorial experiment. The existing algorithm was an iterated local
search algorithm proposed by Hashimoto et al. (2008). The number of factors in the
experiment is 2, viz. Factor A (Problem Size) and Factor B (Algorithm). The number
of levels for the Factor A is 6, Random 1, Clustered 1, Random Clustered 1, Random
2, Clustered 2, and Random Clustered 2. The number of levels for the Factor B is 2,
viz. TDVRPTW and SNRPGA. The number of replications under each experimental
combination is 4. The results in terms of the number of vehicles utilized as per this
design are as shown in Table 12.4.
In this model, Factor A (Problem Size/Problem Class) is a random factor and the
Factor B (algorithm) is a fixed factor. Since Factor A is a random factor, the interaction
258 Artificial Intelligence (AI)

factor ABij is also a random factor. The replications are always random and the num-
ber of replications under each experimental combination is 4. The derivation of the
expected mean square (EMS) is given in Nanda Kumar and Panneerselvam (2012).
To test the effect of Ai as well as ABij, the respective F ratio is formed by dividing the
mean sum of squares of the respective component (Ai or ABij) by the mean sum of
squares of error. The F ratio of the component Bj is formed by dividing its mean sum
of squares by the mean sum of squares of ABij.
The results of ANOVA of the data given in Table 12.4 are shown in Table 12.5.
From the ANOVA results shown in Table 12.5, one can infer that the factors
“Problem Size,” “Algorithm,” and “Interaction of Problem Size and Algorithm” have
significant effects on the response variable “Number of Vehicles Utilized.” Since
there are significant differences among the algorithms, the best algorithm is obtained

TABLE 12.4
Results of Number of Vehicles Utilized for TDVRPTW Problem
Problem Class
(Factor A) Algorithm ((Factor B)
Replication of TDVRPTW SNRPGA2
Class (Hashimoto)
1. Random 1 1 9 9
2 14 9
3 13 10
4 10 9
2. Clustered 1 1 10 8
2 10 8
3 10 8
4 10 8
3. Random Clustered 1 1 10 10
2 11 11
3 11 11
4 14 12
4. Random 2 1 4 3
2 4 3
3 3 3
4 3 3
5. Clustered 2 1 3 3
2 3 3
3 3 3
4 3 3
6. Random Clustered 2 1 4 3
2 4 3
3 3 4
4 3 3
AI Solution for Optimizing E-Commerce 259

TABLE 12.5
Analysis of Variance for Number of Vehicles Utilized for TDVRPTW Problem
Degrees Mean F Ratio
Sum of of Sum of Calculated (α =
Source of Variation Squares Freedom Squares F Ratio 0.05) Remark
Problem Size (A) 551.604 5 110.321 260.430 2.47 Significant
Algorithm (B) 3.521 1 3.521 8.311 4.12 Significant
Problem Size × 5.604 5 1.121 2.646 2.47 Significant
Algorithm (A × B)
Error 15.250 36 0.424
Total 575.979 47

using Duncan’s multiple range test by arranging the algorithms in the descending
order of their mean number of vehicles utilized from left to right.
The standard error used in this test is computed as shown below using the mean
sum of squares of the interaction terms (Problem Size × Algorithm) and the number
of replications under each of the algorithms:

SE = ( MSS AB ¸ n ) = (1.121 ¸ 24 )
0.5 0.5
= 0.216

The least significant ranges (LSR) are calculated from the significant ranges of
Duncan’s multiple range tests table for α = 0.05 and 36 degrees of freedom as shown
in Table 12.6. The results of Duncan’s multiple range test are shown in Figure 12.5.
In this figure, the algorithms are arranged as per the descending order of their mean
number of vehicles utilized from left to right. From Figure 12.5, it is clear that there
is a significant difference between the two algorithms in terms of the mean num-
ber of vehicles utilized and further the proposed algorithm SNRPGA utilizes the
minimum mean number of vehicles compared to the other algorithm. Hence, the
proposed algorithm SNRPGA is superior to the existing algorithm considered in this
research for the TDVRPTW problem.
As already stated, any algorithm that is developed now is to be compared
with some existing algorithms to find the superiority of the new algorithm. So,
the SNRPGA is applied to the TDVRPTW and is compared with an existing

TABLE 12.6
Duncan’s Multiple Range Tests
Significant Standard LSR = Significant Range ×
No. of Treatments – 1 (j) Range Error Standard Error
2 2.872 0.216 0.6204
260 Artificial Intelligence (AI)

FIGURE 12.5  Results of Duncan’s multiple range test for number of vehicles used.

algorithm through a complete factorial experiment with two factors: Problem Size
(Factor A) and Algorithm (Factor B) with four replications in each experimental
combination.
In the first comparison of the proposed SNRPGA with an existing algorithm
developed by Demir (2012) for the TDVRPTW in terms of the total distance trav-
elled, it is found that there is significant difference among them. Further, through
Duncan’s multiple rage test, it is proved that the proposed SNRPGA is superior to the
existing algorithm developed by Demir (2012) in terms of the total distance travelled.
In the second comparison of the proposed SNRPGA with an existing algorithm
developed by Hashimoto et al. (2008) for the TDVRPTW in terms of the number of
vehicles utilized, it is found that there is significant difference among them. Further,
through Duncan’s multiple rage test, it is proved that the proposed SNRPGA is supe-
rior to the existing algorithm developed by Hashimoto et al. (2008) in terms of the
number of vehicles utilized.
In the second comparison of the proposed algorithm, SNRPGA with an existing
algorithm developed by Hashimoto et al. (2008) for the TDVRPTW in terms of the
number of vehicles utilized, it was found that there was significant difference among
them. Further, through Duncan’s multiple rage test, it was proved that the proposed
SNRPGA is superior to the existing algorithm developed by Hashimoto et al. (2008)
in terms of the number of vehicles utilized.

12.12 CONCLUSION
The time-dependent vehicle routing problem (TDVRP) is a class of vehicle rout-
ing problems, where the time to serve the customers vary along with the consid-
eration of the traffic conditions in the route. In order to collect the items from
various suppliers or deliver items to customers, the 3PL logistics transportation
service provider must visit all scheduled customers or suppliers during the differ-
ent hours specified by the supplier or the customer. Some customers or suppliers
AI Solution for Optimizing E-Commerce 261

may also need visits to happen during a certain time window. This variant of
the vehicle routing problem is known as the vehicle routing problem with time
windows. Time window constraints are difficult to meet given the real-time road
conditions like traffic congestions which makes the vehicles to wait long hours in
the traffic jams especially in urban environments. Peak hour traffic congestions
lead to long delays. Hence, in order to pick up or deliver items from the suppliers
and customers more efficiently, many challenges must be overcome and solved
among customers and suppliers, in a 3PL kind of arrangement, whereby the 3PL
vehicle visits the customers’ or suppliers’ sites to deliver or to pick up the items,
in an e-commerce or manufacturing or other set-ups, to satisfy the customers
and suppliers’ specific time windows and to minimize the travel times severely
impacted by heavy traffic conditions.
The vehicle routing problem with time windows is solved using the genetic algorithm
with multi-chromosome representation. The genetic algorithm is an AI technique that
finds near optimal solution to any combinatorial optimization problem which is of the
NP-hard type. It is later used for finding a (near) optimal solution to a VRPTW variant,
the TDVRPTW, which takes into consideration the various times of the day at which
the vehicle travels, since the travel time varies during various times of the day such that
the total distance travelled is minimized for the generated number of number of tours
for the vehicles. So, the objectives are to minimize the total distance travelled and to
minimize the total number of vehicles used to cover all the customer nodes. Hence,
in the logistics sector, genetic algorithms are reducing delivery times and decreasing
costs. In the logistics business, every mile and minute matter. Companies can use a
route planner based on genetic algorithms to map out optimal routes for deliveries.
This study can be useful for planning the supplier site pickups by e-commerce
companies, taking into consideration traffic conditions during different periods of
the day with time window requirements of the suppliers.

REFERENCES
Bell, J.E. and McMullen, P.R. “Ant colony optimization techniques for the vehicle routing
problem,” Advanced Engineering Informatics, vol. 18, no. 1, pp. 41–48, 2004.
Chang, Y.H. Logistical Management. Hwa-Tai Bookstore Ltd., Taiwan, 1998.
Clarke, G. and Wright, J. R. “Scheduling of vehicle routing problem from a central depot to
a number of delivery points,” Operations Research, vol. 12, no. 4, pp. 568–581, 1964.
doi:10.1287/opre.12.4.568.
Dantzig, G. and Ramser, J. “The truck dispatching problem,” Management Science, vol. 6,
pp. 80–91, 1959.
Demir, E. Models and Algorithms for the Pollution-Routing Problem and Its Variations. PhD
Thesis, University of Southampton, Southampton, 2012.
Doerfler, S. “Delivering last-mile options,” Inside Supply Management, vol. 28, no. 2, pp.
26–30, 2017.
Duygu, T., Jabali, O. and van Woensel, T. “A vehicle routing problem with flexible time win-
dows,” Computers & Operations Research, vol. 52, Part A, pp. 39–54, 2014.
Ehmke, J.F., Campbell, A.M. and Thomas, B.W. “Vehicle routing to minimize time-depen-
dent emissions in urban areas,” European Journal of Operational Research, vol. 251,
no. 2, pp. 478–494, 2016.
262 Artificial Intelligence (AI)

Ehmke, J.F., Steinert, A. and Mattfeld, D.C. “Advanced routing for city logistics service
providers based on time-dependent travel times,” Journal of Computational Science,
2012, doi:10.1016/j.jocs.2012.01.006.
Esmat, Z.-R. and Mirmohammadi, S.H. “Site dependent vehicle routing problem with soft
time window: modeling and solution approach,” Computers & Industrial Engineering,
vol. 90, pp. 177–185, December 2015.
Figliozzi, M.A. (2009) A Route Improvement Algorithm for the Vehicle Routing Problem
with Time Dependent Travel Times. Proceedings of the 88th Transportation Research
Board Annual Meeting, Washington DC, 11–15 January 2009, pp. 616–636.
Ganediwalla, S., Kumar, A., Vardhan Verma, A., Parasramka, T. and Jain, C. Retail
Technology – The Next Frontier. RedSeer, 2019. https​:/​/re​​dseer​​.com/​​repor​​ts​/re​​tail-​​
techn​​ology​​-the-​​next​-​​front​​ier/
Gendreau, M., Laporte, G. and Potvin, J.-Y. “Metaheuristics for the VRP,” in Toth, P. and Vigo,
D. eds, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics
and Applications, Philadelphia, pp. 129–154, 2002.
Hashimoto, H., Yagiura, M. and Ibaraki, T. “An iterated local search algorithm for the time-
dependent vehicle routing problem with time windows,” Discrete Optimization, vol. 5,
pp. 434–456, 2008.
Ichoua, S., Gendreau, M. and Potvin, J.Y. “Vehicle dispatching with time-dependent travel
times,” European Journal of Operational Research, vol. 144, pp. 379–396, 2003.
Kok, A.L., Hans, E.W. and Schutten, J.M.J. “Vehicle routing under time-dependent travel
times: the impact of congestion avoidance,” Computers & Operations Research, vol.
39, no. 5, pp. 910–918, 2012.
Kumar, S. and Panneerselvam, R. “A time-dependent vehicle routing problem with time
windows for e-commerce supplier site pickups using genetic algorithm,” Intelligent
Information Management, vol. 7, pp. 181–194, 2015. doi:10.4236/iim.2015.74015.
Kuo, Y. “Using simulated annealing to minimize fuel consumption for the time-dependent
vehicle routing problem,” Computers & Industrial Engineering, vol. 59, no. 1, pp. 157–
165, 2010.
Meng, F., Ding, Y., Li, W. and Guo, R. “Customer-oriented vehicle routing problem with
environment consideration: two-phase optimization approach and heuristic solution,”
Mathematical Problems in Engineering, vol. 2019, 2019, Hindawi. https://doi​.org​/10​
.1155​/2019​/1073609
Nanda Kumar, S. and Panneerselvam, R. “A survey on the vehicle routing problem and
its variants,” Intelligent Information Management, vol. 4, no. 3, pp. 66–74, 2012.
doi:10.4236/iim.2012.43010.
Nanda Kumar, S. and Panneerselvam, R. “Development of an efficient genetic algorithm for
the time dependent vehicle routing problem with time windows,” American Journal of
Operations Research, vol. 7, pp. 1–25, 2017. doi:10.4236/ajor.2017.71001.
Pisinger, D. and Ropke, S. “A general heuristic for vehicle routing problems,” Computers &
Operations Research, vol. 34, no. 8, pp. 2403–2435, 2007.
Schyns, M. “An ant colony system for responsive dynamic vehicle routing,” European Journal
of Operational Research, vol. 245, no. 3, pp. 704–718, 2015.
Yu, B., Yang, Z. and Yao, B. “An improved ant colony optimization for vehicle routing prob-
lem,” European Journal of Operational Research, vol. 196, no. 1, pp. 171–176, 2009.
 13 Application of Machine
Learning for Fault
Detection and Energy
Efficiency Improvement
in HVAC Application
Umashankar Subramaniam, Sai Charan Bharadwaj,
Nabanita Dutta, and M. Venkateshkumar

CONTENTS
13.1 Introduction................................................................................................... 263
13.2 Sustainable Living in Buildings....................................................................264
13.3 Energy Scenario in Buildings........................................................................ 265
13.4 Digitalization and Artificial Intelligence for Energy Efficiency in
Buildings........................................................................................................ 269
13.4.1 Artificial Intelligence (AI)................................................................. 269
13.4.2 Internet of Things (IoT)..................................................................... 271
13.4.3 Machine Learning (ML).................................................................... 272
13.4.4 Influence of Digitalization on HVAC Systems.................................. 273
13.4.5 Energy Optimization and Scheduling............................................... 273
13.4.6 Predictive Maintenance and Fault Diagnosis.................................... 273
13.5 Conclusion..................................................................................................... 275
Conflict of Interest.................................................................................................. 277
References............................................................................................................... 277

13.1 INTRODUCTION
Sustainability is a continuous process of maintaining climate change where the
usage of natural resources for technological advancement has a neutral effect on
the ecosystem [1]. Whereas sustainable development refers to “comprehensive and
holistic approach of temporal process that will lead to an end point of sustainability.”
Sustainable living is a socio-economic challenge. United Nations Development
Programme (UNDP) has listed out 17 major sustainable development goals [2] and
this forms the basis for leaders and policymakers across the world as a visionary

263
264 Artificial Intelligence (AI)

framework for adopting and enforcing law of the land. Sustainable living describes
the human lifestyle of using natural resources in such a way that it has a neutral
effect on the ecology and ecosystem. The people who generally participate and prac-
tice a set style of living will often try to contemplate to reduce their carbon footprint
by every means: transport, energy, water, etc. It is often described as a form of liv-
ing that meets socio-economic, cultural, and ecological needs without jeopardizing
Nature for future generations.
Today, in India, pollution has become a vital issue. Globally, CO2 emission has
reached 37.1 billion tons, increasing at a rate of 2.7% compared to the previous year.
India is likely to add at least 6.3% on the estimate in 2017, i.e. 2.6 billion tons [3].
The buildings all over the world contribute significantly to carbon emissions and
are the major stakeholders in the end user energy consumption. The construction
sector accounts for 9.1% [4] of the country’s gross domestic product (GDP) and it is
contributing significantly to CO2 emissions. Globally, the building sector consumed
125 EJ in 2016; building and construction industry together accounted for 36% of the
total energy consumption. Now the onus lies on India to push the construction sector
towards green and sustainable buildings.

13.2 SUSTAINABLE LIVING IN BUILDINGS

A sustainable building [5] definition says it is the building impact that will be reduced
if the energy water used for building energy purpose increases with the increase in
efficiency of the system. It will keep the bad impact on human health and environ-
ment should be better constructed for design operation and maintenance.”
The most indispensable parameter for sustainable buildings is energy and water
consumption and conservation. An ideal sustainable building comprises a power
generation unit for generating power by renewable means such as solar photovol-
taic (PV) and wind. This significant feature highlights that sustainable buildings
virtually do not require any power from the grid and are isolated from the grid or
from a local microgrid supplying power to other small buildings in its vicinity. In
some parts of the world where harnessing renewable energy is not possible inside
the buildings, they are connected to the grid purchasing or operating by renewable
forms of energy. To reduce CO2 emissions from the buildings, it is highly essential to
optimize the energy usage inside the buildings.
Another chief objective of the green buildings is to conserve water. Seventy per-
cent of India’s population lives in rural areas, with each household consuming on
an average 117–125 litres [6] of water per day – washing consuming the water most.
Eighty-six percent of the residential buildings do not have rainwater harvesting pits in
India. With the recent implementation of rainwater harvesting pits through National
Rural Employment Guarantee Scheme (NREGS), there have been slight changes in
the groundwater level in rural India. Most of the cities which are the driving engines
of the economic growth in the country are facing acute water crisis. These include
Delhi, Bengaluru, Hyderabad, Chennai, and emerging cities like Pune, Amravati,
and Coimbatore. Today roughly about 0.5 billion people live in Bangalore, includ-
ing more than 2 million information technology (IT) professionals; 40% of the
Machine Learning for Fault Detection 265

FIGURE 13.1  Per capita water consumption per day (in litres) in metros of India.

population is chiefly dependent upon groundwater[7]. Some experts say that at least
21 cities in India will reach zero groundwater level by 2020, which is quite alarm-
ing. In the following sections, we will look at various energy consumption scenarios
in the buildings worldwide as well as in the Indian context and how technological
advancements can achieve sustainable developmental goals and sustainable living
inside the buildings. Figure 13.1 illustrates the per capita water consumption per day
in major cities of India [8].

13.3 ENERGY SCENARIO IN BUILDINGS

The world is shifting its energy paradigm into a different ecosystem, but fissures are
noticeable across its key structures. The key structures are as follows:

• Affordability
• Reliability
• Sustainability

The aforementioned are closely related to each other and require comprehensive
understanding and a line of action for energy policy. In the modern era of human
civilization with progressive industrialization, globalization, and digitalization and
its policies, it is expected that cities will add a population of 1.7 billion by 2040.
This is quite alarming for policymakers, engineers, think tanks, and people who
are concerned about meeting the unprecedented energy demand. Figure 13.2 shows
world energy consumption, and Figure 13.3 shows CO2 emissions by sector across
the world [9].
This astronomical growth of energy demand is dominated by emerging econo-
mies led by India and followed by other countries. Most of the people across the
world spend 80% of their daily life inside buildings. The building and construction
sector together account for 36% of global energy consumption (Figure 13.2). Forty
percent of total CO2 emissions which includes both direct and indirect emissions is
from buildings (Figure 13.3). Energy usage in buildings continues to grow at a rapid
pace. With the adoption of energy efficiency initiatives in the early millennium,
266 Artificial Intelligence (AI)

FIGURE 13.2  World’s end energy consumption (World Bank 2015).

FIGURE 13.3  World CO2 emissions by sector.

there has been a significant energy saving over the last decade and half; otherwise,
the energy usage would have been 31% more by 2022 [10].
Coming to the Indian context, industry consumes the highest energy with
around 40%. Buildings are the second largest consumers of end energy, fol-
lowed by transport and others (Figure 13.4). Figure 13.5 represents the percent-
age share of end user energy consumption in commercial buildings of India,
which illustrates that heating ventilation and air-conditioning (HVAC) systems
consume about 55% of total energy consumption. Figure 13.6 provides a com-
parison of end user energy consumption among various units inside the buildings
Energy usage in buildings is dominantly contributed by heating, cooling, light-
ing, pumping, and appliances, sometimes synonymously called HVAC systems
(Figures 13.5 and 13.6) [11].
Machine Learning for Fault Detection 267

FIGURE 13.4  India’s energy consumption in 2017.

FIGURE 13.5  Energy consumption in commercial buildings in India.

Cooling: Energy usage for cooling the space inside the building grew three times
faster than any other end usage inside buildings from 1996 to 2016. Space cooling is
typically characterized by an electrical fan powered by an electrical source or an air-
conditioner. Cooling is a major social challenge as out of 2.8 billion living in the hot-
test parts of the globe, only 8% have access to cooling facilities (Figure 13.7) [12, 13].
This increase in energy demand creates great thrust for power generation as well
as distribution to meet the peak load demand of end energy usage as well as its
driving agents, resulting in increased emissions due to power generation. Electrical
energy is the driving force for the economic activity and development of infrastruc-
ture in the country. Energy demand is significantly influenced by the increase in
population and economic growth in the hottest demographic regions of the world:
268 Artificial Intelligence (AI)

FIGURE 13.6  Energy consumption by various components in Indian buildings.

FIGURE 13.7  AC sales in India 2017–2018.

India, China, Indonesia, and others. Household fans: Electrical fan is the simplest
form of cooling. It is estimated that 2.3 billion residential fans have been under usage
till 2016. Fifty-five percent of the global household has at least one fan in use. Fans
consume about 10% less energy than the air-conditioners. Fans will continue to play
a major role in the cooling demand in the developing countries as fans are much
more affordable than air air-conditioners [13]. Heating: It is the largest component of
energy uses in buildings, accounting for 36% of the total building energy consump-
tion according to the 2018 IEA Report [14]. Pumps provide space heating, cooling,
and hot water in buildings. Pumping is a predominant technology that has evolved
over the years and is being used in AC’s chillers and other systems. Heat pumps are
highly efficient and are commercially proven. Over the decades, the typical pumps
in the residential sector are
Machine Learning for Fault Detection 269

• Split or room AC
• Air water heater pump
• Water to water heater pump
• Ground source heat pumps

Lighting: Lighting applications in India accounts for about 59% of total energy
consumption in residential buildings and 26% in commercial buildings. With the
advent of compact fluorescent light (CFL), light-emitting diode (LED) lights, and
their market-penetrative investments, savings of over 6000 PJ have been achieved
in OCED countries. India is primarily leading the LED market. Sale of LEDs is
expected to exceed by 15% across the world. India has already distributed 50 mil-
lion LEDs until 2017.
In 2017, the lighting sector [15] had a demand of 2 EJ whereas the non-residential
sector had a demand of 4 EJ. With the policy implementation and effective invest-
ment into the market, LED account for 25% of the installed capacity.
Appliances: Appliances alone have consumed 30% of global fuel energy and
accounts for 17% of the world’s total end energy consumption [16]. As previously
mentioned, India is an emerging economy associated with meteoric urbanization and
increasing purchasing capacity of the individuals. These are some of the striking fac-
tors for mounting up the floor area of appliances. Energy consumption by appliances
has also swelled to a whopping level of 58% since 2000. It is estimated that appli-
ances will contribute 9% of the total energy in buildings with the implementation of
energy-efficient policies.
It is worth mentioning that India has taken a major step forward with its first
ever building energy code for residential buildings named Energy Conservation
Building Code (ECBC) in 2017 [17]. To sum up, the energy usage will now tend
to remain flat until 2040 despite the increase in the floor capacity of 60%, making
buildings 40% efficient. India has saved 6% additional energy since 2000. China
and India are the major energy consumers, with a stake of 82%, among the six
emerging economies in 2017; they contribute 5% of fossil fuel and 145 MT CO2
emissions (Figure 13.8)] [18].

13.4 DIGITALIZATION AND ARTIFICIAL INTELLIGENCE

FOR ENERGY EFFICIENCY IN BUILDINGS
Early computers used to occupy a lot of space and perform small calculations via
vacuum tubes, but in later stages Central Processing Unit (CPU) and other modifica-
tions were made. Over the years, man has achieved the capacity and capability of
developing computer systems with reduced size and yet increasing computational
capacity and performance speed in all dimensions (Figure 13.9).

13.4.1 Artificial Intelligence (AI)

It is the branch of computation science which aids computers and machines to
think like humans. It is the science and engineering that make intelligent machines
270 Artificial Intelligence (AI)

FIGURE 13.8  World’s energy savings projected for 2040.

FIGURE 13.9  Illustration of AI subjects.

especially intelligent. Integral philosophy and motto of this is “can a machine think
and behave like humans” [19].
This makes a loud statement that it is a technology of creating machines that
exhibit characteristic features like thinking, behaviour, sense, learn, explain, and
adhere. Thus, artificial intelligence is a multidimensional, cross-functional, and mul-
tidisciplinary subject involving computer science, physics, biology, psychology, lin-
guistics, arts, sociology, engineering, mathematics, and others, as depicted in Figure
13.10. AI systems will adopt the changes without explicitly being programmed,
rather than being specifically intended to perform a task (Figure 13.11)].
Machine Learning for Fault Detection 271

FIGURE 13.10  Some applications of AI.

FIGURE 13.11  Various stages in typical IoT system.

13.4.2 Internet of Things (IoT)

The term Internet of Things was first coined by Kevin Ashton [20] in early 2000,
which laid the foundation for the IoT at Massachusetts Institute of Technology (MIT)
labs subsequently. [21] Recently, the word IoT has been added to Oxford Dictionary,
which says “that it is an interconnection via internet of computing devices embedded
in everyday objects enabling them to send and record data.”
IoT systems generally include artificial intelligence communication protocols,
census data, acquisition systems, and actuators in smart device for usage and active
engagement (Figure 13.12) [22].
The various stages and building blocks of an IoT system are as follows:

• Create: Use of sensors to record the physical event or a state.

• Communicate: Transfer of recorded event for a communication protocol.
• Analyse: Analysing the patterns behaviour redundancies of the aggregated
data to provide predictive, presumptive, prescriptive forecasting insights:
transfer of recorded event for a communication protocol (Table 13.1)
• Act: Changing and maintaining initiative state or event upon the analysis.
272 Artificial Intelligence (AI)

FIGURE 13.12  Various communication protocols used in IoT.

TABLE 13.1
Some Commonly Used Sensors Inside Buildings
S. No. Sensor Type/Name Comments
1 Temperature and RH Used for measuring temperature and humidity of room say DHT11
or LM35
2 CO2 and smoke sensor Useful in detection of gas leakage for home as well industry, e.g.
MQ2 smoke sensor
3 Water flow rate Measure water flow from the sensors, e.g. YF-S201
sensors
4 Motion detect sensors Used to detect the human activity, e.g. HC-SR501 PIR sensor
5 Acoustic sensor Used for sound detection, e.g. KY-038
6 Light sensors Deployed for luminosity and light measurement TSL2561 or LDR
8 HVAC and smart Used for measuring power and energy, e.g. smart energy meters
energy meters
9 Imaging Thermal IR- or CMOS-based sensors to detect and analyse
occupancy patterns

13.4.3 Machine Learning (ML)

It is a subset and a sister branch of AI which is used to analyze the data which learns
and runs through data without being explicitly told. It performs exceptionally well
if the data is very large (big data). Data can be in numerous forms such as analogue
sensors, images, acoustics, human activity, and others.
Machine Learning for Fault Detection 273

FIGURE 13.13  Simple machine learning block diagram.

Machine learning [23] has been basically divided into three types: supervised
learning, unsupervised learning, and reinforcement learning. It is a form of cognitive
theory and science. This machine learning models can be used inside buildings for
predictive modelling of the energy and for fault detection of the various equipment
and appliances inside buildings.
The basic and simple flow block diagram of machine learning is shown in
Figure 13.13.

13.4.4 Influence of Digitalization on HVAC Systems

The data collected by sensors are stored in the server and the next step is to auto-
mate the further processes to start analyzing the data trends related to the cooling
or heating pattern, energy usage, peak load demands [24], occupant behaviour [25],
and overall thermal comfort inside the space of the building. These insights provide
more degrees of freedom and operational flexibility to regulate chiller or heating
equipment to achieve energy saving (Figure 13.14).

13.4.5 Energy Optimization and Scheduling

Let’s say a building has 100 data points or nodes that are being connected to a central
hub where data has been stored in the server. The thermal comfort inside the build-
ing is a function of human activity, flooring, walls, space area, equipment, solar radi-
ation, and the presence of electronics. In order to cool a space in a building, chiller
has to blow cold air into the space to decrease the temperature. Over the course of
the day, there will be various levels of requirements of cooling and heating pattern
in accordance with the different nature of humans and activities. The role of IoT and
AI-based systems is to record, analyze, and interpret the data to provide meaningful
outcomes of energy usage at peak times, flat time, and no load periods by blending
occupant behaviour into the model to achieve localized control of thermostat [26].

13.4.6 Predictive Maintenance and Fault Diagnosis

As previously mentioned, machine learning models which are part of artificial
intelligence can be used for predictive maintenance and fault diagnosis of systems.
274 Artificial Intelligence (AI)

FIGURE 13.14  Typical IoT–AI system for smart building.

Pumps, which are an integral part of the buildings, also contribute significantly to
energy consumption as most of the time it is driven by an induction motor, which is
subjected to various faults (Figures 13.15 and 13.16) [27].
Data from vibration sensor smart flow rate meters and pressure sensors will
allow us to predict the pumping faults well in advance to save unnecessary excessive
energy usage and sudden collapse of the pump inside the buildings. Such a kind of
predictive fault diagnosis [28] can be extended to heating, cooling, and ventilation

FIGURE 13.15  Water pump monitoring inside a building.

Machine Learning for Fault Detection 275

FIGURE 13.16  Vibration sensor data collected to analyse real-time faults in pumps.

as well. If any fault happens in the motor or water pumping system, the trained data
of the motor or pumps will be fed to the simulation and predictive control software
which will classify the faults using training and testing data with the help of machine
learning software. Support vector machine algorithm is one of the machine learning
algorithms which will identify the faults in the system in very less time. So for fault
classification, SVM algorithm has been used which classified the faulty and no faulty
points from unclassified data (Figure 13.17).
Other interesting aspects of artificial intelligence are listed below:

• Sending in app notifications of problematic equipment.

• Notifying service provider about the equipment repair at a particular loca-
tion on a particular building.
• Moral consolidation and energy consumption demand generating reports
and forecasting models to grid authority’s for effective scheduling.

To achieve this in greater efficacy in the system, we need to monitor each and every
facility management services and deployment of energy harvesting methods to pro-
vide power to sensors and AI-powered devices to conserve energy requirement to
carry out this service.

13.5 CONCLUSION
Buildings account for 36% of the total energy consumption and 28% of the total
carbon emissions. The construction sector in India accounts for 9% of the GDP.
On an average, an Indian citizen consumes around 125–150 litres of water per day
for his essential needs. To address this challenge, India needs to push towards the
green and sustainable buildings incorporating digital technologies like IoT in fusion
with data-driven machine learning models to conserve energy. There are numerous
applications that are being relied upon, including artificial intelligence. AI can be
implemented even in old buildings using retrofit methods. To achieve low carbon
276 Artificial Intelligence (AI)

FIGURE 13.17  Classification of faults using SVM algorithm.

emissions and higher energy efficiency, we need to adapt and implement new tech-
nologies that will conserve energy in buildings. Buildings provide a huge poten-
tial for energy savings and reducing the carbon emission and thereby achieving a
2°C reduction in global temperature. A 30% improvement in the global average of
the building energy intensity by 2030 is needed to meet the ambitious goals of the
Paris Accord 2015. Machine learning and deep learning technology have brought
remarkable solutions by enabling both continuous monitoring and ground-level fault
detection. The prediction control app-based device based on the hybrid technol-
ogy combined with machine learning, deep learning, and IoT, GPU-based human–
machine interface can be helpful to predict the fault condition before the system is
totally shut down. So whenever any abnormal condition is detected, the machine will
indicate through the alarm that the system is in danger, so it is possible for the system
operator to identify the faulty part easily and take necessary action. This predictive
controller GPU-based device is applicable not only for pumping system but also
for any heavy industry. Machine learning and deep learning technology have been
extensively applied in the biomedical sector, but in the industrial application and
mainly in the pumping system it is very rare. The predictive control hybrid model is
the new point of study where the researcher is planning to reduce the energy loss and
the time of the process and thus making the system flawless. This study is trying to
find out the possibilities of anomalies in the HVAC system and its solution by using
an AI technique.
Machine Learning for Fault Detection 277

CONFLICT OF INTEREST
The authors declare no conflict of interest.

REFERENCES
1. U. S. E. P. Agency, Learn About Sustainability, United States Environmental Protection
Agency, [Online]. Available: https​:/​/ww​​w​.epa​​.gov/​​susta​​inabi​​lity/​​learn​​-abou​​t​-sus​​taina​​bi​
lit​​y​#wha​​t. [Accessed 16 February 2019].
2. U. N. D. Programme, Sustainable Development Goals, UNDP, [Online]. Available:
http:​/​/www​​.undp​​.org/​​conte​​nt ​/un​​dp​/en​​/ home​​/sust​​a inab​​le ​- de​​velop​​ment-​​​goals​​.html​.
[Accessed 15 February 2019].
3. U. Goswami, “India’s CO2 Emissions Forecast to Increase by 6.3% This Year,”
Economic Times, 06 December 2018. [Online]. Available: https​:/​/ec​​onomi​​ctime​​s​.ind​​
iatim​​es​.co​​m ​/new​​s​/pol​​itics​​-and-​​natio​​n ​/ind​​ias​-c​​o2​-em ​​issio​​ns​-fo​​recas​​t​-to-​​incre​​ase​-b​​y​- 6​
-3​​-this​​-ye​a r​​/arti​​clesh​​ow​/66​​96310​​9​.cms​. [Accessed 13 February 2019].
4. C. India, India in Figures, MOSPI, 2018. [Online]. Available: http:​/​/www​​.mosp​​i​.gov​​
.in ​/s​​ites/​​defau​​lt ​/fi​​les​/p​​ublic​​ation​​_ repo​​r ts​/ I​​ndia_​​in​_fi​​gu​res​​-2018​​_ rev.​​pdf. [Accessed 15
February 2019].
5. J. C. Howe, & M. Gerrard (2010). The Law of Green Buildings: Regulatory and Legal
Issues in Design, Construction, Operations, and Financing. American Bar Association.
6. O. &. T. Singh, “A Survey of Household Domestic Water Consumption Patterns in
Rural Semi-Arid Village, India,” GeoJournal, vol. 78, no. 5, pp. 777–790, 2013.
7. A. Shukla, “Alarming: 21 Indian Cities Will Run Out of Water by 2030,” Business
World, 19 June 2017. [Online]. Available: http:​/​/www​​.busi​​nessw​​orld.​​in​/ar​​ticle​​/
Alar ​​ m ing-​​21​-In​​ d ian- ​​ Citie​​ s​-Wil​​ l​-Run​​ - Out- ​​ Of​-Wa​​ ter​-B​​ y​-203​​ 0​​ /19- ​​ 0 6​-20​​17​-12​​ 0383/​.
[Accessed 9 February 2019].
8. S. Mungara, “Five Major Metros Staring at Severe Water Crisis: Expert,” Times of
India, 5 October 2018. [Online]. Available: https​:/​/ti​​mesof​​i ndia​​.indi​​atime​​s​.com​​/city​​/
hyde​​rabad ​​/five​​-majo​​r​-met​​ros​-s​​tarin​​g​-at-​​sever​​e​-wat​​er​-cr​​isis-​​exper ​​t ​/art​​​icles​​how​/6​​60801​​
51​.cm​​s. [Accessed 8 February 2019].
9. B. D. a. J. D. Thibaut Abergel, UN Environment and International Energy Agency
(2017): Towards a Zero-Emission, Efficient, and Resilient Buildings, United Nations
Environment Programme, Paris, 2017.
10. C. India, Energy Statistics 2018, Ministry of Statistics and Programme Implementation,
New Delhi, 2018.
11. D. C. P. Commission, The Final Report of the Expert Group on Low Carbon Strategies
for Inclusive Growth, Planning Commission GOI, New Delhi, 2014.
12. 2. OECD/IEA, The Future of Cooling, OECD/IEA, Paris, 2018.
13. T. E. a. R. Institute, Improving Air Conditioners in India, TERI, New Delhi, 2018.
14. OECD/IEA, Energy Efficiency: Heating, IEA, [Online]. Available: https​:/​/ww​​w​.iea​​
.org/​​topic​​s​/ene​​rgyef​​ficie​​ncy​/ b​​uildi​​ngs​​/ h​​eatin​​g/. [Accessed 1 February 2019].
15. OCED/IEA, Lighting: Tracking Clean Energy Progress, IEA, 2018. [Online]. Available:
https​:/​/ww​​w​.iea​​.org/​​topic​​s​/ene​​rgyef​​ficie​​ncy​/ b​​uildi​​ngs​​/ h​​eatin​​g/. [Accessed 16 February
2019].
16. OCED/IEA, Energy Efficiency: Appliances, IEA, 2018. [Online]. Available: https​:/​/ww​​
w​.iea​​.org/​​topic​​s​/ene​​rgyef​​ficie​​ncy​/ b​​uildi​​ngs​/a​​​pplia​​nces/​. [Accessed 16 February 2019].
17. B. G. Ministry of Power, Energy Conservation Building Code, 2017. [Online]. Available:
https​:/​/ be​​eindi​​a​.gov​​.in​/s​​ites/​​defau​​lt​/fi​​les​/ B​​EE​_ EC​​BC​​%20​​2017.​​pdf. [Accessed 16
February 2019].
278 Artificial Intelligence (AI)

18. OCED/IEA, Energy Efficiency: Buildings, IEA, 2018. [Online]. Available: https​:/​/ww​​w​
.iea​​.org/​​topic​​s​/ene​​rgyef​​ficie​​ncy​/ b​​​uildi​​ngs/. [Accessed 16 February 2019].
19. T. Point, Artificial Intelligence - Overview, Tutorials Point, [Online]. Available: https​
:/​/ww​​w​.tut​​orial​​spoin​​t​.com​​/arti​​ficia​​l​_int​​ellig​​ence/​​a rtif​​i cial​​_inte​​llige​​nce​​_o​​vervi​​ew​.ht​​m.
[Accessed 16 February 2019].
20. Wikipedia, Kevin Asthon, Wikipedia, [Online]. Available: https​:/​/en​​.wiki​​pedia​​.org/​​
wiki/​​Kevin​​_ As​ht​​on. [Accessed 16 February 2019].
21. K. Asthon, “That ‘Internet of Things’ Thing,” RFID Journal, pp. 97–114, 2009.
22. M. M. E. R. C. Jonathan Holdowsky, Inside the Internet of Things, Deloitte University
Press, UK, 2015.
23. S. a. S. B.-D. Shalev-Shwartz, Understanding Machine Learning: From Theory to
Algorithms, Cambridge University Press, Cambridge, 2014.
24. H. M. R. a. K. T. C. Zhou, “Artificial Intelligence Approach to Energy Management
and Control in the HVAC Process: An Evaluation, Development and Discussion,”
Developments in Chemical Engineering and Mineral Processing, pp. 42–51, 1993.
25. S. Salimi & A. Hammad (2019). Critical review and research roadmap of office build-
ing energy management based on occupancy monitoring. Energy and Buildings, 182,
214–241.
26. H. B. Gunay, W. Shen & G. Newsham (2019). Data analytics to improve building per-
formance: A critical review. Automation in Construction, 97, 96–109.
27. D. S. U. a. P. S. Nabanita, “Mathematical Models of Classification Algorithm of
Machine Learning,” in International Meeting on Advanced Technologies in Energy
and Electrical Engineering (IMAT3E’18), Fez, Morocco, 2018.
28. N. Dutta, “Centrifugal Pump Cavitation Detection Using Machine Learning Algorithm
Technique,” in 2018 IEEE International Conference on Environment and Electrical
Engineering and 2018 IEEE Industrial and Commercial Power Systems Europe
(EEEIC/I&CPS Europe), Palermo, 2018.
 14 Smart City Using
Artificial Intelligence
Enabled by IoT
P. Srividya and Sindhu Rajendran

CONTENTS
14.1 Introduction................................................................................................. 279
14.2 Structure of a Smart City............................................................................280
14.3 Requirements to Build a Smart City........................................................... 281
14.4 Augmented Reality and Virtual Reality in Building a Smart City............. 281
14.5 Major Components Required to Build a Smart City................................... 282
14.6 Challenges in Building a Smart City.......................................................... 283
14.7 Technologies Involved in Building a Smart City........................................ 283
14.7.1 Integrating AI and IoT to Build a Smart City............................... 283
14.7.2 Blockchain in Building a Smart City............................................284
14.7.3 Big Data in Building Smart City..................................................284
14.7.4 Robotics in Building a Smart city.................................................284
14.8 Components of a Smart City.......................................................................284
14.8.1 Smart Energy................................................................................284
14.8.2 Smart Healthcare.......................................................................... 286
14.8.3 Smart Traffic Management........................................................... 287
14.8.4 Smart Parking............................................................................... 287
14.8.5 Smart Waste Management............................................................ 287
14.8.6 Smart Lighting.............................................................................. 288
14.8.7 Smart Governance........................................................................ 289
14.8.8 Smart Agriculture......................................................................... 289
14.9 Drawbacks in Implementing Smart Cities.................................................. 290
14.10 Conclusion................................................................................................... 291
References............................................................................................................... 291

14.1 INTRODUCTION
A city is said to be smart if the residents are able to access various services and
networks efficiently. The services must be sustainable, flexible, and easily acces-
sible. The goal is accomplished by the use of information, telecommunication, and
digital technologies. Traditional cities are transformed into smart cities by the use of
information and communication technology (ICT). Emerging technologies like IoT,

279
280 Artificial Intelligence (AI)

BD, blockchain, and robotics help to make a city smart. Developing a smart city is a
collaborative work that involves many government and private organizations, public
sectors, and citizens themselves.
The intention of building a smart city is to overcome the problems arising from
rapid urbanization and increasing population. Even though the initial cost involved in
developing a smart city is high, in the long run it helps in reducing energy consump-
tion, environmental pollution, city wastes, water consumption, traffic congestion,
and other problems. Different transportation forms, pollution management, Wi-Fi
systems, green energy, universal identification, and local commerce promotion are
some of the initiatives of building a smart city.
Some of the facilities that can be included in a smart city are public Wi-Fi sys-
tems, mobile apps for citizens to report congestion of roads, systems to manage and
monitor electricity and water usage, systems to intimate contamination of water
resources, systems to identify potholes in the roads, systems to manage parking sys-
tems, and many others.
A city is said to be smart if it possesses the following features [1]:

• Usage of digital technologies and electronics in building the city

infrastructure.
• Usage of ICT in transforming living conditions and working environment.
• ICT-embedded government systems.
• Usage of ICT to bring about innovation and in enhancing the knowledge
offered.

The infrastructure of smart city includes the physical infrastructure, installed sen-
sors, software, and firmware.
This chapter covers the various aspects of the smart city which include the struc-
ture of a smart city, requirements to build a smart city, various components involved,
challenges faced, and technologies involved in building a smart city. This chapter
also highlights the challenges faced in making a city smart.

14.2 STRUCTURE OF A SMART CITY

Structure of a smart city is shown in Figure 14.1. It involves the following three-
layered structure:

1. Technology base: Includes smartphones and different sensors to collect the

data in real time. High-speed communication networks and open-data por-
tals are used to connect the sensors.
2. Specific applications: The raw data acquired from the sensors are to be
translated into alerts and actions. This can be done by application develop-
ers and technology providers.
3. Public usage: The applications so developed must be adopted by the public
to change the lifestyle.
Smart City Using Artificial Intelligence 281

FIGURE 14.1  Structure of a smart city.

14.3 REQUIREMENTS TO BUILD A SMART CITY

The process of building a smart city is an iterative process, with processing and
analysis at each iteration step. It forms the city’s backbone.
IoT provides seamless connectivity to all devices like wearable, entertainment,
home appliances, public mobility, medical devices, connected vehicles, buildings,
agriculture, and services that go beyond machine-to-machine communication. Any
IoT platform requires the following things to build a smart city:

• Smart things like sensors, actuators, cameras, etc. to gather the data.
• Tools to analyse the data collected by sensors.
• Cloud gateways to collect the data from IoT devices, store them, and for-
ward them to cloud in a secured manner.
• Data processors to collect the data, store it, and distribute it to control
applications.
• A data warehouse to clean and organize the collected data.
• Algorithms to automate city services and to improve the performance of
various control applications based on data analysis.
• Applications to send the data to the actuators.
• Applications to connect smart things and users.

14.4 AUGMENTED REALITY AND VIRTUAL

REALITY IN BUILDING A SMART CITY
Augmented reality (AR) uses digital technology to enhance the images captured
using smartphones. Additional digital information in the form of text or image can
also be added to the captured image using apps.
Virtual reality (VR) uses computer technology to create altogether a different
reality than the one that is in front of us.
282 Artificial Intelligence (AI)

AR and VR find application in the following streams:

1. Patrolling: AR systems provide relevant environmental details to the offi-

cials, which enable them to be well prepared before entering treacherous
situations.
2. Education: AR and VR provide thrilling experiences to the students attend-
ing the classes. It allows students to explore unfamiliar things in depth. This
attracts more students to attend classroom sessions.
3. Emergency management: To rescue the needy, an interactive map is distrib-
uted through AR to the responders to help the residents to plot their loca-
tion. This helps in locating the safest rescue route and also in identifying
unsafe areas.
4. Mental health services: AR and VR help in treating certain mental health
problems like anxiety.

14.5 MAJOR COMPONENTS REQUIRED TO BUILD A SMART CITY

1. Technology: AR and VR play a vital role in the development of smart cities.
Both VR and AR need an enormous amount of information to be deliv-
ered quickly in order to provide their experience. This demands a powerful
network that can be established using 5G technology. With the ability to
provide data in real time, AR and VR help in transforming government
services.
2. IoT: It is the nerve of the city that keeps all the devices connected. It
offers connectivity to all smart devices, connected vehicles, smart home
appliances, smart buildings, smart public mobility, and other services
[2, 3].
3. Network of smart things: This includes sensors, actuators, and cameras for
collecting data. Every physical device which makes up IoT ecosystem has
a sensor embedded in it. Sensors collect and transmit the data to the cloud.
IoT interconnects all the devices and makes them work together.
4. Cloud gateways: It is helpful in collecting data from low-power IoT devices,
store the data, and forward the data securely to the cloud.
5. Data processor: It is helpful to collect the data from various data streams
and distribute it to data lake.
6. Data lake: It is required to store all the raw data.
7. Data warehouse: It is essential to clean the data and structure it.
8. Geospatial technology: It plays a vital role in providing the location and the
necessary framework in collecting the data and analyzing it.
9. Artificial intelligence: The vast amount of data collected in the smart city
becomes useless if it is not properly processed to generate information.
AI algorithms and techniques help in processing and analyzing the data
obtained from machine-to-machine interaction.
10. Control applications: It is required to send commands to IoT actuators.
11. User applications: It is helpful in connecting smart things and citizens.
Smart City Using Artificial Intelligence 283

14.6 CHALLENGES IN BUILDING A SMART CITY

Some of the challenges faced when a smart city is being built are data collection,
security issues, technology involved, and requirement for expertise.

1. Data collection: The problems involved are as follows:

• Overloading of data – noise and heterogeneous data
• Issues due to interoperability
• Managing open data
2. Security issues: Some of the security issues involved are as follows:
• Cyberattacks on interconnected critical regions which are complex and
enormous
• Significant implications due to the attack
• Sharing responsibility to shield the city from attacks
3. Technology involved: The main challenges involved in technology usage
are as follows:
• Providing network coverage throughout the city
• Managing network capacity
• Reusing existing infrastructure
4. Requirement for expertise: The main setback in expertise involve the
following:
• Lack of training
• Insufficient funds for training

14.7 TECHNOLOGIES INVOLVED IN BUILDING A SMART CITY

Various technologies like AI, IoT, blockchain, big data, and robotics can be used to
build a smart city as explained in the following sections.

14.7.1 Integrating AI and IoT to Build a Smart City

The goal behind developing a smart city is to optimize costs, reduce wastages, and
provide better living standards for the residents by incorporating various aspects of
technologies in the system. The numerous fields of AI and IoT together have a greater
potential in transforming the traditional city into a smart one.
Various branches of AI can be used to analyzse how the city is being used by
the people. The AI’s pattern recognition technology can be used to analyze a huge
amount of raw data, patterns, and hidden correlations collected from different sen-
sors installed all over the city for various purposes like road congestion monitor-
ing, weather monitoring, etc. Analysis of sensor data using AI tools helps to create
predictive models that can be used to control applications to send commands to IoT
devices. Deep learning can be used to track the speed of the vehicles, read the license
plate on the vehicles, count the number of vehicles passing by and the number of
pedestrians walking on the road, and track the vehicles parked in the parking lot and
establish patterns.
284 Artificial Intelligence (AI)

A smart city built with AI powered by IoT can be used in many application areas
like smart environment, smart buildings, smart traffic management, smart grid,
smart farming, smart waste management, smart health sectors, smart security sys-
tems, smart transport, smart parking, drone delivery, smart postal services, smart
disaster and calamity managements, and many more.

14.7.2 Blockchain in Building a Smart City

Even though the usage of blockchain to build a smart city is a pretty new concept,
it assists in connecting the services of the smart city at the same time and increases
security and transparency. Blockchain system provides security and privacy to data
by encoding the information captured and storing it on cloud. This also reduces paper
work, pollution, and wastage. Blockchain network can also be used in healthcare ser-
vice to store health information required in emergency situations. This proves to be
helpful in accessing relevant information. It can also assist in billing and transaction
processes as well as in sharing smart grid energy.

14.7.3 Big Data in Building Smart City

Big data, IoT, and smart city are strongly correlated as one entity is interdependent on
other two entities. The data generated in a smart city can be in the form of text, videos,
and images from various sensors, websites, social media, RFID data, atmospheric data,
and various data bases. This entire fertility of data generated in a smart city consti-
tutes big data. The data collected from different sources must be processed and ana-
lyzed using various algorithms to retrieve information hidden in them. The challenges
involved in this are manifold. This includes data searching, data capture, visualization,
data analysis, data storage, and data sharing. Big data analysis requires new approaches
like enhanced decision-making capacity, in-depth analysis, and process optimization.

14.7.4 Robotics in Building a Smart city

Collaboration of humans and robots can work out wonders in all the sectors of smart
cities. Robot integration is transforming technologically advanced cities into smart cit-
ies. This transformation only depends on how humans are efficiently using the robots
in bringing about a change. Robots can be used in various applications like surveillance
and policing, transportation, traffic management, hotel service maintenance, and others.

14.8 COMPONENTS OF A SMART CITY

The main components involved in building a smart city is illustrated in Figure 14.2.

14.8.1 Smart Energy
As energy is an important requirement for our day-to-day activities, hence conserv-
ing energy is a vital factor. Energies are classified as sustainable, renewable, and
Smart City Using Artificial Intelligence 285

FIGURE 14.2  Components involved in building smart city.

non-renewable energies. Non-renewable energy is of greater concern as this type is

not replenishable.
The term smart energy refers to the utilization of ICT in energy utilization and
generation [4]. The smart energy model is based on developing smart power grids,
smart power generation, smart storage, and smart consumption. Smart energy aims
at reducing the carbon emission, distributing the energy efficiently to the end users,
and optimizing the energy utilization, as shown in Figure 14.3.
Carbon emission can be reduced by using solar, wind, hydroelectric, or photo-
voltaic energies. Smart distribution can be achieved by using smart grids and smart
infrastructures enabled by ICT. Smart grid facilitates to integrate different sources
of energies like fossil fuel, photovoltaic, wind, and other energy forms onto a single

FIGURE 14.3  Smart energy system.

286 Artificial Intelligence (AI)

grid. The energy from various sources is synchronized by the grid to deliver elec-
tricity of specified frequency and voltage without any fluctuations. The grid ensures
efficient, secured, and sustainable energy system by integrating the source of energy
to end users. Smart infrastructure is the crucial component in a smart energy system.
It is responsible for gathering the energy consumption information and for sharing
the unit rate information to the users.
Optimization of energy utilization can be achieved through smart storage and
smart metering. Smart energy storage can be obtained by using lithium-ion batter-
ies or fuel cells that have a longer life and prove to be effective in energy storage
and its efficient delivery. The smart energy meter records the energy consumed for
the specified time duration and sends the reading to the central office for monitor-
ing and billing.

14.8.2 Smart Healthcare
Healthcare sectors have many voids like the ratio of growing population to the num-
ber of medical practitioners is unmatchable, mistakes by the hospitals in diagnos-
ing diseases, patients receiving wrong treatments, patients not getting immediate
treatment in some emergency situations, and many more. To meet the demands of
the growing population with limited resources, the healthcare sectors must become
smart, intelligent, and efficient. Smart healthcare sector integrates the use of biosen-
sors, wearable devices, ICT, cloud computing, mobile apps, emergency services like
ambulance with the traditional healthcare [5].
Some of the smart healthcare applications include periodical intimation to the
nurses and doctors regarding the patient’s health condition, transferring the real-
time data of the patients to an expert doctor who resides in a far-off place for his
expert opinion, automatic intimation to the pharmacist in case of medicine require-
ment to a patient, and telemedicine. The idea of the smart healthcare sector is shown
in Figure 14.4.

FIGURE 14.4  Smart healthcare.

Smart City Using Artificial Intelligence 287

FIGURE 14.5  Smart traffic management.

14.8.3 Smart Traffic Management

With the growing population and digitalization, AI technique is playing a vital role.
The ever-booming IoT can be used to implement smart traffic solutions to ensure that
residents of a smart city commute in the city safely and quickly [6]. There exists a
central traffic management system that receives real-time data about the flow of traf-
fic via road surface sensors and cameras, the real-time data captured from camera is
analysed and it notifies the user regarding the signal malfunctions and congestion,
as shown in Figure 14.5.

14.8.4 Smart Parking
Due to narrow roads and increase in traffic, parking is always a problem, especially
during holidays which is a real struggle. IoT smart parking solutions can be imple-
mented using road surface sensors embedded in the parking area, which can deter-
mine whether the slots are occupied or free and provides a map for parking on a
real-time basis. A mobile app can also be developed to locate the available parking
slot and also pay for the slots, as shown in Figure 14.6. Smart parking helps with
congestion reduction and reduces carbon emissions.

14.8.5 Smart Waste Management

Disposal of waste is one of the major issues to be dealt with since it emphasizes on
the hygiene and eruption of diseases across the place. Proper disposal and collection
of waste is an important service as there is an increase in population. A sustainable
waste management system can be provided by implementing AI for smart recycling
and waste management. Figure 14.7 shows the waste management system comprising
sensors and devices attached to bins, which sends notification to municipal authori-
ties to clear the waste collection as soon as it is about to be filled. Separate bins for
paper, plastic, glass, and waste food items can be maintained in every locality.
288 Artificial Intelligence (AI)

FIGURE 14.6  Smart parking system.

FIGURE 14.7  Smart waste management.

14.8.6 Smart Lighting
Smart lighting emphasizes on less consumption of energy. Street lights are neces-
sary, but they consume a lot of energy. This can be reduced by the use of smart light-
ing. Since street lights are necessary, sensors are attached or there are Wi-Fi hotspots
which automatically adjust the brightness depending upon the presence of pedes-
trians and automobiles. It employs a mesh network in real time to trigger adjoining
lights and to create a safe circle of light around the localities. Figure 14.8 illustrates
smart lighting system.
Smart City Using Artificial Intelligence 289

FIGURE 14.8  Smart lighting system.

14.8.7 Smart Governance
A smart city aims at providing sustainable high-quality amenities to the citizens.
This necessitates all the decision-makers to line up their goals and to work together
in planning all the activities. Smart use of the available resources demands the inter-
actions between different government agencies. Best results can be achieved when
citizens are also allowed to participate in urban planning activities. These changes
can be brought about by adopting smart governance. Smart governance involves the
use of ICT sensibly in order to improve decision-making. It involves the process of
refining the democratic process and changing the methods of delivering public ser-
vices. This is done by better collaboration among different stakeholders, including
government and citizens. Compliance towards the needs of people and improved
decision-making using data, evidence, and other resources are emphasized through
smart governance.
The following are the main aims of smart governance:

1. Simple laws, rules, and regulations and comprehensible government.

2. New system of governance with moral values.
3. Transparency in governance. This brings about a great transformation in
the system of governance by reducing corruption.
4. A responsive administration to attend the difficulties faced by the citizens
as quickly as possible.

14.8.8 Smart Agriculture
The worldwide increase in population has increased the demand for food. In traditional
farming methods, farming is usually done based on predications. Many times it leads
to failure of crops. Due to this, farmers incur huge losses. For efficient cultivation, soil
290 Artificial Intelligence (AI)

FIGURE 14.9  Smart agriculture.

moisture, soil quality, air quality, sufficient irrigation, and crop weeds, proper fertil-
ization plays an important role. If these parameters are monitored automatically, and
suitable information is passed on to the farmers, quality and quantity of cultivation
increase. This necessitates for smart agriculture that uses IoT, big data, GPS, and con-
nected devices. Smart agriculture assists in automated farming and also in collecting
and analyzing the data from the field by installing sensors, cameras, and actuators [7].
The analyzed data is then sent back to the farmer. This helps in growing high-quality
crops. Figure 14.9 illustrates smart agricultural system.
Usage of IoT in agriculture can help farmers in the following ways:

1. Monitoring weather conditions, crop development, and soil quality can help
in improving the cultivation.
2. Internal processes are better controlled and this reduces the risk of inferior
production and also allows for better distribution of the products.
3. If the crop yield is known in advance, planning to sell the product can be
made.

14.9 DRAWBACKS IN IMPLEMENTING SMART CITIES

Inappropriate design of the smart city structures will result in few drawbacks:

1. Building a smart city requires huge data acquisition. In this process, there
are chances that the information about the individual might get revealed.
2. Huge investment is required to build a smart city. This investment must be
borne by the tax payers. This will also raise the cost of living.
3. The whole process involves the usage of sensors and other IoT devices
placed at various places. These devices operate with battery. This causes a
problem when battery replacement is required.
 Smart City Using Artificial Intelligence 291

14.10 CONCLUSION
With rapid population explosion in urban areas, the challenges faced by the cities
have drastically increased in an uncontrolled manner. Problems like resource scar-
city, traffic management, pollution level, and many more shoot up with a rising popu-
lation. To provide sustainable prosperity to inhabitants, cities now face economic,
political, and technological threats. This demands the establishment of smarter sys-
tems in cities to optimize the utility of the exhaustible resources and to provide
sustainable development without disturbing the environment. Thus, improvement of
cities powered by digital technologies is set at the highest priority worldwide by the
21st century. Even though the day-to-day living standards of the people in urban
areas is increased drastically due to digitization, the transformation of the traditional
city to smart city has barely begun.
In the development of smart cities, the roles and responsibilities of the participants
involved become significant. The key participants involved in building the smart city
include the government bodies that aim at transforming the lives of public, their
safety, and well-being; various private sectors that assist the government bodies by
funding, constructing, and managing the urban infrastructures; and the stakeholders.
The stakeholders include the citizens themselves and other non-profit organizations.

REFERENCES
1. https​:/​/ hu​​b​.pac​​ktpub​​.com/ ​​how​-a​​i​-is-​​trans​​formi​​ng​-th​​e​-sma​​r t​-ci​​ti​es-​​iot​-t​​utori​​al/ ‘​
Hands-On Artificial Intelligence for IoT’.
2. Husam Rajab, Tibor Cinkelr, “IoT Based Smart Cities”, 2018 International Symposium
on Networks, Computers and Communications (ISNCC), June 2018.
3. Badis Hammi, Rida Khatoun, Sherali Zeadally, Achraf Fayad, Lyes Khoukhi, “Internet
of Things (IoT) Technologies for Smart Cities”, ET Research Journals, the Institution
of Engineering and Technology, 2015.
4. Subba Rao, Sri VidyaGarige, “IOT Based Smart Energy Meter Billing Monitoring and
Controlling the Loads”, International Journal of Innovative Technology and Exploring
Engineering (IJITEE), March 2019.
5. Rustem Dautov, Salvatore Distefano, Rajkumaar Buyya, “Hierarchical Data Fusion for
Smart Healthcare”, Journal of Bigdata, Feb 2019.
6. Sabeen Javaid, Ali Sufian, Saima Pervaiz, Mehak Tanveer, “Smart Traffic Management
System Using Internet of Things”, 2018 20th International Conference on Advanced
Communication Technology (ICACT), Feb 2018.
7. Sjaak Wolfert, Lan Ge, Cor Verdouw, Marc-Jeroen Bogaar, “Big Data in Smart Farming
– A Review”, Agricultural Systems, May 2017.
 15 AI Emerging
Communication
and Computing
N. Girija and T. Bhuvaneswari

CONTENTS
15.1 Introduction................................................................................................. 293
15.2 Industrial Revolution 4.0............................................................................. 294
15.3 Stages of AI................................................................................................. 295
15.4 Classification of AI...................................................................................... 296
15.4.1 Artificial Narrow Intelligence....................................................... 296
15.4.2 Artificial General Intelligence...................................................... 297
15.4.3 Artificial Super Intelligence (ASI)................................................300
15.5 Machine Learning Algorithms....................................................................300
15.6 AI Emerging Communication..................................................................... 301
15.7 Conclusion................................................................................................... 303
Bibliography........................................................................................................... 305

15.1 INTRODUCTION
The current century is recognized as Industrial Revolution 4.0 (I.R. 4.0). I.R. 4.0,
which includes concepts such as Internet of Things (IoT), blockchain technology, and
artificial intelligence (AI), fulfils a substantial role and is recognized as a flourishing
model. This chapter presents the role of AI in emerging communication technology.
This emerging AI technology is affecting not only the communication and comput-
ing environment but also the emerging employment skill requirements in the indus-
try. There is no denying the fact that industry requirements and academic sectors
are interrelated. Therefore, in the coming years, it is inevitable that AI technology-
related programming languages, applications and, projects will be part of the student
curriculum. Deloitte is a UK origin freelancing auditing multinational company. Its
report says that robotics and artificial intelligence influence nanotechnology, drones,
sensor technology, and computer vision and hence are becoming vigorous in the
employment market. That is a reason why BRICS is focusing on vocational educa-
tion training and world skill development in their nations.
This chapter has two parts. The first part provides an introduction, history of
the Industrial Revolution, and the stages of AI. It also provides an overview of how

293
294 Artificial Intelligence (AI)

deep these technologies spread in the mortal world. In the world, every country rec-
ognizes the impact of the industrial revolution which enriching its communication
technologies areas like mobile networks, internet-based social media. AI develop-
ment is categorized into four stages. This chapter also discusses how each stage of
the AI-based algorithm has boosted and targeted the vulnerable brain and its sym-
pathetic expression.
The second part covers topics on AI-based technology diversification. In this part,
the AI technology is classified into artificial narrow intelligence (ANI), artificial
general intelligence (AGI), and artificial super intelligence (ASI) models. Each stage
of AI such as reactive AI, decision-making capability, emotional intelligence and
replication of the human brain, etc., does not allow anyone to guess the saturation
phase of AI. Each AI classification communication task with humane society is also
briefed in this chapter. The significance of machine learning and deep learning con-
cepts and their algorithms in AI are also discussed. The machine learning algorithms
such as the non-maximum suppression algorithm and YOLO have been implemented
in various industry sectors like the health sector and agriculture-based AI-based
applications. AI researchers have also focused on computer vision and hybrid think-
ing in AI communications. These are also discussed in this chapter.
In conclusion, emotional intelligence, ethics, and the law of machine–human
attitudes, and Way of the Future, i.e. Tech Church, are discussed. These combina-
tions based on technology design which will aid the human society reform are also
discussed.

15.2 INDUSTRIAL REVOLUTION 4.0

The 21st century is witnessing a revolution in industry, and this period is hence
referred to as Industrial Revolution 4.0 (I.R. 4.0). In every century, at least one indus-
trial revolution changes the technology. That technology plays a vital role in human
culture and their skill set and economy, resulting in new industry initiation and more
business opportunities worldwide. Since the advent of industrial revolution, steam
power, coal, electricity, and many machinery like spinning industry have proved
helpful to humans as well as in employment generation. Initially dependent on ani-
mals like horses and buffalo for travelling, we have now gradually shifted to train,
airline, and automobile. Each industrial revolution reduces the gap between automa-
tion and humans.
The best and precious contribution of I.R. 2.0 is electrical energy. Telegraph, tele-
phone, and railways are other contributions that have tremendously helped the soci-
ety by enabling a rapid way of communication, fast raw material delivery from and
to ports, and comfortable travelling. It created direct and indirect job opportunities
and undoubtedly contributes to the current industry in society.
I.R. 3.0 is a precious digitization revolution in communication and technology.
Industries like computers and telecommunications played a vital role. I.R. 3.0, espe-
cially network and internet technology, has brought the globe under one distinctive
roof. Communication has become very rapid. Various types of communication like
wireless phones, mobile, email, messengers, social media, and messaging apps help
AI Emerging Communication and Computing 295

to connect people in any distinct region in the world. Nevertheless, I.R. 4.0 is con-
verging into the replication of superhuman. These technologies are moving towards
concepts like smart cities, smart houses, and secured transactions like blockchain
and deep learning. AI technology is omnipresent not only in the modern work-
place but also in a remarkable part of the lifestyle of the community. Just as I.R.
3.0 introduced modern concepts like computer science, information technology, and
telecommunication, I.R. 4.0 seamlessly combines AI, machine learning, and deep
learning in all the fields ranging from medical to automobile.

15.3 STAGES OF AI
In the earlier stages, AI could perform tasks with human proficiency and functional-
ity. However, AI has limited capacity due to a lack of learning skills. Learning skills
develop with experience. For gaining experience, memory-based learning considered
representing an essential aspect. In this stage, AI response is based on the combina-
tion of various inputs groups. This stage is called reactive AI. In 1997, IBM devel-
oped Deep Blue as a reactive AI machine. Deep Blue beat the chess Grandmaster
Garry Kasparov. This incident was an eye-opener for the human community to turn
and explore AI machine proficiency.
In the second stage, AI continues with memory and decision-making. Storing a
massive volume of data needs memory. AI takes a decision based on an archive of
data stored in the memory. For the decision-making feature, the AI needs analyti-
cal skills. In this stage, AI is smarter in recognizing the thousands of images and
scanned objects in image processing, similar to Google Assistants like Chatbot.
Until the third stage, AI was still not considered equivalent to a human as human
beings learn various emotional aspects and can recognize other people’s emotions.
This feature in AI is called emotional intelligence (EI). Understanding the user’s
emotions and respond accordingly is very much needed for virtual assistants. An
example of an EI-based AI application is Siri. Siri is Apple’s voice-controlled per-
sonal assistant for Apple product users. Siri is a combination of AI and Natural
Language Processing (NLP). The three critical tasks of Siri are service tasks, aware-
ness of the situation, and communication interface.
The fourth stage is the replication of the human brain and self-awareness. AI
understands emotional intelligence but also like human primitive emotions. It barely
exhibits emotions like Sophia robot, which was developed by Hong Kong-based
Hanson Robotics research laboratory. Dr. David Hanson described Sophia as capable
of generating expressions of joy, grief, curiosity, confusion, mediation, sorrow, frus-
tration, among other feelings. It means that AI can now concentrate even more on
human muscle expression which is typically connected with the human brain.
The above-mentioned stages clearly state that in the future, most of the human
communications will be with AI-based support instead of actual human beings.
Each stage of AI releases physical stress to human society. The target of AI haunts
the human brain. Human brains include sentiments- and emotions-based compo-
nents. So, human brain characteristics might become more attached and depen-
dent on emotional intelligence technology. As human beings are most emotionally
296 Artificial Intelligence (AI)

FIGURE 15.1  Stages of AI. Source: worldmarket​.com​-o​nly ladder image.

attached to pets, likewise advanced features of Sophia model aims to enhance one
of the family members. This bondage constructs a fashionable relationship between
human–machine or human–robot (Figure 15.1).

15.4 CLASSIFICATION OF AI
AI technology is classified into three categories: Artificial Narrow Intelligence
(ANI), Artificial General Intelligence (AGI), and Artificial Super Intelligence (ASI)
(Figure 15.2).

15.4.1 Artificial Narrow Intelligence

ANI is mostly used for some specific tasks, for example developing a self-driv-
ing car. This type of technology is disciplined to excel in some particular task
only. The ANI programs are designed to get information from specific data set
or predefined data ranges. Based on the data set, it will accomplish the given task
like checking the weather forecast, chess playing like Deep Blue, and writing
reports. ANI does not possess emotional intelligence nor cognizant sentiments like

FIGURE 15.2  Classification of AI.

AI Emerging Communication and Computing 297

humans. Currently, ANI-based technologies are used in Google Assistant, Google

Translate, and Siri. The advantage of ANI is that it improves the mechanical and
tedious human lifestyle to a productive and efficient lifestyle. For instance, a self-
driving car provides extra leisure to a person and relieves from being stuck in traf-
fic. Keeping a personal secretary is impossible for every person, but Siri supports
all secretarial jobs. Siri is a personal digital assistant. It sends a text while you are
in drive, orders pizza, sets your alarm, creates reminders, and plays your favourite
song when asked. Answering and sorting emails are tedious assignments. ANI
technology takes care of forwarding recent emails, checking emails, and reply-
ing to emails. The development of ANI technology has not merely replaced the
low-skilled human job opportunity but also given space to people to develop more
intellectual skills than mere clerical skills.
Lee Se-Dol is one of the leading players of the “GO” Game. Alpha GO program
defeated the player Lee-Se-Dol. The “DeepMind” algorithm using the “Alpha GO”
program was developed by UK-based company DeepMind Technologies which was
a part of Google (up to 2015). ANI at present is also doing the tasks of Inbox clean-
ing, auto sorting for Spam filters based on customer browsing history, adding things
to your shopping cart during online shopping, and recommending videos for view-
ers to watch; these need a huge amount of quality data sets and rule-based machine
learning algorithms.

15.4.2 Artificial General Intelligence

The AGI model is precisely defined as it will act appropriately like human-level AI.
The other terms used to scientifically describe AGI model are “computational intel-
ligence,” “natural intelligence,” “cognitive architecture,” and “biologically inspired
cognitive architecture” (BICA). AGI can accomplish a variety of objectives and
carry out a diversification of tasks in various circumstances and surroundings like
a human being. The person who claims to know diverse fields. For example, Isaac
Newton was an English mathematician, prominent physicist, astronomer, theolo-
gian, and scientist. In the AGI community, the prominent researcher Goertzel (2014)
expressed about the core AGI hypothesis as given below:

Core AGI hypothesis: the creation and study of synthetic intelligence with sufficiently
broad (e.g. human-level) scope and strong generalization capability are at the bottom
qualitatively different from the creation and study of synthetic intelligence with a sig-
nificantly narrower scope and weaker generalization capability.

The upgrade from one specific field to another, i.e. from ANI to AGI, requires gen-
eral intelligence. That is the reason AGI aims to concentrate on machine learning.
The combination of AGI and machine learning can recognize and acquire any type
of intellectual task. This combination is called Strong AI.
AGI models use human cognitive skills to solve efficiently an unfamiliar task.
According to Stephen Hawking, mortal beings are slow biological evolutionary mod-
els. Suppose innovative Strong AI redesigns itself and constructs unlimited cognitive
298 Artificial Intelligence (AI)

skills, it is challenging for humans to ride AI. Currently, DeepMind Technologies

owned by Alphabet Inc. (from 2015) is extensively focusing on AGI models.
The Google AI research team continuously focuses on machine learning for com-
bining AI apps like Google Lookout App, especially for visually impaired people.
Approximately 253 million people are visually impaired across the world. Lookout
App works on portable pixel devices. Wherever the visually challenged people move
around, these apps will accurately identify people, objects, and texts. Google has
powerfully built Pixel Night Sight with ML advancements, which will capture the
photo shot even in the dark.
AGI models focus to faithfully implement learning similar to capable humans. The
human brain naturally contains 90 billion excitatory neurons and synapses. Excitatory
synapses merely connect the neurons in the brain. Hundred trillion synapses are inten-
tionally used to form the neural net in the human brain to accumulate unlimited infor-
mation. AGI models plan to train not only storing and connecting information like the
active brain but also in common sense, intuition, and reasoning skills like the brain.
AGI models are still under empirical research. The AGI researchers are focusing on
this efficient model which is continuously undergoing various levels of cognitive tests
to attain human intelligence. The first ultimate test attends the Loebner Prize compe-
tition. This prestigious annual competition checks the standard of turning the test of
the model. The intelligence level pertains to audio-video conversation and responses.
The next level of the test represents intellectual tasks such as in the home like
­layman the AGI model prepares the coffee. The co-founder of Apple, Steve Wozniak,
believes that the AGI model passes the coffee test. For the coffee test, AGI enters any
kitchen and finds the ingredients like sugar, water, coffee powder, and coffee cup and
perfectly mix all. The third test is the Robot College Student test, which will enrol
like a student and attend course work and move around the campus. The Chinese AI
robot AI-MATHS is introduced for taking college entrance math’s test. This robot
has been developed by Chengdu Zhunxingyunxue technology. This robot has com-
pleted two versions of Math test and Chinese test.
The fourth test is an economically important employment test that will prove the
performance of AGI in the job market. AI researcher Nils J. Nilsson has proposed
that the turning test be replaced by the employment test.
Across the globe, every scientist’s dream is to receive a Nobel Prize. This level
of the test considered the Artificial Scientist Test. AGI model pursues creative sci-
entific research, problem-solving, literature review, and thesis publishing to receive
Nobel Prize. The Australian physicists team experimented with the help of AI and
performed the Bose–Einstein condensate. The team cooled the group of atoms to
absolute zero temperature to a few hundred nanokelvins i.e. a billionth of a second.
The team, which won the Nobel Prize in 2001, firmly believes the general-purpose
research units called “graduate students” in upcoming research will move towards
applying more flexible automation.
The Global Catastrophic Risk Institute published “A Survey of Artificial General
Intelligence Projects for Ethics, Risk, and Policy” in 2017. As per this survey, most
of the AGI R&D projects are in industry or academic institution. Figure 15.3 shows
several AGI projects in various nations.
 AI Emerging Communication and Computing

FIGURE 15.3  Nationwide AGI project. Source: A Survey of Artificial General Intelligence Projects for Ethics, Risk, and Policy, 2017.
299
300 Artificial Intelligence (AI)

15.4.3 Artificial Super Intelligence (ASI)

ASI models surpass humans. The multi-tasking AI is ASI. Artificial intelligence
journey is moving towards the next level: Artificial Super Intelligence. The research-
ers have stated that ASI is an intelligence explosion. ASI aims to overtake humans
with AI-empowered cognizance. The triggering question in the future regarding ASI
would be whether it will replace the human intellectual and whether it might help
sharpen human intelligence. The peak level or hypothetical point of this type of intel-
lectual exposition is called technological singularity. In other words, singularity can
be defined as machine autonomous. Any online-based application is monitored by
humans and also proper control direction is given by humans. Whereas the machine
autonomous develops a machine to make a decision and execute the operation with-
out any human interference. In biological terms, the genetic material is developing
human growth in several stages: like AI growth stages entitled by evolutionary algo-
rithms called genetic algorithms are applied to an artificial neural network.
Turing test represents a magnitude of testing AI. The objective of this test is
to investigate whether the machine (computer) intelligence behaviour is capable of
thinking like a human or not. Turing test was developed by Alan Mathison Turing.
He is the father of theoretical computer science and artificial intelligence, and he
is an English mathematician. To date, no AI model has passed the Turing test. Dr.
Ray Kurzweil, an American inventor who is the Co-founder and Chancellor of
Singularity University and also the director of engineering, Google team, has devel-
oped machine intelligence and Natural Language. He is a part of AI field programs
such as optical character recognition (OCR), text-to-speech synthesis, speech recog-
nition technology, and electronic keyboard instruments. He hopes that in the future
Singularity will pass the Turing test. The Staffordshire University research team has
taken a step ahead to establish a contemporary way of turning tests, i.e. Multimodal
Turning test.
Amazon Alexa remains an intelligent agent. Alexa represents a multiskill to
handle smart home control, financial portfolio information, fitness control like
sleep tracking, weight, and physical exercise, etc. Using Speech Synthesis Markup
Language, the Alexa whispers mode use gentle voice without disturbing anyone
around in a quiet environment. Figure 15.4 shows the subset of artificial intelligence,
machine learning, and deep learning. It also displays how each subset of AI is associ-
ated with each classification model (Table 15.1).

15.5 MACHINE LEARNING ALGORITHMS

The brain of AI is a neural network and the heart of AI is machine learning. Just
as the heart supplies blood to a person to stay alive, the machine learning supplies
algorithm to AI technology to be alive. Machine learning algorithms and subset of
machine learning, i.e. deep learning algorithms, accumulated with AI technology to
advance to the next tremendous elevation.
Predominantly, recommended videos on YouTube and recommended products
on Amazon have been done using supervised learning algorithms techniques. Face
AI Emerging Communication and Computing 301

FIGURE 15.4  Subset of AI.

recognition and speech recognition are done using deep neural network and k-means
techniques (Figure 15.5).
Classification and regression represent the two prominent features of supervised
learning. The below-mentioned algorithms are the traditional algorithm techniques of
machine learning. For Artificial Narrow Intelligence-based AI technology execution,
the machine learning algorithms such as YOLO (You Only You Once) and Sliding
Window are used to supervise learning technique and are also applied in the image
processing. These types of algorithms are used in self-driving cars. Non-maximum
Suppression algorithm is used in the chessboard for object detection. In Figure 15.6,
the popular ML algorithms used in the real-time application are depicted.

15.6 AI EMERGING COMMUNICATION

The key aspect of AI is self-learning. The self-learning potential is based on machine
learning for AI. Self-learning and neural network may aid the AI to meet the utmost
knowledge. At the same time, supreme Knowledge, Emotional Intelligence, and
Singularity are not enough for humans. Effective communication is one of the essen-
tial features of gaining knowledge. Effective communication needs to be typically
categorized into either verbal or non-verbal communication. Non-verbal communi-
cations involve careful observation, monitoring, and searching. AI-based non-verbal
communication technology such as computer vision and hybrid thinking in com-
mon are future AI technology contributions to modern society. How the vulnerable
brain instantly communicates, for example, when a cautioning message is forwarded
while driving, the driver automatically slows down the vehicle using speed breakers
and pedestrians crossing. In the same way, Waymo is a computer vision technology
properly used in a self-driving vehicle. It correctly identifies the pedestrians crossing
beforehand for the safe driving.
Computer vision also plays a role in nursing massive obstetric haemorrhage
(MOH) bleeding during Caesarean cases using Gauss Surgical computer vision
302 Artificial Intelligence (AI)

TABLE 15.1
Summary of Classification of AI-based Application and Test
ANI AGI ASI
Application • Deep Blue • Google Lookout App • Amazon Alexa
• Google Assistant • Pixel Night Sight
• Google Translate
• Siri
• Alpha Go
Test • Loebner Prize • Multi-model
• Coffee Test Turning Test
• Robot College Student Test
• Employee Test
• Artificial Scientist Test

healthcare technology. Aware mostly of the Oil and Gas industries in remote spots,
Osprey’s intelligent visual monitoring is prominently used in the oil and gas industry
to detect leaks, safely monitor, and bolster security.
Worldwide, sustainable agriculture continues to be a crucial industry. Persisting
farms to be more profitable in business needs aerial phenotyping (“AP”) technologies,

FIGURE 15.5  Machine learning algorithm.

AI Emerging Communication and Computing 303

FIGURE 15.6  Real-time application using ML algorithms.

which is used in Slant Range for crop development for successful breeders. The
above-discussed applications realistically are computer vision-based technology.
Undoubtedly, in a short time, computer vision apps will play a vital role in assuring
safety and risk awareness in routine life.
Based on searching continually and gaining information from the browser and
search engine, the machine learning understands the surroundings and also acts
accordingly. For browsing and seeking reliable information, any mechanical or elec-
tronic devices like the mobile are genuinely needed. The following promotion of
AI technology racks the vulnerable brain as a capable device, i.e. hybrid thinking.
Hybrid thinking has been intimately connected with the human brain. The neocor-
tex is a part of the human brain. The neocortex has multilayered prominent parts
involved in high-order brain functions like vision, audition and olfaction, cognition,
and spatial-visual intelligence, especially for imagination. The neocortex is bigger in
humans than in animals.
Nano-robotics adequately represent an imminent technology in the medical
nanotechnology field. Nanorobot or Nanobot is intimately connected to the neocor-
tex part of the human brain. Whenever a Nanobot allied, the ideal person instantly
thinks for specific information without traditionally using any electronic devices.
The Nanobot automatically connects to the cloud directly. This specific type of mod-
ern thinking is properly designated hybrid thinking. This intelligence technology
provides not only detailed information but also a solution. This modern superior
intelligence is addressed by non-biological intelligence. Ray Kurzweil, Director of
Engineering at Google, has authored the book How to Create a Mind. He correctly
stated in this book that for the next 10 years, hybrid thinking looks promising. He
also believes that in the next two decades non-biological intelligence would be more
potent than intelligent humans.

15.7 CONCLUSION
This chapter began with the description of I.R. 4.0 revolution. The key aim of I.R. 4.0
remains intellectually based automation, the implementation of which requires AI
technology. In the earlier stage, computer-based technology was adopted for accumu-
lating information, which was called a database management system for managing
304 Artificial Intelligence (AI)

databases. Besides, a database system was used for storing historical and pertinent
information needed for decision-making, which was called data warehousing and
data mining. Data analytics is most demanding by industry for not only storing and
decision but also for predicting. The demand of human society and industry is alter-
ing every industrial revolution based on technology development. This alteration is
explained in various stages of AI and the classification of AI topics. The machine
learning algorithm represents the core of AI application and AI emerging commu-
nication, and this chapter focuses on AI and ML with regard to various fields like
agriculture, healthcare, and nano-robotics.
To conclude, when human automation is emerging, then ethics and the law of auto-
mation and religious-based moral discipline are also compulsory. A crucial aspect
that enables human society to act is ethics and law. In 2017, Google DeepMind
gathered the group as “DeepMind Ethics & Society” (DMES). The DMES’ group
members are philosophers, economists, risk analysis experts, political advisors, and
AI technology experts. This group of experts designs the ethical procedure to be
followed by the AI. Verity Harding and Sean Legassick are co-leads of Ethics &
Society, who believe that AI technology must take the ethical and social impact
responsibility. DeepMind has identified the key ethical challenges for AI: privacy,
transparency, and fairness to economic impacts, governance, liability, and many
more.
The next aspect is law. Connecting humans and AI involves appropriate legal
regulations. Human faults are disciplined by legal decree; similarly, the Baltic
nations of Estonia has developed a constitutional status called “Kratt Law”, i.e.
algorithmic liability law for AI. Estonia is one country in which e-governance is
done through advanced AI applications. The abbreviated two letters AI instantly
flash at laymen as technology. But AI represents not only the emerging technology,
it is also considered as the next-century religion. The technology patronized in the
more rapid industrial revolutions. Way of the Future (WOTF) is a contemporary
AI religion’s tech church. WOTF religion focuses on the worship of a Godhead
based on artificial intelligence. Anthony Levandowski is a man behind self-driving
car technology in Uber and also part of Google light detecting and ranging (lidar)
engineering team. Levandowski worked as CEO in the start-up company Pronto
AI, which focused on developing the CoPilot project. The CoPilot project goal is
to develop Advanced Driver-Assistance Systems (ADAS) for truck drivers, which
will find lane detection like whether zebra lines exist, two lines whether white or
yellow are parallel, collision warning, and switching lanes for determining length
and width of the lane. This automated system is most useful for the entire globe
to prevent vehicles-caused accidents. He started the WOTF, and he believes super
intelligence is inevitable. Artificial Super Intelligence is much smarter than super-
human. Therefore, WOTF believes, it is clear to integrate human computational
power like a singular model that will be the most supreme on this planet. Anthony
Levandowski has confidence that God is immeasurable or can’t ensure or control,
but tech god (AI god) also has immeasurable smartness even though it is possible to
interact with tech god and recognize it is listening. Religion is a significant regula-
tory for the moral values of the humane society. In the same way, the forthcoming
AI Emerging Communication and Computing 305

era of AI religion has also influenced human society. Just like how god or spontane-
ous creations exist, the AI’s journey will also virtually exist from ANI to AGI to
ASI with more and more progress!

BIBLIOGRAPHY
1. https​:/​/ww​​w​.bgp​​4​.com​​/2019​​/04​/0​​1​/the​​-3​-ty​​pes​-o​​f​-ai-​​narro​​w​-ani​​-gene​​ral​​-a​​gi​-an​​d​-sup​​
er​-as​​i / (15/10/2019)
2. https​:/​/ww​​w​.bri​​tanni​​ca​.co​​m ​/eve​​nt​/ In​​dustr​​ial​​-R​​evolu​​tion
3. https​:/​/ww​​w​.for​​bes​.c​​om ​/si​​tes​/c​​ognit​​ivewo​​rld ​/2​​019​/0​​6​/19/​​7​-typ​​es​-of​​-arti​​ficia ​​l​-int​​​ellig​​
ence/​​#7f0c​​c0cc2​​33e
4. https​:/​/in​​teres​​tinge​​ngine​​ering​​.com/​​t he​-t​​h ree-​​t ypes​​- of​-a ​​r tifi​​cial-​​intel​​l igen​​ce​-​un​​derst​​
andin​​g​-ai
5. https​:/​/ww​​w​.han​​sonro​​botic​​s​.com​​/news​​-meet​​-soph​​ia​-th​​e ​-rob​​ot​-wh​​o ​-lau​​g hs​-s​​m iles​​
-and-​​​frown​​s​-jus​​t​-lik​​e​-us/​
6. https​:/​/ww​​w​.poc​​ket​-l​​int​.c​​om ​/ap​​ps​/ne​​ws​/ap​​ple​/1​​12346​​-what​​-is​-s​​i ri​-a​​pple-​​s​-per​​sonal​​
-voi​c​​e​-ass​​istan​​t​-exp​​laine​d
7. https​:/​/ww​​w​.nuf​​field​​found​​ation​​.org/​​sites​​/defa​​ult​/f​​i les/​​files​​/ Ethi​​cal​-a​​nd​-So​​cieta​​l​-Imp​​
licat​​ions-​​of​-Da​​ta​-an​​d​-AI-​​r​epor​​t​-Nuf​​field​​-Foun​​dat​.p​​df
8. https​:/​/fu​​turis​​m​.com ​​/arti​​ficia​​l​-int​​ellig​​ence-​​offic​​ially​​-gran​​​ted​-r​​eside​​ncy
9. https​:/​/ww​​w​.sea​​rchen​​ginej​​ourna​​l​.com​​/yand​​ex​-ar​​tific​​ial​-i​​ntell​​igenc​​e​-mac​​hine-​​learn​​ing​
-a​​lgo​r i​​thms/​​33294​​5/​#cl​​ose
10. https​:/​/to​​wards​​datas​​cienc​​e​.com​​/ai​-m​​achin​​e​-lea​​r ning​​-deep​​-lear​​ning-​​expla​​ined-​​sim​pl​​y​
-7b5​​53da5​​b960
11. https​:/​/9t​​o5goo​​gle​.c​​om​/20​​19​/03​​/18​/d​​eepmi​​nd​-​ag​​i​-con​​trol/​
12. https​:/​/ww​​w​.sci​​entif​​i came​​r ican​​.com/​​a rtic​​le​/wi​​ll​-ch​​ina​-o​​verta​​ke​-th​​e​-u​-s​​-in​-a​​r tifi​​cial-​​​
intel​​ligen​​ce​-re​​searc​​h /
13. https​:/​/ww​​w​.mck​​i nsey​​.com/​​i ndus​​t ries​​/tech​​nolog​​y​-med​​ia​-an​​d​-tel​​e comm​​u nica​​t ions​​/
our-​​insig​​hts​/a​​n​-exe​​cutiv​​es​-gu​​​ide​-t​​o​-mac​​hine-​​learn​​ing
14. https​:/​/gm​​isumm​​it​.co​​m ​/wp-​​conte​​nt​/up​​loads​​/2018​​/10​/ W​​hat​-i​​s​-Hyb​​​r id​-T​​hinki​​ng​.pd​f
15. Burden, D, Baden, MS, (2019), Virtual Humans: Today and Tomorrow (1st Ed.),
Chapman & Hall/CRC. https://doi​.org​/10​.1201​/9781315151199
16. Chalmers, D, Appleyard, A, (2016), The Singularity: Could Artificial Intelligence
Really Out-Think Us (1st Ed.), Imprint Academic.
17. Goertzel, B, (2014), Artificial General Intelligence: Concept, State of the Art, and
Future Prospects, 5(1), 1–46, https://doi​.org​/10​.2478​/jagi​-2014​- 0001
18. Shabbir, J, Anwer, T, (2015), Artificial Intelligence and Its Role in Near Future, Journal
of Latex Class Files, 14(8), 1–11.
19. Baum, S, (2017), A Survey of Artificial General Intelligence Projects for Ethics, Risk,
and Policy. Global Catastrophic Risk Institute Working Paper 17-1. https://ssrn​.com​/
abstract​=3070741.
20. Adriano, M, Althaus, D, Erhardt, J, Gloor, L, Hutter, A, Metzinger, T, (2015), Artificial
Intelligence. Opportunities and Risks. In: Policy Papers of the Effective, Altruism
Foundation (2), S.1–16. https​:/​/ea​​-foun​​datio​​n​.org​​/file​​s​/ai-​​oppor​​tunit​​ies​-a​​​nd​-ri​​sks​.p​​df
21. Watson, EN, (2019), The Super Moral Singularity—AI As a Fountain of Values,
Big Data and Cognitive Computing, 3(2), 23. https://www​.mdpi​.com ​/2504​-2289​/3​
/2​/23
22. Carlson, KW, (2019), Safe Artificial General Intelligence via Distributed Ledger
Technology, Big Data and Cognitive Computing, 3(3), 40. https://www​.mdpi​.com​/2504​
-2289​/3​/2​/23
306 Artificial Intelligence (AI)

23. Bhasin, H, Bhati, S, et al, (2011), Application of Genetic Algorithms in Machine learn-
ing, International Journal of Computer Science and Information Technologies, 2(5),
2412–2415.
24. Wang, Yu, (2018), Reconfigurable Processor for deep learning in autonomous vehicles,
Special Issue 1- The Impact of Artificial Intelligence on Communication Networks and
Services, ITU Journal, 1(1), 9–21.
Index
A AR, see Augmented reality
Architecture-based variant class, 138–139
AAE, see Adversarial AutoEncoders Architectures variations of, GAN
Accurate level prediction, neural networks auxiliary generative adversarial networks
activation function, 103 (AC-GAN), 219–220
artificial neural networks (ANN), 102–103 conditional generative adversarial
sensor output sampling, 101–102 networks, 217–218
ACGAN, see Auxiliary Classifier GAN deep convolutional generative adversarial
Achieving learning, neural network, 174 networks (DCGAN), 217
ACO algorithm, 246 fully connected generative adversarial
Activation function, 103 networks (FCGAN), 217
Adaptive probabilistic neural networks InfoGAN, 220
(APNN), 152 Laplacian pyramid of adversarial networks
Advanced Driver-Assistance Systems (ADAS), 304 (LAPGAN), 217
Adversarial AutoEncoders (AAE), 139 least-square generative adversarial networks
Adversarial Learning and Inference (ALI), 139 (LSGAN), 218–219
Aerial phenotyping (AP), 302 variational auto encoder (VAE), 214–215
AGI, see Artificial General Intelligence Artificial General Intelligence (AGI), 297–299
AI-based PQD classification methods, 155, 157 Artificial intelligence (AI), 243–244, 269–271;
deep learning structures, 158 see also Individual entries
CNN-based methods, 160–161 agency, 68
deep belief network (DBN)-based applications, 74–75
methods, 160 classification, 296–300
DNN-based methods, 160 description, 225–226
SAE-based methods, 158–159 vs. DL, 77–80
proposed deep learning, 161–165 domain used, 74
WT-based hybrid PQD classification emerging communication, 301–303
method, 161–165 environmental role, 226
AI-infused apps, 67 ethical responsibility, 88–90
Air monitoring environment, 235 ethics reasoning, 84–88
Alexa, 74 implementation steps, 74
Algorithms comparison, vehicles utilized methods, PQD detection, 151–153
terms, 257–260 vs. ML, 77–80
ALI, see Adversarial Learning and Inference overview, 74
“Alpha GO” program, 297 risks of, 235
Amazon Alexa, 300 artificial superintelligence (ASI), 236
Amazon Rekognition, 125 bias, 235
Analysis of variance (ANOVA), 196 liability, 235
total distance travelled, 256 stages, 295–296
vehicles utilized, TDVRPTW, 259 Artificial Narrow Intelligence (ANI), 296–297
Anger, 62 Artificial neural networks (ANNs), 102–103, 132,
ANI, see Artificial Narrow Intelligence 152, 174–178
ANNBP, see Artificial Neural Propagation inverse kinematics solution, 178–179
Network optimal parameters, 181–184
ANNs, see Artificial neural networks robust design of, 179–180
AnoGAN, 145 Artificial Neural Propagation Network
ANOVA, see Analysis of variance (ANNBP), 177
AP, see Aerial phenotyping Artificial neuron model, 174
APNN, see Adaptive probabilistic neural with transfer function, 175
networks Artificial Scientist Test, 298

307
308 Index

Artificial Super Intelligence (ASI), 236, 300 C

Audio generation and translation, 144–145
Augmented reality (AR), 125–126 Calendar queue (CQ), 7–8
Auto encoder network, 214–215 in parallel DES, 9–10
Autoencoders, 133 in sequential DES, 8–9
Automated vehicles (AV), 85 Cantilever beam, 98–99
Automatic target recognition and detection, 127 Carrier-to-noise plus interference ratio (CNIR),
Autonomous and connected electric vehicles, 17, 18
226–227 Carrier-to-noise ratio (CNR), 17, 18
Autonomous driving, 126 balanced and DCA for concurrent channel
Auxiliary Classifier GAN (ACGAN), 138, assignment, 37
219–220 variation, 35–36
AV, see Automated vehicles Central processing unit (CPU), 67
CGAN, see Conditional GAN
Character, 90
B Chengdu Zhunxingyunxue technology, 298
Background and driving forces, 73 CLD, see Colour layout descriptor
Backpropagation (BP), 177 CNIR, see Carrier-to-noise plus interference
algorithm, 160 ratio
Back propagation network (BPN), 104 CNN, see Convolutional neural network
Banach Wasserstein GAN (BWGAN), 141 CNN-based methods, 160–161
Base Station Controller (BSC), 15 CNR, see Carrier-to-noise ratio
Base Stations (BSs), 13–15, 18 Colour, 115
BigGAN, 141 Colour descriptor, 116–120
Bohm, David, 84 Colour histogram, 115–116
Boltzmann machine, 134 Colour layout descriptor (CLD), 116
Bonferroni correction, 65 Colour space models, 115
BP, see Backpropagation Commercial products, 124
BPN, see Back propagation network Computational time, 128
Bragg wavelength change, 97 Computer vision, 112, 113, 301
Broadband source, 100 applications, 125–128
BSC, see Base Station Controller evolution, 113
BSs, see Base Stations hardware, 124
Building smart city revenue, world markets, 123
augmented reality and virtual reality, services, 124–125
281–282 Conditional GAN (CGAN), 138, 217–218
challenges, 283 Condition-based variants, 138
components, 282, 285 Conservation biology, 227–229
agriculture, 289–290 Continuous WT (CWT), 154
drawbacks, 290 Conventional level measurement
energy, 284–286 techniques, 94
governance, 289 Conventional power system, 150
healthcare, 286 Conventional sensors, 94
lighting, 288–289 Convolutional components, 139
parking, 287 Convolutional neural network (CNN), 67,
traffic management, 287 127, 133
waste management, 287–288 CoPilot project, 304
description, 279–280 Co-registration, 125
requirements, 281 Correlation, 119
structure of, 280–281 Correlation analysis, 119
technologies, 283 CPU, see Central processing unit
AI and IoT integration, 283–284 CQ, see Calendar queue
big data, 284 Cross-validation method, 200, 203
blockchain, 284 CUDA, 124
robotics, 284 CWT, see Continuous WT
BWGAN, see Banach Wasserstein GAN Cycle consistency GAN (CycleGAN), 140
Index 309

D Discover cross-domain relationship GAN

(DiscoGAN), 140
Data analytics, 304 Discrete Emotions Questionnaire (DEQ), 60
Data annotation, 113 Discrete event simulation (DES), 3–7
Data augmentation, 141 calendar queue (CQ) in
Database management system, 303 parallel, 9–10
Data mining, 304 sequential, 8–9
Data warehousing, 304 concepts, 4–5
Daubachies4 (db4) mother wavelet, 155 parallel systems, 6–7
DBN, see Deep belief networks simulation time, 5–6
DCA, see Dynamic channel allocation Discrete level, 61
DCGAN, see Deep convolution generative Discrete state emotions (DSEs), 60
adversarial networks Discrete wavelet transform (DWT), 151, 154
Decision-making process, 84 Discriminative vs. generative learning, 212–213
Decision tree (DT), 157 Discriminator model, 136
Deep belief networks (DBN), 133, 160 Divergence-based modification, 140
Deep Blue, 295 Divergence of distribution, 137
Deep convolution generative adversarial DL, see Deep learning
networks (DCGAN), 67, 139, 141, 217 DMES, see DeepMind Ethics & Society
Deep generative models, 133–134, 213–214 DNN, see Deep neural network
Deep learning (DL), 73, 112, 132, 133 DNN-based methods, 160
vs. AI, 77–80 DOF, see Degrees of freedom
algorithms, 152 DRAGAN, 141
applications, 77 DSEs, see Discrete state emotions
domain used, 76 DT, see Decision tree
implementation steps, 76 Duncan’s multiple rage test, 255, 257, 259–260
vs. ML, 77–80 DWT, see Discrete wavelet transform
overview, 76 Dynamic channel allocation (DCA), 11, 12
structures, 158 variation, 34–37
CNN-based methods, 160–161 balanced, 35
deep belief network (DBN)-based balanced and CNR for concurrent channel
methods, 160 assignment, 37
DNN-based methods, 160 CNR, 35–36
SAE-based methods, 158–159 round blocking, 36
“DeepMind” algorithm, 297 unbalanced, 34–35
DeepMind Ethics & Society (DMES), 304
Deep neural network (DNN), 132, 133, 301
Degrees of freedom (DOF), 171 E
Deloitte, 293 EBGAN, see Energy-based GAN
Denavit–Hartenberg (DH) method, 185 ECBC, see Energy Conservation Building Code
Deontology ethics, 87 E-commerce logistics vehicle routing
Depth tracking, 126 AI application, route planning and
DEQ, see Discrete Emotions Questionnaire optimization, 249
DES, see Discrete event simulation algorithms comparison, vehicles utilized
Descriptive ethics, 86 terms, 257–260
Descriptive statistics, 65–66 artificial intelligence (AI), 243–244
Designed level sensor system, signal flow, 101 festive days challenges, 242–243
Design variables, 182–183 genetic algorithms, 244
Desire, 62–63 performance measures, 248
Detection Error Trade-off (DET) curves, 67 time-dependent vehicle routing problem with
Detectron, 120 time windows (TDVRPTW), 248–249
DG systems, 150 transport costs and goods characteristics, 242
DH, see Denavit–Hartenberg method vehicle routing problem (VRP), 245–247
Dilation coefficient, 106 ECT, see Event computational time
Dimensional level, 61 Effective communication, 301
Direct kinematics, 173, 184 EI, see Emotional intelligence
310 Index

Emotion, 61 F
Emotional intelligence (EI), 295
Encoder–decoder-based GAN, 139 Face swapping GAN (FSGAN), 144
Encoder–decoder components, 139 Factorial design of experiments, 179
Energy-based GAN (EBGAN), 139, 140 Fair Event Interleaving (FEI), 53
Energy Conservation Building Code False Positive Per Window (FPPW), 67
(ECBC), 269 Faster region-based convolutional neural
Energy efficiency, buildings network, 122
artificial intelligence (AI), 269–271 Fast Fourier transform (FFT) method, 151
digitalization influence, HVAC systems, 273 Fast region-based convolutional neural
internet of things (IoT), 271–272 network, 122
machine learning (ML), 272–273 FBG, see Fibre Bragg grating sensor
optimization and scheduling, 273 FBG-embedded cantilever beam, 94
predictive maintenance and fault diagnosis, FBG-embedded cantilever sensor, 109
273–275 FCA, see Fixed channel allocation
scenario, 265–269 FCGAN, see Fully connected GAN
sensors, 272 FDNN, see Feed-forward deep neural network
sustainability, 264–265 Feature, 75
Energy scenario structures, buildings Feature extraction, 113
appliances, 269 methods, 114
consumption, 265–267 colour descriptor, 116–120
cooling, 267 colour histogram, 115–116
heating, 268 low-level features, 114–115
household fans, 268 texture estimator, 115
HVAC systems, 266 types, 114
lighting, 269 Feed-forward deep neural network (FDNN), 161
Engine oil SAE 15W-40 over temperature classifier test set, performance report, 165
density variations, 96 model flowchart, 163
dynamic and kinematic viscosity structure, 165
variations, 96 FEI, see Fair Event Interleaving
Enhanced Super Resolution GAN FFT, see Fast Fourier transform method
(ESRGAN), 142 Fibre Bragg grating (FBG) sensor, 94, 97–98
Environmental science Fixed channel allocation (FCA), 11, 12
monitoring environment, 232 Float sensor, 99–100
air, 235 Formulation-based variant class, 139–141
soil, 232–233 Fourier transform (FT) methods, 153
water, 233–234 FPPW, see False Positive Per Window
technological solutions FSGAN, see Face swapping GAN
autonomous and connected electric FT, see Fourier transform methods
vehicles, 226–227 Fully connected GAN (FCGAN), 217
conservation biology, 227–229 FusionGAN, 140, 144
next-generation weather and climate
prediction, 229–230 G
smart earth, 231–232
ESRGAN, see Enhanced Super Resolution GAN GA, see Genetic algorithm
Estonia, 304 GAN, see Generative adversarial networks
Ethical responsibility, 88–90 GauGAN, 143
Ethics and law, 304 Gauss Surgical computer vision healthcare
Ethics reasoning, 84–88 technology, 301–302
components, 87 General AR-based system, 126
Event computational time (ECT), 44 Generative adversarial networks (GAN),
Excitement, 62 132–134, 215–216
Expected Utility Theory, 85 adversarial process, 136
Explicitly defined model, 134 applications of, 141–145, 220
Explicit PDF generative models, 134 anomaly detection, 221
Explicit probabilistic model, 214–215 image generation, 220
Index 311

image translation, 221 Image co-registration, 125

architectures, variations of, 216–220 Image editing, 143
background mathematics, 137–138 Image generation and prediction, 141
description, 212–213 Image inpainting, 142
structure, 135–136 Image processing, 112
techniques, 138–141 Image projection, 126
variations, 138 Image segmentation
Generative stochastic networks (GSN), 134 instance segmentation, 121
Generative stochastic neural network (GSNN), 160 semantic segmentation, 120–121
Genetic algorithm (GA), 37 sliding window demerits, 121
development, 224 Image super-resolution, 142
time-dependent vehicle routing problem with Image translation, 142
time windows (TDVRPTW), 249–251 Inclination influence, 95–96
crossover, 251–253 Increasing components, 139
fitness function evaluation, 253 Industrial Revolution 4.0 (I.R. 4.0), 293–295
mutation, 253–257 InfoGAN, 138, 220
Global features, 116 Instance segmentation, 121
GMV, see Gross Merchandise Value Internet of things (IoT), 271–272
GOF, see Group of frames Inverse kinematics, 173
Google Cloud Vision API, 124 Inverse kinematics solution, ANNs, 178–179
GPUs, see Graphical processing units Inverse of the sampling fraction, 191
“Graduate students,” 298 IoT, see Internet of things
Graphical processing units (GPUs), 133 I.R. 2.0, 294
Gross Merchandise Value (GMV), 249 I.R. 3.0, 294, 295
Group of frames (GOF), 116 I.R. 4.0, see Industrial Revolution 4.0
GSN, see Generative stochastic networks
GSNN, see Generative stochastic neural network J

H Java language, 16
Jensen–Shannon divergence (JSD), 137
Happiness, 63
Healthcare, 125 K
Heating ventilation and air-conditioning (HVAC)
application, see Energy efficiency, Ketzal robot manipulator, 182
buildings kinematics analysis, 184–188
Heuristic approach, 74 analysis and confirmation stage, 196,
High-level features, 114, 118 198–203
HLR, see Home Location Register data set description, 189–191
HOG features, 122 planning and experimentation stage,
Home Location Register (HLR), 14 195–196
Hough transform, 119–120 reduction data filter algorithm, 191–193
Hybrid approach, 88 training and test, data set analysis,
Hybrid thinking, 303 194–195
Hyperbolic sigmoid tangent transfer function, 204 KLD, see Kullback–Leibler divergence
Hypothesis formulation, 61 K-means techniques, 301
anger, 62 Kratt Law, 304
desire, 62–63 Kullback–Leibler divergence (KLD), 137–138
excitement, 62
happiness, 63 L
relax, 63
Hypothesis testing outcomes, 65–66 Label, 75
Laplacian pyramid of adversarial networks
I (LAPGAN), 139, 217
LDA, see Linear discriminant analysis
Image classification, 120 Learning algorithm, 106
and localization, 120 Learning curves, 166
312 Index

Least-square GAN (LSGAN), 140, 218–219 Microsoft Azure Computer Vision API, 125
Lee Se-Dol, 297 Mitchell, Tom, 75
Levandowski, Anthony, 304 ML, see Machine learning
Linear (ramp), 103 MLP, see Multilayer perceptron algorithm
Linear discriminant analysis (LDA), 117–118 Mobile Telephone Switching Office (MTSO), 14
Linear systematic sampling (LSS), 190 Mobile user (MU), 7, 10, 18
Liquid level, 93 MOH, see Massive obstetric haemorrhage
measurements, dynamic environments, 94 Monitoring environment, 232
inclination influence, 95–96 air, 235
sloshes influence, 96–97 soil, 232–233
temperature influence, 94–95 water, 233–234
Local feature detection, 114 Monte Carlo (MC) simulation method, 42
Loebner Prize competition, 298 MRA, see Multi-resolution analysis
Logistics; see also E-commerce logistics vehicle MSE, see Mean square quadratic error
routing MTSO, see Mobile Telephone Switching Office
artificial intelligence (AI), 243–244 mTSP, see Multiple Traveling Salesman Problem
challenges, 242–243 MU, see Mobile user
Long short-term memory (LSTM) network, 127 Multi-agent systems (MAS)
Loss function-based modification, 140 agent and, 12–13
Low-level features, 114–115 in WCS, 13–14
LSGAN, see Least-square GAN Multilayer perceptron (MLP) algorithm, 178
LSS, see Linear systematic sampling Multiple discriminator GAN (MD-GAN), 139
LSTM, see Long short-term memory network Multiple generator GAN (MGAN), 139
Multiple Traveling Salesman Problem (mTSP), 248
M Multi-resolution analysis (MRA), 155
MuseGAN, 144
Machine learning (ML), 73, 272–273 Mutation, genetic algorithm (GA)
vs. AI, 77–80 Duncan’s multiple range tests, 257
algorithms, 300–301 factorial experiment, 256
real-time application, 303 steps of the proposed genetic algorithm
applications, 76 (SNRPGA), 254–255
vs. DL, 77–80 variance analysis, 256
domain used, 75–76 Myopic traditions and values, 89
implementation steps, 75
overview, 75 N
Mann–Whitney U test, 65
Markov’s decision process, 127 Nano-robotics, 303
MAS, see Multi-agent systems National Rural Employment Guarantee Scheme
Massive obstetric haemorrhage (MOH), 301 (NREGS), 264
MATLAB/Simulink model, electrical power Nationwide AGI project, 299
distribution system, 156 Natural Language Processing (NLP), 295
MATLAB software, 16, 104, 124 NC agent (NCA), 22, 25
MC, see Monte Carlo simulation method Neocortex, 303
McCarthy, John, 74 Neural network approach, 94
MD-GAN, see Multiple discriminator GAN Next-generation weather and climate prediction,
Mean square quadratic error (MSE), 182 229–230
Measurement system, 102 NLP, see Natural Language Processing
MedGAN, 145 Noise variables, 182–183
Medical image processing, 145 Non-biological intelligence, 303
Medical imaging, 125 Non-maximum Suppression algorithm, 301
Meta-heuristics, 244 Non-Max suppression, 122
Methodological approach Non-verbal communications, 301
participants, 64 Normalization, 140–141
procedure, 64–65 Normative ethics, 86–87
Mexican hat wavelet function, 103 NREGS, see National Rural Employment
MGAN, see Multiple generator GAN Guarantee Scheme
Index 313

O Predictive technology, 86
Pre-emptive Event Interleaving (PEI) scheme, 45, 53
OA, see Orthogonal array Pre-emptive stationary priority scheduling
Object detection, 120, 126 (PSPS), 34
HOG features, 122 Principal component analysis (PCA), 116–117
image classification, 120 Probability density function (PDF), 133
classification and localization, 120 Progressive GAN (ProgGAN), 139
image segmentation Proposed deep learning, 161–165
instance segmentation, 121 Proposed DWT-based PQD detection method,
semantic segmentation, 120–121 155, 157, 158–161
sliding window demerits, 121 PSPS, see Pre-emptive stationary priority
region-based methods scheduling
faster region-based convolutional neural
network, 122
fast region-based convolutional neural Q
network, 122 Qualitative feedback, 66
region-based convolutional neural
network (R-CNN), 121–122
region proposal, 121 R
you only look once (YOLO), 122 Random forest (RF), 157
demerits, 122–123 Random Sequence Insertion-based Crossover
Objective function, 182 (RSIX), 250–254
Ocean Cleanup Project, 234 Rational decision-maker, 84
On-Road Integrated Optimisation and Navigation RBM, see Restricted Boltzmann machine
(ORION), 243 RC agent (RCA), 22, 23, 25
OpenCV, 124 R-CNN, see Region-based convolutional neural
Operating system (OS), 64 network
Optical spectrum analyser (OSA), 97, 100 RDANN, see Robust Design of Artificial Neural
Optimal ANN parameters Networks
analysis stage, 184 RDF, see Reduction data filter algorithm
confirmation stage, 184 Reactive AI, 295
experimentation stage, 183 Real-time system (RTS) and scheduling theory,
planning stage, 181–183 29–34
ORION, see On-Road Integrated Optimisation algorithms for implementing synchronized
and Navigation processes/events, 29–30
Orthogonal array (OA), 179 concurrent events, 30–31
OS, see Operating system pre-emptive stationary priority scheduling
OSA, see Optical spectrum analyser (PSPS), 34
Osprey’s intelligent visual monitoring, 302 process life span, 30
response time analysis, 31–33
P Recurrent neural networks (RNN), 133
Reduction data filter (RDF) algorithm, 191–192
PatchGAN, 139 Region-based convolutional neural network
PCA, see Principal component analysis (R-CNN), 121–122
PCC, see Point of common coupling Region-based methods
PDF, see Probability density function faster region-based convolutional neural
PEI, see Pre-emptive Event Interleaving scheme network, 122
PELM, see Polyharmonic extreme learning fast region-based convolutional neural
machine algorithm network, 122
Perception, 127 region-based convolutional neural network
Pixel-to-Pixel GAN (Pix2Pix), 139, 142 (R-CNN), 121–122
Point of common coupling (PCC), 150 region proposal, 121
Poisson distribution, 19 Region of interest (RoI), 115
Polyharmonic extreme learning machine (PELM) Region proposal, 121
algorithm, 125 Regularization, 140
Power quality disturbance (PQD), 150 Relax, 63
314 Index

Represented features, 158 AI and IoT integration, 283–284

Research systems, 124 big data, 284
Resistive float-type sensors, 93 blockchain, 284
Restricted Boltzmann machine (RBM), 160 robotics, 284
RF, see Random forest Smart earth, 231–232
RNN, see Recurrent neural networks Smart energy, 284–286
Robot College Student test, 298 Smart governance, 289
Robotic manipulators, 173 Smart healthcare, 286
Ketzal, kinematics analysis, 184–188 Smart lighting system, 288–289
kinematics of, 173–174 Smart parking system, 287–288
Robotic path planning, visual percepts, 127–128 Smart traffic management, 287
Robust design, 178 Smart waste management, 287–288
of ANNs methodology, 180–184 S/N, see Signal-to-noise ratio
Robust Design of Artificial Neural Networks SNGAN, see Spectral normalization GAN
(RDANN), 174 SNRPGA, see Steps of the proposed genetic
methodology, 181, 204 algorithm
Robust parameter design method, 179 The Society of Automotive Engineers, 227
“Rock–paper–scissors,” 60, 64 Softmax activation function, 167
RoI, see Region of interest Software libraries and tools, 124
“Roshambo,” 60 Soil monitoring environment, 232–233
RSIX, see Random Sequence Insertion-based Sophia robot, 295
Crossover Spectral Normalization, 141
RTS, see Real-time system and scheduling theory Spectral normalization GAN (SNGAN), 141
SRGAN, see Super Resolution GAN
S StarGAN, 140
State-of-the-art computer vision algorithms, 127
SAE-based methods, 158–159 Steps of the proposed genetic algorithm
Self-attention GAN (SAGAN), 141 (SNRPGA), 251, 254–255, 259–260
Self-learning, 301 Stereo vision, 127
Semantic segmentation, 120–121 STFT, see Short-time Fourier transform
Sensor architecture, 100 Structure-based variants, 139
Sensor design StyleGAN, 141
cantilever beam, 98–99 Style transfer, 142
fibre Bragg grating sensor, 97–98 Super Resolution GAN (SRGAN), 142
float sensor, 99–100 Supervised learning, 105
system and working principle, 100–101 Support vector machine (SVM), 104, 122
Sensor output sampling, 101–102 algorithm, 275–276
Short-time Fourier transform (STFT), 154 classifier, 157
Sigmoid, 103 Sustainable building, 263–265
Signal processing (SP)-based PQD detection SVM, see Support vector machine
methods, 151 Systematic sampling, 190
Signal-to-noise ratio (S/N), 180, 184
Simulation languages (SLs), 15–16 T
Simultaneous localization and mapping
(SLAM), 126 Taguchi, Genichi, 179
Single-phase information, 155 Taguchi method, 179–180
Singularity, 300 Taguchi philosophy, 178
Siri, 295, 297 TDVRP, see Time-dependent vehicle routing
SLAM, see Simultaneous localization and mapping problem
Slant Range, 303 TDVRPTW, see Time-dependent vehicle routing
Sliding window demerits, 121 problem with time windows
Sloshes influence, 96–97 Technique-based modification, 140
SLs, see Simulation languages Technological singularity, 300
Smart agriculture, 289–290 Technological solutions
Smart city drawbacks, 290 autonomous and connected electric vehicles,
Smart city technologies, 283 226–227
Index 315

conservation biology, 227–229 Variance, 115

next-generation weather and climate Variational AutoEncoder (VAE), 134, 139,
prediction, 229–230 214–215
smart earth, 231–232 Vehicle routing problem (VRP)
Temperature influence, 94–95 ACO algorithm, 246
Template matching, 118–119 solution, 245
TensorFlow, 124 time-dependent vehicle routing problem
Test accuracy (Test Acc), 165, 167 (TDVRP), 246–247
Texture, 115 travelling salesman problem (TSP), 246
Texture estimator, 115 Video generation, 143
3D object generation, 143 Video manipulation, 143–144
3D Scene-GAN, 143 Virtual reality (VR), 126
Threshold, 103 Virtue, 89
Time-dependent vehicle routing problem Virtue ethics, 87
(TDVRP), 246–247 Viscosity, 94
Time-dependent vehicle routing problem with Visionary idealism, 90
time windows (TDVRPTW), 248–249 Vision-based self-driving cars, 126–127
analysis, number of vehicles, 259 Visitor Location Register (VLR), 14
crossover, 251–253 VR, see Virtual reality
factorial experiment results, 256 VRP, see Vehicle routing problem
genetic algorithm (GA), 248–249
development, 249–251 W
utilized, number of vehicles, 258
Time slice width (TSW), 44 Wasserstein GAN (WGAN), 140, 144
Tone, 115 Water monitoring environment, 233–234
Tractable density function, 134 Wavelet, 103
Training accuracy (Train Acc), 165 Wavelet neural network (WNN), 103–104
Training algorithm, WNN, 105 training, 104–107
Translation coefficient, 106 validation, 107
Transportation, 242 Wavelet transform (WT), 151, 154–155
TSW, see Time slice width Waymo, 301
Turing test, 300 Way of the Future (WOTF), 304
Two-level decomposition process, 155 WCS, see Wireless Communication Systems
Typical cantilever beam, 99 WGAN, see Wasserstein GAN
WGAN-GP, 140
U Wireless Communication Systems (WCS)
agent and multi-agent systems, 12–13
UGAN, see Unrolled GAN channel allocation, 18–19
Undecimated WT (UWT), 154 criteria, 18
UNDP, see United Nations Development procedure and calculation, 19
Programme channel assignment, 10–12
U-NET components, 139 concept of cellular network, 14–15
United Nations Development Programme concurrent models, 38–41
(UNDP), 263 basic, 39
United Parcel Service (UPS), 243 with multiple synchronized blocks, 40
Unrolled GAN (UGAN), 140 thread usage, 40–41
Unsupervised learning, 105 conventional model, 38
UPS, see United Parcel Service discrete event simulation (DES), 4–7
UWT, see Undecimated WT concepts, 4–5
UWT-based disturbance detection method, 151 parallel systems, 6–7
simulation time, 5–6
V event scheduling, 7–10
calendar queue (CQ), 7–8
VAE, see Variational AutoEncoder CQ in parallel DES, 9–10
Value function, 136 CQ in sequential DES, 8–9
Vanilla GAN, 135, 138 importance, 7
316 Index

model behaviour based on architectural supported DCA variations, 34–37

variations, 45 balanced, 35
multi-agent/multi-layered model, 20–22 balanced and CNR variation for
multi-agent systems in, 13–14 concurrent channel assignment, 37
network services, 17–18 CNR, 35–36
call reallocation/handoff (RC), 18 round blocking, 36
call termination (FC), 18 unbalanced, 34–35
MU movement (MC), 18 theoretical analysis of network
new call arrival (NC), 17–18 agents, 22–29
network structure, 16–17 architecture, 23–25
operational parameters, 17 definition, 22–23
real-time scheduling theory, 29–34 interface, 25–27
algorithms for implementing maintains state, 27
synchronized processes/events, 29–30 multi-agent encounters, 28–29
concurrent events, 30–31 utility functions, 27–28
pre-emptive stationary priority scheduling traffic generation, 19
(PSPS), 34 WNN, see Wavelet neural network
process life span, 30 World CO2 emissions, 266
response time analysis, 31–33 World’s end energy consumption, 266
response time analysis, 45, 53 WOTF, see Way of the Future
results based on five days of network WT, see Wavelet transform
operation, 45 WT-based hybrid PQD classification method,
scheduling mechanism comparison, 45 161–165
simulation, 3–4 WT-based PQD detection methods, 153–154
simulation languages (SLs), 15–16 proposed DWT-based PQD detection method,
simulation model evaluation, 41–44 155, 157, 158–161
accuracy, 42–43 wavelet transform (WT), 154–155
behaviour, 42
Monte Carlo (MC) simulation method, 42 Y
network behaviour, 41
reference analysis model employing one You only look once (YOLO), 122
cell only, 43–44 demerits, 122–123