Coevolutionary Fuzzy Modeling

THÈSE No 2XYZ (2002)

Présentée à la Section d’informatique

ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE
pour l’obtention du grade de Docteur ès sciences

par

Carlos Andrés Peña-Reyes

ingeniero electrónico,
Universidad Distrital Francisco José de Caldas, Bogotá, Colombie
de nationalité colombienne

presentée au jury:

Prof. M. Sipper directeur de thèse

Prof. W. Gerstner rapporteur
Prof. E. Sanchez rapporteur
Prof. A. Tettamanzi rapporteur
Prof. M. Tomassini rapporteur

Lausanne, EPFL
2002
Abstract

This thesis presents Fuzzy CoCo, a novel approach for system design, conducive to explain-
ing human decisions. Based on fuzzy logic and coevolutionary computation, Fuzzy CoCo is
a methodology for constructing systems able to accurately predict the outcome of a human
decision-making process, while providing an understandable explanation of the underlying
reasoning.
Fuzzy logic provides a formal framework for constructing systems exhibiting both good
numeric performance (precision) and linguistic representation (interpretability). From a nu-
meric point of view, fuzzy systems exhibit nonlinear behavior and can handle imprecise and
incomplete information. Linguistically, they represent knowledge in the form of rules, a natu-
ral way for explaining decision processes.
Fuzzy modeling—meaning the construction of fuzzy systems—is an arduous task, de-
manding the identification of many parameters. This thesis analyses the fuzzy-modeling prob-
lem and different approaches to coping with it, focusing on evolutionary fuzzy modeling—
the design of fuzzy inference systems using evolutionary algorithms—which constitutes the
methodological base of my approach. In order to promote this analysis the parameters of a
fuzzy system are classified into four categories: logic, structural, connective, and operational.
The central contribution of this work is the use of an advanced evolutionary technique—
cooperative coevolution—for dealing with the simultaneous design of connective and opera-
tional parameters. Cooperative coevolutionary fuzzy modeling succeeds in overcoming several
limitations exhibited by other standard evolutionary approaches: stagnation, convergence to
local optima, and computational costliness.
Designing interpretable systems is a prime goal of my approach, which I study thoroughly
herein. Based on a set of semantic and syntactic criteria, regarding the definition of linguistic
concepts and their causal connections, I propose a number of strategies for producing more
interpretable fuzzy systems. These strategies are implemented in Fuzzy CoCo, resulting in a
modeling methodology providing high numeric precision, while incurring as little a loss of
interpretability as possible.
After testing Fuzzy CoCo on a benchmark problem—Fisher’s Iris data—I successfully
apply the algorithm to model the decision processes involved in two breast-cancer diagnostic
problems: the WBCD problem and the Catalonia mammography interpretation problem. For
the WBCD problem, Fuzzy CoCo produces systems both of high performance and high in-
terpretability, comparable (if not better) than the best systems demonstrated to date. For the
Catalonia problem, an evolved high-performance system was embedded within a web-based
tool—called COBRA—for aiding radiologists in mammography interpretation.
Several aspects of Fuzzy CoCo are thoroughly analyzed to provide a deeper understand-
ing of the method. These analyses show the consistency of the results. They also help derive

iii
iv

a stepwise guide to applying Fuzzy CoCo, and a set of qualitative relationships between some
of its parameters that facilitate setting up the algorithm.
Finally, this work proposes and explores preliminarily two extensions to the method:
Island Fuzzy CoCo and Incremental Fuzzy CoCo, which together with the original CoCo con-
stitute a family of coevolutionary fuzzy modeling techniques. The aim of these extensions is
to guide the choice of an adequate number of rules for a given problem. While Island Fuzzy
CoCo performs an extended search over different problem sizes, Incremental Fuzzy CoCo
bases its search power on a mechanism of incremental evolution.
Résumé

Cette thèse présente Fuzzy CoCo, une nouvelle approche pour la conception de systèmes fa-
vorisant l’explication des décisions humaines. Basée sur la logique floue et sur le calcul
coévolutionniste, Fuzzy CoCo est une méthodologie visant à construire des systèmes capa-
bles de prédire le résultat d’un processus décisionnel humain et de fournir une explication
compréhensible du raisonnement sous-jacent.
La logique floue fournit un cadre formel pour construire des systèmes qui offrent à la
fois une bonne performance numérique (précision), et une représentation linguistique (inter-
prétabilité). D’un point de vue numérique, les systèmes flous sont des systèmes non linéaires
capables de traiter une information imprécise et incomplète. Linguistiquement, ils représen-
tent les connaissances sous forme de règles, ce qui est une façon naturelle d’expliquer des
processus décisionnels.
La modélisation floue—c’est à dire, la conception de systèmes flous—est une tâche dif-
ficile, exigeant l’identification de nombreux paramètres. Cette thèse analyse le problème de
modélisation floue ainsi que des différentes approches existant pour le résoudre, se focalisant
sur la modélisation floue évolutionniste—la conception de systèmes flous en utilisant des al-
gorithmes évolutionnistes—qui constitue la base méthodologique de mon approche. Afin de
favoriser cette analyse, les paramètres d’un système flou sont classifiés en quatre catégories:
logiques, structuraux, connectifs, et opérationnels.
La contribution centrale de ce travail est l’utilisation d’une technique évolutionniste
avancée—la coévolution coopérative—pour faire face à la conception simultanée des
paramètres connectifs et opérationnels. La modélisation floue par coévolution coopérative
réussit à surmonter plusieurs limitations montrées par d’autres approches évolutionnistes:
stagnation, convergence aux optimums locaux, et temps élevé de calcul.
Concevoir des systèmes interprétables est un des buts principaux de mon approche, que
j’étudie complètement. Basé sur un ensemble de critères sémantiques et syntaxiques concer-
nant la définition des concepts linguistiques et leurs liens causals, je propose un certain nombre
de stratégies pour produire des systèmes flous plus facilement interprétables. Ces stratégies
sont implantées dans Fuzzy CoCo, ayant pour résultat une méthodologie de modélisation four-
nissant une précision numérique élevée, tout en gardant une interprétabilité aussi élevée que
possible.
Après avoir essayé Fuzzy CoCo sur un problème benchmark—le problème des iris de
Fisher—j’ai appliqué avec succès l’algorithme pour modéliser les processus de décision im-
pliqués dans deux problèmes de diagnostique de cancer du sein: le problème connu comme
WBCD et le problème d’interprétation de mammographies de Catalogne. Pour le problème
WBCD, Fuzzy CoCo produit des systèmes très performants et hautement interprétables, com-
parables (sinon superieures) aux meilleurs systèmes rapportés jusqu’à présent. Pour le prob-

v
vi

lème de Catalogne, un très bon système évolué a été inclus dans un outil en ligne—appelé
COBRA—aidant des radiologistes à l’interprétation de mammographies.
Plusieurs aspects de Fuzzy CoCo sont minutieusement analysés afin de fournir une
compréhension aprofondie de la méthode. Ces analyses montrent l’uniformité des résultats
obtenus. Sur la base de ces analyses, je propose un guide pour appliquer Fuzzy CoCo, ainsi
qu’un ensemble de rapports qualitatifs entre certains de ses paramètres pour faciliter leur
choix lors de l’utilisation de l’algorithme.
Finalement, ce travail propose et explore, de façon préliminaire, deux extensions à la
méthode: Island Fuzzy CoCo et Incremental Fuzzy CoCo. Combinées avec le CoCo origi-
nal, elles constituent une famille des techniques de modélisation floue coévolutionniste. Le
but de ces extensions est de guider le choix du nombre de règles pour un problème donné;
tandis que Island Fuzzy CoCo exécute une recherche étendue sur différentes tailles du prob-
lème, Incremental Fuzzy CoCo base sa puissance de recherche sur un mécanisme d’évolution
incrémentielle.
Table of Contents

Abstract iii

Résumé v

List of Figures xi

List of Tables xiii

1 Introduction 1
1.1 General Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Proposed solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Fuzzy Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Basic notions of fuzzy sets and fuzzy logic . . . . . . . . . . . . . . 5
1.2.2 Conditional fuzzy statements . . . . . . . . . . . . . . . . . . . . . . 8
1.2.3 Fuzzy inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.4 Fuzzy inference systems . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Evolutionary computation . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.1 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.2 Genetic programming . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.3 Evolution strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.3.4 Evolutionary programming . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.5 Classifier systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2 Evolutionary Fuzzy Modeling 27

2.1 Fuzzy Modeling: The Art of Building Fuzzy Systems . . . . . . . . . . . . . 28
2.1.1 The fuzzy modeling problem . . . . . . . . . . . . . . . . . . . . . . 28
2.1.2 Approaches and techniques . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Evolutionary Fuzzy Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.1 Applying evolution to fuzzy modeling . . . . . . . . . . . . . . . . . 32
2.2.2 Three approaches to behavior and structure learning . . . . . . . . . 33
2.3 Interpretability Considerations . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Semantic criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Syntactic criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.3 Strategies to satisfy semantic and syntactic criteria . . . . . . . . . . 36

vii
viii

2.4 Example: Medical Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.4.1 The Wisconsin breast cancer diagnosis (WBCD) problem . . . . . . 39
2.4.2 A genetic-fuzzy approach to the WBCD problem . . . . . . . . . . . 40
2.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.4.4 Diagnostic confidence . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4.5 Further experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3 Coevolutionary Fuzzy Modeling 51

3.1 Coevolutionary Computation . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Cooperative Coevolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.1 Issues addressed by cooperative coevolution . . . . . . . . . . . . . . 53
3.2.2 Ancestral work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.3 A general model for cooperative coevolution . . . . . . . . . . . . . 54
3.3 Fuzzy CoCo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.1 The algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.2 Elitism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.3 Fitness evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.4 Interpretability considerations . . . . . . . . . . . . . . . . . . . . . 58
3.3.5 Other cooperative coevolutionary approaches to fuzzy modeling . . . 59
3.4 Application Example: The Iris Problem . . . . . . . . . . . . . . . . . . . . 60
3.4.1 Fisher’s Iris data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4.2 Application of Fuzzy CoCo to the Iris problem . . . . . . . . . . . . 62
3.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4 Breast Cancer Diagnosis by Fuzzy CoCo 69

4.1 Breast-biopsy Analysis: The WBCD Problem . . . . . . . . . . . . . . . . . 69
4.1.1 The evolutionary setup . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Mammography Interpretation: The COBRA system . . . . . . . . . . . . . . 74
4.2.1 The Catalonia mammography database . . . . . . . . . . . . . . . . 74
4.2.2 Proposed solution: the COBRA system . . . . . . . . . . . . . . . . 76
4.2.3 Fuzzy CoCo setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

5 Analyzing Fuzzy CoCo 85

5.1 A Stepwise Guide to Applying Fuzzy CoCo . . . . . . . . . . . . . . . . . . 85
5.2 Effects of Some Parameters on Performance . . . . . . . . . . . . . . . . . . 87
5.2.1 Number of rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.2 Population size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.3 Number of cooperators . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2.4 Mutation probability . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.5 Elitism rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.6 Qualitative relationships . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3 Consistency and Quality of Results . . . . . . . . . . . . . . . . . . . . . . . 93
 ix

5.4 Generality of the Designed Fuzzy Systems . . . . . . . . . . . . . . . . . . . 94

5.4.1 Generality of fuzzy rules . . . . . . . . . . . . . . . . . . . . . . . . 94
5.4.2 Generality of the systems designed by Fuzzy CoCo . . . . . . . . . . 95
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

6 Extensions of the Methodology 99

6.1 Island Fuzzy CoCo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.1 The proposed algorithm . . . . . . . . . . . . . . . . . . . . . . . . 100
6.1.2 Preliminary tests of Island Fuzzy CoCo . . . . . . . . . . . . . . . . 102
6.2 Incremental Fuzzy CoCo . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2.1 The proposed algorithm . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2.2 Testing Incremental Fuzzy CoCo . . . . . . . . . . . . . . . . . . . . 108

7 Conclusions and Future Work 113

7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

A Evolutionary Computation in Medicine: An Overview 119

A.1 Using Adaptive Methods in Medicine . . . . . . . . . . . . . . . . . . . . . 119
A.2 Applying Evolutionary Computation to Solve Medical Problems . . . . . . . 120
A.2.1 Data mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
A.2.2 Medical imaging and signal processing . . . . . . . . . . . . . . . . 121
A.2.3 Planning and scheduling . . . . . . . . . . . . . . . . . . . . . . . . 122
A.3 Classified Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
A.3.1 According to the medical task . . . . . . . . . . . . . . . . . . . . . 122
A.3.2 According to the evolutionary technique . . . . . . . . . . . . . . . . 124

Bibliography 127

Curriculum Vitæ 143

List of Figures

1.1 Membership function of a crisp set. . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Membership function of a fuzzy set. . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Common membership functions. . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Common t-norm (AND) operators. . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Common t-conorm (OR) operators. . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Example of a linguistic variable. . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Bidimensional membership function. . . . . . . . . . . . . . . . . . . . . . . 9
1.8 Common implication operators. . . . . . . . . . . . . . . . . . . . . . . . . 11
1.9 Aggregation operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.10 Defuzzification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.11 Basic structure of a fuzzy inference system. . . . . . . . . . . . . . . . . . . 13
1.12 Tourist prediction example: database. . . . . . . . . . . . . . . . . . . . . . 14
1.13 Tourist prediction example: implication. . . . . . . . . . . . . . . . . . . . . 15
1.14 Tourist prediction example: aggregation and defuzzification. . . . . . . . . . 15
1.15 Tourist prediction example: input-output block representation. . . . . . . . . 16
1.16 Tourist prediction example: connectionist representation. . . . . . . . . . . . 18
1.17 Pseudo-code of an evolutionary algorithm. . . . . . . . . . . . . . . . . . . . 20
1.18 Genetic programming parse trees. . . . . . . . . . . . . . . . . . . . . . . . 22
1.19 Crossover in genetic programming . . . . . . . . . . . . . . . . . . . . . . . 23
1.20 A finite state machine with states {Z, T, R}. . . . . . . . . . . . . . . . . . . 24

2.1 Semantically correct fuzzy variable. . . . . . . . . . . . . . . . . . . . . . . 35

2.2 Example of a fuzzy rule and its firing range. . . . . . . . . . . . . . . . . . . 36
2.3 Grid partition of the input space. . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Strategies to define a complete rule base. . . . . . . . . . . . . . . . . . . . . 38
2.5 WBCD problem: proposed diagnosis system. . . . . . . . . . . . . . . . . . 40
2.6 Input fuzzy variables for the WBCD problem. . . . . . . . . . . . . . . . . . 41
2.7 Example of a genome for a single-rule system. . . . . . . . . . . . . . . . . . 43
2.8 Summary of results of 120 evolutionary runs. . . . . . . . . . . . . . . . . . 44
2.9 The best evolved, fuzzy diagnostic system with three rules. . . . . . . . . . . 45
2.10 The best evolved, fuzzy diagnostic system with two rules. . . . . . . . . . . . 46
2.11 The best evolved, fuzzy diagnostic system with one rule. . . . . . . . . . . . 46
2.12 The best evolved, fuzzy diagnostic system with active rules encoding malig-
nant cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

xi
xii

3.1 Potter’s cooperative coevolutionary system. . . . . . . . . . . . . . . . . . . 55

3.2 Pseudo-code of Fuzzy CoCo. . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Fitness evaluation in Fuzzy CoCo. . . . . . . . . . . . . . . . . . . . . . . . 58
3.4 Iris data distribution in the P L–P W subspace. . . . . . . . . . . . . . . . . 60
3.5 Fuzzy controller used to solve Fisher’s iris data problem. . . . . . . . . . . . 61
3.6 Fuzzy classifier used to solve Fisher’s iris data problem. . . . . . . . . . . . . 61
3.7 Input fuzzy variables for the iris problem. . . . . . . . . . . . . . . . . . . . 62
3.8 Summary of results of 145 Fuzzy CoCo runs searching for controller-type sys-
tems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.9 The best evolved, controller-type system with three rules. . . . . . . . . . . . 66
3.10 Summary of results of 144 Fuzzy CoCo runs searching for classifier-type sys-
tems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.11 The best evolved, classifier-type system with three rules. . . . . . . . . . . . 68

4.1 Summary of results of 495 evolutionary runs. . . . . . . . . . . . . . . . . . 72

4.2 The best evolved, fuzzy diagnostic system with seven rules. . . . . . . . . . . 73
4.3 The best evolved, fuzzy diagnostic system with two rules. . . . . . . . . . . . 74
4.4 The COBRA system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 User interface: reading form. . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.6 User interface: biopsy recommendation. . . . . . . . . . . . . . . . . . . . . 78
4.7 Input fuzzy variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.8 Summary of results of 65 evolutionary runs. . . . . . . . . . . . . . . . . . . 82

5.1 Evolution of fitness as a function of the maximum number of rules. . . . . . . 88

5.2 Evolution of fitness as a function of the population size. . . . . . . . . . . . . 89
5.3 Evolution of fitness as a function of the number of cooperators. . . . . . . . . 90
5.4 Evolution of fitness as a function of the mutation probability. . . . . . . . . . 91
5.5 Evolution of fitness as a function of the elitism rate. . . . . . . . . . . . . . . 92
5.6 Evolution of the best, the worst, and the average runs of a given Fuzzy CoCo
setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.7 Improving the generality of a fuzzy system. . . . . . . . . . . . . . . . . . . 96

6.1 Island Fuzzy CoCo. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

6.2 Pseudo-code of Island Fuzzy CoCo. . . . . . . . . . . . . . . . . . . . . . . 101
6.3 Results, island by island, for the Iris classifier problem. . . . . . . . . . . . . 104
6.4 Results, island by island, for the WBCD problem. . . . . . . . . . . . . . . . 105
6.5 Effect of migration on evolutionary dynamics. . . . . . . . . . . . . . . . . . 106
6.6 Pseudo-code of Incremental Fuzzy CoCo. . . . . . . . . . . . . . . . . . . . 107
6.7 Summary of results of 32 evolutionary runs. . . . . . . . . . . . . . . . . . . 109
6.8 Evolution of classification performance in Incremental Fuzzy CoCo. . . . . . 109
6.9 Comparison of Incremental Fuzzy CoCo with Fuzzy CoCo. . . . . . . . . . . 110
List of Tables

2.1 Parameter classification of fuzzy inference systems. . . . . . . . . . . . . . . 29

2.2 WBCD problem: parameter encoding of an individual’s genome. . . . . . . . 42
2.3 Summary of results of 120 evolutionary runs. . . . . . . . . . . . . . . . . . 44
2.4 Results of the best systems evolved by the fuzzy-genetic approach. . . . . . . 45
2.5 Results of evolutionary runs in which the variables-per-rule constraint has been
removed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.6 Results of evolutionary runs in which the default diagnosis is benign. . . . . . 49
2.7 Summary of top 45 evolutionary runs. . . . . . . . . . . . . . . . . . . . . . 50

3.1 Genome encoding of parameters for both species. . . . . . . . . . . . . . . . 63

3.2 Fuzzy CoCo set-up. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Comparison of the best controller-type systems. . . . . . . . . . . . . . . . . 66
3.4 Comparison of the best classifier-type systems. . . . . . . . . . . . . . . . . 67

4.1 WBCD problem: genome encoding of parameters for both species. . . . . . . 70

4.2 Fuzzy CoCo set-up for the WBCD problem. . . . . . . . . . . . . . . . . . . 71
4.3 WBCD problem: comparison of the best systems . . . . . . . . . . . . . . . 72
4.4 COBRA system: variables corresponding to a patient’s clinical data. . . . . . 75
4.5 COBRA system: variables corresponding to radiologic features. . . . . . . . 75
4.6 Genome encoding for membership-function species. . . . . . . . . . . . . . 79
4.7 Genome encoding for rules species. . . . . . . . . . . . . . . . . . . . . . . 79
4.8 Fuzzy CoCo set-up for the COBRA system. . . . . . . . . . . . . . . . . . . 80
4.9 Diagnostic performance measures. . . . . . . . . . . . . . . . . . . . . . . . 80
4.10 Results of the best systems evolved. . . . . . . . . . . . . . . . . . . . . . . 82
4.11 Diagnostic performance of two selected evolved systems. . . . . . . . . . . . 83
4.12 Results of ten-fold cross-validation. . . . . . . . . . . . . . . . . . . . . . . 83

5.1 Fuzzy CoCo set-up used to analyze the effects of some parameters. . . . . . . 88
5.2 Qualitative relationships between Fuzzy CoCo parameters. . . . . . . . . . . 92
5.3 Comparison of Fuzzy CoCo with knn. . . . . . . . . . . . . . . . . . . . . . 94
5.4 Generality of the three-rule fuzzy controller. . . . . . . . . . . . . . . . . . . 96
5.5 Generality of the three-rule fuzzy classifier. . . . . . . . . . . . . . . . . . . 97
5.6 Generality of the seven-rule WBCD fuzzy system. . . . . . . . . . . . . . . . 98

6.1 Fuzzy CoCo set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

6.2 Results of Island Fuzzy CoCo for the Iris classifier problem. . . . . . . . . . 103

xiii
xiv

6.3 Results of Island Fuzzy CoCo for the WBCD problem. . . . . . . . . . . . . 104
6.4 Fuzzy CoCo set-up in Incremental Fuzzy CoCo for the WBCD problem. . . . 108
6.5 Comparison of Incremental Fuzzy CoCo with Fuzzy CoCo. . . . . . . . . . . 110

A.1 Classified bibliography: summary. . . . . . . . . . . . . . . . . . . . . . . . 123

Chapter 1

Introduction

Human thinking and in particular our capacity to make decisions has long interested scien-
tists from many disciplines: philosophy, medicine, psychology, mathematics, and engineering,
among others. Philosophers are interested in the motivations and the implications of decisions,
physicians and psychologists in the different mechanisms leading to them, and mathematicians
and engineers in obtaining a model that permits the reproduction of this capacity. The central
aim of my thesis is to propose a methodology to attain this latter goal: modeling and repro-
ducing human decision making processes.

1.1 General Context

1.1.1 Problem description
The general problem that motivates my thesis is the development of an approach to automat-
ically construct systems which predict, as accurately as possible, the outcome of a human
decision-making process while providing an understandable explanation of a possible reason-
ing leading to it. Such systems should, as much as possible, deal with the features of the
information involved in human decisions: imprecise, incomplete, ambiguous, and subjective.
Human decision-making processes involve many factors, from factual information to in-
tuitionistic relations, that are hard to explain, even for the person making the decisions (whom
we call the expert). Acquiring, or transmitting expertise (i.e., the knowledge required to make
decisions in a given domain) involves one or more of the following knowledge-processing
tasks: formalization, ordered presentation, experiment-driven verification and validation, and
continuous actualization based on experience. This is a long process, and a human can take
up to several years to reach a satisfactory level of expertise in critical domains. Nowadays,
with the increasing amount of available information and the development of new technolo-
gies, many domains require the shortening of acquiring, transmitting, and updating knowledge.
Computer-based systems may provide the tools necessary to do this.
Present-day databases contain a huge amount of data concerning human decisions that
should be used to model decision-making processes. These data are, however, just a collection
of recorded facts that do not contain by themselves any information or knowledge useful to
explain or to predict the decisions. There exist many methods that, based on these data, can

1
2 1. Introduction

build systems to predict the outcome of a given decision. Albeit useful and widely used, these
methods and systems lack the explanatory power required to transmit knowledge to humans.
It thus becomes necessary to produce systems that, besides proposing accurate predic-
tions, also provide human-understandable explanations of the decisions made. Note that the
goal herein is not to model the actual human reasoning process, but only to express the knowl-
edge behind a decision in a manner conducive to human understanding. As with many other
human activities, decision-making involves information that is inherently imprecise, often am-
biguous, and sometimes incomplete. The systems developed should be able to deal with such
information, and still provide accurate predictions and understandable explanations.
Many human tasks may benefit from, and sometimes require, decision explanation sys-
tems. Among them one can cite client attention, diagnosis, prognosis, and planning. In client
attention, the person or system in charge usually requires vast knowledge to answer simple
questions. Diagnostic tasks involve the interpretation of many facts to identify the condition
of a system. Prediction implies the analysis of actual conditions to predict the future devel-
opment of a system. Planning tasks involve deciding on actions that, once performed, would
drive a system to a desired condition. There exists a domain that concerns all these tasks and
several others, where decisions must be made and accompanied by explanations: medicine.
Indeed, in medical practice it is customary to perform patient examination, diagnosis of risks
and diseases, prognosis of a disease development, and treatment planning. Medicine is thus a
domain where explanatory systems might bring many advantages.
Building explanatory systems is a difficult task. An extended approach, used to build
so-called expert systems, is based on knowledge gathered directly from experts. One applies a
process, known as knowledge engineering, which starts by collecting raw knowledge from di-
verse experts, continuing by systematically organizing and formalizing this knowledge, finally
producing an explanatory system capable of providing sufficiently accurate predictions. Such
knowledge construction is a lengthy task that involves many people from diverse domains, and
long sessions of knowledge gathering, formalization, and validation with experts. Despite the
fact that it is costly and time-consuming, knowledge engineering is still the best alternative to
designing large and hierarchic explanatory systems (i.e., involving hundreds of input variables,
and tens of chained, hierarchical decisions).
Other approaches for building explanatory systems, in which this thesis is interested, take
advantage of available data to support many design stages, extracting the knowledge embed-
ded in the data and representing such knowledge in a manner accessible to humans. These
approaches render knowledge engineering more automatic, as the role of the expert is reduced
to delimiting the problem and validating the soundness and the coherence of the extracted
knowledge.

1.1.2 Proposed solution

The solution I propose in this thesis, called Fuzzy CoCo, is a novel approach that combines
two methodologies—fuzzy systems and coevolutionary algorithms—so as to automatically
produce accurate and interpretable systems. The approach is based on two elements: (1) a
system model capable of providing both accuracy and human understandability, and (2) an
algorithm that allows to build such a model from available information.
To represent the knowledge, I chose fuzzy systems, which may attain good numeric
1.1 General Context 3

performance and provide a scheme to represent knowledge in a way that resembles human
communication and reasoning. Fuzzy systems exhibit some characteristics that render them
adequate to solving the problem tackled in this thesis:

1. They represent knowledge in the form of rules, a natural way to explain decision pro-
cesses.

2. They express concepts with linguistic labels, close to human representation (e.g., “high
fever” instead of “temperature higher than 39.3 degrees”).

3. Such linguistic representation (i.e., concepts and rules) is accompanied by a precise

numeric equivalent that is adequate for managing information available in a numeric
way.

4. Fuzzy systems are adequate to model nonlinear behaviors, exhibited by almost all natural
processes.

5. Fuzzy systems have proven to be universal approximators, meaning that provided

enough rules and concepts, they can approximate any numeric behavior with the desired
precision level (at the cost of reduced linguistic expressiveness).

The construction of fuzzy models of large and complex systems is a hard task, demand-
ing the identification of many parameters. To better understand this problem—i.e., the fuzzy
modeling problem—I propose a classification of fuzzy parameters into four classes: logic,
structural, connective, and operational. This classification serves as a conceptual framework
to decompose the fuzzy modeling problem, to understand how existing modeling techniques
deal with this problem, and to propose novel techniques to solve it efficiently.
As a general methodology for constructing fuzzy systems I use evolutionary computa-
tion, a set of computational techniques based on the principle of natural selection. Evolu-
tionary algorithms are widely used to search for adequate solutions in complex spaces that
resist analytical solutions. Specifically, I use an advanced evolutionary technique, cooperative
coevolution, which deals particularly well with requirements of complexity and modularity,
while exhibiting reasonable computational cost.
The search for interpretability in evolutionary fuzzy modeling is represented by several
constraints taken into account when designing the evolutionary algorithm. However, there
is no well-established definition for interpretability of fuzzy systems serving to define these
constraints. Based on some works that have attempted to define objective criteria to reinforce
interpretability, I define two groups of criteria—semantic and syntactic—and propose some
strategies to satisfy them.
The resulting approach, Fuzzy CoCo, is a fuzzy modeling technique, based on coopera-
tive coevolution, conceived to provide high numeric precision (accuracy), while incurring as
little a loss of linguistic descriptive power (interpretability) as possible. In Fuzzy CoCo, two
coevolving, cooperative species are defined: rules and membership functions (which define
linguistic concepts). The interpretability of the resulting systems is reinforced by applying the
strategies proposed with this aim. Fuzzy CoCo proves to be very efficient in designing highly
accurate and interpretable systems to solve hard problems, in particular modeling medical
diagnostic decision processes.
4 1. Introduction

1.1.3 Outline of the thesis

The methodology proposed in this thesis is based on two domains: fuzzy logic and evolu-
tionary computation. The rest of this chapter is dedicated to presenting some basic notions
of these two fields, necessary for the subsequent chapters. Section 1.2 presents fuzzy systems
and Section 1.3 introduces evolutionary computation.
The fusion of these two domains gives rise to evolutionary fuzzy modeling, the sub-
ject presented in Chapter 2. This chapter starts (Section 2.1) presenting in detail the fuzzy-
modeling problem, including a novel classification of fuzzy system parameters, that provides
some key ideas important to the conception of Fuzzy CoCo. After presenting the applica-
tion of evolutionary computation to fuzzy modeling in Section 2.2, the chapter continues with
an original discussion on interpretability issues in Section 2.3. The subject is illustrated in
Section 2.4 with an example application to a real-world medical problem.
Chapter 3 presents Fuzzy CoCo, a novel cooperative coevolutionary approach to fuzzy
modeling. After introducing coevolutionary algorithms (Section 3.1) and cooperative coevo-
lution (Section 3.2), Section 3.3 presents in detail the algorithm. Section 3.4 concludes the
chapter by presenting a simple application of Fuzzy CoCo to a flower classification problem.
Chapter 4 presents the application of Fuzzy CoCo to modeling the decision processes
involved in two breast-cancer diagnostic problems. First, the Wisconsin breast cancer diagno-
sis problem (Section 4.1) illustrates the capabilities of the method, comparing it with recent
works. The systems found to solve the second problem, the Catalonia mammography database
(Section 4.2), have led to the design of an operational online breast-cancer risk assessor.
Several aspects of Fuzzy CoCo are thoroughly analyzed in Chapter 5. These aspects are:
application (Section 5.1), effect of parameters on performance (Section 5.2), consistency and
quality (Section 5.3), and local generality (Section 5.4), a non-conventional measure applied
to single fuzzy systems. Generality is complementary and closely related to generalization
tests that are applied to modeling methods and not to the systems it produces.
Chapter 6 sketches two extensions to Fuzzy CoCo: Island Fuzzy CoCo, described in
Section 6.1, and Incremental Fuzzy CoCo, delineated in Section 6.2. The goal of proposing
these two extensions is to simplify the task of finding an adequate size of the rule base, a
critical user-defined parameter in Fuzzy CoCo. Each extension is accompanied by preliminary
tests performed to explore its behavior.
Finally, Chapter 7 presents some conclusions, summarizes the goals achieved in the
present work, and recaps the original contributions made. The chapter ends with several con-
siderations on possible future work derived from this thesis.
The body of the thesis is followed by an appendix that presents an overview of evolution-
ary computation in medicine, written up for a special issue of the journal “Artificial Intelligence
in Medicine” [127].

1.2 Fuzzy Systems

A fuzzy inference system—or simply a fuzzy system—is a rule-based system that uses fuzzy
logic, rather than Boolean logic, to reason about data. Fuzzy logic is a computational paradigm
that provides a mathematical tool for representing and manipulating information in a way that
resembles human communication and reasoning processes [186]. It is based on the assumption
1.2 Fuzzy Systems 5

that, in contrast to Boolean logic, a statement can be partially true (or false), and composed of
imprecise concepts. For example, the expression “I live near Geneva,” where the fuzzy value
“near” applied to the fuzzy variable “distance,” in addition to being imprecise, is subject to
interpretation. The foundations of Fuzzy Logic were established in 1965 by Lotfi Zadeh in
his seminal paper about fuzzy sets [193]. While research and applications of fuzzy sets have
grown to cover a wide spectrum of disciplines [185], the present thesis concentrates on its use
in rule-based systems.
In this section I present some concepts of fuzzy logic and fuzzy systems necessary for my
work. In the next subsection I introduce basic notions of fuzzy sets and fuzzy logic. Then, in
Section 1.2.2 I briefly present concepts related to propositional fuzzy logic: linguistic variables
and conditional fuzzy statements. Next, I describe the steps involved in the fuzzy inference
process in Section 1.2.3 to finally present, in Section 1.2.4, what a fuzzy inference system is
and explain its functioning based on a simple example.
For more detailed introductions to fuzzy logic and fuzzy inference systems, the reader is
referred to [101, 185, 186]

1.2.1 Basic notions of fuzzy sets and fuzzy logic

1.2.1.1 Fuzzy sets
Recall that an ordinary, or crisp, set A in a universe of discourse U can be defined by listing
all its members or by defining conditions to identify the elements x ∈ A (i.e., A = {x | x
meets some condition}). The characteristic function, generally called membership function,
associated with A is a mapping µA : U → {0, 1} such that for any element x of the universe,
µA (x) = 1, if x is a member of A and µA (x) = 0, if x is not a member of A. The example
in Figure 1.1 shows the membership function characterizing the crisp set A = {x | 20 ≤ x ≤
26}).
Membership
A
1

0
20 26

Figure 1.1 Membership function of a crisp set. The crisp set is defined as A = {x | 20 ≤ x ≤ 26}.

Fuzzy sets are a generalization of crisp sets. A fuzzy set F defined on a universe of dis-
course U is characterized by a membership function µF (x) which takes values in the interval
[0, 1] (i.e., µA : U → [0, 1]). Note that the term membership function makes more sense in
the context of fuzzy sets as it stresses the idea that µF (x) denotes the degree to which x is a
member of the set F . The operation that assigns a membership value µ(x) to a given value x
is called fuzzification. The example in Figure 1.2 shows the membership function of the fuzzy
set F = {x | x is between 20 and 26} (i.e., the fuzzy set representing approximately the same
concept than the crisp set of Figure 1.1), together with the fuzzification of an example value.
6 1. Introduction

Membership
F
1
0.67

19
0
17 20 26 29

Figure 1.2 Membership function of a fuzzy set. The fuzzy set is defined as F = {x | x is between 20
and 26}. In the figure, an example value 19 is fuzzified as (i.e. it is assigned the membership value)
µF (19) = 0.67

Membership functions might formally take any arbitrary form as they express only
an element-wise membership condition. However, they usually exhibit smooth, monotonic
shapes. This is due to the fact that membership functions are generally used to represent
linguistic units described in the context of a coherent universe of discourse (i.e., the closer
the elements, the more similar the characteristics they represent, as is the case for variables
with physical meaning). The most commonly used membership functions are triangular,
trapezoidal, and bell-shaped (see Figure 1.3).

(a) (b) (c)

Figure 1.3 Common membership functions. (a) triangular, (b) trapezoidal, and (c) bell-shaped.

1.2.1.2 Fuzzy logic

I mentioned before that, in this thesis, I concentrate on the use of fuzzy rule-based systems.
The rules of such a system are expressed as logical implications (i.e., in the form of if . . . then
statements). We can relate our previous discussion about fuzzy sets with our need for a formal
basis for fuzzy logic taking advantage of the fact that “it is well established that propositional
logic is isomorphic to set theory under the appropriate correspondence between components
of these two mathematical systems. Furthermore, both of these systems are isomorphic to a
Boolean algebra” [101]. Some of the most important equivalences between these isomorphic
domains are:
1.2 Fuzzy Systems 7

Sets Logic Algebra

Membership Truth Value
Member (∈) True (T) 1
Non-member (∈) False (F) 0


Intersection ( ) AND (∨) Product (×)

Union ( ) OR (∧) Sum (+)
Complement (¯) NOT (∼) Complement (
)
A simple way to express this equivalence is to formulate the membership relation x ∈ A
in the form of the proposition “x is A”; e.g., Today ∈ Warm where Warm = {x | x is be-
tween 20 and 26 degrees} is equivalent to the proposition “Today is warm”. The difference
between crisp and fuzzy interpretations of this proposition lies in the definition of the mem-
bership function µWarm . Indeed, referring to the crisp and fuzzy sets shown in Figures 1.1 and
1.2, respectively, and assuming that today is 19 degrees, the truth value of “Today is warm”
would be respectively 0 and 0.67. Note that in the same way as fuzzy membership functions
generalize the concept of crisp membership, fuzzy logic generalizes Boolean logic.

1.2.1.3 Operations
In this subsection I present the extension of the most commonly used crisp operations in the
fuzzy domain. This extension imposes as a prime condition that all those fuzzy operations
which are extensions of crisp concepts must reduce to their usual meaning when the fuzzy sets
reduce themselves to crisp sets (i.e., when they have only 1 and 0 as membership values). In
most cases there exist several possible operators satisfying this condition, some of them having
a sole theoretic interest. For the purpose of this thesis, I present below the fuzzy operators most
commonly used in the frame of fuzzy inference systems. For the following definitions, assume
A and B are two fuzzy subsets of U; x denotes an arbitrary element of U.

Operators for intersection/AND operations (µA∩B (x) = µA (x) ∧ µB (x)). Also known as
t-norm operators, the most common—minimum, product, and bounded product, illustrated in
Figure 1.4—are defined as follows:

minimum : min(µA (x), µB (x))

product : µA (x) · µB (x)
bounded product : max(0, µA (x) + µB (x) − 1)

A B A B A B

(a) (b) (c)

Figure 1.4 Common t-norm (AND) operators. (a) minimum, (b) product, and (c) bounded product.
8 1. Introduction

Operators for union/OR operations (µA∪B (x) = µA (x) ∨ µB (x)). Also known as t-conorm
operators, the most common—maximum, probabilistic sum, and bounded sum, exemplified in
Figure 1.5—are defined as follows:

maximum : max(µA (x), µB (x))

probabilistic sum : µA (x) + µB (x) − µA (x) · µB (x)
bounded sum : min(1, µA (x) + µB (x))

A B A B A B

(a) (b) (c)

Figure 1.5 Common t-conorm (OR) operators. (a) maximum, (b) probabilistic sum, and (c) bounded
sum.

complement/NOT operator (µA (x) = µ∼A (x)). The so-called fuzzy complement operator is
almost universally used in fuzzy inference systems. It is defined as follows:

fuzzy complement : 1 − µA (x)

1.2.2 Conditional fuzzy statements

1.2.2.1 Linguistic variables
In the example presented in the last section, the fuzzy set of Figure 1.2 was used to represent
the concept Warm, referring to the day’s temperature. We can also propose fuzzy sets to
represent Cold and Hot. Temperature is then called a linguistic variable, while Warm, Cold,
and Hot are their possible linguistic values.
The concept of linguistic variable was introduced and formalized by Zadeh [195]. A
linguistic variable, also called fuzzy variable, is characterized by its name tag, a set of linguistic
values (also known as fuzzy values or labels), and the membership functions of these labels.
Figure 1.6 shows the linguistic variable Temperature mentioned above.

1.2.2.2 Fuzzy conditions

As mentioned before, a proposition like “Today is warm” is assigned a truth value according to
the membership function characterizing the fuzzy value Warm. This proposition is called con-
ditional fuzzy statement or fuzzy condition. However, fuzzy inference systems rarely depend
on a single variable to define their state. Instead of that, sentences of the type “Today is warm
and partially sunny” are commonly used. Such a double fuzzy condition receives as truth value
the result of the AND operation of the two single fuzzy conditions “Today is warm” and “To-
day is partially sunny”. This corresponds to the Cartesian product of the fuzzy sets Warm and
1.2 Fuzzy Systems 9

Cold Warm Hot

1
0.67

0.33

19 Temperature
0
17 20 26 29

Figure 1.6 Example of a linguistic variable: Temperature has three possible linguistic values, labeled
Cold, Warm, and Hot, plotted above as degree of membership versus input value. In the figure, the
example input value 19 degrees is assigned the membership values µCold (19) = 0.33, µWarm (19) =
0.67, and µHot (19) = 0.

Partially Sunny, defined for the variables Temperature and Sunshine respectively. Figure 1.7
shows the bidimensional membership function characterizing the fuzzy condition “Today is
warm and partially sunny”. Note that fuzzy conditions can also be constructed using OR and
NOT operators (e.g., “Today is hot or not cloudy”).

1
W PS
0.8

0.6

0.4

0.2

0 100
35
30
25 50
Temperature 20
15 Sunshine
10 0

Figure 1.7 Bidimensional membership function. The fuzzy condition “Today is warm and partially
sunny” is characterized by the AND operation—the minimum operator in this case—applied on the
fuzzy values Warm (W) and Partially Sunny (PS) of variables Temperature and Sunshine respectively.

1.2.3 Fuzzy inference

To infer is defined as “to conclude from facts and reasoning”. This definition presents three
elements of the inference process: facts, the raw material; reasoning, the transformation en-
gine; and conclusions, the final product. However it lacks a fourth element: knowledge, the
fuel feeding the inference engine. In fuzzy inference, these four elements are also present:
10 1. Introduction

fuzzy facts are transformed in fuzzy conclusions by a fuzzy reasoning mechanism; knowledge
is represented by fuzzy rules. This section describes three processes participating in the fuzzy
reasoning mechanism: implication, aggregation, and defuzzification.

1.2.3.1 Implication
In a fuzzy inference system , the fuzzy rules have the form:
if (input fuzzy condition) then (output fuzzy assignment),
where output fuzzy assignment designates a fuzzy value resulting from applying a causal im-
plication operator to the input and output fuzzy sets. The term causal implication, also known
as engineering implication [101], refers to the fact that it preserves cause-effect relationships.
The input condition and the output assignment are called, respectively antecedent and conse-
quent. To understand the concept of causal implication, let us define the rule
if u is A then v is C,
where u ∈ U and v ∈ V (U and V are, respectively, the input and output universes of dis-
course). The implication operation A → C for a given input value x is characterized by the
fuzzy set C
= µA→C (x) defined on the output universe V . The most commonly used im-
plication operators, minimum and product, are defined as follows (Figure 1.8 illustrates these
operators):

minimum C
= min(µA (x), C)

product C
= µA (x) · C

A rule that does not fire, i.e., whose antecedent is absolutely false (µA (x) = 0), produces an
empty implication. Consequently, this rule does not participate in the inference of the whole
system.

1.2.3.2 Aggregation
The rules of a fuzzy inference system are related by an also connective, which implies that
all the rules are evaluated. Due to the overlap between fuzzy values, several rules might fire
simultaneously with different truth levels, each proposing an output fuzzy set. In general,
fuzzy systems are expected to propose a single decision instead of a number of different fuzzy
sets. To solve this, these fuzzy sets are integrated, by means of an aggregation operation, to
obtain a single fuzzy set that describes the output of the fuzzy system. Generally, t-conorms
(i.e., OR operators) are used as aggregation operators. Figure 1.9 illustrates the aggregation
operation and shows its results using maximum and bounded-sum operators.

1.2.3.3 Defuzzification
Although a single output fuzzy set—as that obtained by aggregating rule outputs—contains
useful qualitative information, most of the time the output of a fuzzy system must be a crisp
value. The process that produces a crisp output from a fuzzy set is called defuzzification.
Many defuzzification methods have been proposed in the literature; however, the two most
1.2 Fuzzy Systems 11

A C

µA(x)

x u v
(a) (b)

C C

µA(x) C’ µA(x) C’

v v
(c) (d)

Figure 1.8 Common implication operators. Given the rule “if u is A then v is C”, the figure shows:
(a) the input fuzzy value (antecedent) A with a given input element x, (b) the output fuzzy value (conse-
quent) C, and the resulting implication C
= µA→C (x) applying (c) minimum and (d) product impli-
cation operators.

commonly used methods are the Center of Areas (COA), also called center of gravity or cen-
troid, and the Mean of Maxima (MOM). Following, I define these two methods, whose results
are exemplified in Figure 1.10.
Given an output fuzzy set Y = µY (v) defined in the universe V of the variable v, the
defuzzified output y is given by the expressions:

Center of Areas COA


v · µY (v) dv
yCOA = V

V µY (v) dv

If the output universe V is discrete, the integral expressions are replaced by the corresponding
sums:

v · µY (v)
yCOA = V

V µY (v)

Mean of Maxima MOM

yinf = min(z | µY (z) = max(µY (v)))

ysup = max(z | µY (z) = max(µY (v)))
yinf + ysup
yMOM =
2
12 1. Introduction

L M H
1 1
M2
L1
M3
H4
0 0
v v
(a) (b)

Y’
1 1
Y

0 0
v v
(c) (d)

Figure 1.9 Aggregation operation. The output variable v of a system has three fuzzy values, L, M, and
H, presented in (a). In a given instant, four fuzzy rules are fired with activation levels of 0.4, 0.6, 0.2,
and 0.1, producing the fuzzy sets L1 , M2 , M3 , and H4 , respectively, as shown in (b). The aggregation
operation integrates these fuzzy sets applying a t-conorm operator at each point of the output universe.
The figure shows the resulting fuzzy sets when using maximum (c) and bounded-sum (d) operators.
Note that the latter operator ascribes greater importance to the number of active rules with identical
(or similar) consequent than the former does.

1.2.4 Fuzzy inference systems

Generally speaking, the term fuzzy system applies to any system whose operation is mainly
based on fuzzy theory concepts such as reasoning, arithmetic, algebra, topology, or program-
ming, among others. However, in the frame of this thesis I use the terms fuzzy inference
system and fuzzy system to design a rule-based system that uses: (1) linguistic variables, (2)
fuzzy sets to represent the linguistic values of such variables, (3) fuzzy rules to describe causal
relationships between them, and (4) fuzzy inference to compute the response of the system to
a given input.

1
Y

0 yCOA yMOM v

Figure 1.10 Defuzzification. This process produces a crisp value y from a given fuzzy set Y . The
figure shows the resulting values obtained using two common methods: yCOA for the Center of Areas
and yMOM for the Mean of Maxima.
1.2 Fuzzy Systems 13

1.2.4.1 General structure

The basic structure of a fuzzy system consists of four main components, as depicted in Fig-
ure 1.11: (1) a knowledge base, which contains both an ensemble of fuzzy rules, known as the
rule base, and an ensemble of membership functions known as the database; (2) a fuzzifier,
which translates crisp inputs into fuzzy values; (3) an inference engine that applies a fuzzy rea-
soning mechanism to obtain a fuzzy output; and (4) a defuzzifier, which translates this latter
output into a crisp value. These components perform the aforementioned processes necessary
to fuzzy inference.

Knowledge base

Database

Rule base

Crisp Inference Crisp

Fuzzifier Defuzzifier
Input Fuzzy Engine Fuzzy Output
Input Output

Figure 1.11 Basic structure of a fuzzy inference system.

Below, I illustrate the functioning of a fuzzy inference system by means of a simple

example.

The tourist prediction system. A simple fuzzy system is proposed to predict the number
of tourists visiting a resort (for more details on building fuzzy systems, see Section 2.1). The
prediction is based on two weather-related input variables: (1) the temperature, measured in
degrees and (2) the sunshine, expressed as a percentage of the maximum expected sunshine.
The system outputs an estimated amount of tourists expressed in terms of percentage of the
resort’s capacity. The example presented here is developed for a slightly cold, partially sunny
day: the observed temperature and sunshine are 19 degrees and 60%, respectively.

Knowledge base. The knowledge describing the system’s behavior is represented by the
membership functions defining the linguistic variables (the “database” unit in Figure 1.11)
and by a set of rules (the “rule-base” unit in figure 1.11).

Database. Three linguistic variables are defined. Two inputs: Temperature, and
Sunshine, and one output: Tourists. Each variable has three membership functions:
Cold, Warm, and Hot for the first input, Cloudy, Partially Sunny, and Sunny for the sec-
ond input, and Low, Medium, and High for the output. Figure 1.12 shows these variables.
14 1. Introduction

Cold Warm Hot Cloudy PartSunny Sunny Low Medium High

0.67 0.8
0.33
0.2

17 20 26 29 30 50 100 30 50 100
Temperature Sunshine Tourists

Figure 1.12 Tourist prediction example: database. The three variables Temperature, Sunshine, and
Tourists have three membership functions each one. The figure shows the fuzzification of two given
input values: 19 degrees for Temperature and 60% for Sunshine.

Rule base. The following three rules describe the behavior of the prediction system:
Rule 1: if (Temperature is Hot) or (Sunshine is Sunny) then (Tourists is High)
Rule 2: if (Temperature is Warm) and (Sunshine is Partially Sunny) then (Tourists
is Medium)
Rule 3: if (Temperature is Cold) or (Sunshine is Cloudy) then (Tourists is Low)

Fuzzifier. This unit computes the membership values of each input variable in accordance
with the fuzzy values defined in the database (see Figure 1.12).

Temperature Sunshine
µCold (19) = 0.33, µCloudy (60) = 0,
µWarm (19) = 0.67, µPartSunny (60) = 0.8,
µHot (19) = 0. µSunny (60) = 0.2.

Inference engine. This unit interprets the rules contained in the rule base. The inference is
performed in three steps: (1) antecedent activation, (2) implication, and (3) aggregation.

Antecedent activation. The inference engine takes the membership values of the input
fuzzy conditions of each rule—computed previously by the fuzzifier—and then applies
the fuzzy operators indicated to obtain the rule’s truth level.
Rule 1: if (Temperature is Hot) or (Sunshine is Sunny)

µRule 1 = µHot (19) ∨ µSunny(60)

= max(0, 0.2)
= 0.2

Rule 2: if (Temperature is Warm) and (Sunshine is Partially Sunny)

µRule 2 = µWarm (19) ∧ µPartSunny(60)

= min(0.33, 0.8)
= 0.33

Rule 3: if (Temperature is Cold) or (Sunshine is Cloudy)

µRule 3 = µCold (19) ∨ µCloudy(60)

1.2 Fuzzy Systems 15

= max(0.67, 0)
= 0.67

Implication. The inference engine then applies the implication operator to each rule in
order to obtain the fuzzy output values. In this example, I apply the minimum implication
operator. Figure 1.13 shows the fuzzy sets resulting from this operation.

Low Medium High Low Medium High Low Medium High

0 50 100 0 50 100 0 50 100

Rule 1 Rule 2 Rule 3

Figure 1.13 Tourist prediction example: implication. The inference engine applies the minimum
implication operator to each rule in order to obtain the fuzzy output values. Rule truth values are
respectively µRule 1 = 0.2, µRule 2 = 0.33, and µRule 3 = 0.67.

Aggregation. Finally, the inference engine aggregates the three fuzzy sets obtained in
the implication stage in a single fuzzy output. In this example, I apply maximum as
aggregation operator. Figure 1.14 shows the fuzzy set resulting of this operation.

Defuzzifier. As mentioned before, an output fuzzy set as that shown in Figure 1.14 carries
important qualitative information. However, the tourist prediction problem requires a crisp
value quantifying the expected percentage of tourists. The defuzzifier of this example com-
putes this value applying the COA defuzzification method (Section 1.2.3). The output of the
tourist prediction system for the given input conditions—i.e. a temperature of 19 degrees and
a sunshine of 60%—is thus approximately 32%.

µ Tourists(19,60)

0 50 100
tourists(19,60) = 32

Figure 1.14 Tourist prediction example: aggregation and defuzzification. Aggregation integrates the
fuzzy sets of Figure 1.13 into a single output fuzzy set. This fuzzy set, called here µTourists (19, 60), rep-
resents the fuzzy (qualitative) prediction of the expected amount of tourists. The defuzzification process
translates the output fuzzy value, µTourists (19, 60), into a crisp value, Tourists(19, 60), quantifying this
expected amount. In this example, the final output value is approximately 32%.

The fuzzy system presented above, can be seen as a 2-input, 1-output block as that shown
in Figure 1.15
16 1. Introduction

temperature (T)
tourists(T,S)
sunshine (S)

Figure 1.15 Tourist prediction example: input-output block representation. The tourist-prediction
block has two inputs—temperature (T) and sunshine (S)—and one output—tourists(T, S).

1.2.4.2 Types of fuzzy systems

There exists three main types of fuzzy systems that differ in the way they define the conse-
quents of their rules: Mamdani, TSK, and Singleton fuzzy systems. I sketch below their main
characteristics:

Mamdani fuzzy systems. These fuzzy systems have fuzzy sets as rule consequents, as those
presented in this section. They are called Mamdani fuzzy systems as they were first proposed
by Mamdani [94, 95]. Their main advantage is high interpretability, due to the fact that output
variables are defined linguistically and each rule consequent takes on a value from a set of
labels with associated meaning. However, they also have some drawbacks, among which the
most important are their lack of accuracy and their high computational cost. The accuracy of
the system is usually limited by the rigidity of linguistic values. The high computational cost
is related to the operations required to compute the center of areas of the fuzzy sets, i.e., two
integrals or sums and one division for each inference.

TSK fuzzy systems. These systems derive their name from Takagi, Sugeno, and Kang who
proposed an alternative type of fuzzy systems [168, 170], in which the consequent of each
fuzzy rule is represented by a function—usually linear—of the input variables. Thanks to
their functional description, TSK-type systems usually exhibit greater accuracy than Mamdani
systems. However, the interpretability of their rules is significantly reduced as they no longer
represent linguistic concepts.
Coming back to the tourist-prediction example, we can convert the system into a TSK sys-
tem by replacing the three linguistic consequents—Low, Medium, and High—with carefully
calculated, corresponding linear functions of the input variables. The three rule consequents
thus become:

Rule 1: if . . . then Tourists = f1 (T, S) = 2T + 0.8S − 40

Rule 2: if . . . then Tourists = f2 (T, S) = T + S − 23
Rule 3: if . . . then Tourists = f3 (T, S) = 0.5T + 0.3S

With this consequents, the implication step propose the following predictions:

Rule 1 Rule 2 Rule 3

Truth level (µRule ) 0.2 0.33 0.67
Estimated tourists f1 (19, 60) = 38% f2 (19, 60) = 52% f3 (19, 60) = 26%
1.2 Fuzzy Systems 17

The aggregation step does not modify these values. Then, the defuzzifier receives this discrete
output set to produce the following output:
3
fi (19, 60) · µRule i
tourists(19, 60) = i=1
3
i=1 µRule i
0.2 × 38% + 0.33 × 52% + 0.67 × 26%
=
0.2 + 0.33 + 0.67
= 35%

Singleton fuzzy systems. The rule consequents of this type of systems are constant values.
Singleton fuzzy systems can be considered as a particular case of either Mamdani or TSK fuzzy
systems. In fact, a constant value is equivalent to both a singleton fuzzy set—i.e., a fuzzy set
that concentrates its membership value in a single point of the universe—and a linear function
in which the coefficients of the input variables value 0. The singleton representation constitutes
a trade-off between the interpretability offered by Mamdani systems with their meaningful
labels and the accuracy provided by TSK systems with their linear functions. Moreover, thanks
to the discrete representation of the output variable, the defuzzification process demands less
computation than in the other two cases.
In the tourist-prediction example, the Mamdani-type system is easily converted into a
singleton-type system replacing the fuzzy values Low, Medium, and High for their correspond-
ing center of areas: 0%, 50%, and 100%. With these new consequents, the implication step
propose thus:

Rule 1 Rule 2 Rule 3

Truth level (µRule ) 0.2 0.33 0.67
Estimated tourists High = 100% Medium = 52% Low = 26%

that is defuzzified as follows:

3
fi (19, 60) · µRule i
tourists(19, 60) = i=1
3
i=1 µRule i
0.2 × 100% + 0.33 × 50% + 0.67 × 0%
=
0.2 + 0.33 + 0.67
= 31%

1.2.4.3 Connectionist representation

Several authors [71, 89] have used an alternative representation of fuzzy systems, similar to
connectionist models, in which the elements of the fuzzy system are represented by functional
blocks connected according to the knowledge contained in the system and usually organized in
five layers representing different steps of the reasoning process. The five steps represented by
the layers are: fuzzification (membership functions), antecedent activation (logical operators),
implication, aggregation, and defuzzification.
To illustrate this representation, Figure 1.16 shows the connectionist model of the tourist-
prediction system.
18 1. Introduction

µHot
Rule 1
Temp. µWarm
OR Minimum

µCold Rule 2
Tourists
AND Minimum Maximum COA

µSunny Rule 3

Sun µPartSun OR Minimum

µCloudy

(1) (2) (3) (4) (5)

Figure 1.16 Tourist prediction example: connectionist representation. Five layers of connected func-
tional blocks represent the steps and the elements of the fuzzy inference process: (1) fuzzification (mem-
bership functions), (2) antecedent activation (logical operators), (3) implication, (4) aggregation, and
(5) defuzzification.

1.3 Evolutionary computation

The domain of evolutionary computation involves the study of the foundations and the appli-
cations of computational techniques based on the principles of natural evolution. Evolution
in nature is responsible for the “design” of all living beings on earth, and for the strategies
they use to interact with each other. Evolutionary algorithms employ this powerful design
philosophy to find solutions to hard problems.
Generally speaking, evolutionary techniques can be viewed either as search methods,
or as optimization techniques. As written by Michalewicz [104]: “Any abstract task to be
accomplished can be thought of as solving a problem, which, in turn, can be perceived as a
search through a space of potential solutions. Since usually we are after ‘the best’ solution, we
can view this task as an optimization process.”
The first works on the use of evolution-inspired approaches to problem solving date back
to the late 1950s [16, 17, 44, 49, 50]. Independent and almost simultaneous research conducted
by Rechenberg and Schwefel on evolution strategies [147,148,157,158], by Holland on genetic
algorithms [62, 64], and by Fogel on evolutionary programming [47, 48] triggered the study
and the application of evolutionary techniques.
Three basic mechanisms drive natural evolution: reproduction, mutation, and selection.
These mechanisms act on the chromosomes containing the genetic information of the individ-
ual (the genotype), rather than on the individual itself (the phenotype). Reproduction is the
process whereby new individuals are introduced into a population. During sexual reproduc-
tion, recombination (or crossover) occurs, transmitting to the offspring chromosomes that are
1.3 Evolutionary computation 19

a melange of both parents’ genetic information. Mutation introduces small changes into the
inherited chromosomes; it often results from copying errors during reproduction. Selection is
a process guided by the Darwinian principle of survival of the fittest. The fittest individuals
are those best adapted to their environment, which thus survive and reproduce.
Evolutionary computation makes use of a metaphor of natural evolution. According to
this metaphor, a problem plays the role of an environment wherein lives a population of indi-
viduals, each representing a possible solution to the problem. The degree of adaptation of each
individual (i.e., candidate solution) to its environment is expressed by an adequacy measure
known as the fitness function. The phenotype of each individual, i.e., the candidate solution
itself, is generally encoded in some manner into its genome (genotype). Like evolution in
nature, evolutionary algorithms potentially produce progressively better solutions to the prob-
lem. This is possible thanks to the constant introduction of new “genetic” material into the
population, by applying so-called genetic operators that are the computational equivalents of
natural evolutionary mechanisms.
There are several types of evolutionary algorithms, among which the best known are ge-
netic algorithms, genetic programming, evolution strategies, and evolutionary programming;
though different in the specifics they are all based on the same general principles. The archety-
pal evolutionary algorithm proceeds as follows: An initial population of individuals, P (0), is
generated at random or heuristically. Every evolutionary step t, known as a generation, the
individuals in the current population, P (t), are decoded and evaluated according to some
predefined quality criterion, referred to as the fitness, or fitness function. Then, a subset of
individuals, P
(t)—known as the mating pool—is selected to reproduce, with selection of in-
dividuals done according to their fitness. Thus, high-fitness (“good”) individuals stand a better
chance of “reproducing,” while low-fitness ones are more likely to disappear.
Selection alone cannot introduce any new individuals into the population, i.e., it cannot
find new points in the search space. These points are generated by altering the selected popu-
lation P
(t) via the application of crossover and mutation, so as to produce a new population,
P

(t). Crossover tends to enable the evolutionary process to move toward “promising” regions
of the search space. Mutation is introduced to prevent premature convergence to local optima,
by randomly sampling new points in the search space. Finally, the new individuals P

(t) are
introduced into the next-generation population, P (t + 1); although a part of P (t) can be pre-
served, usually P

(t) simply becomes P (t + 1). The termination condition may be specified
as some fixed, maximal number of generations or as the attainment of an acceptable fitness
level. Figure 1.17 presents the structure of a generic evolutionary algorithm in pseudo-code
format.
Evolutionary techniques exhibit a number of advantages over other search methods as
they combine elements of directed and stochastic search. First, they usually need a smaller
amount of knowledge and fewer assumptions about the characteristics of the search space.
Second, they can more easily avoid getting stuck in local optima. Finally, they strike a good
balance between exploitation of the best solutions, and exploration of the search space. The
strength of evolutionary algorithms relies on their population-based search, and on the use of
the genetic mechanisms described above. The existence of a population of candidate solutions
entails a parallel search, with the selection mechanism directing the search to the most promis-
ing regions. The crossover operator encourages the exchange of information between these
search-space regions, while the mutation operator enables the exploration of new directions.
20 1. Introduction

begin EA
t:=0
Initialize population P (t)
while not done do
Evaluate P (t)
P
(t) = Select[P (t)]
P

(t) = ApplyGeneticOperators[P
(t)]
P (t + 1) = Merge[P

(t),P (t)]
t:=t+1
end while
end EA

Figure 1.17 Pseudo-code of an evolutionary algorithm.

The application of an evolutionary algorithm involves a number of important considera-

tions. The first decision to take when applying such an algorithm is how to encode candidate
solutions within the genome. The representation must allow for the encoding of all possible
solutions while being sufficiently simple to be searched in a reasonable amount of time. Next,
an appropriate fitness function must be defined for evaluating the individuals. The (usually
scalar) fitness must reflect the criteria to be optimized and their relative importance. Represen-
tation and fitness are thus clearly problem-dependent, in contrast to selection, crossover, and
mutation, which seem prima facie more problem-independent. Practice has shown, however,
that while standard genetic operators can be used, one often needs to tailor these to the problem
as well.
We noted above that there are several types of evolutionary algorithms. This classifica-
tion is due mainly to historical reasons and the different types of evolutionary algorithms are
in fact quite similar. One could argue that there is but a single general evolutionary algorithm,
or just the opposite–that “there are as many evolutionary algorithms as the researchers work-
ing in evolutionary computation” [137]. The frontiers among the widely accepted classes of
evolutionary algorithms have become fuzzy over the years as each technique has attempted
to overcome its limitations, by imbibing characteristics of the other techniques. To design an
evolutionary algorithm one must define a number of important parameters, which are precisely
those that demarcate the different evolutionary-computation classes. Some important param-
eters are: representation (genome), selection mechanism, crossover, mutation, size of popu-
lations P
and P

, variability or fixity of population size, and variability or fixity of genome
length.
The rest of this chapter presents the major evolutionary methods, emphasizing the specific
properties of each one along with the most typical choices of parameters. These methods are:
Genetic algorithms (Section 1.3.1), genetic programming (Section 1.3.2), evolution strategies
(Section 1.3.3), and evolutionary programming (Section 1.3.4). Finally, Section 1.3.5 intro-
duces a somewhat different evolutionary technique, known as classifier systems, which offers
on-line learning capabilities. A detailed discussion on theory and on advanced topics of evo-
lutionary computation, which is available in many books [4, 104, 106, 171, 177], goes beyond
1.3 Evolutionary computation 21

the reach of this chapter.

1.3.1 Genetic algorithms

Proposed by John Holland in the 1960s [62, 64], genetic algorithms are the best known class
of evolutionary algorithms. They are used so extensively that often the terms genetic algo-
rithms and evolutionary computation are used interchangeably (though, as noted, they should
be considered distinct).
There is a clear distinction between the solution being tested, the “adult individual” or
phenotype, and its representation—the genome or genotype. Traditionally, the genome is a
fixed-length binary string. With such a data structure it is possible to represent solutions to
virtually any problem. However, so that the genetic algorithm may converge to good solutions,
the representation must be carefully designed to minimize redundancy (i.e., several genotypes
encoding the same phenotype) and to avoid invalid representations (i.e., a genotype encoding
a phenotype which is not a possible solution to the problem at hand).
With genetic algorithms, the population size is constant and individuals are decoded and
evaluated at each generation. Individuals are then selected according to their fitness. Many
selection procedures are currently in use, one of the simplest being fitness-proportionate se-
lection, where individuals are selected with a probability proportional to their fitness. This
ensures that the expected number of times an individual is chosen is approximately propor-
tional to its relative performance in the population. Selection produces a mating pool of the
same size as the original population (P
 = P ).
Crossover is performed with probability pc (the “crossover probability” or “crossover
rate”) between two selected individuals, called parents, by exchanging parts of their genomes
to form two new individuals, called offspring. In its simplest form, known as one-point
crossover, substrings are exchanged after a randomly selected crossover point. The mutation
operator is carried out by flipping bits at random, with some (usually small) probability pm .
The crossover and mutation operators preserve the size of the population, i.e., P

 = P
.
Genetic algorithms are by far the most popular evolutionary technique (though genetic
programming is rapidly “gaining” on them). This is due in part to their conceptual simplicity,
the ease of programming entailed, and the small number of parameters to be defined (apart
from the genomic representation and the fitness function, parameters include mainly popula-
tion size, crossover and mutation probabilities, and termination condition).
There are several variations of the simple genetic algorithm [177], with different selection
mechanisms (e.g., ranking, tournament, and elitism), crossover operators (e.g., multi-point
crossover), and mutation operators (e.g., adaptive mutation). There exist a profusion of books
concerning these and other advanced topics related with genetic algorithms [104, 106, 177].

1.3.2 Genetic programming

John Koza developed a variant of genetic algorithms called genetic programming [81, 82]. In
this approach, instead of encoding possible solutions to a problem as a fixed-length character
string, they are encoded as computer programs. To wit, the individuals in the population are
programs that— when executed—are the candidate solutions (phenotypes) to the problem.
22 1. Introduction

Programs in genetic programming may be expressed in any language in principle. How-

ever, to guarantee that evolution be able to generate valid, executable programs, it is necessary
to restrict the choice of language. Thus, programs are expressed as parse trees, rather than as
lines of code, i.e., using a functional language rather than a procedural one. The set of possible
internal (non-terminal) nodes of these parse trees, called the function set (F ), is composed of
user-defined functions. The terminal nodes, which form the terminal set (T ), are usually either
variables or constants. The syntactic closure property requires that each function in F be able
to accept as arguments any other returned function value and any value and data type in the
terminal set T . This property prevents the proliferation of illegal programs due to crossover
and mutation. As an example consider a basic arithmetic language whose function and termi-
nal set are defined as follows: F = {+, −, ∗, /} and T = {A, B, C, 2}. Figure 1.18 shows two
examples of parse trees.

*
/

A /

* +
B +

A B 2 C
2 C

a) b)

Figure 1.18 Genetic programming parse trees, representing the following programs in LISP-like
syntax: a) (/(∗AB)(+2C)) and b) (∗A(/B(+2C))). Both programs implement the expression
AB/(2 + C). It is important to note that though LISP is the language chosen by Koza to implement
genetic programming, it is not the only possibility. Any language capable of representing programs as
parse trees is adequate. Moreover, machine language has been used as well [118].

Evolution in genetic programming proceeds along the general lines of the generic evolu-
tionary algorithm (Figure 1.17), with the genetic operators adapted to the tree representation.
Reproduction is performed in both asexual and sexual ways. Asexual reproduction, or cloning,
allows some of the fittest individuals to survive into the next generation; this is equivalent to
so-called elitist selection in genetic algorithms. Sexual reproduction, i.e., crossover, starts out
by selecting a random crossover point in each parent tree and then exchanging the subtrees
“hanging” from these points, thus producing two offspring trees (Figure 1.19). Mutation in
genetic programming is considered as a secondary genetic operator and is applied much less
frequently [82]. It is performed by removing a subtree at a randomly selected point and insert-
ing at that point a new random subtree.
One important issue in genetic programming is related to the size of the trees. Under
the influence of the crossover operator, the depth of the trees can quickly increase, leading
to a fitness plateau. The presence of huge programs in the population also has direct conse-
quences vis-a-vis computer memory and evaluation speed. Most implementations of genetic
programming include mechanisms to prevent trees from becoming too deep. However, these
1.3 Evolutionary computation 23

/ * / *

A / A *
* + / +
B + A B
A B 2 C B + 2 C
2 C
2 C

Figure 1.19 Crossover in genetic programming. The two shadowed subtrees of the parent trees are
exchanged to produce two offspring trees. Note that the two parents, as well as the two offspring, are
typically of different size.

mechanisms also present a disadvantage, in that they reduce the genetic diversity contained in
larger trees. There exist a number of books consacred to other advanced topics and applications
of genetic programming, which is a field in constant expansion [6].

1.3.3 Evolution strategies

Evolution strategies were introduced by Ingo Rechenberg [147, 148] and Hans-Paul Schwef-
fel [157, 158] in the 1960s as a method for solving parameter-optimization problems. In its
most general form the phenotype of an individual is a vector x containing the candidate values
of the parameters being optimized. The genotype of each individual is a pair of real-valued
vectors v = {x, σ}, where x is the above phenotypic vector (the genotype-phenotype dis-
tinction is thus somewhat degenerate with evolution strategies), and σ is a vector of standard
deviations used to apply the mutation operator. The inclusion of the σ vector in the genome
allows the algorithm to self-adapt the mutation operator while searching for the solution.
Somewhat different than the generic evolutionary algorithm (Figure 1.17), selection is
performed after the genetic operators have been applied. The standard notations in this domain,
(µ,λ)–ES and (µ + λ)–ES, denote algorithms in which a population of µ parents generates λ
offspring. The next generation is created by selecting the fittest µ individuals. In the case
of (µ,λ)–ES only theλ offspring are considered for selection, thus limiting the “life” of an
individual to one generation, while in the (µ + λ)–ES the µ parents are also considered for
selection.
Mutation is the major genetic operator in evolution strategies. It also plays the role of
a reproduction operator given that the mutated individual is viewed as an offspring for the
selection operator to work on. In its most general form, mutation modifies a genotype v =
{x, σ} by first randomly altering σ , and then modifying x according to the new values provided
by σ . This operation produces a new individual v
= {x
, σ
}, where x
= x + N(0, σ
).
N(0, σ
) denotes a vector of independent random Gaussian values with mean 0 and standard
deviations σ
.
The crossover (or recombination) operator generates an offspring from a number of par-
ents (usually two). There are two types of crossover operators: discrete and intermediate. In
discrete recombination each component of v, i.e., each pair of scalars (xi , σi ), is copied from
one of the parents at random. In intermediate recombination, the offspring values are a linear
combination of all the parent vectors participating in the recombination process.
The earliest evolution strategies were (1+1)–ES [147, 157], involving a single parent–
24 1. Introduction

0/n 1/n

Z R

1/q 0/m 0/q 1/p

Figure 1.20 A finite state machine with states {Z, T, R}. The input symbols belong to the set {0, 1},
whereas the output alphabet is the set {m, n, p, q}. The edges representing the state transitions are
labeled a/b, where a represents the input symbol triggering the transition, and b represents the output
symbol. For example, when the machine is in state R and the input is 0 it switches to state T and outputs
q. A double circle indicates the initial state.

single offspring search. Mutation was the only genetic operator, and the standard deviation
vector σ was constant or modified by some deterministic algorithm. Later, recombination was
added as evolution strategies were extended to encompass populations of individuals.
A good source for further information on evolution strategies is the book by Schwe-
fel [159].

1.3.4 Evolutionary programming

Lawrence Fogel proposed evolutionary programming [47, 48] as a means to develop artificial
intelligence. He argued that intelligent behavior requires both the ability to predict changes
in an environment, and a translation of the predictions into actions appropriate for reaching a
goal. In its most general, the environment is described as a sequence of symbols taken from
a finite alphabet. With its knowledge of the environment the evolving entity is supposed to
produce an output symbol that is related in some way to the next symbol appearing in the
environment. The output symbol should maximize a payoff function, which measures the
accuracy of the prediction. Finite state machines were selected to represent individuals in
evolutionary programming as they provide a meaningful representation of behavior based on
interpretation of symbols.
A finite state machine is a machine possessing a finite number of different internal states.
When stimulated by an input symbol the machine undergoes a transition (i.e., a change in
the internal state) and produces an output symbol. The behavior of the finite state machine
is described completely by defining the (output symbol,next state) pair for each (input sym-
bol,current state) pair. Figure 1.20 shows an example of a three-state machine.
Evolutionary programming maintains a population of finite state machines, each one rep-
resenting a particular candidate behavior for solving the problem at hand. The fitness of an
individual is calculated by presenting sequentially to the finite state machine the symbols in
the environment and observing the predicted output. The quality of the prediction is quantified
1.3 Evolutionary computation 25

according to the given payoff function. Once the individual has been exposed to the whole
sequence of symbols, its overall performance (e.g., average payoff per symbol) is used as the
fitness value.
Like with evolution strategies, evolutionary programming first generates offspring and
then selects the next generation. There is no sexual reproduction (crossover), but rather each
parent machine is mutated to produce a single offspring. There are five possible mutation oper-
ators: change of an output symbol, change of a state transition, addition of a state, deletion of a
state, and change of the initial state. The two latter operations are not allowed when the parent
machine has only one state. Mutation operators and the number of mutations per offspring are
chosen with respect to a probability distribution. The offspring are then evaluated in the same
way as their parents, and the next generation is selected from the ensemble of parents and
offspring. This process is iterated until a new symbol is required in the environment. The best
individual obtained up to this moment provides the prediction, the new symbol is added to the
environment, and the algorithm is executed again. Note that as opposed to most evolutionary-
computation applications where fitness is fixed from the outset, evolutionary programming
inherently incorporates a dynamic fitness, i.e., the environment changes in time. Fogel’s book
[45] is a good reference on evolutionary programming.

1.3.5 Classifier systems

Classifier systems, presented by Holland [63, 64], are evolution-based learning systems, rather
than a “pure” evolutionary algorithm. They can be thought of as restricted versions of clas-
sical rule-based systems, with the addition of input and output interfaces. A classifier system
consists of three main components: (1) the rule and message system, which performs the in-
ference and defines the behavior of the whole system, (2) the apportionment of credit system,
which adapts the behavior by credit assignment, and (3) the genetic algorithm, which adapts
the system’s knowledge by rule discovery.
The rule and message system includes four subcomponents: the message list, the classi-
fier list, the input interface, and the output interface. The input interface, also known as the
detector, translates information from the system’s environment into one or more finite-length
binary messages, which are posted to the finite-length message list. These messages may
then activate one or more matching classifiers from the classifier list. A classifier is a rule of
the form if condition then message, where condition is a finite-length string , and the mes-
sage, posted to the message list, may then activate other classifiers or trigger a system’s action
through the output interface, also called the effector (the alphabet of classifiers includes the
symbol # that plays the role of a wild-card character).
The apportionment-of-credit algorithm adapts the behavior of the classifier system by
modifying the way existing classifiers are used. Unlike traditional rule-based systems, clas-
sifier systems use parallel rule activation. This characteristic allows the system to accelerate
the inference process and to coordinate several actions simultaneously. However, with such a
competitive approach the system must determine the importance (strength) of rules in order
to combine them to make an overall decision. Although there are several ways to accomplish
this, the bucket-brigade algorithm continues to be the most popular [51]. It is a parallel, lo-
cal, credit-assignment-based reinforcement learning algorithm, which may be viewed as an
“information market,” where the right to trade information is bought and sold by classifiers.
26 1. Introduction

Each matched classifier makes a bid proportional to its strength. Rules that have accumulated
a large “capital” (i.e., strength) are preferred over other rules.
The genetic algorithm adapts the classifier system by introducing new classifiers (rules).
There exist two approaches for the application of evolutionary techniques in the design of
rule-based systems in general: the Michigan approach and the Pittsburgh approach; these two
approaches are also applied to classifier systems. In the Michigan approach, each individual
represents a single rule, and the classifier list is represented by the entire population. The
strengths calculated by the bucket-brigade algorithm are used as fitness function to evaluate
the quality of each classifier. In the Pittsburgh approach, the genetic algorithm maintains a
population of candidate classifier lists, with each individual representing an entire list. A good
introduction to classifier systems is given by Goldberg [51].
Chapter 2

Evolutionary Fuzzy Modeling

The Merriam-Webster online dictionary (www.m-w.com) defines model as “a system of pos-

tulates, data, and inferences presented as a mathematical description of an entity or state of
affairs.” Designing models of complex real-world systems and processes is essential in many
fields of science and engineering. The models developed can be used, among others, to ex-
plain the behavior of a system, to predict the system’s future development, and to keep it under
control.
The traditional approach to building models—known as white-box modeling—assumes
that everything about the system is known a priori, expressed either mathematically or ver-
bally. In another common approach—black-box modeling—a model is constructed entirely
from data using no additional a-priori knowledge. For example, in artificial neural networks,
a structure is chosen for the network and the parameters (e.g., the connection weights and
threshold values) are tuned to fit the observed data as best as possible. Such parameters are
not human-interpretable and do not offer any insight about the modeled system.
A third, intermediate approach, called grey-box modeling [90], takes into account cer-
tain prior knowledge of the modeled system to provide the black-box models with human-
interpretable meaning. In this way, the modeling process efficiently uses the available a priori
knowledge and concentrates its efforts in estimating what is still unknown.
Fuzzy modeling techniques, namely, the construction of fuzzy rule-based inference sys-
tems, can be viewed as grey-box modeling because they allow the modeler to extract and
interpret the knowledge contained in the model, as well as to imbue it with a-priori knowl-
edge. However, the construction of fuzzy models of large and complex systems—with a large
number of intricately related input and output variables—is a hard task demanding the identifi-
cation of many parameters. One way to solve this problem is to use a nature-inspired method:
evolution.
In this chapter, section 2.1 provides an overview of fuzzy modeling, section 2.2 presents
the application of artificial evolution to the fuzzy modeling problem, section 2.3 discusses
some aspects related to interpretability requirements, and in section 2.4 a real-world problem—
the Wisconsin breast cancer diagnosis (WBCD) problem—serves to illustrate the application
of evolutionary fuzzy modeling.

27
28 2. Evolutionary Fuzzy Modeling

2.1 Fuzzy Modeling: The Art of Building Fuzzy Systems

The principles of fuzzy modeling were outlined by Zadeh in [194] where he proposed a new
approach that “provides an approximate and yet effective means of describing the behavior
of systems which are too complex or too ill-defined to admit use of precise mathematical
analysis.” The models proposed by Zadeh present three distinguishable features: (1) the use of
linguistic variables in place or in addition to numerical variables, (2) the description of simple
relations between variables by conditional fuzzy statements, and (3) the characterization of
complex relations by fuzzy algorithms. Current fuzzy modeling techniques still follow these
principles.
An important issue in designing fuzzy models, which is a difficult and extremely ill-
defined process, involves the question of providing a methodology for their development. i.e.,
a set of techniques for obtaining the fuzzy model from information and knowledge about the
system. The next section presents some generalities about the fuzzy modeling problem, fol-
lowed by a description of some approaches to solve it in section 2.1.2

2.1.1 The fuzzy modeling problem

Fuzzy modeling is the task of identifying the parameters of a fuzzy inference system so that a
desired behavior is attained [185]. Note that, due to linguistic and numeric requirements, the
fuzzy-modeling process has generally to deal with an important trade-off between the accu-
racy and the interpretability of the model. In other words, the model is expected to provide
high numeric precision while incurring as little a loss of linguistic descriptive power as pos-
sible. With the direct approach a fuzzy model is constructed using knowledge from a human
expert [185]. This task becomes difficult when the available knowledge is incomplete or when
the problem space is very large, thus motivating the use of automatic approaches to fuzzy mod-
eling. One of the major problems in fuzzy modeling is the curse of dimensionality, meaning
that the computation requirements grow exponentially with the number of variables.
Following the connectionist representation of a fuzzy system, presented in Figure 1.16,
the parameters of a fuzzy inference system can be classified into the four categories presented
below. Table 2.1, summarizes this classification.

1. Logical parameters. Functions and operators which define the type of transformations
undergone by crisp and fuzzy quantities during the inference process. They include
the shape of the membership functions, the fuzzy logic operators applied for AND, OR,
implication, and aggregation operations, and the defuzzyfication method.

2. Structural parameters. Related mainly with the size of the fuzzy system. Includes the
number of variables participating in the inference, the number of membership functions
defining each linguistic variable, and the number of rules used to perform the inference.

3. Connective parameters. Related with the topology of the system, these parameters
define the connection between the different linguistic instances. They include the an-
tecedents, the consequents, and the weights of the rules.
2.1 Fuzzy Modeling: The Art of Building Fuzzy Systems 29

4. Operational parameters. These parameters define the mapping between the linguistic
and the numeric representations of the variables. They characterize the membership
functions of the linguistic variables.

Table 2.1 Parameter classification of fuzzy inference systems.

Class Parameters
Reasoning mechanism
Logical Fuzzy operators
Membership function types
Defuzzification method
Relevant variables
Structural Number of membership functions
Number of rules
Antecedents of rules
Connective Consequents of rules
Rule weights
Operational Membership function values

In fuzzy modeling, logical parameters are usually predefined by the designer based on
experience and on problem characteristics. Typical choices for the reasoning mechanism are
Mamdani-type, Takagi-Sugeno-Kang (TSK)-type, and singleton-type (Section 1.2.4) [185].
Common fuzzy operators are minimum, maximum, product, bounded product, bounded sum,
and probabilistic sum(Section 1.2.1. The most common membership functions are triangular,
trapezoidal, and bell-shaped. As for defuzzification, several methods have been proposed with
the Center Of Area (COA) and the Mean Of Maxima (MOM) being the most popular [101,185]
(Section 1.2.3).
Structural, connective, and operational parameters may be either predefined, or obtained
by synthesis or search methodologies. Generally, the search space, and thus the computational
effort, grows exponentially with the number of parameters. Therefore, one can either invest
more resources in the chosen search methodology, or infuse more a priori, expert knowledge
into the system (thereby effectively reducing the search space). The aforementioned trade-
off between accuracy and interpretability is usually expressed as a set of constraints on the
parameter values, thus complexifying the search process.

2.1.2 Approaches and techniques

The first fuzzy modeling works were very similar to, and inspired by, the knowledge-
engineering methods used in expert systems. They implemented Zadeh’s ideas by trying to
build a fuzzy model directly from the expert knowledge in what we call the direct approach.
The increasing availability of input-output data of the modeled processes, which is not
specifically used to determine the structure or the parameters of the fuzzy model in the direct
approach, together with the inherent difficulty to collect expert’s knowledge, motivated the
30 2. Evolutionary Fuzzy Modeling

use of more automatic approaches to fuzzy modeling, in which only a part of the fuzzy model
is built from a priori knowledge.
There exist a great number of fuzzy modeling methods differing in the search strategy
they apply and in the amount of parameters they can search for—related directly with the part
of the system they require to be pre-defined. Below, I briefly describe the direct approach to
fuzzy modeling (Section 2.1.2.1) as well as other approaches based on classic identification
algorithms (Section 2.1.2.2), on constructive learning methods (Section 2.1.2.3), and on bio-
inspired techniques (Section 2.1.2.4).

2.1.2.1 The direct approach to fuzzy modeling

In this approach, the system is first linguistically described, based on the expert’s a priori
knowledge. It is then translated into the formal structure of a fuzzy model following the steps
proposed by Zadeh [194]:

1. Selection of the input, state, and output variables (structural parameters);

2. Determination of the universes of discourse (structural parameters);

3. Determination of the linguistic labels into which these variables are partitioned (struc-
tural parameters);

4. Definition of the membership functions corresponding to each linguistic label (opera-

tional parameters);

5. Definition of the rules that describe the model’s behavior (connective parameters);

6. Selection of an adequate reasoning mechanism (logic parameters);

7. Evaluation of the model adequacy.

Unfortunately, there is no general methodology for the implementation of the direct approach,
which is more an art of intuition and experience than precise theory. This approach has been,
however, successfully used since the first fuzzy system applications [94, 95] to present-day
research [15, 43, 196].
One simple, and rather intuitive, improvement of the direct approach is the use of quan-
titative input-output information to update the membership-function values and/or the rule
weights in order to fine-tune the knowledge contained in the fuzzy model [113].

2.1.2.2 Approaches based on classic identification algorithms

A fuzzy model is a special type of nonlinear model. In this context, fuzzy modeling may be
done applying classic non-linear identification methods. These methods deals with an itera-
tive, convergent, estimation of a set of numeric parameters, which are applied to a, usually
pre-defined, model structure in order to approximate an expected behavior. In these fuzzy
modeling approaches, the general structure of the fuzzy system (i.e., logic and structural pa-
rameters) is pre-defined, while the rest of the system (i.e., connective and operational parame-
ters) is estimated.
2.1 Fuzzy Modeling: The Art of Building Fuzzy Systems 31

The simplest methods apply linear least-squares parameter estimation as they assume
that the parameters appear in a linear fashion into the model. Such linearity assumption lim-
its their applicability in fuzzy modeling and asks for the development of methods applying
nonlinear least-squares parameter estimation [90]. Recent works using this approach, apply
identification methods such as orthogonal least-squares [163], gradient descent [34], quasi-
Newton [155], Levenberg-Marquardt [35], or auto-regressive (AR) modeling [21].

2.1.2.3 Constructive learning approaches

In this approach, the a priori expert knowledge serves to direct the search process instead of
being used to directly construct a part of, or the whole, fuzzy system. After an expert-guided
definition of the logic parameters and of some of the structural parameters (mainly relevant
variables and their universes of discourse), a sequence of learning algorithms is applied so
as to progressively construct an adequate final fuzzy model. Most of the methods belonging
to this class begin by identifying a large fuzzy system—even systems with one rule for each
training case—satisfying certain performance criteria. They then apply a pruning strategy to
reduce the size of the system while keeping an acceptable performance. Recent examples of
this kind of approaches are presented by Espinosa and Vandewalle [37] and by Jin [75]. Other
methods, as for example that of Rojas et al. [150], start with simple fuzzy systems and then
iteratively increase the system’s complexity, by adding new rules and membership functions,
until a specified threshold of performance or of size is reached.

2.1.2.4 Bio-inspired approaches: neuro-fuzzy and evolutionary-fuzzy

As mentioned before, artificial neural networks, evolutionary algorithms, and fuzzy logic be-
long to the same family of bio-inspired methodologies. Indeed, they model in different extents
natural processes such as evolution, learning, or reasoning. The dynamic and continuously
growing research on these subjects, have allowed to identify the strengths and weaknesses of
each methodology, motivating a relatively recent trend to combine them in order to take advan-
tage of their complementarities. In fuzzy modeling, such combinations have originated hybrid
techniques known as neuro-fuzzy systems and evolutionary fuzzy modeling. As these latter
are presented in more detail in Section 2.2, I concentrate on the former.
Three main streams can be identified in the research on hybrid neural-fuzzy systems:

• Fuzzy-rule extraction from neural networks. This approach attempts to extract, in the
form of fuzzy rules, the knowledge embedded in trained neural networks [32, 107, 161].
The main drawback of these techniques is that the access to the knowledge requires a
previous rule-extraction phase.

• Neuro-fuzzy systems. These are fuzzy inference systems implemented as neural net-
works, taking advantage of their structural similarity (see Section 1.2.4). The main
advantage of this kind of representation is that such hybrid systems can be optimized
via powerful, well-known neural-network learning algorithms. ANFIS [71] is a well
known neuro-fuzzy system consisting of a six-layer generalized network with super-
vised learning. Most of the current research on this area is derived from the original
neuro-fuzzy concept, either in new flavors (i.e., by changing the network structure or the
32 2. Evolutionary Fuzzy Modeling

learning strategy) [19, 119, 173], or in adaptation of existing methods to face new hard
problems [91]. The main drawback of this approach is that the methods are intended to
maximize accuracy, neglecting human interpretability. In many applications this is not
acceptable.

• Interpretability-oriented neuro-fuzzy systems. A recent family of neuro-fuzzy systems

are constructed respecting certain interpretability-related constraints (see Section 2.3) to
keep permanent readability of the system during the learning process. One of the first
steps towards interpretable neuro-fuzzy systems is represented by the suite of methods
NEFCON, NEFCLASS, and NEFPROX [111, 112], based on a three-layer neuro-fuzzy
architecture whose synaptic weights are constrained to respect the integrity of the fuzzy
linguistic variables.

2.2 Evolutionary Fuzzy Modeling

Evolutionary algorithms are used to search large, and often complex, search spaces. They have
proven worthwhile on numerous diverse problems, able to find near-optimal solutions given
an adequate performance (fitness) measure. Fuzzy modeling can be considered as an opti-
mization process where part or all of the parameters of a fuzzy system constitute the search
space. Works investigating the application of evolutionary techniques in the domain of fuzzy
modeling first appeared more than a decade ago [79, 80]. These focused mainly on the tun-
ing of fuzzy inference systems involved in control tasks (e.g., cart-pole balancing, liquid-level
system, and spacecraft rendezvous operation). Evolutionary fuzzy modeling has since been
applied to an ever-growing number of domains, branching into areas as diverse as chemistry,
medicine, telecommunications, biology, and geophysics. Alander [1] and Cordón and Her-
rera [26] presented detailed bibliographies on evolutionary fuzzy modeling up to 1996.

2.2.1 Applying evolution to fuzzy modeling

Depending on several criteria—including the available a priori knowledge about the system,
the size of the parameter set, and the availability and completeness of input-output data—
artificial evolution can be applied in different stages of the fuzzy-parameter search. Three of
the four categories of fuzzy parameters in Table 2.1 can be used to define targets for evolu-
tionary fuzzy modeling: structural, connective, and operational parameters. As noted before,
logical parameters are usually predefined by the designer based on experience. The aforemen-
tioned categories lead to the definition of three levels of fuzzy modeling: knowledge tuning,
behavior learning, and structure learning, respectively. They are delineated below:

• Knowledge tuning (operational parameters). The evolutionary algorithm is used to

tune the knowledge contained in the fuzzy system by finding membership-function val-
ues. An initial fuzzy system is defined by an expert. Then, the membership-function
values are encoded in a genome, and an evolutionary algorithm is used to find systems
with high performance. Evolution often overcomes the local-minima problem present in
gradient descent-based methods. One of the major shortcomings of knowledge tuning is
its dependency on the initial setting of the knowledge base.
2.2 Evolutionary Fuzzy Modeling 33

• Behavior learning (connective parameters). In this approach, the membership func-

tions are defined using expert knowledge or an identification technique (e.g. fuzzy clus-
tering) is sufficient in order to define the membership functions; this determines, in fact,
the maximum number of rules [185]. The evolutionary algorithm is used to find either
the rule consequents, or an adequate subset of rules to be included in the rule base.
In the first case, the genome encodes directly the rule consequents. By not assigning
consequents from the output set and creating a selection pressure by means of a care-
fully designed fitness function, one may obtain smaller rule bases. A faster approach
encodes within the genome only a fixed set of user-defined rule templates. Selection of
the specific rules is made by indexing them or by fully defining the antecedents and the
consequents of each rule.
As the membership functions are fixed and predefined, this approach lacks the flexibility
to modify substantially the system behavior. Furthermore, as the number of variables and
membership functions increases, the curse of dimensionality becomes more pronounced
and the interpretability of the system decreases rapidly.

• Structure learning (structural parameters). In many cases, the available informa-

tion about the system is composed almost exclusively of input-output data, and specific
knowledge about the system structure is scant. In such a case, evolution has to deal
with the simultaneous design of rules, membership functions, and structural parame-
ters. Some methods use a fixed-length genome encoding a fixed number of fuzzy rules
along with the membership-function values. In this case the designer defines structural
constraints according to the available knowledge of the problem characteristics. Other
methods use variable-length genomes to allow evolution to discover the optimal size of
the rule base. The strong interdependency among the parameters involved in this form
of learning may slow down, or even prevent altogether, the convergence of the genetic
algorithm.

2.2.2 Three approaches to behavior and structure learning

Both behavior and structure learning can be viewed as rule-base learning processes with dif-
ferent levels of complexity. They can thus be assimilated within other methods from machine
learning, taking advantage of experience gained in this latter domain. In the evolutionary-
algorithm community there are two major approaches for evolving such rule systems: the
Michigan and the Pittsburgh approaches [104]. A more recent method has been proposed
specifically for fuzzy modeling: the iterative rule learning approach [59]. These three ap-
proaches are outlined below.

• The Michigan approach. Each individual represents a single rule. The fuzzy inference
system is represented by the entire population. Since several rules participate in the
inference process, the rules are in constant competition for the best action to be proposed,
and cooperate to form an efficient fuzzy system. The cooperative-competitive nature of
this approach renders difficult the decision of which rules are ultimately responsible for
good system behavior. It necessitates an effective credit-assignment policy to ascribe
fitness values to individual rules.
34 2. Evolutionary Fuzzy Modeling

• The Pittsburgh approach. The evolutionary algorithm maintains a population of can-

didate fuzzy systems, each individual representing an entire fuzzy system. Selection and
genetic operators are used to produce new generations of fuzzy systems. Since evalu-
ation is applied to the entire system, the credit-assignment problem is eschewed. This
approach facilitates to include additional optimization criteria in the fitness function,
thus affording the implementation of multi-objective optimization. The main shortcom-
ing of this approach is its computational cost, since a population of full-fledged fuzzy
systems has to be evaluated each generation.
• The iterative rule learning approach. As in the Michigan approach, each individ-
ual encodes a single rule. An evolutionary algorithm is used to find a single rule, thus
providing a partial solution. The evolutionary algorithm is then used iteratively for the
discovery of new rules, until an appropriate rule base is built. To prevent the process
from finding redundant rules (i.e., rules covering the same input subspace), a penaliza-
tion scheme is applied each time a new rule is added. This approach combines the speed
of the Michigan approach with the simplicity of fitness evaluation of the Pittsburgh ap-
proach. However, as with other incremental rule-base construction methods, it can lead
to a non-optimal partitioning of the antecedent space.

2.3 Interpretability Considerations

As mentioned before, the fuzzy-modeling process has to deal with an important trade-off be-
tween the accuracy and the interpretability of the model. The model is expected to provide
high numeric precision while incurring as little a loss of linguistic descriptive power as possi-
ble. Currently, there exist no well-established definitions for interpretability of fuzzy systems,
mainly due to the subjective nature of such a concept. However, some works have attempted
to define objective criteria that facilitate the automatic modeling of interpretable fuzzy sys-
tems [56, 175].
The fuzzy system of Figure 1.11 processes information in three stages: the input interface
(fuzzifier), the processing stage (inference engine), and the output interface (defuzzifier). The
interface deals with linguistic variables and their corresponding labels. These linguistic vari-
ables define the semantics of the system. The inference process is performed using fuzzy rules
that define the connection between input and output fuzzy variables. These fuzzy rules define
the syntax of the fuzzy system. Fuzzy modelers must thus take into account both semantic
and syntactic criteria to obtain interpretable systems. Below, I present some criteria that rep-
resent conditions driving fuzzy modeling toward human-interpretable systems together with
strategies to satisfy them.

2.3.1 Semantic criteria

The notion of linguistic variable formally requires associating a meaning to each fuzzy la-
bel [195]. Hence, each membership function should represent a linguistic term with a clear
semantic meaning. For example, in Figure 2.1, the fuzzy variable Triglycerides has three
meaningful labels: Normal, High, and Very High. The following semantic criteria describe a
set of properties that the membership functions of a fuzzy variable should possess in order to
2.3 Interpretability Considerations 35

facilitate the task of assigning linguistic terms [125, 129, 134]. The focus is on the meaning of
the ensemble of labels instead of the absolute meaning of each term in isolation.
Membership
Normal High Very High
1
0.75

0.25
Triglycerides
0 250 [mg/dL]
P1 =200 P2 =400 P3 =1000

Figure 2.1 Semantically correct fuzzy variable: Triglycerides has three possible fuzzy values, labeled
Normal, High, and Very High, plotted above as degree of membership versus input value. The values
Pi , setting the trapezoid and triangle apices, define the membership functions. In the figure, an example
input value 250 mg/dl is assigned the membership values µN ormal (250) = 0.75, µHigh (250) = 0.25,
and µV eryHigh (250) = 0. Note that µN ormal (250) + µHigh (250) + µV eryHigh (250) = 1.

• Distinguishability. Each linguistic label should have semantic meaning and the fuzzy set
should clearly define a range in the universe of discourse of the variable. In the example
of Figure 2.1, to describe variable Triglycerides we used three meaningful labels: Nor-
mal, High, and Very High. Their membership functions are defined using parameters
P1 , P2 , and P3 .
• Justifiable number of elements. The number of linguistic labels—i.e., the number of
membership functions of a variable—should be compatible with the number of concep-
tual entities a human being can handle. This number, that should not exceed the limit of
7 ± 2 distinct terms, is related directly with the expertise of the human interacting with
the system. In the example of Figure 2.1, while a patient would not feel comfortable
with more than the three labels defined, a physician should certainly handle more labels.
• Coverage. Any element from the universe of discourse should belong to at least one of
the fuzzy sets. That is, its membership value must be different than zero for at least one
of the linguistic labels. More generally, a minimum level of coverage may be defined,
giving rise to the concept of strong coverage. Referring to Figure 2.1, we see that any
value along the x-axis belongs to at least one fuzzy set; no value lies outside the range
of all sets.
• Normalization. Since all labels have semantic meaning, then, for each label, at least one
element of the universe of discourse should have a membership value equal to one. In
Figure 2.1, we observe that all three sets Normal, High, and Very High have elements
with membership value equal to 1.
• Complementarity. For each element of the universe of discourse, the sum of all its mem-
bership values should be equal to one (as in the example in Figure 2.1). This guarantees
uniform distribution of meaning among the elements.
36 2. Evolutionary Fuzzy Modeling

2.3.2 Syntactic criteria

A fuzzy rule relates one or more input-variable conditions, called antecedents, to their corre-
sponding output fuzzy conclusions, called consequents. The example rule presented in Fig-
ure 2.2, associates the conditions High and Middle of the input variables Triglycerides and
Age, respectively, with the conclusion Moderate of the output variable Cardiac risk. The lin-
guistic adequacy of a fuzzy rule base lies in the interpretability of each rule as well as in that
of the whole set of rules. The following syntactic criteria define some conditions which—if
satisfied by the rule base—reinforce the interpretability of a fuzzy system [56].
Tryglicerides

VeryHigh

R
High

Normal

Age

Young Middle Old

Figure 2.2 Example of a fuzzy rule and its firing range. The rule if Triglycerides is High and Age
is Middle then Cardiac risk is Moderate, marked as R, is (partially) fired by input values into the
dashed-line rectangle (i.e. µ(R) > 0). The solid-line rectangle denotes the region where µ(R) ≥ 0.5.

• Completeness. For any possible input vector, at least one rule should be fired to prevent
the fuzzy system from getting blocked.
• Rule-base simplicity. The set of rules must be as small as possible. If, however, the rule
base is still large, rendering hard a global understanding of the system, the number of
rules that fire simultaneously for any input vector must remain low in order to furnish a
simple local view of the behavior.
• Single-rule readability. The number of conditions implied in the antecedent of a rule
should be compatible with the aforementioned number of conceptual entities a human
being can handle (i.e., number of entities ≤ 7 ± 2).
• Consistency. If two or more rules are simultaneously fired, their consequents should not
be contradictory, i.e., they should be semantically close.

2.3.3 Strategies to satisfy semantic and syntactic criteria

The criteria presented above, intended to assess interpretability of a fuzzy system, define a
number of restrictions on the definition of fuzzy parameters. Semantic criteria limit the choice
2.3 Interpretability Considerations 37

of membership functions, while syntactic criteria bind the fuzzy rule base. I present below
some strategies to apply these restrictions when defining a fuzzy model.

• Linguistic labels shared by all rules. A number of fuzzy sets is defined for each variable,
which are interpreted as linguistic labels and shared by all the rules [56]. In other words,
each variable has a unique semantic definition. This results in a grid partition of the input
space as illustrated in Figure 2.3. To satisfy the completeness criterion, we normally use
a fully defined rule base, meaning that it contains all the possible rules, The example
system shown in Figure 2.4a contains all nine possible rules of the form if Triglycerides
is label and Age is label then Cardiac risk is . . . . Label sharing by itself facilitates
but does not guarantee the semantic integrity of each variable. More conditions are
necessary.

Tryglicerides

VeryHigh

High

Normal
Age

Young Middle Old

Figure 2.3 Grid partition of the input space. In this example, two semantically correct input variables,
each with three labels, divide the input space into a grid of nine regions.

• Normal, orthogonal membership functions. The membership functions of two succes-

sive labels must be complementary (i.e., their sum must be equal to one) in their over-
lapping region, whatever form they have [37, 175]. Moreover, in such regions each label
must ascend from zero to unity membership values [125, 129]. The variables presented
in Figures 2.1 and 2.3, satisfy these requirements.

• Don’t-care conditions. A fully defined rule base, as that shown in Figure 2.4a, becomes
impractical for high-dimension systems. The number of rules in a fully defined rule base
exponentially increases as the number of input variables increases (e.g., a system with
five variables, each with three labels, would contain 35 = 243 rules). Moreover, given
that each rule antecedent contains a condition for each variable, the rules might be too
lengthy to be understandable, and too specific to describe general circumstances.
To tackle these two problems some authors use “don’t-care” as a valid input label [69,
125,129]. Variables in a given rule that are marked with a don’t-care label are considered
as irrelevant. For example, in the rule base shown in Figure 2.4b two rules, RA and RB ,
containing don’t-care labels cover almost half of the input space. The rule RA :

if Triglycerides is don’t-care and Age is Old then Cardiac risk is Moderate,

38 2. Evolutionary Fuzzy Modeling

covers the space of three rules (i.e., R3 , R6 , and R9 in Figure 2.4a) and is interpreted as:
if Age is Old then Cardiac risk is Moderate.
In the same way, the rule RB is interpreted as:
if Triglycerides is VeryHigh then Cardiac risk is . . . .
Although don’t-care labels allow a reduction of rule-base size, their main advantage is
the improvement of rule readability.
Tryglicerides
VeryHigh R7 R8 R9 RB RB

High R4 R5 R6 R4 R5 RA R5 RA

Normal R1 R2 R3 R1 R2 R0

Age
Young Middle Old
a) b) c)

Figure 2.4 Strategies to define a complete rule base. a) Fully defined rule base. b) don’t-care labels.
c) Default rule. By definition, the activation of the (fuzzy) default rule is µ(R0 ) = 1 − max(µ(Ri )),
with i = {1, 2, 3, . . .}. The rectangles denote the region where µ(Ri ) ≥ 0.5.

• Default rule. In many cases, the behavior of a system exhibits only a few regions of
interest, which can be described by a small number of rules (e.g., R5 , RA , and RB in
Figure 2.4b). To describe the rest of the input space, a simple default action, provided
by the default rule, would be enough [178]. The example in Figure 2.4c shows that
the default rule, named R0 , covers the space of rules R1 , R2 , and R4 . By definition, a
default condition is true when all other rule conditions are false. In a fuzzy context, the
default rule is as true as all the others are false. Consequently, the activation degree of
the default rule, µ(R0 ), is thus given by µ(R0 ) = 1 − max(µ(Ri )), where µ(Ri ) is the
activation degree of the i-th rule.
• Linguistic fitness. Some linguistic criteria can be reinforced by taking them into account
to compute the fitness value of a given fuzzy system. Size factors, related with sim-
plicity and readability, such as the number of rules effectively used or the number of
conditions implied in the rule antecedents can be easily quantified and included in the
fitness function.

2.4 Example: Medical Diagnosis

A major class of problems in medical science involves the diagnosis of disease, based upon
various tests performed upon the patient. When several tests are involved, the ultimate diag-
nosis may be difficult to obtain, even for a medical expert. This has given rise, over the past
2.4 Example: Medical Diagnosis 39

few decades, to computerized diagnostic tools, intended to aid the physician in making sense
out of the welter of data.
A prime target for such computerized tools is in the domain of cancer diagnosis. Specif-
ically, where breast cancer is concerned, the treating physician is interested in ascertaining
whether the patient under examination exhibits the symptoms of a benign case, or whether her
case is a malignant one.
A good computerized diagnostic tool should possess two characteristics, which are often
in conflict. First, the tool must attain the highest possible performance, i.e., diagnose the
presented cases correctly as being either benign or malignant. Second, it would be highly
beneficial for such a diagnostic system to be interpretable. This means that the physician is not
faced with a black box that simply spouts answers (albeit correct) with no explanation; rather,
we would like for the system to provide some insight as to how it derives its outputs.
Thanks to their linguistic representation and their numeric behavior, fuzzy systems can
provide both characteristics. Moreover, an evolutionary fuzzy modeling approach, carefully
setup following the strategies presented in Section 2.3.3, can deal with the trade-off between
performance and interpretability. Following, I describe the Wisconsin breast cancer diagnosis
(WBCD) problem. Section 2.4.2 then describes a genetic-fuzzy approach to the WBCD prob-
lem, while Section 2.4.3 delineates the results. In Section 2.4.4, I discuss the issue of obtaining
a confidence measure of the system’s output, going beyond a mere binary, benign-malignant
classification. Finally, Section 2.4.5, briefly describes two further experiments that I carried
out.

2.4.1 The Wisconsin breast cancer diagnosis (WBCD) problem

This section presents the medical-diagnosis problem which is the object of my study, and the
fuzzy system I propose to solve it.
Breast cancer is the most common cancer among women, excluding skin cancer. The
presence of a breast mass1 is an alert sign, but it does not always indicate a malignant cancer.
Fine needle aspiration (FNA)2 of breast masses is a cost-effective, non-traumatic, and mostly
non-invasive diagnostic test that obtains information needed to evaluate malignancy.
The WBCD database [102] is the result of the efforts made at the University of Wisconsin
Hospital for accurately diagnosing breast masses based solely on an FNA test [96]. Nine
visually assessed characteristics of an FNA sample considered relevant for diagnosis were
identified, and assigned an integer value between 1 and 10. The measured variables are as
follows:

1 Clump thickness (v1 );

2 uniformity of cell size (v2 );
3 uniformity of cell shape (v3 );
4 marginal adhesion (v4 );
5 single epithelial cell size (v5 );
1
Most breast cancers are detected as a lump or mass on the breast, by self-examination, by mammography, or
by both [97].
2
Fine needle aspiration is an outpatient procedure that involves using a small-gauge needle to extract fluid
directly from a breast mass [97].
40 2. Evolutionary Fuzzy Modeling

6 bare nuclei (v6 );

7 bland chromatin (v7 );
8 normal nucleoli (v8 );
9 mitosis (v9 ).
The diagnostics in the WBCD database were furnished by specialists in the field. The
database itself consists of 683 cases, with each entry representing the classification for a
certain ensemble of measured values:

case v1 v2 v3 · · · v9 diagnostic
1 5 1 1 ··· 1 benign
2 5 4 4 ··· 1 benign
.. .. .. .. .. . ..
. . . . . .. .
683 4 8 8 ··· 1 malignant
Note that the diagnostics do not provide any information about the degree of benignity or
malignancy.
There are several studies based on this database. Bennet and Mangasarian [10] used
linear programming techniques, obtaining a 99.6% classification rate on 487 cases (the reduced
database available at the time). However, their solution exhibits little understandability, i.e.,
diagnostic decisions are essentially black boxes, with no explanation as to how they were
attained. With increased interpretability in mind as a prior objective, a number of researchers
have applied the method of extracting Boolean rules from neural networks [160–162, 169].
Their results are encouraging, exhibiting both good performance and a reduced number of
rules and relevant input variables. Nevertheless, these systems use Boolean rules and are not
capable of furnishing the user with a measure of confidence for the decision made. My own
work on the evolution of fuzzy rules for the WBCD problem has shown that it is possible
to obtain diagnostic systems exhibiting high performance, coupled with interpretability and a
confidence measure [123–126].

2.4.2 A genetic-fuzzy approach to the WBCD problem

The solution scheme I propose for the WBCD problem is depicted in Figure 2.5. It consists
of a fuzzy system and a threshold unit. The fuzzy system computes a continuous appraisal
value of the malignancy of a case, based on the input values. The threshold unit then outputs a
benign or malignant diagnostic according to the fuzzy system’s output.

Input Fuzzy Subsystem Appraisal Threshold Subsystem Diagnostic

malignant

benign

Figure 2.5 Proposed diagnosis system. Note that the fuzzy subsystem displayed to the left is in fact the
entire fuzzy inference system of Figure 1.11.

In order to evolve the fuzzy model we must make some preliminary decisions about the
fuzzy system itself and about the genetic algorithm encoding.
2.4 Example: Medical Diagnosis 41

2.4.2.1 Fuzzy system parameters

My previous knowledge about the WBCD problem represents valuable information to be used
for my choice of fuzzy parameters (Table 2.1). When defining the setup, I took into consid-
eration the following three results concerning the composition of potential high-performance
systems: (1) small number of rules; (2) small number of variables; and (3) monotonicity of
the input variables. Some fuzzy models forgo interpretability in the interest of improved per-
formance. Where medical diagnosis is concerned, interpretability is the major advantage of
fuzzy systems. This motivates to take into account the interpretability criteria presented in
Section 2.3 to define constraints on the fuzzy parameters. Referring to Table 2.1, I delineate
below the fuzzy-system setup:
• Logical parameters: singleton-type fuzzy systems; min-max fuzzy operators; orthogo-
nal, trapezoidal input membership functions (see Figure 2.6); weighted-average defuzzi-
fication.
• Structural parameters: two input membership functions (Low and High; see Figure 2.6);
two output singletons (benign and malignant); a user-configurable number of rules. The
relevant variables are one of the evolutionary objectives.
Membership
Low High
1

0 Variable

P d

Figure 2.6 Input fuzzy variables for the WBCD problem. Each fuzzy variable has two possible fuzzy
values labeled Low and High, and orthogonal membership functions, plotted above as degree of mem-
bership versus input value. P and d define the start point and the length of membership function edges,
respectively.

• Connective parameters: the antecedents of the rules are searched by the evolutionary
algorithm. The consequents of the rules are predefined, the algorithm finds rules for the
benign diagnostic, while the default-rule consequent is the malignant diagnostic. All
rules have unitary weight.
• Operational parameters: the input membership-function values are to be found by the
evolutionary algorithm. For the output singletons I used the values 2 and 4, for benign
and malignant, respectively.

2.4.2.2 The evolutionary setup

The problem, at this stage, consists of searching for three fuzzy-system parameters: input
membership functions, antecedents of rules, and relevant variables. I applied a Pittsburgh-like
42 2. Evolutionary Fuzzy Modeling

approach, using a simple genetic algorithm [177] to search for individuals whose genomes,
encoding these three parameters, are defined as follows:

• Membership-function parameters. There are nine variables (v1 – v9 ), each with two
parameters P and d, defining the start point and the length of the membership-function
edges, respectively (Figure 2.6).

• Antecedents. The i-th rule has the form:

if (v1 is Ai1 ) and . . . and (v9 is Ai9 ) then (output is benign),

where Aij represents the membership function applicable to variable vj . Aij can take on
the values: 1 (Low), 2 (High), or 0 or 3 (Don’t Care).

• Relevant variables are searched for implicitly by letting the algorithm choose Don’t care
labels as valid antecedents; in such a case the respective variable is considered irrelevant
(See Section 2.3.3).

Table 2.2 delineates the parameter encoding and Figure 2.7 shows a sample genome en-
coding a whole fuzzy system.

Table 2.2 Parameter encoding of an individual’s genome. Total genome length is 54 + 18Nr , where
Nr denotes the number of rules (Nr is set a priori to a value between 1–5, and is fixed during the
genetic-algorithm run).

Parameter Values Bits Qty Total bits

P {1,2,. . . ,8} 3 9 27
d {1,2,. . . ,8} 3 9 27
A {0,1,2,3} 2 9 × Nr 18 × Nr

To evolve the fuzzy inference system, I used a genetic algorithm with a fixed popu-
lation size of 200 individuals, and fitness-proportionate selection (Section 1.3). The algo-
rithm terminates when the maximum number of generations, Gmax , is reached (I set Gmax =
2000 + 500 × Nr , i.e., dependent on the number of rules used in the run), or when the increase
in fitness of the best individual over five successive generations falls below a certain threshold
(in these experiments, I used threshold values between 2 × 10−7 and 4 × 10−6 ).
My fitness function combines three criteria: (1) Fc : classification performance, computed
as the percentage of cases correctly classified; (2) Fe : the quadratic difference between the
continuous appraisal value (in the range [2, 4]) and the correct discrete diagnosis given by the
WBCD database (either 2 or 4); and (3) Fv : the average number of variables per active rule.
The fitness function is given by F = Fc − αFv − βFe , where α = 0.05 and β = 0.01 (these
latter values were derived empirically). Fc , the ratio of correctly diagnosed cases, is the most
important measure of performance. Fv measures the interpretability, penalizing systems with
a large number of variables per rule (on average). Fe adds selection pressure towards systems
with low quadratic error.
2.4 Example: Medical Diagnosis 43

P1 d1 P2 d2 P3 d3 P4 d4 P5 d5 P6 d6 P7 d7 P8 d8 P9 d9
...
4 3 1 5 2 7 1 1 1 6 3 7 4 6 7 1 1 5

A11 A12 A13 A14 A15 A16 A17 A18 A19

...
0 1 3 3 2 3 1 3 1

(a)
Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 4 1 2 1 1 3 4 7 1
d 3 5 7 1 6 7 6 1 5
Rule base
Rule 1 if (v2 is Low) and (v5 is High) and (v7 is Low) and (v9 is Low) then
(output is benign)
Default else (output is malignant)
(b)
Figure 2.7 Example of a genome for a single-rule system. (a) Genome encoding. The first 18 positions
encode the parameters P and d for the nine variables v1 –v9 . The rest encode the membership function
applicable for the nine antecedents of each rule. (b) Interpretation. Database and rule base of the
single-rule system encoded by (a). The parameters P and d are interpreted as illustrated in Figure 2.6.

2.4.3 Results
The evolutionary experiments performed fall into three categories, in accordance with the data
repartitioning into two distinct sets: training and test (or evaluation). The three experimental
categories are: (1) the training set contains all 683 cases of the WBCD database, while the test
set is empty; (2) the training set contains 75% of the WBCD cases, and the test set contains
the remaining 25% of the cases; (3) the training set contains 50% of the WBCD cases, and
the test set contains the remaining 50% of the cases. In the last two categories, the choice of
training-set cases is done randomly, and is performed anew at the outset of every evolutionary
run. The number of rules per system was also fixed at the outset, to be between one and five,
i.e., evolution seeks a system with an a priori given number of rules (the choice of number of
rules per system determines the final structure of the genome, as presented in Table 2.2).
A total of 120 evolutionary runs were performed, all of which found systems whose
classification performance exceeds 94.5%. In particular, considering the best individual per
run (i.e., the evolved system with the highest classification success rate), 78 runs led to a
fuzzy system whose performance exceeds 96.5%, and of these, 8 runs found systems whose
performance exceeds 97.5%; these results are summarized in Figure 2.8. Table 2.3 presents the
average performance obtained by the genetic algorithm over all 120 evolutionary runs, divided
according to the three experimental categories discussed above. A more detailed account of
these results is presented in Table 2.7 at the end of this chapter, which lists the top evolved 45
systems.
Table 2.4 shows the results of the best systems obtained with the fuzzy-genetic approach.
The number of rules per system was fixed at the outset to be between one and five, i.e., evo-
lution seeks a system with an a priori given number of rules. A comparison of these systems
with other approaches is presented in Section 4.1.2.
44 2. Evolutionary Fuzzy Modeling

Figure 2.8 Summary of results of 120 evolutionary runs. The histogram depicts the number of systems
exhibiting a given performance level at the end of the evolutionary run. The performance considered is
that of the best individual of the run, measured as the overall percentage of correctly classified cases
over the entire database.

Table 2.3 Summary of results of 120 evolutionary runs, divided according to the three experimental
categories discussed in the text (i.e., the three classes which differ in the training-set to test-set ratio).
The table lists the average performance over all 120 runs, where the averaging is done over the best
individual of each run. The performance value denotes the percentage of cases correctly classified.
Three such performance values are shown: (1) performance over the training set; (2) performance
over the test set; and (3) overall performance, considering the entire database. In addition, the average
number of variables per rule is also shown.

Training/test Performance Variables

ratio Training set Test set Overall per rule
100% / 0% – – 96.97% 3.32
75% / 25% 97.00% 96.02% 96.76% 3.46
50% / 50% 97.71% 94.73% 96.23% 3.41

I next describe three of my top-performance systems, which serve to exemplify the solu-
tions found by my evolutionary approach. The first system, delineated in Figure 2.9, consists
of three rules (note that the default rule is not counted as an active rule). Taking into account
all three criteria of performance—classification rate, number of rules per system, and average
number of variables per rule— this system can be considered the top one over all 120 evolu-
tionary runs. It obtains 98.2% correct classification rate over the benign cases, 97.07% correct
classification rate over the malignant cases,3 and an overall classification rate (i.e., over the
entire database) of 97.8%.
3
The WBCD database contains 444 benign cases and 239 malignant cases.
2.4 Example: Medical Diagnosis 45

Table 2.4 Results of the best systems evolved by the fuzzy-genetic approach. Shown below are the
classification performance values of the top systems obtained by these approaches, along with the
average number of variables-per-rule. Results are divided into five classes, in accordance with the
number of rules-per-system, going from one-rule systems to five-rule ones.

Rules-per-system Performance variables-per-rule

1 97.07% 4
2 97.36% 3
3 97.80% 4.7
4 97.80% 4.8
5 97.51% 3.4

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 3 5 2 2 8 1 4 5 4
d 5 2 1 2 4 7 3 5 2
Rule base
Rule 1 if (v3 is Low) and (v7 is Low) and (v8 is Low) and (v9 is Low) then
(output is benign)
Rule 2 if (v1 is Low) and (v2 is Low) and (v3 is High) and (v4 is Low) and
(v5 is High) and (v9 is Low) then (output is benign)
Rule 3 if (v1 is Low) and (v4 is Low) and (v6 is Low) and (v8 is Low) then
(output is benign)
Default else (output is malignant)
Figure 2.9 The best evolved, fuzzy diagnostic system with three rules. It exhibits an overall classifica-
tion rate of 97.8%, and an average of 4.7 variables per rule.

A thorough test of this three-rule system revealed that the second rule (Figure 2.9) is never
actually used, i.e., is fired by none of the input cases. Thus, it can be eliminated altogether from
the rule base, resulting in a two-rule system (also reducing the average number of variables
per rule from 4.7 to 4).

Can the genetic algorithm automatically discover a two-rule system, i.e., without recourse
to any post-processing (such as that described in the previous paragraph)? My results have
shown that this is indeed the case—one such solution is presented in Figure 2.10. It obtains
97.3% correct classification rate over the benign cases, 97.49% correct classification rate over
the malignant cases, and an overall classification rate of 97.36%.

Finally, Figure 2.11 delineates the best one-rule system found through my evolutionary
approach. It obtains 97.07% correct classification rate over the benign cases, 97.07% correct
classification rate over the malignant cases, and an overall classification rate of 97.07%.
46 2. Evolutionary Fuzzy Modeling

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 1 1 1 6 2 3
d 5 3 2 7 4 1
Rule base
Rule 1 if (v1 is Low) and (v3 is Low) then (output is benign)
Rule 2 if (v2 is Low) and (v5 is Low) and (v6 is Low) and (v8 is Low) then
(output is benign)
Default else (output is malignant)
Figure 2.10 The best evolved, fuzzy diagnostic system with two rules. It exhibits an overall classifica-
tion rate of 97.36%, and an average of 3 variables per rule.

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 4 4 2 2
d 3 1 5 7
Rule base
Rule 1 if (v1 is Low) and (v2 is Low) and (v6 is Low) and (v8 is Low) then
(output is benign)
Default else (output is malignant)

Figure 2.11 The best evolved, fuzzy diagnostic system with one rule. It exhibits an overall classification
rate of 97.07%, and a rule with 4 variables.

2.4.4 Diagnostic confidence

So far, we have been using the evolved fuzzy systems to ultimately produce a binary classifi-
cation value—benign or malignant—with no finer gradations. Going back to Figure 2.5, we
note that the diagnostic system comprises in fact two subsystems: the first subsystem consists
of the (evolved) fuzzy system, which, upon presentation of an input (in this case, a WBCD
database entry) proceeds to produce a continuous appraisal value; this value is then passed
along to the second subsystem—the threshold unit—which produces the final binary output
(benign or malignant). The first subsystem (the fuzzy system) is the one evolved in this ap-
proach. The threshold subsystem simply outputs malignant if the appraisal value is below a
fixed threshold value, and outputs benign otherwise. The threshold value is assigned by the
user through knowledge of the problem at hand.
To gain an intuitive understanding of how a classification value is computed, let me sketch
a simple example. Referring to the system of Figure 2.5, assume that the following values are
presented as inputs (these represent case #145 of the WBCD database):

v1 v2 v3 v4 v5 v6 v7 v8 v9
Value 4 3 1 1 2 1 4 8 1

The membership value of each variable is then computed in accordance with the (evolved)
database of Figure 2.9:
2.4 Example: Medical Diagnosis 47

v1 v2 v3 v4 v5 v6 v7 v8 v9
µLow 0.8 1 1 1 1 1 1 0.4 1
µHigh 0.2 0 0 0 0 0 0 0.6 0

This completes the fuzzification phase (the “fuzzifier” unit of Figure 1.11). Having com-
puted these membership values, the inference engine (Figure 1.11) can now go on to compute
the so-called truth value of each rule. This truth value is computed by applying the fuzzy AND
operator (Section 1.2) to combine the antecedent clauses (the membership values) in a fuzzy
manner; this results in the output truth value, namely, a continuous value which represents the
rule’s degree of activation. Thus, a rule is not merely either activated or not, but in fact is ac-
tivated to a certain degree—represented by a value between 0 and 1. In this example, the rule
activation values are as follows (remember that we “chucked out” rule 2, since it was found to
never fire):

Rule 1 Rule 3 Default

Truth value 0.4 0.4 0.6

The inference engine (Figure 1.11) now goes on to apply the aggregation operator (Sec-
tion 1.2), combining the continuous rule activation values to produce a fuzzy output with a
certain truth value (the point marked “fuzzy output” in Figure 1.11). The defuzzifier then
kicks in (Figure 1.11), producing the final continuous value of the fuzzy inference system; this
latter value is the appraisal value that is passed on to the threshold unit (Figure 2.5). In this
example the appraisal value is 2.86.
In general, the appraisal value computed by my evolved fuzzy systems is in the range
[2, 4]. I chose to place the threshold value at 3, with inferior values classified as benign, and
superior values classified as malignant . Thus, in this example, the appraisal value of 2.86 is
classified as benign—which is correct.
This case in the WBCD database produces an appraisal value (2.86) which is among the
closest to the threshold value. Most other cases result in an appraisal value that lies close to
one of the extremes (i.e., close to either 2 or 4). Thus, in a sense, we can say that we are
somewhat less confident where this case is concerned, with respect to most other entries in
the WBCD database; specifically, the appraisal value can accompany the final output of the
diagnostic system, serving as a confidence measure. This demonstrates a prime advantage
of fuzzy systems, namely, the ability to output not only a (binary) classification, but also a
measure representing the system’s confidence in its output.

2.4.5 Further experiments

In this section I describe two further experiments carried out; these are aimed at searching for
systems with yet better performance than obtained hitherto—though possibly at the expense
of some other aspect of the resulting system.
As noted in Section 2.1, fuzzy systems offer a major advantage in terms of (possible)
linguistic integrity, i.e., interpretability by humans. With this goal in mind, the experiments
described previously were constrained: I limited both the number of rules per system, as well
as the number of variables per rule (See Section 2.3). This latter constraint was incorporated by
48 2. Evolutionary Fuzzy Modeling

favoring systems with few variables-per-rule via the Fv coefficient: lower Fv , meaning fewer
variables-per-rule, entails higher overall fitness.
Can higher-performance systems be obtained by eliminating the Fv factor (albeit at the
cost of reduced interpretability due to more complicated rules)? This was the aim of the first
of the two experiments described herein. I eliminated not only the Fv measure but also the
Fe factor, the resulting fitness function thus containing solely Fc . My intent was to provide
selection pressure for but one goal: overall classification performance. With this aim in mind
I was also more “generous” with the number of rules per system: whereas previously this was
set to a (fixed) value between one and five, herein I set this value to be between three and
seven.
I performed a total of 31 evolutionary runs, the results of which are summarized in Ta-
ble 2.5. Note that my previously best system (Figure 2.9) obtained 97.8% overall classification
performance, while Table 2.5 shows an evolved system with a 98.24% classification rate. This
latter system is thus able to correctly classify 3 additional cases. This small improvement in
performance carries, however, a price: the slightly better system comprises four rules with an
average of 5.8 variables per rule, whereas the previous one (Figure 2.9) contains but three rules
with an average of 4.7 variables per rule; I have thus traded off interpretability for performance.
The judgment of whether this is worthwhile or not is entirely dependent on the human user. It
is noteworthy that this choice (interpretability versus performance) can be easily implemented
in this approach.

Table 2.5 Results of evolutionary runs in which the variables-per-rule constraint has been removed.
Results are divided into five classes, in accordance with the number of rules-per-system, going from
three-rule systems to seven-rule ones. I performed 5–7 runs per class, totaling 31 runs in all; shown
below are the resulting best systems as well as the average per class. Results include the overall
classification performance and the average number of variables-per-rule in parentheses.

Rules-per-system Best system Average

3 97.66% (5.3) 97.49% (5.4)
4 98.24% (5.8) 97.63% (5.4)
5 97.95% (6) 97.63% (5.6)
6 98.10% (6.2) 97.77% (5.4)
7 97.95% (5) 97.88% (5.2)
Total 98.24% (5.8) 97.68% (5.4)

As explained in Section 2.4.2, the active rules diagnose benignity, with the default diagno-
sis being malignancy; this means that the if conditions have benign as a consequent, whereas
the else condition has malignant as a consequent. The second experiment sought to find out
what would happen if this were reversed, i.e., could better systems be evolved with benignity as
the default diagnosis? Table 2.6 delineates the results of 27 evolutionary runs. While I did not
improve upon the results of the malignancy-default systems of Section 2.4.3, I did note a ten-
dency toward a smaller number of variables-per-rule. The highest-performance system found
in this experiment is fully specified in Figure 2.12. It comprises five rules with an average of
2.8 variables-per-rule, exhibiting the same overall performance (97.8%) as the three-rule, 4.7
2.4 Example: Medical Diagnosis 49

average-variables-per-rule system of Figure 2.9. This nicely illustrates the tradeoff between
these two parameters: number of rules per system and average number of variables per rule.

Table 2.6 Results of evolutionary runs in which the default diagnosis is benign (rather than malignant).
Results are divided into five classes, in accordance with the number of rules-per-system, going from
one-rule systems to five-rule ones. I performed 4–6 runs per class, totaling 27 runs in all; shown below
are the resulting best systems as well as the average per class. Results include the overall classification
performance and the average number of variables-per-rule in parentheses.

Rules-per-system Best system Average

1 94.73% (2) 94.44% (2)
2 96.93% (1.5) 96.34% (1.8)
3 97.36% (2) 96.57% (1.7)
4 97.07% (1.8) 97.00% (2)
5 97.80% (2.8) 96.52% (2.2)
Total 97.80% (2.8) 97.68% (1.9)

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 4 2 2 8 4 2 2 5 6
d 8 5 5 1 8 6 3 4 5
Rule base
Rule 1 if (v2 is High) and (v7 is High) then (output is malignant)
Rule 2 if (v2 is High) and (v3 is High) and (v4 is Low) and (v8 is High) and
(v9 is Low) then (output is malignant)
Rule 3 if (v3 is High) and (v6 is High) then (output is malignant)
Rule 4 if (v3 is Low) and (v5 is High) and (v8 is High) then (output is malig-
nant)
Rule 5 if (v1 is High) and (v6 is High) then (output is malignant)
Default else (output is benign)
Figure 2.12 The best evolved, fuzzy diagnostic system with active rules encoding malignant cases. It
exhibits an overall classification rate of 97.80%, and an average of 2.8 variables per rule.
2. Evolutionary Fuzzy Modeling

Table 2.7 Summary of top 45 evolutionary runs (of 120) described in Section 2.4.3, divided according to the three experimental categories discussed
in the text (i.e., the three classes which differ in the training-set to test-set ratio). For each of the 45 evolved systems, the table lists its fitness value,
its performance, and its average number of variables-per-rule. As explained in Section 2.4.3, the performance value denotes the percentage of cases
correctly classified. Three such performance values are shown: (1) performance over the training set; (2) performance over the test set; and (3)
overall performance, considering the entire database.

100% / 0% 75% / 25% 50% / 50%

performance

performance

performance
variables

variables

variables
training

training
fitness

fitness

fitness
Rules

test

test
0.9548 96.78 3 0.9553 97.46 95.91 97.07 4 0.9637 97.66 95.6 96.63 3
1 0.9548 96.78 3 0.9597 97.27 95.32 96.78 3 0.9607 97.37 95.01 96.19 3
0.9533 96.63 3 0.9557 96.88 96.49 96.78 3 0.9607 97.37 95.01 96.19 3
0.9576 97.36 3.5 0.9598 97.27 97.66 97.36 3 0.9603 97.95 95.89 96.93 4
2 0.9593 97.22 3 0.9548 97.07 96.49 96.93 3.5 0.9579 97.08 96.77 96.93 3
0.9578 97.07 3 0.9648 97.46 94.74 96.78 2.5 0.9636 97.95 94.43 96.19 3.5
0.9548 97.8 4.67 0.9577 97.27 97.66 97.36 3.33 0.9659 97.66 95.89 96.78 2.67
3 0.9554 97.66 4.33 0.9557 97.27 97.08 97.22 3.67 0.9546 97.37 95.89 96.63 4
0.9594 97.22 3 0.9577 97.27 95.91 96.93 3.33 0.9626 97.95 95.01 96.49 3.67
0.9543 97.8 4.75 0.952 97.27 98.25 97.51 4.25 0.9755 99.12 95.6 97.36 3.5
4 0.9529 97.51 4.5 0.9563 97.07 96.49 96.93 3.25 0.971 98 95.89 97.36 3.75
0.9594 97.22 3 0.9591 97.66 94.15 96.78 3.75 0.965 98.25 95.31 96.78 3.75
0.9599 97.51 3.4 0.9587 97.66 97.08 97.51 3.8 0.9586 97.66 96.77 97.22 3.8
5 0.9483 97.36 5 0.9575 97.66 97.08 97.51 4 0.9756 98.83 94.72 96.78 3
0.9584 97.36 3.4 0.9561 97.27 96.49 97.07 3.6 0.9623 98.25 95.01 96.63 4.2
50
Chapter 3

Coevolutionary Fuzzy Modeling

In the simplified models of evolution discussed in Section 1.3, we consider individuals be-
longing to a single species—i.e., sharing the same genetic encoding, and reproducing with
each other. We assume this species evolves in isolation, in an almost unchanging environment.
In nature, species live in the niches afforded by other species, modifying themselves and the
environment and being affected by such modifications.
Over time, the evolution of many species has been influenced by interactions with other
species. Species that have mutually influenced one another’s evolution are said to have coe-
volved [145]. Flowers coevolved with the insects that pollinated them and fed on their nectar
in a mutualist relationship where reproductive success (fitness) of one species is beneficial for
other species’ survival. On the other hand, predator-prey interaction constitutes an example
of competitive coevolution where the survival of individuals of one species requires the death
of individuals from other species. Although species-specific coevolution is easily identifiable
(e.g., yucca plants and the so-called yucca moths cannot reproduce without the other) it is rare,
contrary to the widespread diffuse coevolution. In diffuse coevolution, species are influenced
by a wide variety of predators, parasites, preys, and mutualists.
Coevolution has served as inspiration to propose a family of evolutionary algorithms ca-
pable of surmounting some of the limitations encountered by evolutionary computation. These
coevolutionary algorithms deal particularly well with increasing requirements of complexity
and modularity while keeping computational cost bounded.
In this chapter I present Fuzzy CoCo, an original approach which applies cooperative
coevolution to tackle the fuzzy-modeling problem. I begin by presenting some general notions
of coevolutionary computation in the next section. Then, in Section 3.2, I discuss cooperative
coevolution. In Section 3.3 I present in detail Fuzzy CoCo, my coevolutionary fuzzy modeling
approach. I finally illustrate in Section 3.4 some of the capabilities of Fuzzy CoCo by applying
it to a well-known classification problem.

3.1 Coevolutionary Computation

Inspired by natural coevolution, artificial coevolution refers to the simultaneous evolution of
two or more species with coupled fitness (i.e., the fitness of one individual depends on the
fitness of individuals of other species and/or on its interaction with them) [151]. Such coupled

51
52 3. Coevolutionary Fuzzy Modeling

evolution provides some advantages over non-coevolutionary approaches that render coevolu-
tion an interesting alternative when confronting certain problems. Among these advantages,
we can mention [77, 121, 140]:

• Coevolution favors the discovery of complex solutions whenever complex solutions are
required.

• Coevolution helps preserve genetic diversity.

• Coevolution favors computationally efficient fitness functions.

• Coevolution is suitable for parallel implementation.

Another multiple-population evolutionary approach is the island model [40,116] in which

isolated populations—the islands—evolve separately with occasional migration between them.
Such migration implies genetic exchange among individuals from different populations, thus
violating the interspecific genetic isolation required by strict coevolution. There exist, how-
ever, some island-model approaches in which individuals from different islands share their
fitness, becoming fitness-dependent on one another as in the coevolutionary paradigm, leading
to specialized subpopulations which are not genetically isolated. The fitness-sharing schema
might be either explicit or implicit [27].
Simplistically speaking, one can say that coevolving species can either compete (e.g.,
to obtain exclusivity on a limited resource) or cooperate (e.g., to gain access to some hard-
to-attain resource). Below I present the main concepts involved in competitive coevolution.
Cooperative coevolution is presented in more detail in Section 3.2 as it is the approach on
which I base my work.
In a competitive-coevolutionary algorithm, the fitness of an individual is based on direct
competition with individuals of other species, which in turn evolve separately in their own
populations. Increased fitness of one of the species implies a diminution in the fitness of the
other species. This evolutionary pressure tends to produce new strategies in the populations
involved so as to maintain their chances of survival. This “arms race” ideally increases the
capabilities of each species until they reach an optimum.
One of the competitive relationships that have inspired artificial coevolution is parasitism.
In this case, the term host refers to the individual whose fitness is under evaluation, and para-
sites refer to the individuals that are testing the host (trying to weaken it, not to kill it). This
approach has been used to evolve sorting networks [61] and neural classifiers [121]. In other
competitive approaches, two different species competing for the access to a limited resource
(usually complementary fitness value) develop survival strategies in order to perform better
than each other. This approach has been used to evolve artificial game players for Nim and
three-dimensional tic-tac-toe [151], and to model predator-prey interaction in evolutionary
robotics experiments [117].

3.2 Cooperative Coevolution

In nature, many species have developed cooperative interactions with other species to improve
their survival. Such cooperative—also called mutualistic or symbiotic—coevolution can be
3.2 Cooperative Coevolution 53

found in organisms going from cells (e.g., eukaryotic organisms resulted probably from the
mutualistic interaction between prokaryotes and some cells they infected) to superior animals
(e.g., African tick birds obtain a steady food supply by cleaning parasites from the skin of
giraffes, zebras, and other animals), including the common mutualism between plants and
animals (e.g., pollination and seed dispersion in change of food) [145].
Cooperative coevolutionary algorithms involve a number of independently evolving
species which together form complex structures, well-suited to solve a problem. The fitness
of an individual depends on its ability to collaborate with individuals from other species. In
this way, the evolutionary pressure stemming from the difficulty of the problem favors the
development of cooperative strategies and individuals. As in nature the species are genetically
isolated because they evolve in separate populations, because their genomes are genetically
incompatible, or both.

3.2.1 Issues addressed by cooperative coevolution

Single-population evolutionary algorithms often perform poorly—manifesting stagnation,
convergence to local optima, and computational costliness—when confronted with problems
presenting one or more of the following features:

1. The sought-after solution is complex.

2. The problem or its solution is clearly decomposable.

3. The genome encodes different types of values.

4. Strong interdependencies among the components of the solution.

5. Components-ordering drastically affects fitness.

Cooperative coevolution addresses effectively these issues, consequently widening the

range of applications of evolutionary computation [140, 143]. Indeed, the coevolution of sev-
eral species facilitates the existence of specialized individuals dealing with a subtask of a
complex problem. The decomposition of the problem can be either user-designed in the case
of clearly decomposable problems or emergent otherwise. As each species evolve separately,
one can tailor the evolutionary algorithm controlling evolution to the specificities of the sub-
task, including a representation adapted to the type of values involved in the corresponding
species. Finally, cooperative coevolution can be robust in face of the strong effects that in-
terdependencies and component-ordering might have on fitness. In fact, as each individual
has the possibility to cooperate several times, the chance that a good component disappears
due to a destructive genetic operation on another component is lower than in single-population
evolution.

3.2.2 Ancestral work

To my knowledge, cooperative coevolution was applied for the first time by Husbands and
Mill [68] to a highly generalized version of the manufacturing planning problem. This problem
involves finding an optimal way to interleave a number of concurrent process plans which share
54 3. Coevolutionary Fuzzy Modeling

resources. The optimization problem consists of simultaneously optimizing the individual

process plans and the overall schedule. In their approach, Husbands and Mill evolve one
species per process plan, together with an additional species of “Arbitrators” in charge of
making decisions to deal with conflicts. The final fitness of each individual is calculated in
two stages. The first stage involves local criteria and the second stage takes into account
interactions between populations.
Potter and De Jong [140,141,143] developed a general model for cooperative coevolution
in which a number of populations explore different decompositions of the problem. They ex-
plored their approach by applying it to different complex problems such as function optimiza-
tion [141], neural network design [142], and the learning of sequential decision rules [144].
Section 3.2.3 detail this framework as it forms the basis of my own approach.
Paredis [121] applied cooperative coevolution to problems which involved finding simul-
taneously the values of a solution and their adequate order. In his approach, a population of
solutions coevolves alongside a population of permutations performed on the genotypes of the
solutions. Paredis used as test problem the solution of n random linear equations with different
levels of coupling between the variables (i.e., with different difficulty levels).
Eriksson [36] applied cooperative coevolution to optimize the parameters controlling the
inventory in a factory. This problem involves finding a set of parameters defining policies to
control the stock of several items composing the inventory. Eriksson investigated two different
ways to define the coevolving species. In the first representation, each individual encodes the
values of one parameter for all the items, while in the second representation, each individual
encodes all the parameters of a single inventory item. His results show that the latter represen-
tation exhibits some advantages over the former as it expresses better the interdependencies
between the parameters.
Moriarty [109] used a cooperative coevolutionary approach to evolve neural networks.
Each individual in one species corresponds to a single hidden neuron of a neural network
and its connections with the input and output layers. This population coevolves alongside a
second one whose individuals encode sets of hidden neurons (i.e., individuals from the first
population) forming a neural network.

3.2.3 A general model for cooperative coevolution

As mentioned before, Potter and De Jong [140,143] developed a general model for cooperative
coevolution. Their hypothesis was that explicit notions of modularity are necessary in order to
evolve complex structures in the form of interacting coadapted subcomponents.
Their model has the following characteristics:
1. Each species represents a subcomponent of a potential solution.
2. Complete solutions are obtained by assembling representative members of each of the
species (populations).
3. The fitness of each individual depends on the quality of (some of) the complete solutions
it participated in, thus measuring how well it cooperates to solve the problem.
4. The evolution of each species is controlled by a separate, independent evolutionary al-
gorithm.
3.3 Fuzzy CoCo 55

5. Given an ensemble of conditions, the number of species should itself be adapted by a

mechanism of birth and death of species.

Figure 3.1 shows the general architecture of Potter’s cooperative coevolutionary frame-
work, and the way each evolutionary algorithm computes the fitness of its individuals by com-
bining them with selected representatives from the other species. The representatives can be
selected via a greedy strategy as the fittest individuals from the last generation.

Species 1 Species 2
Individual
EA to be E A Species 3
evaluated
E A Species 4

Population Population EA

Population
representatives representatives
Population
Merge
representatives

representatives

fitness fitness
evaluation

Figure 3.1 Potter’s cooperative coevolutionary system. The figure shows the evolutionary process
from the perspective of Species 1. The individual being evaluated is combined with one or more repre-
sentatives of the other species so as to construct several solutions which are tested on the problem. The
individual’s fitness depends on the quality of these solutions.

Results presented by Potter and De Jong [143] show that their approach addresses ad-
equately issues like problem decomposition and interdependencies between subcomponents.
The cooperative coevolutionary approach performs as good as, and sometimes better than,
single-population evolutionary algorithms. Finally, cooperative coevolution usually requires
less computation than single-population evolution as the populations involved are smaller, and
convergence—in terms of number of generations—is faster.

3.3 Fuzzy CoCo

Fuzzy CoCo is a Cooperative Coevolutionary approach to fuzzy modeling wherein two coe-
volving species are defined: database (membership functions) and rule base. This approach
is based primarily on the framework defined by Potter and De Jong [140, 143] (Section 3.2).
Fuzzy CoCo is conceived to allow a high degree of freedom in the type of fuzzy systems it
can design in order to allow the user to manage the trade-off between performance and inter-
pretability.
A fuzzy modeling process needs usually deal with the simultaneous search for opera-
tional and connective parameters (Table 2.1). These parameters provide an almost complete
56 3. Coevolutionary Fuzzy Modeling

definition of the linguistic knowledge describing the behavior of a fuzzy system, and of the
values mapping this symbolic description into a real-valued world (a complete definition also
requires structural parameters, such as relevant variables and number of rules). Thus, fuzzy
modeling can be thought of as two separate but intertwined search processes: (1) the search
for the membership functions (i.e., operational parameters) that define the fuzzy variables, and
(2) the search for the rules (i.e., connective parameters) used to perform the inference.
Fuzzy modeling presents several features discussed in Section 3.2 which justify the ap-
plication of cooperative coevolution: (1) The required solutions can be very complex, since
fuzzy systems with a few dozen variables may call for hundreds of parameters to be defined.
(2) The proposed solution—a fuzzy inference system—can be decomposed into two distinct
components: rules and membership functions. (3) Membership functions are represented by
continuous, real values, while rules are represented by discrete, symbolic values. (4) These two
components are interdependent because the membership functions defined by the first group
of values are indexed by the second group (rules).
Consequently, in Fuzzy CoCo, the fuzzy modeling problem is solved by two coevolving,
cooperating species. Individuals of the first species encode values which define completely all
the membership functions for all the variables of the system. For example, with respect to the
variable Triglycerides shown in Figure 2.1, this problem is equivalent to finding the values of
P1 , P2 , and P3 .
Individuals of the second species define a set of rules of the form:
if (v1 is A1 ) and . . . and (vn is An ) then (output is C),
where the term Ai indicates which one of the linguistic labels of fuzzy variable v is used by
the rule. For example, a valid rule could contain the expression
if . . . and (T emperature is W arm) and . . . then . . .
which includes the membership function W arm whose defining parameters are contained in
the first species.

3.3.1 The algorithm

The two evolutionary algorithms used to control the evolution of the two populations are in-
stances of a simple genetic algorithm [177]. Figure 3.2 presents the Fuzzy CoCo algorithm
in pseudo-code format. The genetic algorithms apply fitness-proportionate selection to choose
the mating pool (essentially, probabilistic selection according to fitness), and apply an elitist
strategy with an elitism rate Er to allow some of the best individuals to survive into the next
generation. The elitism strategy extracts ES individuals—the so-called elite—to be reinserted
into the population after evolutionary operators have been applied (i.e., selection, crossover,
and mutation). Note that the elite is not removed from the population, participating thus in the
reproduction process. Standard crossover and mutation operators are applied [104]: crossover
between two genomes is performed with probability Pc by selecting at random (with uniform
probability) a single crossover point and exchanging the subsequent parts to form two new off-
spring; if no crossover takes place (with probability 1 − Pc ) the two offspring are exact copies
of their parents. Mutation involves flipping bits in the genome with probability Pm per bit.
The condition under which the algorithm terminates is usually satisfied either when a given
threshold fitness is attained, or when the maximum number of generations, Gmax , is reached.
3.3 Fuzzy CoCo 57

begin Fuzzy CoCo

g:=0
for each species S
Initialize populations PS (0)
Evaluate population PS (0)
end for
while not done do
for each species S
g:=g+1
ES (g) = Elite-select[PS (g − 1)]
PS
(g) = Select[PS (g − 1)]
PS

(g) = Crossover[PS
(g)]
PS


(g) = Mutate[PS

(g)]
PS (g) = PS


(g) + ES (g)
Evaluate population PS (g)
end for
end while
end Fuzzy CoCo

Figure 3.2 Pseudo-code of Fuzzy CoCo. Two species coevolve in Fuzzy CoCo: membership functions
and rules. The elitism strategy extracts ES individuals to be reinserted into the population after evo-
lutionary operators have been applied (i.e., selection, crossover, and mutation). Selection results in a
reduced population PS
(g) (usually, the size of PS
(g) is PS
 = PS  − ES ). The line “Evaluate
population PS (g)” is elaborated in Figure 3.3.

3.3.2 Elitism

I introduced elitism to avoid the divergent behavior of Fuzzy CoCo, observed in preliminary
trial runs. Contrary to Juillé’s statement about premature convergence of cooperative coevolu-
tion [77]—due in his case to the fact that he was searching for solution-test pairs, two inher-
ently competitive species—non-elitist versions of Fuzzy CoCo often tend to lose the genetic
information of good individuals found during evolution, consequently producing populations
with mediocre individuals scattered throughout the search space. This is probably due to the
relatively small size of the populations which renders difficult the preservation (exploitation)
of good solutions while exploring the search space.
The introduction of simple elitism produces an undesirable effect on Fuzzy CoCo’s per-
formance: populations converge prematurely even with reduced values of the elitism rate Er .
To offset this effect without losing the advantages of elitism, it is necessary to increase the mu-
tation probability Pm by an order of magnitude so as to improve the exploration capabilities
of the algorithm. As the dispersion effect is less important when Fuzzy CoCo is allowed to
manage relatively large populations, the values of both Er and Pm should be reduced in such
case.
58 3. Coevolutionary Fuzzy Modeling

3.3.3 Fitness evaluation

A more detailed view of the fitness evaluation process is depicted in Figure 3.3. An individual
undergoing fitness evaluation establishes cooperations with one or more representatives of the
other species, i.e., it is combined with individuals from the other species to construct fuzzy
systems. The fitness value assigned to the individual depends on the performance of the fuzzy
systems it participated in (specifically, either the average or the maximal value).
Representatives, or cooperators, are selected both fitness-proportionally and randomly
from the last generation in which they were already assigned a fitness value (see Figure 3.2). In
Fuzzy CoCo, Ncf cooperators are selected probabilistically according to their fitness, favoring
the fittest individuals, thus boosting the exploitation of known good solutions. The other Ncr
cooperators are selected randomly from the population to represent the diversity of the species,
maintaining in this way exploration of the search space.

Fitness Cooperators
111
000 1111
0000
000
111 0000
1111
0000
1111
g−1 0000
1111

1111
0000
0000
1111
0000
1111 111
000
0000
1111 000
111
000
111
0000
1111
Selected Cooperators Fitness
g
cooperators
Species 1 Evaluation Species 2
Generation Species 1 Species 2 Environment
(a) (b)
Figure 3.3 Fitness evaluation in Fuzzy CoCo. (a) Several individuals from generation g − 1 of each
species are selected both randomly and according to their fitness to be the representatives of their
species during generation g; these representatives are called “cooperators.” (b) During the evalua-
tion stage of generation g (after selection, crossover, and mutation—see Figure 3.2), individuals are
combined with the selected cooperators of the other species to construct fuzzy systems. These systems
are then evaluated on the problem domain and serve as a basis for assigning the final fitness to the
individual being evaluated.

3.3.4 Interpretability considerations

As mentioned before, Fuzzy CoCo allows a high degree of freedom in the type of fuzzy sys-
tems it can design, letting the user determine the accuracy-interpretability trade-off. When
the interest is to preserve as much as possible the interpretability of the evolved systems, the
fuzzy model should satisfy the semantic and syntactic criteria presented in Section 2.3. The
mentioned strategies—label sharing, orthogonal membership functions, don’t-care conditions,
and default rule—must guide both the design of the fuzzy inference system and the defini-
tion of both species’ genomes. Note that the existence of one separated species representing
membership functions in Fuzzy CoCo, already implements the label-sharing strategy.
Besides, one or more of the linguistic criteria may participate in the fitness function as a
way to reinforce the selection pressure towards interpretable systems. The example presented
3.3 Fuzzy CoCo 59

in the next section illustrates both the application of Fuzzy CoCo and the way interpretability
criteria are introduced in it.

3.3.5 Other cooperative coevolutionary approaches to fuzzy modeling

Apart from Fuzzy CoCo, the use of cooperative coevolution to develop fuzzy models is scarce
and the few existing works limit themselves to a brief description and do not develop a struc-
tured approach nor study the characteristics of the proposed algorithms. Because of that, I can
can cautiously claim that Fuzzy CoCo is the first consistent work in cooperative coevolutionary
fuzzy modeling.
The earliest work reporting on coevolution for building fuzzy systems does not fit strictly
the conditions to be considered as coevolutionary. Indeed, in their approach, Włoszek and Do-
mański [182, 183], describe an island-based evolutionary system implementing in each island
a Michigan-type modeling process (Section 2.2.2). After each evolutionary step, they allow
the populations to exchange rules using an aggressive migration policy—up to 25% of the rule
base.
In 1997, Hopf [66] applied Potter’s coevolutionary model [140,141] to build the rule base
of a fuzzy system. In his approach, Hopf implements a species for each rule composing the
rule base. A candidate rule base is thus formed by an individual of each species. The pre-
defined membership functions are distributed regularly over each variable’s universe of dis-
course. This behavior-learning approach (Section 2.2.1) is applied to approximate an abstract
function obtaining results that compare favorably with those of a simple genetic algorithm.
Jun, Joung, and Sim [78] applied a method, similar to Fuzzy CoCo in that it implements
two species encoding, respectively, rules and membership functions. They applied different
evolutionary algorithms to control each species evolution: (1) a discrete alphabet representa-
tion with (µ + λ) selection and no crossover operator for the rule species, and (2) a simple
genetic algorithm with binary representation for the membership-function species. They ap-
plied their approach to design a fuzzy controller for the navigation of two simulated mobile
robots which “find their goal positions in relatively short time and avoid obstacles success-
fully”. The main difference between their approach and Fuzzy CoCo is the fitness evaluation
stage and the associated computational cost. While the evaluation of each individual in Fuzzy
CoCo is based on a reduced number of cooperating relations (i.e., Ncf + Ncr ); each individual
in their approach must cooperate with every individual in the other population to obtain its
fitness. The cost of this meticulous exploration is a huge computational load.
Two more recent works have applied cooperative coevolution to design fuzzy controllers.
Both Jeong and Oh [74] and Juang, Lin, and Lin [76] implement a species-per-rule strategy in
which each individual represents a single rule defined by its own input and output membership
functions. As usual in fuzzy-control applications, where the number of involved variables is
low, these two works neglect interpretability in favor of precision. Note that Juang, Lin, and
Lin select only random cooperators (i.e., Ncf = 0), while Jeong and Oh select only the fittest
individual as cooperators (i.e., Ncf = 1 and Ncr = 0).
60 3. Coevolutionary Fuzzy Modeling

3.4 Application Example: The Iris Problem

Applying Fuzzy CoCo requires the definition of parameters of its two main components: (1)
the fuzzy system and (2) the cooperative coevolutionary algorithm (Section 5.1 discusses in
more detail the application of Fuzzy CoCo). Some of the fuzzy-system parameters must be
pre-defined according to the available a-priori knowledge, and the remaining parameters are
considered targets for the evolutionary algorithm. As for the cooperative-coevolution param-
eters, the designer needs to define the maximum number of generations, the population size,
the number of cooperators, and the crossover, mutation, and elitism rates. In the fuzzy systems
evolved in the present chapter, I applied Fuzzy CoCo with the same evolutionary parameters
for both species. To illustrate the application of Fuzzy CoCo, I present in the following sub-
sections the evolution of two different fuzzy systems to solve a hard classification problem.

3.4.1 Fisher’s Iris data

Fisher’s Iris data is a well-known classification problem consisting of feature measurements
for the speciation of iris flowers [13,41]. There are three classes corresponding to three species
of iris: setosa, versicolor, and virginica, where each flower can be identified according to four
continuous attributes measured in centimeters: (1) sepal length (SL); (2) sepal width (SW ),
(3) petal length (P L), and (4) petal width (P W ). The 150 database entries include 50 sample
cases for each of the three species. The goal is to classify irises into one of the three classes in
accordance with the four inputs. To illustrate the distribution of the iris data, Figure 3.4 depicts
all the 150 entries in the bidimensional subspace defined by variables P L and P W .

2.5

1.5
PW

0.5

0
1 2 3 4 5 6 7
PL

Figure 3.4 Iris data distribution in the P L–P W subspace. The three classes: setosa, versicolor, and
virginica correspond to marks x, o, and +, respectively. Note that in both axes, almost all the versicolor
cases are in between setosa and virginica cases. This fact can be exploited to facilitate the design of
classification systems.
3.4 Application Example: The Iris Problem 61

Fisher’s iris data has been widely used to test classification and modeling algorithms,
recently including fuzzy models [65, 67, 152, 164, 184]. I propose herein two types of fuzzy
logic-based systems to solve the iris data classification problem: (1) fuzzy controller-type (as
used by Shi et al. [164] and Russo [152]), and (2) fuzzy classifier-type (as used by Hong
and Chen [65], Hung and Lin [67], and Wu and Chen [184]). Both types consist of a fuzzy
inference system and a selection unit.
In the fuzzy controller (see Figure 3.5) the fuzzy subsystem computes a single continuous
value estimating the class to which the input vector belongs. Note that each class is assigned
a numeric value; based on the iris data distribution, I assigned values 1, 2, and 3 to the classes
setosa, versicolor, and virginica, respectively (such an assignment makes sense only under
the assumption that versicolor is an intermediate species in between setosa and virginica, see
Figure 3.4). The selection unit approximates this value to the nearest class value using a stair
function.

Fuzzy Subsystem Stair−function

Input Class Subsystem Class

vector estimation

Figure 3.5 Fuzzy controller used to solve Fisher’s iris data problem. The fuzzy subsystem is used to
compute a continuous value that describes the class to which a given input vector belongs (the three
classes: setosa, versicolor, and virginica correspond to values 1, 2, and 3, respectively). The stair-
function approximates the computed value to the nearest class value.

In the fuzzy classifier (see Figure 3.6) the fuzzy inference subsystem computes a con-
tinuous membership value for each class. The selection unit chooses the most active class,
provided that its membership value exceeds a given threshold (which I set to 0.5).

Fuzzy Subsystem µ (setosa) Maximum and

Input Threshold Subsystem Class
µ(versicolor)
vector µ(virginica)

Figure 3.6 Fuzzy classifier used to solve Fisher’s iris data problem. The fuzzy subsystem is used to
compute the membership value of a given input vector for each of the three classes: setosa, versicolor,
and virginica. The maximum-and-threshold subsystem chooses the class with the maximum membership
value, provided this value exceeds a given threshold.

The two fuzzy subsystems thus differ in the number of output variables: a single out-
put (with values {1,2,3}) for the controller-type and three outputs (with values {0,1}) for the
classifier-type. In general, controller-type systems take advantage of data distribution while
classifier-type systems offer higher interpretability because the output classes are independent;
these latter systems are harder to design.
62 3. Coevolutionary Fuzzy Modeling

3.4.2 Application of Fuzzy CoCo to the Iris problem

I used prior knowledge about the iris data problem and about some of the fuzzy systems pro-
posed in the literature for its solution to guide my choice of fuzzy parameters. In addition, I
took into account the interpretability criteria presented in Section 2.3 to define constraints on
the fuzzy parameters. Referring to Table 2.1, I delineate below the fuzzy system’s set-up:

• Logical parameters: singleton-type fuzzy systems; min-max fuzzy operators; orthog-

onal, trapezoidal-triangular input membership functions (see Figure 3.7); weighted-
average defuzzification.

• Structural parameters: three input membership functions (Low, Medium, and High);
three output singletons for the controller-type system and two output singletons for the
classifier-type system; a user-configurable number of rules. The relevant variables are
one of Fuzzy CoCo’s evolutionary objectives.

• Connective parameters: the antecedents and the consequent of the rules are searched by
Fuzzy CoCo. The algorithm also searches for the consequent of the default rule. All
rules have unitary weight.

• Operational parameters: the input membership function values are to be found by Fuzzy
CoCo. For the output singletons I used the values 1, 2, and 3 for the controller-type
system, and the values 0 and 1 for the classifier-type system.

Fuzzy CoCo thus searches for four parameters: input membership-function values, rele-
vant input variables, and antecedents and consequents of rules. The genomes of the two species
are constructed as follows:

• Species 1: Membership functions. There are four variables (SL, SW , P L, and P W ),

each with three parameters P1 , P2 , and P3 , defining the membership-function edges
(Figure 3.7).

µ
Low Medium High
1

0
P1 P2 P3

Figure 3.7 Input fuzzy variables for the iris problem. Each fuzzy variable has three possible fuzzy
values labeled Low, Medium, and High, and orthogonal membership functions, plotted above as
degree of membership versus input value. Parameters P1 , P2 , and P3 define the membership-function
edges.

• Species 2: Rules (Controller-type systems). The i-th rule has the form:
if (SL is AiSL ) and . . . and (P W is AiP W ) then (output is C i ),
3.4 Application Example: The Iris Problem 63

Aij can take on the values: 1 (Low), 2 (Medium), 3 (High), or 0 (Don’t Care). C i can take
on the values: 1 (setosa), 2 (versicolor), or 3 (virginica). Relevant variables are searched
for implicitly by allowing the algorithm to choose the Don’t Care label (i.e., Aij = 0)
as valid antecedents; in such case the respective variable is considered irrelevant, and is
removed from the rule. The default rule is defined by its consequent parameter C 0 .

• Species 2: Rules (Classifier-type systems). The i-th rule has the form:
if (SL is AiSL ) and . . . and (P W is AiP W )
i i i
then {(setosa is Cset ), (versicolor is Cver ), (virginica is Cvir )},
Aij can take on the values: 1 (Low), 2 (Medium), 3 (High), or 0 (Other). Cji can take
on the values: 0 (No), or 1 (Yes). Relevant variables are searched for implicitly by
allowing the algorithm to choose non-existent membership functions (i.e., Aij = 0) as
valid antecedents; in such case the respective variable is considered irrelevant, and is
0
removed from the rule. The default rule is defined by its consequent parameters Cset ,
0 0
Cver , and Cvir .

Table 3.1 delineates the parameters encoding both species’ genomes, which together de-
scribe an entire fuzzy system.

Table 3.1 Genome encoding of parameters for both species. Genome length for Species 1 (membership
functions) is 60 bits. Genome length for Species 2 (rules) is 10 × Nr + 2 for the controller-type system
and 11 × Nr + 3 for the classifier-type system, where Nr denotes the number of rules. Values Vmin and
Vmax for parameters Pi are defined according to the ranges of the four variables SL, SW , P L, and
PW.

Species 1: Membership functions

Parameter Values Bits Qty Total bits
Pi [Vmin − Vmax ] 5 3×4 60

Species 2: Rules (Controller-type)

Parameter Values Bits Qty Total bits
Aij {0,1,2,3} 2 4 × Nr 8 × Nr
Ci {1,2,3} 2 Nr + 1 2 × (Nr + 1)
Total Genome Length 10 × Nr + 2

Species 2: Rules (Classifier-type)

Parameter Values Bits Qty Total bits
Aij {0,1,2,3} 2 4 × Nr 8 × Nr
Cji {0,1} 1 3 × (Nr + 1) 3 × (Nr + 1)
Total Genome Length 11 × Nr + 3

Table 3.2 delineates values and ranges of values of the evolutionary parameters. The
algorithm terminates when the maximum number of generations, Gmax , is reached (I set
Gmax = 500 + 100 × Nr , i.e., dependent on the number of rules used in the run), or when the
64 3. Coevolutionary Fuzzy Modeling

increase in fitness of the best individual over five successive generations falls below a certain
threshold (10−4 in my experiments). Note that mutation rates are relatively higher than with
a simple genetic algorithm, which is typical of coevolutionary algorithms [140, 143]. This is
due in part to the small population sizes and to elitism.

Table 3.2 Fuzzy CoCo set-up. Population size was fixed to 60 for controller-type systems and to 70 for
classifier-type systems.

Parameter Values
Population size PS  {60,70}
Maximum generations Gmax 500 + 100 × Nr
Crossover probability Pc 1
Mutation probability Pm {0.02,0.05,0.1}
Elitism rate Er {0.1,0.2}
“Fit” cooperators Ncf 1
Random cooperators Ncr {1,2}

My fitness function combines three criteria: (1) Fc : classification performance, computed

as the percentage of cases correctly classified; (2) Fmse : a value dependent on the mean square
error (mse), measured between the continuous values of the outputs and the correct classifica-
tion given by the iris data set (Fmse = 1 − mse); and (3) Fv : a rule-length dependent fitness
equal 0 when the average number of variables per active rule is maximal and equal 1 in the
hypothetical case of zero-length rules. The fitness function, defined as follows, combines these
three measures:

Fc × Fmse
β
if Fc < 1
F =
(Fc − αFv ) × Fv if Fc = 1,
β

where α = 1/150 and β = 0.3. Fc , the ratio of correctly classified samples, is the most
important measure of performance. Fmse adds selection pressure towards systems with low
quadratic error, in which misclassifications are closer to “crossing the line” and becoming cor-
rect classifications. Fv measures the interpretability, penalizing systems with a large number
of variables per rule (on average). Fv penalization is only applied to perfect classifiers as the
number of variables in the iris data problem is low and 100% classification rate can be attained.
The value β was set small enough to penalize systems exhibiting a large quadratic error. The
value α was calculated to allow Fv to penalize rule multiplicity, but without decreasing fitness
to a value below lower-performance systems.

3.4.3 Results
In this section I present the fuzzy systems evolved using Fuzzy CoCo for the two setups de-
scribed above. I compare my systems with those presented in recently published articles,
which thus represent the state-of-the-art. I detail some high-performance systems obtained for
each problem in order to illustrate the type of systems found by Fuzzy CoCo.
3.4 Application Example: The Iris Problem 65

3.4.3.1 Controller-type systems

I performed a total of 145 evolutionary runs, searching for controller-type systems with 2,
3, and 4 rules, all of which found systems whose classification performance exceeds 97.33%
(i.e., the worst system misclassifies only 4 cases). The average classification performance
of these runs was 98.98%, corresponding to 1.5 misclassifications. 121 runs led to a fuzzy
system misclassifying 2 or less cases, and of these, 4 runs found perfect classifiers. Figure 3.8
summarizes my results.

90
Accumulated number of evolved systems

2 rules
80 3 rules
4 rules
70

60

50

40

30

20

10

0
0 1 2 3 4
Number of misclassifications

Figure 3.8 Summary of results of 145 Fuzzy CoCo runs searching for controller-type systems. The
histogram depicts the total number of systems for a given number of misclassifications, at the end of the
Fuzzy CoCo run.

Table 3.3 compares my best controller-type systems with the top systems obtained by two
other evolutionary fuzzy modeling approaches. Shi et al. [164] used a simple genetic algorithm
with adaptive crossover and adaptive mutation operators. Russo’s FuGeNeSys method [152]
combines evolutionary algorithms and neural networks to produce fuzzy systems. The search
power of this latter method lies mainly in artificial evolution, while neural-based learning tech-
niques are applied only to improve the performance of promising (high-performance) systems.
The main drawback of these two methods is the low interpretability of the generated systems.
As they do not define constraints on the input membership-function shapes, almost none of the
semantic criteria presented in Section 2.3 are respected.
As evident in Table 3.3, the evolved fuzzy systems described in this section surpass or
equal those obtained by the two other approaches in terms of performance, while maintaining
high interpretability. Thus, my approach not only produces systems exhibiting high perfor-
mance, but also ones with less rules and less antecedents per rule (which systems are thus
more interpretable).
Fuzzy CoCo found controller-type systems with 3 and 4 rules exhibiting perfect perfor-
mance (no misclassifications). Among these, I consider as best the system with fewest rules
and variables. Figure 3.9 presents one such three-rule system, with an average of 1.7 variables
per rule.
66 3. Coevolutionary Fuzzy Modeling

Table 3.3 Comparison of the best controller-type systems evolved by Fuzzy CoCo with the top fuzzy sys-
tems obtained by Russo’s FuGeNeSys method [152] and with those obtained using a single-population
evolutionary approach by Shi et al. [164]. Shown below are the classification rates of the top systems
obtained by these approaches, along with the average number of variables per rule in parentheses.
Results are divided into four classes, in accordance with the number of rules per system, going from
two-rule systems to five-rule ones. The highlighted system is the top-performance one, detailed in
Figure 3.9. A dash implies that the relevant system was not designed.
Rules per Shi et al. [164] FuGeNeSys [152] Fuzzy CoCo
system best best average best
2 – – 98.71% (1.9) 99.33% (2)
3 – – 99.10% (1.3) 100% (1.7)
4 98.00% (2.6) – 99.12% (1.3) 100% (2.5)
5 – 100% (3.3) – –

Database
SL SW PL PW
P1 5.68 3.16 1.19 1.55
P2 6.45 3.16 1.77 1.65
P3 7.10 3.45 6.03 1.74
Rule base
Rule 1 if (P L is High) then (output is virginica)
Rule 2 if (SW is Low) and (P W is Low) then (output is virginica)
Rule 3 if (SL is Medium) and (P W is Medium) then (output is setosa)
Default else (output is setosa)
Figure 3.9 The best evolved, controller-type system with three rules. It exhibits a classification rate
of 100%, and an average of 1.7 variables per rule. Note that there is no rule classifying explicitly
for versicolor class. The system interprets as versicolor cases where rules classifying both setosa and
virginica classes have similar activation levels. For example, if the activation level is 0.5 for both
classes 1 (setosa) and 3 (virginica), then the defuzzifier will output 2 (versicolor).

3.4.3.2 Classifier-type systems

I performed a total of 144 evolutionary runs, searching for controller-type systems with 2,
3, and 4 rules, all of which found systems whose classification performance exceeds 95.33%
(i.e., the worst system misclassifies 7 cases). The average classification performance of these
runs was 97.40%, corresponding to 3.9 misclassifications. 104 runs led to a fuzzy system mis-
classifying 5 or less cases, and of these, 13 runs found systems with a single misclassification.
Figure 3.10 summarizes these results.
Table 3.4 compares my best classifier-type systems [128] with the top systems obtained
by three other fuzzy modeling approaches. Hong and Chen [65] and Wu and Chen [184] pro-
posed constructive learning methods (Section 2.1.2.3) to progressively construct their fuzzy
systems. The stages of their methods are very similar: find relevant attributes, build initial
membership functions, simplify the attributes and the membership functions, and derive deci-
sion rules; the two works differ in the strategies implemented to perform each stage. Thanks
to attribute-selection and simplification stages, these two approaches are able to find systems
3.4 Application Example: The Iris Problem 67

35

Accumulated number of evolved systems

2 rules
3 rules
30
4 rules

25

20

15

10

0
1 2 3 4 5 6 7
Number of misclassifications

Figure 3.10 Summary of results of 144 Fuzzy CoCo runs searching for classifier-type systems. The
histogram depicts the total number of systems for a given number of misclassifications, at the end of the
Fuzzy CoCo run.

with either a few [184] or simple rules [65]. They do not, however, constrain the input mem-
bership functions, thus rendering the obtained systems less interpretable. Hung and Lin [67]
proposed a neuro-fuzzy hybrid approach to learn classifier-type systems. As their learning
strategy hinges mainly on the adaptation of the connection weights, their systems exhibit low
interpretability.
The evolved fuzzy systems described herein surpass or equal those obtained by these
three approaches in terms of both performance and interpretability. As evident in Table 3.4,
my approach not only produces systems exhibiting higher performance, but also ones with less
rules and less antecedents per rule (which are thus more interpretable).

Table 3.4 Comparison of the best classifier-type systems evolved by Fuzzy CoCo with systems obtained
applying constructive learning methods proposed by Hong and Chen [65] and by Wu and Chen [184],
and with those obtained by Hung and Lin’s neuro-fuzzy approach [67]. Shown below are the clas-
sification rates of the top systems obtained by these approaches, along with the average number of
variables per rule in parentheses. Results are divided into four classes, in accordance with the number
of rules per system, going from two-rule systems to eight-rule ones. The highlighted system is the top-
performance one, detailed in Figure 3.11. A dash implies that the relevant system was not designed.

Rules per Hong and Wu and Hung and Fuzzy CoCo

system Chen [65] Chen [184] Lin [67]
best average average average best
2 – – – 96.47% (2.1) 98.00% (1.5)
3 – 96.21% (4) – 97.51% (2.4) 99.33% (2.3)
4 – – 97.40% (4) 98.21% (2.3) 99.33% (2)
8 97.33% (2) – – – –

Fuzzy CoCo found classifier-type systems with 3 and 4 rules exhibiting the highest clas-
sification performance to date (i.e., 99.33%, corresponding to 1 misclassification). I consider
68 3. Coevolutionary Fuzzy Modeling

as most interesting the system with the smallest number of conditions (i.e., the total number of
variables in the rules). Figure 3.11 presents one such system. This three-rule system presents
an average of 2.3 variables per rule, corresponding to a total of 7 conditions.

Database
SL SW PL PW
P1 4.65 2.68 4.68 0.39
P2 4.65 3.74 5.26 1.16
P3 5.81 4.61 6.03 2.03
Rule base
Rule 1 if (P W is Low)
then {(setosa is Yes), (versicolor is No), (virginica is No) }
Rule 2 if (P L is Low) and (P W is Medium)
then{(setosa is No), (versicolor is Yes), (virginica is No)}
Rule 3 if (SL is High) and (SW is Medium) and (P L is Low) and (P W is High)
then{(setosa is No), (versicolor is Yes), (virginica is No)}
Default else{(setosa is No), (versicolor is No), (virginica is Yes)}
Figure 3.11 The best evolved, classifier-type system with three rules. It exhibits a classification rate
of 99.33%, and an average of 2.3 variables per rule.

3.5 Summary
In this chapter I presented a novel approach to fuzzy modeling based on cooperative coevo-
lution. The method, called Fuzzy CoCo, involves the existence of two separated coevolving
species: rules and membership functions. A greedy fitness evaluation strategy, based on the
use of selected cooperators from each species, allows the method to balance adequately the
exploration and the exploitation of the search while keeping bounded the computational cost.
I propose Fuzzy CoCo as a methodology for modeling fuzzy systems and have conceived
it to allow a high degree of freedom in the type of fuzzy systems it can design. Fuzzy CoCo can
be used to model Mamdani-type, TSK-type, and singleton-type fuzzy models2.1. The rules
can contain an arbitrary number of antecedents (i.e., zero, one, or many) for the same variable.
The designer is free to choose the type of membership functions used for each variable and the
way they are parametrized. The membership functions can be defined either as shared by all
fuzzy rules, or per-rule. Fuzzy CoCo is thus highly general and generic.
The configurability of Fuzzy CoCo facilitates the management of the interpretability-
accuracy trade-off. To satisfy interpretability criteria, the user must impose conditions on
the input and output membership functions as well as on the rule definition. In Fuzzy CoCo
these conditions are translated both into restrictions on the choice of fuzzy parameters and into
criteria included in the fitness function.
I also illustrated the use of Fuzzy CoCo by applying it to a well-known classification
problem: the Iris problem. Two types of fuzzy systems—controller-type and classifier-type—
were successfully modelled to solve the problem. Further applications of Fuzzy CoCo to hard
medical diagnosis problems are presented in the next chapter.
Chapter 4

Breast Cancer Diagnosis by Fuzzy CoCo

Although computerized tools for medical diagnosis have been developed since the early 60s,
their number and capabilities have grown impressively in the last years due, mainly, to the
availability of medical data and increased computing power. Most of these systems are con-
ceived to provide high diagnostic performance. However, interest has recently shifted to sys-
tems capable of providing, besides a correct diagnosis, insight on how the answer was ob-
tained. Thanks to their linguistic representation and their numeric behavior, fuzzy systems can
provide both performance and explanataion.
In this chapter, I apply Fuzzy CoCo to model the decision processes involved in two
breast-cancer diagnostic problems. First, I describe in Section 4.1, the application of Fuzzy
CoCo to the Wisconsin Breast Cancer Diagnostic problem (already presented in Section 2.4).
Then, in Section 4.2, I present the application of Fuzzy CoCo to the design of COBRA, a
fuzzy-based tool to assess mammography interpretation.

4.1 Breast-biopsy Analysis: The WBCD Problem

The Wisconsin Breast Cancer Diagnosis (WBCD) problem, already introduced in Section 2.4,
involves classifying a presented case of putative cancer as to whether it is benign or malignant.
It admits a relatively high number of variables and consequently a large search space. Recall
that the WBCD database consists of 683 cases of breast biopsy evaluations. Each case has
nine input attributes and a binary diagnostic output. The solution I propose for this problem
consists of a fuzzy appraisal system and a threshold unit (see Figure 2.5). The goal is to evolve
a fuzzy model that describes the diagnostic decision, while exhibiting both good classifica-
tion performance and high interpretability. A detailed description of the WBCD problem is
presented in Section 2.4.1.

4.1.1 The evolutionary setup

The system proposed to solve the WBCD problem, consisting of a fuzzy system and a thresh-
old unit, is presented in Section 2.4.2. While Fuzzy CoCo might use the same fuzzy-system
parameters defined for the fuzzy-genetic approach, the evolutionary setup must be redefined
within the Fuzzy CoCo framework.

69
70 4. Breast Cancer Diagnosis by Fuzzy CoCo

Fuzzy CoCo is used to search for four parameters: input membership-function values,
relevant input variables, and antecedents and consequents of rules. These search goals are more
ambitious than those defined for the fuzzy-genetic approach (Section 2.4.2) as the consequents
of rules are added to the search space. The genomes of the two species are constructed as
follows:

• Species 1: Membership functions. There are nine variables (v1 – v9 ), each with two
parameters, P and d, defining the start point and the length of the membership-function
edges, respectively (Figure 2.6).

• Species 2: Rules. The i-th rule has the form:

if (v1 is Ai1 ) and . . . and (v9 is Ai9 ) then (output is C i ),

Aij can take on the values: 1 (Low), 2 (High), or 0 or 3 (Don’t Care). C i bit can take on
the values: 0 (Benign) or 1 (Malignant). Relevant variables are searched for implicitly
by letting the algorithm choose Don’t care labels as valid antecedents; in such a case the
respective variable is considered irrelevant.

Table 4.1 delineates the parameter encoding for both species’ genomes, which together
describe an entire fuzzy system. Note that in the fuzzy-genetic approach (Section 2.4.2) both
membership functions and rules were encoded in the same genome, i.e., there was only one
species.

Table 4.1 Genome encoding of parameters for both species. Genome length for membership functions
is 54 bits. Genome length for rules is 19 × Nr + 1, where Nr denotes the number of rules.

Species 1: Membership functions

Parameter Values Bits Qty Total bits
P {1,2,. . . ,8} 3 9 27
d {1,2,. . . ,8} 3 9 27
Total Genome Length 54

Species 2: Rules
Parameter Values Bits Qty Total bits
A {0,1,2,3} 2 9 × Nr 18 × Nr
C {1,2} 1 Nr + 1 Nr + 1
Total Genome Length 19 × Nr + 1

To evolve the fuzzy inference system, I applied Fuzzy CoCo with the same evolutionary
parameters for both species. Table 4.2 delineates the values and ranges of values used for these
parameters. The algorithm terminates when the maximum number of generations, Gmax , is
reached (I set Gmax = 1000 + 100 × Nr , i.e., dependent on the number of rules used in the
run), or when the increase in fitness of the best individual over five successive generations falls
below a certain threshold (10−4 in my experiments).
4.1 Breast-biopsy Analysis: The WBCD Problem 71

Table 4.2 Fuzzy CoCo set-up for the WBCD problem.

Parameter Values
Population size P s [30-90]
Maximum generations Gmax 1000 + 100Nr
Crossover probability Pc 1
Mutation probability Pm [0.02-0.3]
Elitism rate Er [0.1-0.6]
“Fit” co-operators Ncf 1
Random co-operators Ncr {1,2,3,4}

The fitness function combines two criteria: 1) Fc —classification performance, computed

as the percentage of cases correctly classified, and 2) Fv —the maximum number of variables
in the longest rule. The fitness function is given by F = Fc − αFv , where α = 0.0015. Fc , the
percentage of correctly diagnosed cases, is the most important measure of performance. Fv
measures the interpretability, penalizing systems with a large number of variables in their rules.
The value α was calculated to allow Fv to occasion a fitness difference only among systems
exhibiting similar classification performance. The fitness value assigned to an individual is
the maximum of the fitness values obtained by the Nc fuzzy systems it participated in (where
Nc = Ncf + Ncr ).
I stated earlier that cooperative coevolution reduces the computational cost of the search
process. In order to measure this cost I calculated the maximum number of fuzzy-system
evaluations performed by a single run of Fuzzy CoCo. Each generation, the P s individuals
of each population are evaluated Nc times (where Nc = Ncf + Ncr ). The total number of
fuzzy-system evaluations per run is thus 2 × Gmax × P s × Nc . This value ranged from
5.28 × 105 evaluations for a one-rule system search, up to 8.16 × 105 evaluations for a seven-
rule system (using typical parameter values: P s = 80, Ncf = 1, and Ncr = 2). The number
of fuzzy-system evaluations required by the single-population approach was, on the average,
5 × 105 for a one-rule system and 11 × 105 for a seven-rule system 2.4. This shows that Fuzzy
CoCo produces markedly better results using similar computational resources.

4.1.2 Results
A total of 495 evolutionary runs were performed, all of which found systems whose classifi-
cation performance exceeds 96.7%. In particular, considering the best individual per run (i.e.,
the evolved system with the highest classification success rate), 241 runs led to a fuzzy system
whose performance exceeds 98.0%, and of these, 81 runs found systems whose performance
exceeds 98.5%.; these results are summarized in Figure 4.1.
Table 4.3 compares the best systems found by Fuzzy CoCo with the top systems obtained
by the fuzzy-genetic approach (Section 2.4) [125, 130] and with the systems obtained by Se-
tiono’s NeuroRule approach [161] (note that the results presented by these two works were
the best reported to date for genetic-fuzzy and neuro-Boolean rule systems, respectively, and
that they were compared with other previous approaches such as [160, 162, 169]). The evolved
72 4. Breast Cancer Diagnosis by Fuzzy CoCo

160

120
Number of systems

80

40

0
94 95 96 97 98 99
Classification performance

Figure 4.1 Summary of results of 495 evolutionary runs. The histogram depicts the number of systems
exhibiting a given performance level at the end of the evolutionary run. The performance considered is
that of the best individual of the run, measured as the overall percentage of correctly classified cases
over the entire database.

fuzzy systems described herein can be seen to surpass those obtained by other approaches in
terms of performance, while still containing simple, interpretable rules. As shown in Table 4.3,
I obtained higher-performance systems for all rule-base sizes but one, i.e., from two-rule sys-
tems to seven-rule ones, while all my one-rule systems perform as well as the best system
reported by Setiono.

Table 4.3 Comparison of the best systems evolved by Fuzzy CoCo with the top systems obtained using
single-population evolution [125] and with those obtained by Setiono’s NeuroRule approach [161].
Shown below are the classification performance values of the top systems obtained by these approaches,
along with the number of variables of the longest rule in parentheses. Results are divided into seven
classes, in accordance with the number of rules per system, going from one-rule systems to seven-rule
ones.
Rules per Neuro-Rule [161] Single population Fuzzy CoCo
system GA [125]a
best best average best
1 97.36% (4) 97.07% (4) 97.36% (4) 97.36% (4)
2 – 97.36% (4) 97.73% (3.9) 98.54% (5)
3 98.10% (4) 97.80% (6) 97.91% (4.4) 98.54% (4)
4 – 97.80% (-) 98.12% (4.2) 98.68% (5)
5 98.24% (5) 97.51% (-) 98.18% (4.6) 98.83% (5)
6 – 98.10% (9) 98.18% (4.3) 98.83% (5)
7 – 97.95% (8) 98.25% (4.7) 98.98% (5)
a
Data extracted from Tables 2.4 and 2.5
4.1 Breast-biopsy Analysis: The WBCD Problem 73

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 2 1 1 1 6 1 3 5 2
d 7 8 4 8 1 4 8 4 1
Rule base
Rule 1 if (v1 is Low) and (v3 is Low) then (output is benign)
Rule 2 if (v4 is Low) and (v6 is Low) and (v8 is Low) and (v9 is Low) then
(output is benign)
Rule 3 if (v1 is Low) and (v3 is High) and (v5 is High) and (v8 is Low) and
(v9 is Low) then (output is benign)
Rule 4 if (v1 is Low) and (v2 is High) and (v4 is Low) and (v5 is Low) and
(v8 is High) then (output is benign)
Rule 5 if (v2 is High) and (v4 is High) then (output is malignant)
Rule 6 if (v1 is High) and (v3 is High) and (v6 is High) and (v7 is High) then
(output is malignant)
Rule 7 if (v2 is High) and (v3 is High) and (v4 is Low) and (v5 is Low) and
(v7 is High) then (output is malignant)
Default else (output is malignant)

Figure 4.2 The best evolved, fuzzy diagnostic system with seven rules. It exhibits an overall classifica-
tion rate of 98.98%, and its longest rule includes 5 variables.

I next describe two of my top-performance systems, which serve to exemplify the solu-
tions found by Fuzzy CoCo. The first system, delineated in Figure 4.2, presents the highest
classification performance evolved to date. It consists of seven rules with the longest rule
including 5 variables. This system obtains an overall classification rate (i.e., over the entire
database) of 98.98%.

In addition to the above seven-rule system, evolution found systems with between 2 and
6 rules exhibiting excellent classification performance, i.e., higher than 98.5% (Table 4.3).
Among these systems, I consider as the most interesting the system with the smallest number
of conditions (i.e., total number of variables in the rules). Figure 4.3 presents one such two-
rule system, containing a total of 8 conditions, and which obtains an overall classification rate
of 98.54%; its longest rule has 5 variables.

The improvement attained by Fuzzy CoCo, while seemingly slight (0.5-1%) is in fact
quite significant. A 1% improvement implies 7 additional cases which are classified correctly.
At the performance rates in question (above 98%) every additional case is hard-won. Indeed,
try as I did with the fuzzy-genetic approach—tuning parameters and tweaking the setup—I ar-
rived at a performance impasse. Fuzzy CoCo, however, readily churned out better-performance
systems, which were able to classify a significant number of additional cases; moreover, these
systems were evolved in less time.
74 4. Breast Cancer Diagnosis by Fuzzy CoCo

Database
v1 v2 v3 v4 v5 v6 v7 v8 v9
P 3 1 3 4 5 7 2
d 8 3 1 2 2 4 1
Rule base
Rule 1 if (v1 is Low) and (v3 is Low) and (v5 is Low) then (output is benign)
Rule 2 if (v1 is Low) and (v4 is Low) and (v6 is Low) and (v8 is Low) and (v9
is Low) then (output is benign)
Default else (output is malignant)
Figure 4.3 The best evolved, fuzzy diagnostic system with two rules. It exhibits an overall classification
rate of 98.54%, and a maximum of 5 variables in the longest rule.

4.2 Mammography Interpretation: The COBRA system

This section presents the design of a tool based on Fuzzy CoCo denominated “COBRA: Cat-
alonia online breast-cancer risk assessor.” COBRA is designed to aid radiologists in the inter-
pretation of mammography to decide whether to perform a biopsy on a patient or not. Mam-
mography remains the principal technique for detecting breast cancer. Despite its undoubtable
value in reducing mortality, mammography’s positive predictive value (PPV) is low: only be-
tween 15 and 35% of mammographic-detected lesions are cancerous [120,139]. The remaining
65 to 85% of biopsies, besides being costly and time-consuming, cause understandable stress
on women facing the shadow of cancer. A computer-based tool that assists radiologists during
mammographic interpretation would contribute to increasing the PPV of biopsy recommenda-
tions.

4.2.1 The Catalonia mammography database

The Catalonia mammography database, which is the object of this section, was collected at
the Duran y Reynals hospital in Barcelona. It consists of 15 input attributes and a diagnos-
tic result indicating whether or not a carcinoma was detected after a biopsy. The 15 input
attributes include three clinical characteristics (Table 4.4) and two groups of six radiologic
features, according to the type of lesion found in the mammography: mass or microcalcifica-
tions (Table 4.5).
A radiologist fills out a reading form for each mammography, assigning values for the
clinical characteristics and for one of the groups of radiologic features. Then, the radiologist
interprets the case using a five-point scale: (1) benign; (2) probably benign; (3) indeterminate;
(4) probably malignant; (5) malignant. According to this interpretation a decision is made on
whether to practice a biopsy on the patient or not. The Catalonia database contains data cor-
responding to 227 cases, all of them sufficiently suspect to justify a biopsy recommendation.
For the purpose of this study, each case was further examined by three different readers—for a
total of 681 readings—but only diverging readings were kept. The actual number of readings
in the database is 516, among which 187 are positive (malignant) cases and 329 are negative
(benign) cases.
4.2 Mammography Interpretation: The COBRA system 75

Table 4.4 Variables corresponding to a patient’s clinical data.

v1 Age [28-82] years
v2 Menstrual history 1 Premenopausal
2 Postmenopausal
v3 Family history 1 None
2 Second familiar
3 First familiar
4 Contralateral
5 Homolateral

Table 4.5 Variables corresponding to radiologic features. There are two groups of variables that
describe the mammographic existence of microcalcifications and masses.
Microcalcifications Mass
v4 Disposition v10 Morphology
1 Round 1 Oval
2 Indefinite 2 Round
3 Triangular or Trapezoidal 3 Lobulated
4 Linear or Ramified 4 Polilobulated
v5 Other signs of group form 5 Irregular
1 None v11 Margins
2 Major axis in direction of nipple 1 Well delimited
3 Undulating contour 2 Partially well delimited
4 Both previous 3 Poorly delimited
v6 Maximum diameter of group 4 Spiculated
[3-120] mm v12 Density greater than parenchyma
v7 Number 1 Not
1 <10 2 Yes
2 10 to 30 v13 Focal distortion
3 >30 1 Not
v8 Morphology 2 Yes
1 Ring shaped v14 Focal asymmetry
2 Regular sharp-pointed 1 Not
3 Too small to determine 2 Yes
4 Irregular sharp-pointed v15 Maximum diameter
5 Vermicular, ramified [5-80] mm
v9 Size irregularity
1 Very regular
2 Sparingly regular
3 Very irregular
76 4. Breast Cancer Diagnosis by Fuzzy CoCo

4.2.2 Proposed solution: the COBRA system

The solution scheme I propose, the so-called COBRA system, is depicted in Figure 4.4. It
is composed of four elements: a user interface, a reading form, a database, and a diagnostic
decision unit which is the core of the system. This latter consists of a fuzzy system and a
threshold unit. Based on the 15 input attributes collected with the reading form, the fuzzy
system computes a continuous appraisal value of the malignancy of a case. The threshold unit
then outputs a biopsy recommendation according to the fuzzy system’s output. The threshold
value used in this system is 3, which corresponds to the “indeterminate” diagnostic. Fuzzy
CoCo is applied to design the fuzzy system in charge of appraising malignancy [131, 132].
The web-based user interface was developed to provide online access to the system (avail-
able at the URL: http://lslwww.epfl.ch/˜cobra). In the developed tool, the user fills the read-
ing form (see a snapshot in Figures 4.5). The COBRA provides, in addition to the final
biopsy recommendation, information about the appraisal value computed by the fuzzy sub-
system and about the rules involved in the decision (Figure 4.6 shows a snapshot of COBRA’s
output). The tool can be also used to train novel radiologists as the reading form can ac-
cess previously diagnosed cases contained in the database. The tool is available at the URL:
http://lslwww.epfl.ch/˜cobra.

Web−based user interface

Reading form
Reading Fuzzy system Malignancy
Threshold unit
Biopsy

input appraisal Proposal

Diagnostic decision unit

Database

Figure 4.4 The COBRA system comprises a user interface, a reading form—used to collect the pa-
tient’s data from either the user or the database, a database containing selected cases, and a diagnostic
decision unit which is the core of the system. In the decision unit the fuzzy system estimates the malig-
nancy of the case and the threshold unit outputs a biopsy recommendation.

4.2.3 Fuzzy CoCo setup

4.2.3.1 Fuzzy-parameter setup
I used prior knowledge about the Catalonia database to guide my choice of fuzzy parame-
ters. In addition, I took into account the interpretability criteria presented in Section 2.3 to
define constraints on the fuzzy parameters. Referring to Table 2.1, I delineate below the fuzzy
system’s set-up:
4.2 Mammography Interpretation: The COBRA system 77

Figure 4.5 User interface: reading form. The snapshot illustrates the reading form through which the
COBRA system collects data about a case.

• Logical parameters: singleton-type fuzzy systems; min-max fuzzy operators; orthogo-

nal, trapezoidal input membership functions (see Figure 4.7); weighted-average defuzzi-
fication.

• Structural parameters: two input membership functions (Low and High; see Figure 4.7);
two output singletons (benign and malignant); a user-configurable number of rules. The
relevant variables are one of Fuzzy CoCo’s evolutionary objectives. Low and High lin-
guistic labels may be further replaced by labels having medical meaning according to
the specific context of each variable.

• Connective parameters: the antecedents and the consequent of the rules are searched by
Fuzzy CoCo. The algorithm also searches for the consequent of the default rule. All
rules have unitary weight.

• Operational parameters: the input membership-function values are to be found by Fuzzy

CoCo. The values of P1 and P2 (Figure 4.7) are restricted to the universe of each vari-
able. For the output singletons I used the values 1 and 5, for benign and malignant,
respectively.
78 4. Breast Cancer Diagnosis by Fuzzy CoCo

Figure 4.6 User interface: biopsy recommendation. The snapshot shows a diagnostic recommendation
for a given patient. Note that besides recommending to practice a biopsy, COBRA gives information
about the appraisal value (indeterminate, 2.8), about the rules involved in the decision (rules 7, 8, and
default), and about their truth value (0.33, 0.4, and 0.6, respectively).

4.2.3.2 Genome encodings

Fuzzy CoCo thus searches for four parameters: input membership-function values, relevant
input variables, and antecedents and consequents of rules. To encode these parameters into
both species’ genomes, which together describe an entire fuzzy system, it is necessary to take
into account the heterogeneity of the input variables as explained below.

• Species 1: Membership functions. The fifteen input variables (v1 – v15 ) present three
different types of values: continuous (v1 ,v6 , and v15 ), discrete (v3 – v5 and v7 – v11 ), and
binary (v2 and v12 – v14 ). It is not necessary to encode membership functions for binary
variables as they can only take on two values. The membership-function genome en-
codes the remaining 11 variables—three continuous and eight discrete—each with two
parameters P1 and P2 , defining the membership-function apices (Figure 4.7). Table 4.6
delineates the parameters encoding the membership-function genome.
4.2 Mammography Interpretation: The COBRA system 79

Membership
Low High
1

0 Variable
P1 P2
Figure 4.7 Input fuzzy variables. Each fuzzy variable has two possible fuzzy values labeled Low and
High, and orthogonal membership functions, plotted above as degree of membership versus input value.
P1 and P2 define the membership-function apices.

Table 4.6 Genome encoding for membership-function species. Genome length is 106 bits.
Variable type Qty Parameters Bits Total bits
Continuous 3 2 7 42
Discrete 8 2 4 64
Total Genome Length 106

• Species 2: Rules. The i-th rule has the form:

if (v1 is Ai1 ) and . . . and (v15 is Ai15 ) then (output is C i ),

where Aij can take on the values: 1 (Low), 2 (High), or 0 or 3 (Don’t care). C i can take
on the values: 1 (benign) or 2 (malignant). As mentioned before, each database case
presents three clinical characteristics and six radiologic features according to the type
of lesion found: mass or microcalcifications (note that only a few special cases contain
data for both groups). To take advantage of this fact, the rule-base genome encodes,
for each rule, 11 parameters: the three antecedents of the clinical-data variables, the
six antecedents of one radiological-feature group, an extra bit to indicate whether the
rule applies for mass or microcalcifications, and the rule consequent. Furthermore, the
genome contains an additional parameter corresponding to the consequent of the default
rule. Relevant variables are searched for implicitly by allowing the algorithm to choose
Don’t care labels as valid antecedents (Aij = 0 or Aij = 3); in such a case the respective
variable is considered irrelevant, and removed from the rule. Table 4.7 delineates the
parameters encoding the rules genome.
Table 4.7 Genome encoding for rules species. Genome length is (20 × Nr ) + 1 bits, where Nr denotes
the number of rules.
Parameters Qty Bits Total bits
Clinical antecedents 3 × Nr 2 6 × Nr
Radiologic antecedents 6 × Nr 2 12 × Nr
Rule-type selector Nr 1 Nr
Consequents Nr + 1 1 Nr + 1
Total Genome Length (20 × Nr ) + 1
80 4. Breast Cancer Diagnosis by Fuzzy CoCo

4.2.3.3 Evolutionary parameters

Table 4.8 delineates values and ranges of values of the evolutionary parameters. The algorithm
terminates when the maximum number of generations, Gmax , is reached (I set Gmax = 700 +
200 × Nr , i.e., dependent on the number of rules used in the run), or when the increase in
fitness of the best individual over five successive generations falls below a certain threshold
(10−4 in these experiments).

Table 4.8 Fuzzy CoCo set-up for the COBRA system.

Parameter Values
Population size Np 90
Maximum generations Gmax 700 + 200Nr
Crossover probability Pc 1
Mutation probability Pm {0.005,0.01}
Elitism rate Er {0.1,0.2}
“Fit” cooperators Ncf 1
Random cooperators Ncr 1

4.2.3.4 Fitness function

The fitness definition takes into account medical diagnostic criteria. The most commonly em-
ployed measures of the validity of diagnostic procedures are the sensitivity and specificity, the
likelihood ratios, the predictive values, and the overall classification (accuracy) [18]. Table 4.9
provides expressions for four of these measures which are important for evaluating the perfor-
mance of my systems. Three of them are used in the fitness function, the last one is used in
Section 4.2.4 to support the analysis of the results. Besides these criteria, the fitness function
provides extra selective pressure based on two syntactic criteria: simplicity and readability
(see Section 2.3).

Table 4.9 Diagnostic performance measures. The values used to compute the expressions are: True
positive (TP): the number of positive cases correctly detected, true negative (TN): the number of neg-
ative cases correctly detected, false positive (FP): the number of negative cases diagnosed as positive,
and false negative (FN): the number of positive cases diagnosed as negative.
TP
Sensitivity
TP + FN
TN
Specificity
TN + FP
TP + TN
Accuracy
TP + TN + FP + FN
TP
Positive predictive value (PPV)
TP + FP
4.2 Mammography Interpretation: The COBRA system 81

The fitness function combines the following five criteria: 1) Fsens : sensitivity, or true-
positive ratio, computed as the percentage of positive cases correctly classified; 2) Fspec : speci-
ficity, or true-negative ratio, computed as the percentage of negative cases correctly classified
(note that there is usually an important trade-off between sensitivity and specificity which ren-
ders difficult the satisfaction of both criteria); 3) Facc : classification performance, computed
as the percentage of cases correctly classified; 4) Fr : rule-base size fitness, computed as the
percentage of unused rules (i.e., the number of rules that are never fired and can thus be re-
moved altogether from the system); and 5) Fv : rule-length fitness, computed as the average
percentage of don’t-care antecedents—i.e., unused variables—per rule. This order also repre-
sents their relative importance in the final fitness function, from most important (Fsens ) to least
important (Fr and Fv ).
The fitness function is computed in three steps—basic fitness, accuracy reinforcement,
and size reduction—as explained below:

1. Basic fitness. Based on sensitivity and specificity, it is given by

Fsens + αFspec
F1 = ,
1+α

where the weight factor α = 0.3 reflects the greater importance of sensitivity.

2. Accuracy reinforcement. Given by

F1 + βFacc
F2 = ,
1+β

where β = 0.01. Facc = Facc when Facc > 0.7; Facc = 0 elsewhere. This step slightly
reinforces the fitness of high-accuracy systems.

3. Size reduction. Based on the size of the fuzzy system, it is given by

F2 + γFsize
F = ,
1 + 2γ

where γ = 0.01. Fsize = (Fr + Fv ) if Facc > 0.7 and Fsens > 0.98; Fsize = 0 elsewhere.
This step rewards top systems exhibiting a concise rule set, thus directing evolution
toward more interpretable systems.

4.2.4 Results
A total of 65 evolutionary runs were performed, all of which found systems whose fitness
exceeds 0.83. In particular, considering the best individual per run (i.e., the evolved system
with the highest fitness value), 42 runs led to a fuzzy system whose fitness exceeds 0.88, and
of these, 6 runs found systems whose fitness exceeds 0.9; these results are summarized in
Figure 4.8.
Table 4.10 shows the results of the best systems obtained. The maximum number of rules
per system was fixed at the outset to be between ten and twenty-five.
82 4. Breast Cancer Diagnosis by Fuzzy CoCo

25

20

15

10

0
0.82 0.83 0.84 0.85 0.86 0.87 0.88 0.89 0.9 0.91 0.92

Figure 4.8 Summary of results of 65 evolutionary runs. The histogram depicts the number of systems
exhibiting a given fitness value at the end of the evolutionary run. The fitness considered is that of the
best individual of the run.

Table 4.10 Results of the best systems evolved. Results are divided into four classes, in accordance
with the maximum number of rules-per-system, going from 10-rule systems to 25-rule ones. Shown
below are the fitness values of the top systems as well as the average fitness per class, along with the
number of rules which effectively used by the system (Reff ) and the average number of variables per
rule (Vr ). The performance of the highlighted systems is presented in more detail in Table 4.11.
Maximum number Best individual Average per class
of rules Fitness Reff Vr Fitness Reff Vr
10 0.8910 9 2.22 0.8754 9.17 2.52
15 0.8978 12 2.50 0.8786 12.03 2.62
20 0.9109 17 2.41 0.8934 14.15 2.59
25 0.9154 17 2.70 0.8947 15.78 2.76

As mentioned before, my fitness function includes two syntactic criteria to favor the evo-
lution of good diagnostic systems exhibiting interpretable rule bases (see Section 2.3). Con-
cerning the simplicity of the rule base, rules that are encoded in a genotype but that never fire
are removed from the phenotype (the final system), rendering it more interpretable. Moreover,
to improve readability, the rules are allowed (and encouraged) to contain don’t-care conditions.
The relatively low values of Reff and Vr in Table 4.10 confirm the reinforced interpretability of
the evolved fuzzy systems.
Table 4.11 shows the diagnostic performance of two selected evolved systems. The first
system, which is the top one over all 65 Fuzzy CoCo runs, is a 17-rule system exhibiting a
sensitivity of 99.47% (i.e., it detects all but one of the positive cases), and a specificity of
68.69% (i.e., 226 of the 329 negative cases are correctly detected as benign). The second sys-
tem is the best found when searching for ten-rule systems. The sensitivity and the specificity
of this 9-rule system are, respectively, 98.40% and 64.13%. As mentioned before, the usual
4.2 Mammography Interpretation: The COBRA system 83

positive predictive value (PPV) of mammography ranges between 15 and 35%. As shown in
Table 4.11, Fuzzy CoCo increases this value beyond 60%—64.36% for the 17-rule system—
while still exhibiting a very high sensitivity.

Table 4.11 Diagnostic performance of two selected evolved systems. Shown below are the sensitiv-
ity, the specificity, the accuracy, and the positive predictive value (PPV) of two selected systems (see
Table 4.9). In parentheses are the values, expressed in number of cases, leading to such performance
measures. The 17-rule system is the top system. The 9-rule system is the best found when searching for
ten-rule systems.
17-rule 9-rule
Sensitivity 99.47% (186/187) 98.40% (184/187)
Specificity 68.69% (226/329) 64.13% (211/329)
Accuracy 79.84% (412/516) 76.55% (395/516)
PPV 64.36% (186/289) 60.93% (184/302)

To assess the generalization capabilities of the fuzzy systems obtained, I apply ten-fold
cross-validation [181]. For each ten-fold test the data set is first partitioned into ten equal-
sized sets, then each set is in turn used as the test set while Fuzzy CoCo trains on the other
nine sets. Table 4.12 presents the average results obtained performing six cross-validation runs
for several rule-base sizes. The percent difference between training and test sets is relatively
low, even for large rule bases, indicating low overfitting. A further discussion on generalization
in fuzzy systems is presented in Section 5.4.

Table 4.12 Results of ten-fold cross-validation. Results are divided into three classes, in accordance
with the maximum number of rules-per-system, going from 3-rule systems to 30-rule ones. Shown below
is the average fitness per class obtained in both training and test sets (the fitness represents the average
of the average fitnesses for the six cross-validation runs), along with the difference between training
and test performances, expressed both in absolute value and percentage.
Maximum number Training set Test set Difference Percentage
of rules
3 0.8269 0.7865 0.0404 4.89
7 0.8511 0.8036 0.0474 5.58
10 0.8712 0.8030 0.0682 7.83
15 0.8791 0.8125 0.0666 7.58
20 0.8814 0.8244 0.0569 6.47
30 0.8867 0.8104 0.0763 8.60

In summary, Fuzzy CoCo was able to evolve high-performance systems for two hard
problems, with all systems exhibiting high interpretability.
Chapter 5

Analyzing Fuzzy CoCo

In this thesis I proposed a novel coevolutionary approach to fuzzy modeling—Fuzzy CoCo—

which I then applied to the solution of several hard problems. The experience acquired while
exploring the usage and limitations of the method have permitted the extraction of guiding
criteria to facilitate the use of Fuzzy CoCo. It is necessary, however, to perform further anal-
ysis to better understand the effect that some of the Fuzzy-CoCo parameters may have on its
performance. It is also convenient to take a second look at the results obtained, in order to
verify the consistency and the generality of the systems designed with Fuzzy CoCo.
This chapter is organized as follows: The next section presents a guide to applying Fuzzy
CoCo. Section 5.2, after analyzing the effect of parameters on the performance of Fuzzy
CoCo, proposes some qualitative relationships that may facilitate setting up the algorithm.
Next, in Section 5.3, I discuss the consistency and the quality of the results obtained by Fuzzy
CoCo for the different problems presented. Section 5.4 presents the concept of local generality
and analyzes this aspect of the best systems found by Fuzzy CoCo. Finally, I summarize the
chapter in Section 5.5.

5.1 A Stepwise Guide to Applying Fuzzy CoCo

Applying Fuzzy CoCo requires the definition of parameters of its two main components: (1)
the fuzzy system and (2) the cooperative coevolutionary algorithm.

1. Fuzzy-system parameters. In Section 2.1 I classified the parameters of a fuzzy sys-

tem into four categories: logical, structural, connective, and operational. Based on this
classification, the following four-level procedure serves to define the fuzzy-system pa-
rameters:

(a) Define the logical parameters. As noted in Section 2.1, logical parameters are
predefined by the designer based on experience and on problem characteristics.
The user must define the type of fuzzy system (e.g., singleton-type), the operators
used for AND, OR, implication, and aggregation operations (e.g., min-
max operators), the type of membership functions (e.g., orthogonal, trapezoidal
ones), and the defuzzyfication method (e.g., COA).
(b) Choose the structural parameters.

85
86 5. Analyzing Fuzzy CoCo

• A set of relevant variables should be defined. Usually, this set includes all the
available variables, as Fuzzy CoCo can be set to automatically reduce their
number.
• Fuzzy CoCo requires predefining the number of membership functions. Al-
though this number could be set to a relatively high value to then let Fuzzy
CoCo seek automatically an efficient subset of the membership functions, this
strategy must be used carefully. The increase in the size of the search space
may prevent Fuzzy CoCo from converging towards good results.
• The number of rules is fixed (by the designer) for a given Fuzzy-CoCo run. A
discussion about this number is presented in Section 5.2.1.
(c) Encode the connective parameters into the rules genome. The rules may be either
complete (i.e., containing at least one antecedent for each variable) or incomplete
(i.e., using don’t-care labels). The antecedents in rules may be connected merely
by the AND operator, or may contain also OR and NOT operators. Fuzzy CoCo
thus offers the designer the freedom of choosing any type of rule, given that there
exists a proper way to encode it. If the problem requires a good interpretability, the
syntactic criteria presented in Section 2.3 must be taken into account to constrain
the definition of the rules genome.
(d) Encode the operational parameters into the membership-function genome. The
membership functions can be of arbitrary form. The only condition imposed by
Fuzzy CoCo is that all possible labels implied by the rules species should be de-
fined. Besides, to reinforce the interpretability of the system, the semantic criteria
presented in Section 2.3 should be used to define some restrictions on the definition
of the membership functions.

Note that Fuzzy CoCo is a methodology for improving the performance and speed of the
fuzzy-modeling process. It cannot correct on its own wrong decisions made during the
definition of the fuzzy-system parameters. Thus, the designer need posses knowledge of
the problem or a good evaluation heuristics.

2. Coevolutionary parameters. Once both genomes—rules and membership-functions—

are encoded, the coevolutionary parameters presented below must be set according
to a number of criteria, the two most important being computational costliness and
exploration-exploitation tradeoff. A discussion concerning the qualitative relations
among the ranges of the coevolutionary parameters is presented in Section 5.2.6.

(a) Population size PS . Fuzzy CoCo requires smaller populations than a simple ge-
netic algorithm; typically, 50 to 80 percent smaller. This markedly reduces the
computational cost. In the WBCD example, typical population sizes in Fuzzy
CoCo are 40 to 80, while the standard fuzzy-genetic approach uses 200 individ-
uals. A deeper analysis of the effects of PS  is presented in Section 5.2.2.
(b) Number of cooperators Nc . Typical values range from 1 to 6 (1 to 3 “fit” coopera-
tors and 0 to 4 random cooperators). The number of “fit” cooperators Ncf directly
affects exploitation, while the number of random cooperators Ncr directly affects
5.2 Effects of Some Parameters on Performance 87

exploration. Both of them affect the computational cost. The effects of this param-
eter are analyzed in Section 5.2.3.
(c) Crossover probability Pc . There is no special consideration concerning the value
of Pc in Fuzzy CoCo. Standard values—0.5 to 1—are used [106].
(d) Mutation probability Pm . As discussed in Section 3.3, due to an exploration-
exploitation tradeoff with the elitism rate, Pm values in Fuzzy CoCo are
usually an order of magnitude higher than in a simple genetic algorithm.
While the value of Pm proposed by Potter and DeJong (Pm = 1/Lg where
Lg = length(genome)) [143] can be applied with relatively large populations,
it has to be increased by up to 10 times when Fuzzy CoCo is applied with small
populations. See Section 5.2.4 for an analysis of the effects of Pm on Fuzzy CoCo
performance.
(e) Elitism Rate Er . Typical values for Er are between 0.1 and 0.4, where larger
values are required in systems with few rules and small populations. Er encourages
exploitation of solutions found. Section 5.2.5 analyzes the effects of this parameter
on the performance of Fuzzy CoCo.
(f) Maximum number of generations Gmax . Due to the speed gain offered by Fuzzy
CoCo, the value Gmax , related directly to computational cost, can be up to five
times smaller than single-population algorithms. For example, for a 5-rule sys-
tem in the WBCD problem, while Fuzzy CoCo runs 1500 generations, the fuzzy-
genetic approach runs 4500 generations.

5.2 Effects of Some Parameters on Performance

Fuzzy CoCo requires the designer to define a number of interdependent parameters that affect
directly or indirectly the performance of the algorithm. In this section I analyze the effects that
five of these parameters have on the algorithm’s performance and on its dynamics—i.e., the
way in which this performance evolves. The five parameters analyzed are: number of rules,
population size, number of cooperators, mutation probability, and elitism rate.
For the purpose of this analysis, I use a simple ad hoc experiment based on the WBCD
problem (Section 4.1). The experiment consists of 125 Fuzzy-CoCo runs in which the sole
criterion maximized is the classification performance (i.e., the fitness function does not include
any interpretability criteria). In order to investigate the effect of each parameter, Fuzzy CoCo
is tested with five different values of the concerned parameter. The fitness reported in each
case is the average of five runs.
Table 5.1 shows the values used in this experiment. Note that as some of these parameters
may affect the number of fuzzy systems evaluated per generation, I used the cumulative num-
ber of fitness evaluations instead of the number of generations, to represent the computational
effort. Consequently, the parameter Gmax (maximum number of generations) is substituted by
the maximum number of fitness evaluations Fmax . This value, set to 4 × 105 , is lower than
those used by Fuzzy CoCo in Section 4.1, as this analysis does not intend to find better systems
than those already obtained.
88 5. Analyzing Fuzzy CoCo

Table 5.1 Fuzzy CoCo set-up for the WBCD problem used to analyze the effects of some parameters.
Parameter Default value Tested values
Number of rules Nr 5 {2, 3, 5, 8, 12}
Population size PS  70 {20, 30, 50, 80, 120}
Crossover probability Pc 1 —
Mutation probability Pm 0.05 {0.001, 0.005, 0.01, 0.05, 0.1}
Elitism rate Er 0.1 {0.03, 0.1, 0.2, 0.4, 0.6}
“Fit” cooperators Ncf 1 —
Random cooperators Ncr 1 {0, 1, 2, 3, 4}
5
Maximum fitness evaluations Fmax 4 × 10 —

The first parameter analyzed—number of rules—principally affects the size and the shape
of the search space. The next two parameters, population size and number of cooperators,
define the number of fuzzy systems evaluated per generation. Then, I analyze the effects of
mutation probability and elitism rate. Finally, I derive some qualitative relationships between
these parameters.

5.2.1 Number of rules

Figures 5.1 (a) and (b) show the evolution of Fuzzy CoCo performance for different numbers
of rules. One can observe that, even though larger systems evolve faster initially, at about 5000
fitness evaluations all the systems exhibit similar performance. From that point onward larger
systems continue to increase their fitness while smaller ones tend to stagnate (Figure 5.1.b).
Note that 12-rule systems evolve slowly and perform worst than slightly smaller ones.
(a) (b)
0.99 0.99
2 rules
3 rules
5 rules
0.98 8 rules
0.985 12 rules

0.97

0.98

0.96

2 rules 0.975
0.95 3 rules
5 rules
8 rules
12 rules
0.94 0.97
3 4 5 4 5
10 10 10 10 10

Figure 5.1 Evolution of fitness as a function of the maximum number of rules. The figures show the
average fitness of Fuzzy CoCo runs for different values of Nr . The inset box in (a)—corresponding to
a steadier evolution of the fitness—is enlarged in (b). The abscissa represents the computational effort
measured in number of fitness evaluations.

For a given problem, searching for a compact fuzzy rule base is harder than searching
5.2 Effects of Some Parameters on Performance 89

for a slightly larger system, even if the genome is larger for this latter search. The reason
for this apparent contradiction is that a less compact fuzzy system—i.e., with more rules—
can cover a larger part of the problem space. However, if evolution seeks too many rules, the
fitness landscape becomes too “flat” (intuitively, an abundance of low-performance hills, rather
than a few high-performance mountains), thus rendering the search more difficult. This idea
is reinforced by the systems found for the COBRA problem (Table 4.10), which effectively
utilize between 63% and 91% of the number of rules encoded into the genome (respectively,
15.78 rules for 25-rule runs and 9.13 for 10-rule runs). It seems clear that for each problem
there exists a range of ideal rule-base sizes (between 4 and 7 for the WBCD problem, and
about 17 for the COBRA system). Besides trial and error, I am aware of no current method to
determine this range.

5.2.2 Population size

As mentioned before, coevolutionary algorithms involve smaller populations than single-
population algorithms. One clear advantage of such smaller populations is the fast evaluation
of each generation. However, under these conditions the population might not have enough
individuals to perform both exploration of the search space and exploitation of good solutions.
Consequently, the search can stagnate after a limited number of generations. Figure 5.2,
which shows the evolution of the fitness for different population sizes, corroborates this
surmise. Indeed, populations of 20 individuals evolve fast (around 3000 fitness evaluations)
to fitness values close to their maximum, while larger populations take longer to reach such
performance levels. Observation of the steadier evolution stage (Figure 5.2.b) shows that
the largest population (i.e., 120 individuals) outperforms the group formed by medium-size
populations (30, 50, and 80 individuals). 20-individual populations almost stagnate at this
stage of evolution.

(a) (b)
0.99 0.99
Pop= 20
Pop= 30
Pop= 50
0.98 Pop= 80
0.985 Pop=120

0.97

0.98

0.96

Pop= 20 0.975
0.95 Pop= 30
Pop= 50
Pop= 80
Pop=120
0.94 0.97
3 4 5 4 5
10 10 10 10 10

Figure 5.2 Evolution of fitness as a function of the population size. The figures show the average
fitness of Fuzzy CoCo runs for different values of PS . The inset box of (a)—corresponding to a
steadier evolution of the fitness—is enlarged in (b). The abscissa represents the computational effort
measured in number of fitness evaluations.
90 5. Analyzing Fuzzy CoCo

The above analysis suggests that when applying Fuzzy CoCo to a given problem the
use of small populations would provide a fast, accurate estimate of attainable fitness values.
This can be useful for coarse-tuning the fitness function and the genetic operators. Medium-
size populations would then be used to fine-tune the algorithm. The final search should be
performed using large populations in order to provide the algorithm with the diversity required
to adequately explore the search space.

5.2.3 Number of cooperators

As noted in Section 5.1, the number of selected fit and random cooperators (Ncf and Ncr ),
besides affecting the computational cost of the algorithm, are parameters related, respectively,
to exploitation and exploration capabilities of Fuzzy CoCo. However, in contrast to elitism and
mutation, these two parameters do not affect directly the genetic pool of the next generation,
only affecting it indirectly by permitting Fuzzy CoCo to evaluate more accurately individual
fitness. Figure 5.3 shows the evolution of fitness for different numbers of cooperators. As in the
case of population size, systems with few fitness evaluations per generation (i.e., with one or
two cooperators) evolve fast and provide solutions with “acceptable” fitness in approximately
5000 fitness evaluations. In the long term (i.e., more than 105 evaluations—see Figure 5.3.b)
all the systems perform similarly. Only the systems using five cooperators perform slightly
better.

(a) (b)
0.99 0.99
Nc=1
Nc=2
Nc=3
0.98 Nc=4
0.985 Nc=5

0.97

0.98

0.96

Nc=1 0.975
0.95 Nc=2
Nc=3
Nc=4
Nc=5
0.94 0.97
3 4 5 4 5
10 10 10 10 10

Figure 5.3 Evolution of fitness as a function of the number of cooperators. The figures show the
average fitness of Fuzzy CoCo runs for different values of Nc = Ncf + Ncr . The inset box of (a)—
corresponding to a steadier evolution of the fitness—is enlarged in (b). The abscissa represents the
computational effort measured in number of fitness evaluations.

Thus, concerning the number of cooperators, it seems that a light setup, with few co-
operators, might be used in the earliest stages of the design of a solution based on Fuzzy
CoCo. Once the fitness function has been adjusted the search can be performed using more
cooperators. Note, however, that the present analysis did not evaluate the combined effect of
simultaneously having a large population and a high number of cooperators, both of which are
5.2 Effects of Some Parameters on Performance 91

computationally costly strategies.

5.2.4 Mutation probability

Figure 5.4 shows the effect that mutation probability has on the evolution of the fitness. The
initial evolution (i.e., up to 20000 fitness evaluations) does not depend much on the mutation.
In a later stage of evolution (Figure 5.4.b), moderate values of mutation probability—i.e.,
0.005 and 0.01, corresponding approximately to 0.5/Lg and 1/Lg , where Lg is the genome’s
length—seem to be the most adequate.

(a) (b)
0.99 0.99
Pm=0.001
Pm=0.005
Pm=0.01
0.98 Pm=0.05
0.985 Pm=0.1

0.97

0.98

0.96

Pm=0.001 0.975
0.95 Pm=0.005
Pm=0.01
Pm=0.05
Pm=0.1
0.94 0.97
3 4 5 4 5
10 10 10 10 10

Figure 5.4 Evolution of fitness as a function of the mutation probability. The figures show the average
fitness of Fuzzy CoCo runs for different values of Pm . The inset box of (a)—corresponding to a steadier
evolution of the fitness—is enlarged in (b). The abscissa represents the computational effort measured
in number of fitness evaluations.

5.2.5 Elitism rate

Observing the initial evolution of the fitness for different elitism rates (Figure 5.5.a) one can see
that, during the first 10000 evaluations, Fuzzy CoCo runs using an elitism rate of 10% perform
better, while runs with an elitism rate of 3% (corresponding to an elite with 2 individuals)
perform significantly worse. This latter trend is confirmed in the long term (Figure 5.5.a),
whereas the use of an elitism rate of 40% appears as a slightly better alternative than the
others.

5.2.6 Qualitative relationships

The analysis presented above describes the effects that some parameters have on the perfor-
mance of Fuzzy CoCo for a given problem. However, these effects are evaluated by modify-
ing, each time, only the involved parameter. Many more tests would be necessary to assess the
combined effects of two or more parameters (e.g., to determine an adequate combination of
92 5. Analyzing Fuzzy CoCo

(a) (b)
0.99 0.99
Er=0.03
Er=0.1
Er=0.2
0.98 Er=0.4
0.985 Er=0.6

0.97

0.98

0.96

Er=0.03 0.975
0.95 Er=0.1
Er=0.2
Er=0.4
Er=0.6
0.94 0.97
3 4 5 4 5
10 10 10 10 10

Figure 5.5 Evolution of fitness as a function of the elitism rate. The figures show the average fitness of
Fuzzy CoCo runs for different values of Er . The inset box of (a)—corresponding to a steadier evolution
of the fitness—is enlarged in (b). The abscissa represents the computational effort measured in number
of fitness evaluations.

values of Pm and Er ) in order to characterize completely the performance of the method. The
same exhaustive analysis should be performed for different problems so as to identify whether
or not these effects depend on the problem under study.
Instead of such an arduous analysis, I have derived some qualitative relationships between
various parameters of Fuzzy CoCo. Shown in Table 5.2, these relationships are based on the
simulations carried out on the different problems presented in this thesis.

Table 5.2 Qualitative relationships between Fuzzy CoCo parameters. Delineated below are criteria to
guide the choice of the number of cooperators Nc , the mutation probability Pm (expressed as a function
of the genome’s length, Lg ), and the elitism rate Er , all three as a function of the desired number of
rules and of the desired population size (expressed in terms of percentage of the typical population size
of a single-population algorithm). For example, if the user wishes to employ a large population with
few rules then she should set Nc , Pm , and Er to values within the ranges specified in the upper-right
quadrant of the table.

Number of rules Population size

Small (20%–40%) Large (40%–70%)
Nc many (4–6) some or many (3–6)
Few (2–4) Pm large ((8 − 10)/Lg ) medium ((4 − 6)/Lg )
Er high (0.4–0.6) medium–high (0.3–0.6)
Nc few or some (2–5) few (1–3)
Many (4–10) Pm small or medium ((2 − 5)/Lg ) small ((1 − 2)/Lg )
Er medium (0.2–0.4) low (0.1–0.2)
5.3 Consistency and Quality of Results 93

5.3 Consistency and Quality of Results

Fuzzy CoCo has proven capable of finding good solutions to hard problems. Such good results
emerge consistently as witnessed by the small differences observed between the best systems
reported and the average results, as well as by the distribution of the results shown in the cor-
responding histograms (see sections 3.4.3, 4.1.2, 4.2.4). This fact is exemplified in Figure 5.6,
showing the fitness evolution of the best, the worst, and the average runs of a given Fuzzy
CoCo setup.

0.99

0.98

0.97

0.96

Best run
Average
Worst run
0.95
0 100 200 300 400 500 600 700 800 900 1000

Figure 5.6 Evolution of the best, the worst, and the average runs of a given Fuzzy CoCo setup. The
curves correspond to 40 evolutionary runs, searching for 7-rule systems for the WBCD problem. The
difference between the best and the worst runs is around 1%.

The consistency of the results does not provide information concerning the quality of
the systems obtained by Fuzzy CoCo. This is usually done by comparing the results with
previously published works. Because previous results for the Catalonia database are only
partially available, I have opted for the use of a standard test that provides a rough estimation
of the attainable performance: An intuitive, simple method known as k-nearest-neighbor (knn).
The knn method [181] is a classification technique based on the memorization of a train-
ing database. Given an unclassified input case, the system searches the memory for the k
training cases that most strongly resemble the new one. The resemblance is usually measured
through a distance metric—e.g., Manhattan or Euclidean distances. The closest training cases
are called the nearest neighbors. The unknown case is assigned the class of the majority of its
k-nearest neighbors.
Table 5.3 shows the performance obtained applying knn to the problems presented in this
thesis and their comparison with the results obtained by Fuzzy CoCo. The results for knn are
computed using “leave-one-out” cross-validation (i.e., the case being evaluated is left out and
the knn algorithm is applied in the remaining cases) [181].
94 5. Analyzing Fuzzy CoCo

Table 5.3 Comparison of Fuzzy CoCo with knn. Knn is used to roughly estimate an attainable fitness
value. Shown below are the classification performance values obtained for each problem, along with
the number of misclassifications in parentheses. The values for the COBRA system correspond to the
basic fitness F1 described in Section 4.2 that depends on sensitivity and specificity.
Problem knn Fuzzy CoCo
prediction Rules Runs Average Best
Iris (controller) 0.9733 (4) 3 49 0.9910 (1.35) 1.0000 (0)
Iris (classifier) 0.9733 (4) 3 48 0.9751 (3.74) 0.9933 (1)
WBCD 0.9736 (18) 7 40 0.9825 (11.95) 0.9898 (7)
Cobra 0.7441 25 15 0.8947 0.9154

These comparisons allow to relativize the results obtained for a given problem. For the Iris
and the WBCD problem, we can expect classification performances over 97.33%. However,
this does not imply that slightly higher performance values can be easily attained, as they
correspond to new correctly-classified cases that are hard to obtain, usually accepting increased
complexity of the solution. On the other hand, fitness values obtained for the COBRA system,
close to 0.89, while seemingly low in comparison with those of the WBCD problem, are in
fact good results that correspond to an increase of more than 200 correctly-diagnosed cases
with respect to knn results. Note that other machine-learning methods can be used to estimate
the attainable fitness values.

5.4 Generality of the Designed Fuzzy Systems

The generalization capability of fuzzy systems is usually assessed at a global level through
well-known generalization tests (e.g., 10-fold cross-validation, as that presented for the Cat-
alonia database in Section 4.2.4). However, little emphasis is put on the local generality of
those fuzzy systems—i.e., the generality of their rules—although this latter may affect the
global generalization capability. Indeed, overfitting in fuzzy systems is due to very specific
rules that describe exceptional conditions instead of general patterns.
In this section I analyze local generality and the effect that some fuzzy parameters can
have on it. I propose that the interpretability considerations presented in Section 2.3 may
reinforce the local generality of fuzzy systems. I then analyze the generality of the fuzzy
systems designed with Fuzzy CoCo to solve the Iris problem and the WBCD problem.

5.4.1 Generality of fuzzy rules

A fuzzy rule is said to be general with respect to a given input space, if it covers a non-
negligible portion of this space. Note that generality depends on the significance—in terms
of information—of the region covered and not on its extension, although significance and
extension are usually related. When the input space is presented discretely, as is the case for
the problems considered in this thesis, the generality of a rule is expressed in terms of the
number of instances it covers. Given that this concept is related only with the input space, the
generality of a rule does not imply its adequacy to describe a specific mapping between the
5.4 Generality of the Designed Fuzzy Systems 95

input and the output spaces. Moreover, high performance is often attained at the expense of
local generality.
Even though generality is defined at the rule level, some parameters of the whole fuzzy
system may affect its generality as discussed below:
• Membership functions per variable. The number of membership functions defining a
linguistic variable, also known as granularity, influences the fineness of the partition
and hence the capacity of the rules to target many reduced portions of the input space.
In general, a reduced number of membership functions favors the existence of general
rules.
• Variables per rule. Intuitively, the less variables in a rule’s premise, the more extended
may be the portion of the input space it covers. As explained in Section 2.3, several long
rules are needed to cover the same space than a short rule does.
• Rules per system. The coverage that a fuzzy system can attain for a given input space,
depends on the cumulative coverage of its rules. If the system disposes of only a few
rules, in order to guarantee an adequate coverage these rules will tend to be general.
Good performance requires good coverage. One can say that a small fuzzy system can-
not do both: perform well and overlearn.
Note that generality is a meaningful concept both linguistically and numerically speak-
ing. Because of this, the strategies proposed to reinforce the interpretability of fuzzy systems
(Section 2.3.3) also favor the emergence of general rules. That is explained below:
• The use of common shared labels prevents each rule from defining its own membership
functions. This avoids the existence of erratic rules tuned to very specific conditions.
• Don’t-care conditions promote the existence of rules with few variables in their premise.
• The default rule is, by definition, a general rule. It provides coverage for wide (and
sometimes disjoint) regions that would require many specific rules to describe them.
The example presented in Figure 5.7 illustrates how these strategies allow to improve the
generality of the rule base while preserving classification performance. The rule in the center
of the input space, albeit being specific in terms of extension, is quite general as it represents
one third of the cases contained in the hypothetic database. The rule containing a don’t-care
condition, marked A covers the space of three more-specific rules. The default rule, marked
D, covers the space of several very-specific rules including coverage for a region without any
instance.

5.4.2 Generality of the systems designed by Fuzzy CoCo

To design the fuzzy systems presented in this thesis, I used Fuzzy CoCo to search for both high
performance and high interpretability. Generality was not explicitly used as design-driving cri-
terion. As mentioned before, the strategies used to favor interpretability may also result in the
design of general rules. But if generality constitutes a priority when designing fuzzy systems
for a given problem, the criteria presented herein must be used to constrain the representation
of the fuzzy system, the fitness function, or both. Below I analyze the generality of the rules
of the best systems presented in previous chapters.
96 5. Analyzing Fuzzy CoCo

D A
3 9 20

18 50 35 35 50 65

3 0 10

a) b)

Figure 5.7 Improving the generality of a fuzzy system. The figure illustrates how a rule containing
a don’t-care condition and a default rule (marked A and D, respectively) increases the generality of
the system. The numbers in the rules represent the cases covered by the rule. The colors represent the
output class proposed by each rule. Note that the classification performance of the system is preserved.

5.4.2.1 The three-rule fuzzy controller for the Iris problem

Recall that the best controller-type system that solves the Iris problem—presented in Fig-
ure 3.9—has three short rules (1.7 variables per rule). It classifies correctly all the cases in the
database. Table 5.4 shows the activation figures of the three rules of this system, which serve
to analyze their generality.

Table 5.4 Generality of the three-rule fuzzy controller. The activation profile of each rule consists
of: number of firing instances, winning instances, instances where it fires alone, average activation
level, and maximum activation. Note that the average activation level is computed using only the firing
instances. The Iris database has 150 instances.

Rule Number of instances Firing level

fired winner alone average maximum
Rule 1 102 63 3 0.71 1
Rule 2 34 32 0 0.97 1
Rule 3 4 0 0 0.45 0.53
Default 115 61 48 0.63 1

Two of the active rules (i.e., rules 1 and 2) and the default rule are clearly general as they
are involved in the classification of many instances. In contrast, rule 3 is more specific as it
is fired only by 4 instances. However, despite its specificity, rule 3 never fires alone and it is
never the winner rule. Note that the four instances for which rule 3 fires are hard to classify.
Indeed, they lie close to the border between the classes versicolor and virginica in the input
space.
5.5 Summary 97

5.4.2.2 The three-rule fuzzy classifier for the Iris problem

The best classifier-type system found to solve the Iris problem, presented in Figure 3.11, has
three short rules (2.3 variables per rule). It classifies correctly all but one instance in the
database. The three rule-activation figures used to analyze the generality of the system are
presented in Table 5.5.

Table 5.5 Generality of the three-rule fuzzy classifier. The activation profile of each rule consists
of: number of firing instances, number of winning instances, number of instances where it fires alone,
average activation level, and maximum activation. Note that the average activation level is computed
using only the firing instances. The Iris database has 150 instances.

Rule Number of instances Firing level

fired winner alone average maximum
Rule 1 60 50 41 0.85 1
Rule 2 74 45 0 0.55 0.96
Rule 3 47 2 0 0.20 0.50
Default 109 53 32 0.54 1

It is clear from their firing profiles that all the rules are general. Even rule 3, whose
premise is very specific—it defines conditions for all four input variables (see Figure 3.11)—
fires for 47 instances.

5.4.2.3 The seven-rule system for the WBCD problem

The best evolved system for the WBCD problem (see Figure 4.3) consists of 7 rules with an
average of 3.4 variables per rule. This system classifies correctly 98.98% of the 683 instances
in the WBCD database (i.e., it misclassifies 7 instances). Rule generality is analyzed through
the activation figures presented in Table 5.6.
All but one of the active rules and the default rule are clearly general as they are in-
volved in the classification of many instances. In contrast, rule 3 is more specific (it fires for
only 12 instances) and relatively long (its premise contains 5 variables). However, despite its
specificity, rule 3 never fires alone and it is the winner rule for only two instances.

5.5 Summary
In this chapter I presented some analyses that should guide future users of Fuzzy CoCo in
their experimental design. First, I presented the steps required to define all the parameters
of the algorithm. Secondly, I proposed some qualitative criteria to define the values of the
most important parameters, based on an analysis of their effect on Fuzzy CoCo performance.
Then, to analyze the quality of the results independent of the availability of previous results, I
proposed the use of a simple machine-learning method (knn in this case) to roughly estimate
an attainable performance. Such an estimation allows to relativize the real performance of a
system for a given problem. Finally, I introduced the concept of local generality as a measure
98 5. Analyzing Fuzzy CoCo

Table 5.6 Generality of the seven-rule WBCD fuzzy system. The activation profile of each rule consists
of: number of firing instances, number of winning instances, number of instances where it fires alone,
average activation level, and maximum activation. Note that the average activation level is computed
using only the firing instances. The WBCD database has 683 instances.

Rule Number of instances Firing level

fired winner alone average maximum
Rule 1 482 202 8 0.74 1
Rule 2 438 386 1 0.91 1
Rule 3 12 3 0 0.43 0.75
Rule 4 52 2 0 0.22 0.63
Rule 5 239 106 2 0.47 1
Rule 6 183 48 0 0.39 0.88
Rule 7 115 21 0 0.29 0.75
Default 305 149 8 0.43 1

that can assess, or even reinforce, the global generalization capabilities of the fuzzy systems
designed using Fuzzy CoCo.
Chapter 6

Extensions of the Methodology

As mentioned in Section 5.1, Fuzzy CoCo requires the user to define a maximum number
of rules for a given run. However, for each problem there exists a range of ideal rule-base
sizes, that is hard to determine. The user is thus obliged to find this range by trial and error. I
propose in this chapter two extensions to Fuzzy CoCo, intended to simplify the task of finding
an adequate size of the rule base.
The first extension, called Island Fuzzy CoCo, is based on the Island model [172, 180].
It takes advantage of the exploration performed separately by concurrent instances of Fuzzy
CoCo, where each instance is set to search for systems of different sizes. Island Fuzzy CoCo
is presented in Section 6.1. Section 6.2 presents the second extension, inspired by the iter-
ative rule learning approach [59], and called Incremental Fuzzy CoCo. In this method, the
number of rules of the sought-after system increases each time that evolution satisfies certain
criteria. In this way, the search for more complex systems starts on the basis of some “good”
individuals.

6.1 Island Fuzzy CoCo

Often, when solving a fuzzy modeling problem, several instances of Fuzzy CoCo (in many
cases running simultaneously on separated machines) are set to search for systems with differ-
ent numbers of rules. Observing many of these runs, I remarked two interesting trends showing
that some individuals from a given run should be of interest for the population of other runs of
different complexity (either simpler or more complex).:
• In simple runs (i.e., searching for small systems), occasionally systems appear with
performance similar to those of more complex runs, but with the advantage of being
smaller and sometimes with shorter rules.

• In complex runs (i.e., searching for large systems), systems often appear that effectively
utilize less rules than the maximum allowed. Such systems exhibit fitness values similar
or superior to those of the simpler runs corresponding to their effective number of rules.
Island Fuzzy CoCo, the approach proposed herein, is similar to the so-called island model
where several (sub)populations, called islands or demes, evolving separately most of the time,
occasionally exchange individuals according to a certain migration policy [172, 180]. Below, I

99
100 6. Extensions of the Methodology

sketch Island Fuzzy CoCo and describe two preliminary tests performed to explore its behav-
ior.

6.1.1 The proposed algorithm

Several instances of Fuzzy CoCo—called islands—with different levels of complexity run
concurrently. Contrary to the classic island model, in which the genomes of the individuals of
all the islands are compatible, in Island Fuzzy CoCo individuals from different islands cannot
reproduce directly as their genomes encode a different number of rules. A controlled migratory
flux, adapted to the model, allows individuals to pass from an island to another as illustrated
in Figure 6.1.

FCC m Migration

FCC n FCC q

FCC p Island

Figure 6.1 Island Fuzzy CoCo. Several instances of Fuzzy CoCo, called here FCCc , run concurrently.
The index c represents the number of rules of the systems evolving in the island. The exchange of
individuals (migration) is controlled by a migration policy, which in this example allows exchange of
individuals between all the islands.

Migration introduces three new groups of parameters to be defined: the migration topol-
ogy, i.e., a set of allowed migration paths; the migration interval (Gm ), i.e., the number of
generations between exchanges; and the migration size (Nm ), i.e., the number of individuals
that migrate between the islands. In other words, from time to time, each island proposes
interesting candidates to some other islands and chooses a number of immigrants among the
candidates proposed to it. Figure 6.2 presents the Island Fuzzy CoCo algorithm in pseudo-code
format.
The main criterion to select the Nm migrants for a given destination island is the relative
performance with respect to the actual performance of the population in the destination island.
Two possible criteria, for an island, to select its Nm immigrants from a pool of candidates (MD )
are: (1) relative performance with respect to the actual average performance in the island and
(2) keep a balanced immigration either by choosing the same number of immigrants from
each origin island, or by establishing a maximum number of migrants per origin island. Even
6.1 Island Fuzzy CoCo 101

begin Island Fuzzy CoCo

g=0
for each island I
Initialize Fuzzy CoCo population PI (0)
end for
while not done do
for each island I
Evaluate population
g=g+1
if migration interval Gm elapsed
for each destination island D
MDI = Select-emigrants[PD (g-1)]
MDI = Adapt-genome[MDI ]
send MDI to D
end for
PI
(g − 1) = Reduce[PI (g − 1)]
PI
(g) = Fuzzy-CoCo evolve PI
(g − 1)
MI
= Select-immigrants[MI ]
PI (g) = PI
(g) + MI
(g)
else
PI (g) = Fuzzy-CoCo evolve PI (g − 1)
end if
Evaluate population PI (g)
end for
end while
end Island Fuzzy CoCo

Figure 6.2 Pseudo-code of Island Fuzzy CoCo. In each island an instance of Fuzzy CoCo runs for
a given number of rules. Every Gm generations, each island propose Nm emigrants to other islands
m immigrants from the pool of candidates proposed to it (MI =
(called here MD I ) and chooses N
 j
j (MI )). “Fuzzy-CoCo evolve” corresponds to the evolutionary operators applied in Fuzzy CoCo
(see Figure 3.2).

though each island may use its own, specific fitness criteria to evaluate its individuals, migrants
must be selected on the basis of a common fitness measure (e.g., performance).
Given that the complexity of individuals in the origin and the destination islands are
different, the genome of the migrants must be adapted: eliminating unused rules in downsizing
migrants or duplicating active rules for enlarging migrants. Special care must be taken to
preserve as much as possible the fitness of the migrants after these operations.
Fuzzy CoCo involves two coevolving species, but migration is decided according to the
fitness computed for entire fuzzy systems consisting of a pair (individual-cooperator). Be-
sides, the complexity of the system, which is a key concept in Island Fuzzy CoCo, is based only
on individuals from the rules species. There exist thus at least three possible ways to choose
102 6. Extensions of the Methodology

the migrants: (1) migrants are chosen separately for each species according to their fitness in-
dependent of their relation with the other species; (2) migrants are chosen for each species, but
emigrate together with their cooperators (i.e., selected members of the other species), and (3)
migrants are selected only from the rule species and migrate together with their cooperators.

6.1.2 Preliminary tests of Island Fuzzy CoCo

I present herein some tests of Island Fuzzy CoCo on the Iris-classifier and the WBCD prob-
lems, for which adequate setups of Fuzzy CoCo are already known. The main goal of these
tests is to explore the role of two of the new migration parameters: interval and size. I first
used the Iris classifier problem (Section 3.4) obtaining good results, but the low number of
rules required for this problem and the consequent low volume of data, render difficult the
extraction of conclusions. I thus used the WBCD problem (Section 4.1) that offers a wider
range of complexity. (Note: Part of the Iris-classifier tests were performed in the frame of a
student project by Yves Blatter.)

6.1.2.1 Experimental setup.

Setting up Island Fuzzy CoCo requires the definition of the following parameters:
• Number of islands and number of rules of their individuals. Based on the previous expe-
rience, I set 4 islands for the Iris classifier problem, whose individuals encode systems
going from two to five rules. For the WBCD problem, individuals distributed in seven
islands encoded systems going from one rule to seven rules.
• Fuzzy CoCo setup. All the islands use the same setup. Table 6.1 delineates the values
used for the Fuzzy CoCo parameters in each problem.

Table 6.1 Fuzzy CoCo set-up used for all the islands for the Iris classifier and the WBCD problems.
Parameter Iris WBCD
Population size P s 70 50
Maximum generations Gmax 1000 1500
Crossover probability Pc 1 1
Mutation probability Pm 0.1 0.1
Elitism rate Er 0.2 0.2
“Fit” cooperators Ncf 1 1
Random cooperators Ncr 1 1

• Migration topology. All the possible migratory paths are allowed.

• Migration interval Gm and migration size Nm . These are the parameters under study.
To test the migration effect I used three values for each parameter (for a total of 9 pos-
sible combinations). The values for the Iris classifier problem are Gm = {1, 10, 100}
generations and Nm = {5, 20, 50} individuals. For the WBCD problems the values are
Gm = {1, 5, 50} generations and Nm = {2, 5, 20} individuals.
6.1 Island Fuzzy CoCo 103

• Emigrant selection. The pool of possible emigrants to a given island is composed of indi-
viduals whose fitness value is greater than the actual average fitness of the destination is-
land. For each destination island, up to Nm emigrants are selected fitness-proportionally
from each species. Note that migrants carry with them their cooperators. However, if
several individuals carry the same cooperator, only one copy of the cooperator is placed
in the migrating group.

• Immigrant selection. The pool of possible immigrants to an island consists of all the
candidates proposed to it that exhibit a fitness value greater than the actual average fit-
ness on the island. Each island selects up to Nm immigrants using fitness-proportionate
selection. No mechanism to balance immigration is applied.

• Fitness function. All the islands use the same fitness function for both evolution and
migration purposes. The Iris classifier islands simply use the classification performance
Fc . The fitness function of the WBCD islands includes an interpretability-related term.
It is given by F = Fc − αFv , where α = 0.0015, Fc is the classification performance,
and Fv penalizes systems with a large number of variables in their rules.

6.1.2.2 Results for the Iris classifier problem.

Ten different tests were performed: nine combinations of migration parameters and a control
test without migration. Ten runs were performed for each case, for a total of 100 runs (each
with four islands). Table 6.2 presents the average fitness for each test. All the tests succeed in
finding systems exhibiting the maximum fitness (i.e., Fc = 99.33%).

Table 6.2 Results of Island Fuzzy CoCo for the Iris classifier problem. Shown below are the average,
over ten runs, of the maximum fitness obtained for each combination of the migration parameters:
migration interval Gm and migration size Nm expressed in generations and individuals, respectively.
Gm Nm
5 20 50
1 98.59% 98.53% 98.60%
10 98.73% 98.67% 98.80%
100 98.73% 98.93% 98.73%
None 98.83%

From the analysis of these global results—i.e., extracting the best system found at each
run—one can not see that migration affords a real advantage with respect to simple Fuzzy
CoCo (i.e., without migration). Only those runs that allow 20 migrants every 100 generations
perform slightly better. However, as Island Fuzzy CoCo searches simultaneously for different
system sizes, it is necessary to analyze the results for each island. Figure 6.3 compares, island
by island, the results obtained using the best migration policy with those obtained without
migration.
Migration appears to have a benefical effect for those islands containing large systems.
However, the low number of islands required for this problem renders difficult the extraction
of conclusions.
104 6. Extensions of the Methodology

99

98.5

98

97.5

97

96.5
Migration
No migration
96
2 3 4 5

Figure 6.3 Results, island by island, for the Iris classifier problem. The figure shows the average
fitness obtained, for each number of rules, using the best migration policy (i.e., Gm = 100 generations
and Nm = 20 individuals), compared with those obtained without migration.

6.1.2.3 Results for the WBCD problem.

Ten different tests were performed—i.e., nine combinations of migration parameters and a
control test without migration. Five runs were performed for each case, for a total of 50 runs
(each with seven islands). Tables 6.3.a and b present, respectively, the average and the best
fitness for each test.

Table 6.3 Results of Island Fuzzy CoCo for the WBCD problem. Shown below are (a) the average
over ten runs and (b) the maximum fitness, obtained for each combination of the tested migration
parameters: migration interval Gm and migration size Nm expressed in generations and individuals,
respectively.

(a) Average fitness (b) Best fitness

Gm Nm Gm Nm
2 5 20 2 5 20
1 97.93% 97.73% 97.98% 1 98.14% 97.82% 98.14%
5 97.90% 97.94% 97.98% 5 97.99% 97.99% 98.14%
50 98.00% 98.01% 98.06% 50 98.12% 98.12% 98.41%
None 97.98% None 98.12%

From these global results, one can see that only those runs using either Gm = 50 or
Nm = 20 find systems with similar performance to those found without migration. Runs
combining both values find the best systems. To evaluate the effect that migration has on each
island, Figure 6.4 compares the results of the best migration policy with those obtained without
migration.
These results confirm, as observed in the Iris-classifier test, that an adequate migration
6.1 Island Fuzzy CoCo 105

(a) Average fitness (b) Best fitness

98.4 98.4

98.1 98.1

97.8 97.8

97.5 97.5
Migration Migration
No migration No migration
97.2 97.2
1 2 3 4 5 6 7 1 2 3 4 5 6 7

Figure 6.4 Results, island by island, for the WBCD problem. The figure shows the average fitness
obtained, for each number of rules, using the best migration policy (i.e., Gm = 50 generations and
Nm = 20 individuals), compared with those obtained without migration.

policy may have a clearly positive effect on the performance of islands containing complex
individuals. In this case four-rule to seven-rule systems.
Migration also affects the dynamics of evolution in each island. Figure 6.5 shows the
evolution of performance in different islands with and without migration. Note that even for
small systems, where migration does not improve the final performance, the fitness evolve at
least as fast as in the absence of migration.

6.1.2.4 Analysis of results

The tests presented above constitute a first exploration of the method. One can, however,
advance some conclusions that should be verified with further tests.
• Migration has positive effects, with respect to simple Fuzzy CoCo, on the global search
capabilities and on the evolutionary dynamics; in particular, for islands containing sys-
tems with a high number of rules.
• Island Fuzzy CoCo exhibits slightly lower search power for smaller systems. This may
be due to the use of identical setups for all the islands, tailored to high-complexity is-
lands.
• The setup of Island Fuzzy CoCo may be difficult. First, because of the new migratory
parameters that enlarge the number of parameters to be defined by the user. And second,
because each island should require specific setups, both migratory and coevolutionary,
to perform optimally. However, the heuristics, presented in Section 5.2.6, to set Fuzzy
CoCo up may alleviate this task.
• Due to migration requirements, Island Fuzzy CoCo is slightly heavier in computational
resources than the corresponding multiple search with Fuzzy CoCo. However, as the re-
sults are obtained in less generations, for equivalent computational efforts, Island Fuzzy
CoCo performs better than Fuzzy CoCo.
106 6. Extensions of the Methodology

Fitness curves for 2 rules Fitness curves for 3 rules

0.98 0.98

0.975 0.975

M =50; M =20 M =50; M =20

P S P S
No migration No migration
0.97 0.97
500 1000 1500 500 1000 1500

Fitness curves for 6 rules Fitness curves for 7 rules

0.98 0.98

0.975 0.975

MP=50; MS=20 MP=50; MS=20

No migration No migration
0.97 0.97
500 1000 1500 500 1000 1500

Figure 6.5 Effect of migration on evolutionary dynamics. The figure shows the average fitness in
different islands with and without migration. The abscissa represents the number of generations.

• Island Fuzzy CoCo should be implemented in parallel platforms, with each island exe-
cuted on a separate processor. The communication between the islands is sparse as it is
reduced to information about some migrants (i.e., genome and fitness value) and about
the global performance of island populations.

• The principle of Island Fuzzy CoCo, i.e., the use of islands containing individuals with
different levels of complexity, may be used for other optimization methods dealing with
variable complexity (e.g., the evolution of finite state machines [103] and the evolution
of fuzzy systems with different input-space dimensions [149]).

6.2 Incremental Fuzzy CoCo

Island Fuzzy CoCo—presented in Section 6.1—tackles the problem of finding an adequate size
of the rule base by performing concurrent evolutionary searches with different sizes. It is better
than Fuzzy CoCo for designing large systems, probably due to the information gathered in
low-complexity islands and shared through migration. It is, however, costly in computational
resources.
I propose thus an alternative, computationally cheaper, method inspired by the iterative
rule learning approach (see Section 2.2.2) [59], in which a single instance of Fuzzy CoCo
is used to search for fuzzy systems whose complexity increases as evolution advances. In
contrast to iterative rule learning methods, individuals in Incremental Fuzzy CoCo represent
entire fuzzy systems instead of one-rule systems. Below, I delineate Incremental Fuzzy CoCo
6.2 Incremental Fuzzy CoCo 107

and describe a first test performed to explore its potential search capabilities..

6.2.1 The proposed algorithm

Incremental Fuzzy CoCo starts as a simple instance of Fuzzy CoCo, used to search for small
fuzzy systems (i.e., with a reduced number of rules, usually one). This instance runs until
evolution satisfies a given criterion. At this point, part of the evolved population is used to
seed the initial population of a new instance of Fuzzy CoCo. This new instance, that is set up
to search for larger systems, runs until a new criterion is satisfied and a new instance with larger
individuals is launched. The process is repeated until a termination criterion is satisfied. Due
to the change of complexity, the genomes must be adapted before their use in a new instance
of Fuzzy CoCo. Figure 6.6 presents the Island Fuzzy CoCo algorithm in pseudo-code format.

begin Incremental Fuzzy CoCo

R = Rmin
Initialize Fuzzy CoCo populations for R rules PR (0)
while not done do
while increment criterion not satisfied do
Run Fuzzy CoCo: PR = F CC(PR (0))
PR (0) = PR
end while
QR = Select-seed[PR ]
R
= R + Rinc
Q
R = Adapt-genome[QR ]
PR
(0) = complete Q
R with random individuals with R
rules
R = R

end while
end Incremental Fuzzy CoCo

Figure 6.6 Pseudo-code of Incremental Fuzzy CoCo. A simple Fuzzy CoCo evolves a population of
systems of size R. Each time an increment criterion is satisfied, the sought-after complexity is increased
and part of the population is used to seed a new instance of Fuzzy CoCo.

The criteria used to decide to increase the complexity may be one or more of the follow-
ing: number of generations elapsed, average performance of the entire population or of a se-
lected elite, stagnation of evolution, or explicit user interaction. A part of the actual population
is used to seed the new, complexity-increased, population. Due to the increased complexity
in the new population, the genome of the selected seed must be adapted by duplicating some
active rules. Special care must be taken to ensure that fitness is preserved after this operation.
The remaining individuals in the initial population are generated randomly. In this way it is
possible to keep the flexibility of the search while launching the new search with known “good
individuals.”
108 6. Extensions of the Methodology

6.2.2 Testing Incremental Fuzzy CoCo

I present herein a first test of Incremental Fuzzy CoCo performed on the WBCD database. The
main goal of this test is to verify the potential search capabilities of the method.

6.2.2.1 Experimental setup.

Setting up Incremental Fuzzy CoCo requires the definition of the following parameters:

• Range of the number of rules. The algorithm starts with one-rule systems and works its
way up to ten-rule systems.

• Increment criterion. Complexity is increased after a given number of generations. The

number of generations allowed for each instance depends on the number of rules: it
starts at 200 generations for one-rule systems and goes up to 400 generations for ten-
rule systems.

• Fuzzy CoCo setup. All the instances of Fuzzy CoCo use the same setup. Table 6.4
delineates the values used for the Fuzzy CoCo parameters.

Table 6.4 Fuzzy CoCo set-up for all the instances in Incremental Fuzzy CoCo for the WBCD problem.

Parameter Value
Population size P s 100
Maximum generations Gmax 1000
Crossover probability Pc 1
Mutation probability Pm 0.15
Elitism rate Er 0.1
“Fit” cooperators Ncf 1
Random cooperators Ncr 1

• Number of seeding individuals. The best five percent of the evolved population—the
elite—is used to seed the new initial population. For the rule species, the genomes are
adapted to the new size by adding a new rule. Each individual of the selected elite is
used to generate three new individuals: the first is obtained by duplicating one of the
active rules, the other two by adding a random rule. The remaining individuals (i.e.,
85% of the population) are randomly initialized. In this way, the first 5% are known to
perform well, the following 10% explore promising regions of the new enlarged search
space, while the remaining 85% explore new regions.

6.2.2.2 Results.
Thirty-two runs were performed, all but one of which found systems whose classification
performance exceeds 98.0%. In particular, considering the best individual per run (i.e., the
evolved system with the highest classification success rate), 20 runs led to a fuzzy system
6.2 Incremental Fuzzy CoCo 109

whose performance exceeds 98.5%, and of these, 2 runs found systems whose performance
exceeds 98.7%.; these results are summarized in Figure 6.7. The best system found obtains
an overall classification rate of 98.83%. The average performance over the 32 runs is 98.50%.
(Note: The results presented here were obtained by Olivier Rutti in the course of a student
project.)

20

15

10

0
97.5 98 98.5 99

Figure 6.7 Summary of results of 32 evolutionary runs. The histogram depicts the number of systems
exhibiting a given classification performance level at the end of the evolutionary run.

Figure 6.8 shows the evolution of the classification performance during Incremental
Fuzzy CoCo runs.

(a) Average run (b) Best run

99 99

98.5

98 98

1 2 3 4 5 6 7 8 9 10 97.5 1 2 3 4 5 6 7 8 9 10

97 97

96.5

96 96
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

Figure 6.8 Evolution of classification performance in Incremental Fuzzy CoCo. The figures show
the classification performance for: (a) the average over 32 runs, and (b)the best run. The abscissa
represents the number of generations elapsed. The numbers 1 to 10 represent the number of rules of the
systems evolved in the corresponding interval.

Taking the results obtained by Incremental Fuzzy CoCo at the end of the search for seven-
rule systems (corresponding to 1650 generations), we can compare them with the results of
Fuzzy CoCo when searching for seven-rule systems presented in Section 4.1.2. Table 6.5
110 6. Extensions of the Methodology

presents this comparison, while Figure 6.9 extends the comparison to other rule base sizes.
The best systems found by Incremental Fuzzy CoCo are worse than those found by Fuzzy
CoCo. However, the average performance of the former is better, and its results are obtained
with less generations than those of the latter.

Table 6.5 Comparison of seven-rule systems evolved by Incremental Fuzzy CoCo with those obtained
using Fuzzy CoCo.
Generations Average Best
Incremental Fuzzy CoCo 1650 98.43% 98.54%
Simple Fuzzy CoCo 1700 98.25% 98.98%

(a) Average fitness (b) Best fitness

98.9 98.9

98.7 98.7

98.5 98.5

98.3 98.3

98.1 98.1
Incremental Fuzzy CoCo Incremental Fuzzy CoCo
Simple Fuzzy CoCo Simple Fuzzy CoCo
97.9 97.9
3 4 5 6 7 3 4 5 6 7

Figure 6.9 Comparison of systems evolved by Incremental Fuzzy CoCo with those obtained using
Fuzzy CoCo. The figure shows the classification performance obtained at the end of the search for each
rule-base size.

6.2.2.3 Analysis of results

The test presented above constitutes a very first approach to the method. The results suggest,
however, some interesting features that deserve further exploration.

• Incremental Fuzzy CoCo seems to exhibit better repeatability than Fuzzy CoCo, as sug-
gested by the narrow distribution of the results (see Figure 6.7). This is true not only for
the global results but also for different rule-base sizes as shown in Figure 6.9. However,
this approach fails to find top systems as good as those found by Fuzzy CoCo. One of
the possible reasons is, again, that all the instances of Fuzzy CoCo were identically set
up.

• The use of a fixed number of generations to define the increment of the complexity
should be a weak criterion as the population can either converge toward mediocre per-
formances, if this number is too high, or be stopped before a good exploration of the
6.2 Incremental Fuzzy CoCo 111

search space is performed, if this number is too low. This criterion affects directly the
diversity and the quality of the initial population of the following instance.

• Intuitively, the seeding strategy (5% elite, 10% modified elite) is adequate, but the rate of
seeding individuals needs to be tuned, as it should be excessive if the elite is not diverse
enough.

• This method requires a deeper study to better understand the effects of the parameters
(i.e., Fuzzy CoCo, increment criterion, and seeding strategy) on its performance. The
goal of the study should be to find a setup that improves the quality of the best systems
while preserving as much as possible the repeatability of the results.

• Even if this goal is not attained, Incremental Fuzzy CoCo can be useful, as it is, for:

– searching automatically for an adequate range of sizes of the rule base, and
– estimating attainable performance values for a given problem, and for different
number of rules.
Chapter 7

Conclusions and Future Work

7.1 Summary
I presented a novel approach for system design—Fuzzy CoCo—based on fuzzy logic and
coevolutionary computation, which is conducive to explaining human decisions. My algorithm
has been able to find accurate and interpretable systems for hard, real-world problems. The
analysis of Fuzzy CoCo and of the systems it produced shows, among other features, the
consistency of the results. I also proposed two extensions to the method, which deserve further
exploration.
Evolutionary fuzzy modeling—i.e., the design of fuzzy inference systems using evolu-
tionary algorithms—constitutes the methodological base of my work. In Chapter 2 I studied
extant evolutionary fuzzy modeling approaches and some criteria and considerations involved
in this task. I emphasized an aspect usually neglected in most approaches: interpretability. In
particular, I presented some strategies to satisfy semantic and syntactic criteria that reinforce
the interpretability of the systems produced. To illustrate these concepts, I applied a basic
fuzzy-genetic approach to solve a medical diagnostic problem: the WBCD problem. The
systems obtained were the best explanatory systems presented at the time for this problem.
The aforementioned study brought to the fore some limitations of evolutionary fuzzy
modeling, but, at the same time, it provided clues on how to overcome them. Based on these
clues, I proposed the use of cooperative coevolution to surmount the problem of dealing with
different types of parameters in the same genome. This is the origin of Fuzzy CoCo, pre-
sented in detail in Chapter 3, together with a simple application example: the Iris classification
problem. The systems obtained for this problem surpassed previous fuzzy modeling results.
Fuzzy CoCo was then used in Chapter 4 to model the decision processes involved in two
breast-cancer diagnostic problems: the WBCD problem and the Catalonia mammography in-
terpretation problem. For the WBCD problem, Fuzzy CoCo produced markedly better results
using less or similar computational resources than the fuzzy-genetic approach. For the Catalo-
nia problem, an evolved system was embedded within a web-based tool—called COBRA—for
aiding radiologists in mammography interpretation.
In order to attain a deeper understanding of Fuzzy CoCo, I performed in Chapter 5 several
analyses regarding the performance of the methodology and of the systems it produces. These
analyses involve aspects like the application of the method, the effects that some parameters
have on performance, the consistency and the quality of the systems designed using Fuzzy

113
114 7. Conclusions and Future Work

CoCo, as well as their local generality. Finally, I proposed two extensions: Island Fuzzy CoCo
and Incremental Fuzzy CoCo, which together with the original CoCo constitute a family of
coevolutionary fuzzy modeling techniques. The extensions are intended to guide the choice
of an adequate number of rules for a given problem—a critical, hard-to-define parameter of
Fuzzy CoCo. The encouraging preliminary results obtained with these extensions motivate
further investigation.

7.2 Original Contributions

To the best of my knowledge, this thesis is the first work to analyze and develop in-depth
the concept of coevolutionary fuzzy modeling. While this in itself constitutes an original
achievement, it is not the only contribution of the present research to the state of the art. In
this section, I will outline the most interesting original contributions.
• In Section 2.1, which studies the fuzzy modeling problem, I proposed a novel classifi-
cation of the parameters of fuzzy systems into four classes: logic, structural, connective,
and operational. This classification was used to decompose the fuzzy modeling problem
and to analyze how existing modeling techniques dealt with it. These analyses provided
some key ideas, important to the conception of Fuzzy CoCo. In particular, they served
to show that most of the time the fuzzy modeling problem reduces to the design of con-
nective and operational parameters, which are very different in nature. This fact led to
the idea of applying cooperative coevolution.

• The interpretability considerations presented in Section 2.3 contain several original el-
ements. First, observing that most of the interpretability criteria presented in the litera-
ture were mainly oriented toward constraining the definition of membership functions,
I grouped them under the label semantic criteria, as they affect the coherence of the
linguistic concepts. Then, I identified and proposed some other criteria regarding the
rule base, which are called syntactic criteria, as they affect the (causal) connection be-
tween linguistic concepts. Finally, I proposed a set of modeling strategies that, when
applied, should reinforce the linguistic integrity—both semantic and syntactic—of the
fuzzy systems produced. Note that these considerations are valid for any fuzzy modeling
technique.

• Chapter 3 introduced Fuzzy CoCo, a novel cooperative coevolutionary approach to fuzzy

modeling. The idea of applying cooperative coevolution arose from the observation that
the fuzzy modeling problem should be decomposed into (at least) two separated but
related search processes: one for connective, the other for operational parameters. Co-
operative coevolution succeeds in overcoming some limitations exhibited by standard
evolutionary fuzzy modeling: stagnation, convergence to local optima, and computa-
tional costliness. Besides performance improvement, Fuzzy CoCo was designed with
interpretability being a prime goal. This was evidenced in the different applications of
Fuzzy CoCo, where I applied the proposed interpretability strategies.

• The systems obtained to solve the Catalonia mammography problem (Section 4.2),
served as a base for the development of the COBRA system—the Catalonia On-
7.3 Future Work 115

line Breast-cancer Risk Assessor—a web-based tool for aiding radiologists in

mammography interpretation.

• Finally, I proposed two extensions to Fuzzy CoCo, intended to simplify the task of find-
ing an adequate size of the rule base. Two key elements of the first extension, Island
Fuzzy CoCo, are: (1) the existence of evolutionary islands containing individuals of dif-
ferent sizes, and (2) a migration mechanism that adapts the size of the migrants to render
them compatible with their destination island. The second extension, Incremental Fuzzy
CoCo, bases its search power on a mechanism of incremental evolution. In contrast with
iterative rule learning methods, the search does not set a good system early on. Good
systems are only used to signal promising search regions as they seed the new initial
population. These methods represent two new members of a family of coevolutionary
fuzzy modeling techniques based on Fuzzy CoCo.

7.3 Future Work

Throughout this thesis I have suggested a number of possible directions for future research
within the domain of coevolutionary fuzzy modeling, and particularly for the development of
tools based on Fuzzy CoCo. In this section, I will briefly expand on some ideas.

Deepening and improving the methodology

In all the applications of Fuzzy CoCo described in this thesis, the evolutionary algorithms for
both species, membership functions and rules, were set up identically. However, a number of
questions concerning this choice could be addressed in the future:

• Studying the effects of the evolutionary parameters on the performance of each species
to determine specific optimal setups.

• Investigating other types of evolutionary algorithms for each species, for example, real-
coded ones—such as evolution strategies—for membership-function species, or tree-
based representations for rules species.

Exploration of Island Fuzzy CoCo

The encouraging performance of this algorithm in the preliminary tests (Section 6.1) invites
further exploration. Future work on this algorithm should concentrate, among others, on the
following issues:

• Tuning the setup, as the existence of migration between islands may change the effect of
evolutionary parameters on performance. This means revisiting the analysis presented
in Section 5.2.

• Defining a mechanism for dynamically finding the range of number of rules to perform
the search, e.g., stopping islands that stagnate in low performance values or creating
islands for larger individuals if performance suggests this being viable.
116 7. Conclusions and Future Work

• Comparing the performance of Island Fuzzy CoCo with that of simple Fuzzy CoCo (i.e.,
with zero migration) on harder problems. The island version shows higher performance
improvement for islands containing systems with a high number of rules. Given that
harder problems usually require larger systems, the algorithm seems adequate to solve
them.

• Parallel implementation. As mentioned before, the algorithm is well suited for paral-
lelization, implementing each island on a separate processor. There is no need for cen-
tral control, and the communication between the islands is reduced to information about
some migrants (their genome and fitness values) and about the global performance of
island populations.

Further development of Incremental Fuzzy CoCo

Although this algorithm was explored only preliminarily, it obtained some promising results
(Section 6.2). Future work on this algorithm could include:

• The setup of each instance of Fuzzy CoCo should be adapted to the features of the
systems it evolves. Parameters such as population size, mutation probability, elitism
rate, and number of cooperators should change as the complexity of the encoded systems
increases.

• This method requires a more thorough study to better understand the effects on perfor-
mance of its specific new parameters (i.e. increment criteria and seeding strategy).

• In the same way the complexity of the search is increased, there should exist some
criteria to decrease this complexity. For example, if the effective number of rules used
by (most of) the best systems is lower than the currently allowed maximum number of
rules.

Development of a fuzzy modeling toolbox

Several of the contributions afforded by this thesis could constitute elements for developing a
fuzzy modeling toolbox based on the Fuzzy CoCo family.

• Incremental Fuzzy CoCo can serve for early exploration of the problem space. It can
estimate both attainable performance values and an adequate range of sizes of the rule
base.

• Island Fuzzy CoCo could explore in-depth the range of complexities found in order to
validate the estimation obtained before.

• Simple Fuzzy CoCo could perform specific searches for systems with a given, user-
defined number of rules.

• An extended version of the user interface developed for the COBRA system could serve
for the interaction with both the fuzzy modeler and the end user.
7.3 Future Work 117

Integration within knowledge-engineering environments

As mentioned in Chapter1, knowledge engineering is the best alternative to designing large and
hierarchic explanatory systems. However, this process usually involves many smaller modules
whose design could well take advantage of the main modeling features of Fuzzy CoCo: high-
performance, data-based, and interpretability-oriented.
Appendix A

Evolutionary Computation in Medicine:

An Overview

“Evolution is the natural way to program.”

Thomas Ray

A.1 Using Adaptive Methods in Medicine

The Merriam-Webster OnLine Dictionary (www.m-w.com) defines medicine as “the science
and art dealing with the maintenance of health and the prevention, alleviation, or cure of dis-
ease.” In aiming to fulfill their defined mission, medical professionals are confronted daily with
problems from diverse walks of (medical) life, which all exhibit an underlying commonality:
Searching for a good solution among a (usually huge) space of possible solutions. Whether
trying to pry out signs of malignant cancer in cell biopsies or looking for irregularities in EEG
signals, the basic problem is that of sifting through a welter of candidate solutions to find as
best a solution as possible.
As in any other area of modern life, computers are omnipresent in medicine, from the
hospital accounting computer to the high-end MRI scanner. In particular, computers are used
as tools to abet medical professionals in resolving search problems. Numerous techniques
have been applied over the past few decades to solve medical problems: expert systems, ar-
tificial neural networks, linear programming, and database systems are but a sampling of the
approaches used. One of the relative newcomers in medicine is the approach known as evolu-
tionary computation, which is the object of focus in this paper.
The idea of applying the biological principle of natural evolution to artificial systems,
introduced more than four decades ago, has seen impressive growth in the past few years.
Known as evolutionary algorithms or evolutionary computation, these techniques are common
nowadays, having been successfully applied to numerous problems from different domains,
including optimization, automatic programming, circuit design, machine learning, economics,
ecology, and population genetics, to mention but a few. In particular, evolutionary algorithms
have been applied to problems in medicine.
This annex provides an overview of evolutionary computation in medicine. The next sec-
tion describes how evolutionary algorithms are applied to solve medical problems, including

119
120 A. Evolutionary Computation in Medicine: An Overview

diagnosis, prognosis, imaging, signal processing, planning, and scheduling, while section A.3
provides an extensive bibliography, classified both according to the medical task addressed
and according to the evolutionary technique used.

A.2 Applying Evolutionary Computation to Solve Medical

Problems
Most (if not all) medical decisions can be formulated as a search in some appropriate space.
For example, a pathologist analyzing biopsies to decide whether they are malignant or not,
is searching in the space of all possible cell features for a set of features permitting him to
provide a clear diagnosis [125, 129]. A radiologist planning a sequence of radiation doses is
searching for the best treatment in the space of all possible treatments [191].
Medical search spaces are usually very large and complex. Medical decisions are based
on clinical tests which provide huge amounts of data. Based on these data one must ultimately
make a single decision (e.g., benign or malignant). Given the tight interdependency among
the domain variables, and the inherent non-linearity of most real-world problems, neighboring
points in the search space may have widely differing qualities, turning the search into a com-
plex task. Indeed, due to this complexity, several medical problems are used as benchmarks to
test and compare machine learning techniques [102, 105].
Evolutionary computation provides powerful techniques for searching complex spaces.
As stated in Section 1.3, evolutionary techniques exploit mechanisms of natural evolution to
search efficiently in a given space. Their intrinsic parallelism diminishes the risk of the search
being trapped in a local optimum.
The construction of accurate models of medical decision from extant knowledge is a hard
task. On the one hand, the models involve too many non-linear and uncertain parameters to be
treated analytically. On the other hand, medical experts are usually not available, or simply do
not collaborate in translating their experience into a usable decision tool.
Nevertheless, there is a large number of accessible medical databases. Currently, medical
results are electronically stored and accumulated in databases so as to serve both as a record
of patients’ history, and as a source of medical knowledge. The amount of available data is
increasing continuously, and therefore its exploitation requires the use of more sophisticated
computational processing tools.
Evolutionary computation is applied in medicine to perform several types of tasks. When-
ever a decision is required in medicine, it is usually possible to find a niche for evolutionary
techniques. The tasks performed by evolutionary algorithms in the medical domain can be
divided into three groups: (1) data mining, mainly as applied to diagnosis and prognosis, (2)
medical imaging and signal processing, and (3) planning and scheduling.

A.2.1 Data mining

Data mining, also known as knowledge discovery, is the process of finding patterns, trends, and
regularities by sifting through large amounts of data [39]. Data mining involves the analysis
of data stored in databases to discover associations or patterns, to segment (or cluster) records
based on similarity of attributes, and to create predictive (or classification) models.
A.2 Applying Evolutionary Computation to Solve Medical Problems 121

There are two major approaches to data mining: supervised and unsupervised. In the
supervised approach, specific examples of a target concept are given, and the goal is to learn
how to recognize members of the class using the description attributes. In the unsupervised
approach, a set of examples is provided without any prior classification, and the goal is to
discover underlying regularities and patterns, most often by identifying clusters or subsets of
similar examples.
Clinical databases have accumulated large amounts of data on patients and their medical
conditions. The clinical history of a patient generates data that goes beyond the disease being
treated. This information, stored along with that of other patients, constitutes a good place to
look for new relationships and patterns, or to validate proposed hypotheses.
The range of applications of data mining in medicine is very wide, with the two most
popular applications being diagnosis and prognosis. Diagnosis is the process of selectively
gathering information concerning a patient, and interpreting it according to previous knowl-
edge, as evidence for or against the presence or absence of disorders [92]. In a prognostic
process, a patient’s information is also gathered and interpreted, but the objective is to predict
the future development of the patient’s condition. Due to the predictive nature of this process,
prognostic systems are frequently used as tools to plan medical treatments [93].
The role played by data mining in the context of diagnosis and prognosis is the discovery
of the knowledge necessary to interpret the gathered information. In some cases this knowl-
edge is expressed as probabilistic relationships between clinical features and the proposed
diagnosis or prognosis. In other cases a rule-based representation is chosen so as to provide
the physician with an explanation of the decision. Finally, in yet other cases, the system is
designed as a black-box decision maker that is totally unconcerned with the interpretation of
its decisions.
Evolutionary computation is usually applied in medical data mining as a parameter finder.
Evolutionary techniques search for the parameter values of the knowledge representation set
up by the designer so that the mined data are optimally interpreted. For example, evolutionary
algorithms can search for the weights of a neural network, the membership function values of
a fuzzy system, or the coefficients of a linear regressor.

A.2.2 Medical imaging and signal processing

Much of the medical data are expressed as images or other types of temporal signals. Many
exams, such as magnetic resonance or mammography, provide as result a group of images.
Sometimes the number of images is high and the important information is distributed among
all of them. Other types of exams, like electroencephalography (EEG), provide results in the
form of time-variant signals. In these tests, the information is hidden within the temporal
features. The fields of signal and image processing have developed tools to deal with such
data. Techniques such as filtering, compression, segmentation, and pattern recognition allow
to extract the desired features from signals.
Signal processing in medicine is subject to several important constraints. First, the num-
ber of signals to be processed is high, and often tightly interdependent. Second, signals are
unique, in the sense that the circumstances under which they were obtained are normally not
repeatable. Third, given the characteristics of their sources, medical signals are often very
noisy. Finally, in some cases information about the signals is required in real time in order to
122 A. Evolutionary Computation in Medicine: An Overview

take crucial decisions.

Evolutionary-computation techniques are used in different ways in medical imaging and
signal processing. In some cases they are applied to improve the performance of signal-
processing algorithms (e.g., filters or compressors) by finding their optimal parameters. Other
works use evolutionary algorithms directly to extract useful information from the welter of
data.

A.2.3 Planning and scheduling

Evolutionary computation is frequently applied in problems of planning and scheduling. These
tasks involve the assignment of resources subject to several constraints. In a planning problem,
several resources (or operations) are needed in order to accomplish a task, and the planning
process tries to define the sequence in which they are applied. In a scheduling problem one or
several resources have to be distributed among several tasks, and the problem is to find a dis-
tribution (spatial and temporal) for this assignment so that all tasks are executed. Evolutionary
techniques are well suited to solve these kinds of problems thanks to their ability to search
complex spaces.
In medicine, planning and scheduling are widely used in clinical and hospital-
administration applications. Clinical applications involve mainly planning problems; for
example, in 3D radiotherapy it is necessary to plan carefully the doses applied for several
different radioactive beams, respecting clinical and geometrical constraints [38]. Hospital
administrators deal with problems of personnel, patient, and resource management and
control. An example is the problem of scheduling a patient for undergoing different medical
procedures and seeing different physicians, optimizing both patient waiting time and aparatus
utilization [136].

A.3 Classified Bibliography

This section contains a bibliography of articles dealing with evolutionary computation in
medicine, classified both according to the medical task and according to the evolutionary tech-
nique. I have chosen to concentrate mainly on articles in archival journals so as to limit the
explosive number of references. Table A.1 summarizes these two classifications.

A.3.1 According to the medical task

1. Data Mining.

(a) Diagnosis [7–9, 11, 12, 14, 46, 52, 54, 70, 85, 88, 114, 122, 125, 138, 146, 153, 188].
Diagnosis is the process of selectively gathering information concerning a patient,
and interpreting it according to previous knowledge, as evidence for or against
the presence or absence of disorders (Section A.2.1). The papers in this category
apply evolutionary algorithms to solve numerous diagnostic problems, including:
patient’s general condition evaluation, location of primary tumor, detection of hya-
line membrane disease in preterm newborn infants, detection of breast cancer cells
 A.3 Classified Bibliography
Medical task Data Mining Medical imaging Planning and
and
Evolutionary technique Diagnosis Prognosis signal processing scheduling

Table A.1 Classified bibliography: summary.

Unidimensional, binary [7–9,70,85,88, [2, 22, 25, 33, [5, 20, 24, 29, 31, 42, [187, 189–191]
125, 138, 153] 72, 73, 85, 86, 53, 58, 60, 83, 108,
98,99,110,165, 115, 154, 167, 174,
166] 176, 179, 197]
Genetic Multidimensional [28, 30, 55] [23, 136, 192]
algorithms Real-valued [54] [3, 100, 156] [38, 57, 87]
Rule-encoding [11, 12] [12]
Indexed [84]
Genetic programming [11, 52, 114]
Evolution strategies [7–9] [135]
Evolutionary programming [46, 114, 188]
Classifier systems [14, 122, 146] [14]
Hybrid Evolutionary-fuzzy [7–9, 70, 125] [86] [176]
systems Evolutionary-neural [7–9, 46, 122, [22, 25, 33, [24, 31, 58]
188] 72, 73, 98, 110,
166]

123
124 A. Evolutionary Computation in Medicine: An Overview

in peripheral blood flow, lymphography classification, brain tumor classification,

risk evaluation of heart and coronary artery diseases, profiling of risk factors in
diabetes and female urinary incontinence, and the detection of different diseases
from clinical tests (infarcts from radiography, arrythmias from electrocardiogra-
phy, neuromuscular disorders from electromyography, sleep profiling from elec-
troencephalography, and breast cancer from mammography).
(b) Prognosis [2,14,22,25,33,72,73,85,86,98,99,110,165,166]. Prognosis is the pro-
cess of selectively gathering information concerning a patient, and interpreting it
according to previous knowledge, in order to predict the future development of the
patient’s condition (Section A.2.1). The papers in this category involve prognos-
tic applications of evolutionary computation, including: breast cancer recurrency,
donor compatibility for transplants in highly sensitized patients, outcome of duo-
denal ulcer, hypoxic resistance on jet pilots, outcome of intensive-care patients,
after-surgery response for patients with lung cancer, prediction of depression after
mania, prevision of tractolimus blood level in liver tansplantation patients, and pa-
tient’s survival estimation in different types of cancer (malignant skin melanoma,
lung cancer, colorectal cancer, and gestational trophoblastic tumours).

2. Medical imaging and signal processing [3, 5, 20, 24, 28–31, 42, 53, 55, 58, 60, 83, 84, 100,
108,115,154,156,167,174,176,179,197]. The fields of medical imaging and signal pro-
cessing have developed tools to deal with huge amounts of data expressed as images or
other types of temporal signals. All but one of the papers in this category deal with prob-
lems related to clinical tests, including: thorax radiography, retinal and cardiac angiog-
raphy, computarized tomography, magnetoencephalography, ultrasound imaging, elec-
troencephalography, electrocardiography, radiographic cephalography, laser profilome-
try, and many applications of both mammography and magnetic resonance. The “odd
paper out” presents an application of surgery assistance [5].

3. Planning and scheduling [12, 23, 38, 57, 87, 135, 136, 187, 189–192]. Planning and
scheduling involve the assignment of resources to accomplish one or more tasks subject
to several constraints (Section A.2.3). The papers in this category use evolutionary
techniques to solve problems such as: allocation of hospital resources, electrical carotid
sinus nerve stimulation, radiologist allocation, three-dimensional radiation therapy
treatment planning, dosimetric preplanning and treatment planning of permanent
prostate implants, patient scheduling in highly constrained situations, and stereotactic
radiosurgery planning.

A.3.2 According to the evolutionary technique

1. Genetic algorithms. As mentioned in Section 1.3, genetic algorithms are the best-known
class of evolutionary algorithms and their use is so extensive that often the terms genetic
algorithms and evolutionary computation are used interchangeably. The main differ-
ence between genetic algorithms and other evolutionary algorithms is the representation
(genome), which I use as a criterion to further subdivide this class:
A.3 Classified Bibliography 125

(a) Unidimensional, binary genome [2,5,7–9,20,22,24,25,29,31,33,42,53,58,60,70,

72,73,83,85,86,88,98,99,108,110,115,125,138,153,154,165–167,174,176,179,
187, 189–191, 197]. This is the most popular representation in genetic algorithms
because it is simple to use and applicable to almost any problem [171].
(b) Multidimensional genome [23, 28, 30, 55, 136, 192]. In many cases the nature of
the problem (e.g., bidimensional medical imaging problems) suggests the use of
matrices or multidimensional arrays to represent the candidate solutions.
(c) Real-valued genome [3,38,54,57,87,100,156]. The binary representation has some
drawbacks when applied to high-precision numerical problems [104]. In particu-
lar, for parameter-optimization problems with variables from continous domains,
a real-valued representation is sometimes more efficient (and more natural) than a
binary one.
(d) Rule-encoding genome [11, 12]. Rule-based systems are usually encoded in ge-
netic algorithms using a binary representation. However, some works, such as the
two papers of this subcategory, represent directly in the genome the rules of the
knowledge base.
(e) Indexed representation [84]. In this representation the genome is encoded using
a finite-length alphabet (non-binary) . Each symbol of the alphabet indexes an
element, and the presence of a symbol in the genotype indicates the presence of
the element in the phenotype. In the referenced paper the indexed elements are
“leafs” of a known tree representation (note that this is still a genetic-algorithm
string representation and not a genetic-programming tree).

2. Genetic programming [11, 52, 114]. In genetic programming solutions are encoded as
computer programs rather than as fixed-length character strings (Section 1.3.2).
3. Evolution strategies [7–9, 135]. Evolution strategies are well suited for parameter-
optimization problems. They use mainly the mutation operator. A major characteristic
of evolution strategies is that mutation values are evolved along with the parameters
being optimized (Section 1.3.3).
4. Evolutionary programming [46, 114, 188]. In evolutionary programming, individuals
are represented by finite state machines, which provide a meaningful representation of
behavior based on interpretation of environmental symbols (Section 1.3.4).
5. Classifier systems [14, 122, 146]. Classifier systems are evolution-based learning sys-
tems. They can be viewed as restricted versions of classical rule-based systems that add
interaction with the exterior thanks to input and output interfaces (Section 1.3.5).
6. Hybrid approaches

(a) Evolutionary-fuzzy systems [7–9, 70, 86, 125, 176]. In evolutionary-fuzzy systems,
the capability of expressing knowledge in a linguistic, “human-friendly” way of-
fered by fuzzy logic, is combined with the power of evolutionary algorithms to
search and optimize. Thus one obtains systems with both high performance as
well as high interpretability.
126 A. Evolutionary Computation in Medicine: An Overview

(b) Evolutionary-neural systems [7–9, 22, 24, 25, 31, 33, 46, 58, 72, 73, 98, 110, 122, 166,
188]. With evolutionary-neural systems evolution and learning strategies work in
concert to attain adaptation.
Bibliography

[1] J. T. Alander. An indexed bibliography of genetic algorithms with fuzzy logic. In

Pedrycz [133], pages 299–318.

[2] S. S. Anand, A. E. Smith, P. W. Hamilton, J. S. Anand, J. G. Hughes, and P. H. Bar-

tels. An evaluation of intelligent prognostic systems for colorectal cancer. Artificial
Intelligence in Medicine, 15(2):193–214, February 1999.

[3] M. A. Anastasio, H. Yoshida, R. Nagel, R. M. Nishikawa, and K. Doi. A genetic

algorithm-based method for optimizing the performance of a computer-aided diagno-
sis scheme for detection of clustered microcalcifications in mammograms. Medical
Physics, 25(9):1613–20, September 1998.

[4] T. Bäck, U. Hammel, and H.-P. Schwefel. Evolutionary computation: Comments on the
history and current state. IEEE Transactions on Evolutionary Computation, 1(1):3–17,
April 1997.

[5] S. Baluja and D. Simon. Evolution-based methods for selecting point data for object
localization: Applications to computer-assisted surgery. Applied Intelligence, 8(1):7–
19, January 1998.

[6] Wolfgang Banzhaf, Peter Nordin, Robert E. Keller, and Frank D. Francone. Genetic
Programming – An Introduction; On the Automatic Evolution of Computer Programs
and its Applications. Morgan Kaufmann, dpunkt.verlag, January 1998.

[7] R. Baumgart-Schmitt, R. Eilers, and W. M. Herrmann. On the use of neural network

techniques to analyse sleep EEG data. Second communication: Training of evolutionary
optimized neural networks on the basis of multiple subjects data and the application of
context rules according to rechtschaffen and Kales. Somnologie, 1(4):171–183, 1997.

[8] R. Baumgart-Schmitt, W. M. Herrmann, and R. Eilers. On the use of neural network

techniques to analyze sleep EEG data. Third communication: Robustification of the
classificator by applying an algorithm obtained from 9 different networks. Neuropsy-
chobiology, 37(1):49–58, 1998.

[9] R. Baumgart-Schmitt, W. M. Herrmann, R. Eilers, and F. Bes. On the use of neural

network techniques to analyse sleep EEG data. First communication: Application of
evolutionary and genetic algorithms to reduce the feature space and to develop classifi-
cation rules. Neuropsychobiology, 36(4):194–210, 1997.

127
128 Bibliography

[10] K. P. Bennett and O. L. Mangasarian. Neural network training via linear programming.
In P. M. Pardalos, editor, Advances in Optimization and Parallel Computing, pages 56–
57. Elsevier Science, 1992.

[11] A. S. Bickel and R. W. Bickel. Tree structured rules in genetic algorithms. In John J.
Grefenstette, editor, Genetic Algorithms and their Applications: Proceedings of the sec-
ond International Conference on Genetic Algorithms, pages 77–81, MIT, Cambridge,
MA, USA, July 1987. Lawrence Erlbaum Associates.

[12] A. S. Bickel and R. W. Bickel. Determination of near-optimum use of diagnostic

resources using the “GENES” genetic algorithm shell. Computers in Biology and
Medicine, 20(1):1–13, 1990.

[13] C. L. Blake and C. J. Merz. UCI repository of machine learning databases, 1998.

[14] P. Bonelli and A. Parodi. An efficient classifier system and its experimental compari-
son with two representative learning methods on three medical domains. In R. Belew
and L. Booker, editors, Proceedings of the Fourth International Conference on Genetic
Algorithms, pages 288–295, San Mateo, CA, July 1991. Morgan Kaufman.

[15] C. Bonivento, A. Davalli, and C. Fantuzzi. Tuning of myoelectric prostheses using fuzzy
logic. Artificial Intelligence in Medicine, 21(1–3):221–225, January–March 2001.

[16] G. E. P. Box. Evolutionary operation: A method for increasing industrial productivity.

Applied Statistics, 6(2):81–101, 1957.

[17] G. E. P. Box and J. S. Hunter. Condensed calculations for evolutionary operation pro-
grams. Technometrics, 1:77–95, 1959.

[18] H. Brenner. Measures of differential diagnostic value of diagnostic procedures. Journal

of Clinical Epidemiology, 49(12):1435–1439, December 1996.

[19] D. Chakraborty and N.R. Pal. Integrated feature analysis and fuzzy rule-based system
identification in a neuro-fuzzy paradigm. IEEE Transactions on Systems, Man, and
Cybernetics, Part B: Cybernetics, 31(3):391–400, June 2001.

[20] H.-P. Chan, B. Sahiner, K. L. Lam, N. Petrick, M. A. Helvie, M. M. Goodsitt, and D. D.

Adler. Computerized analysis of mammographic microcalcifications in morphological
and texture feature spaces. Medical Physics, 25(10):2007–19, October 1998.

[21] B.-S. Chen, S.-C. Feng, and K.-C. Wang. Traffic modeling, prediction, and congestion
control for high-speed networks: A fuzzy AR approach. IEEE Transactions on Fuzzy
Systems, 8(5):491–508, October 2000.

[22] H.-Y. Chen, T. C. Chen, D. I. Min, G. W. Fischer, and Y.-M. Wu. Prediction of
tacrolimus blood levels by using the neural network with genetic algorithm in liver
transplantation patients. Therapeutic Drug Monitoring, 21(1):50–56, February 1999.
Bibliography 129

[23] Y. Chen, M. Narita, M. Tsuji, and S. Sa. A genetic algorithm approach to optimization
for the radiological worker allocation problem. Health Physics, 70(2):180–6, February
1996.

[24] Y.-T. Chen, K.-S. Cheng, and J.-K. Liu. Improving cephalogram analysis through
feature subimage extraction. IEEE Engineering in Medicine and Biology Magazine,
18(1):25–31, January-February 1999.

[25] B. D. Clark and S. W. Leong. Crossmatch prediction of highly sensitized patients.

Clinical Transplants, pages 435–55, 1992.

[26] O. Cordón, F. Herrera, and M. Lozano. On the combination of fuzzy logic and evolu-
tionary computation: A short review and bibliography. In Pedrycz [133], pages 33–56.

[27] P. Darwen and X. Yao. Every niching method has its niche: Fitness sharing and implicit
sharing compared. In H.-M. Voigt, W. Ebeling, I. Rechenberg, and H.-P. Schwefel,
editors, Parallel Problem Solving from Nature – PPSN IV, pages 398–407, Berlin, 1996.
Springer.

[28] K. K. Delibasis, P. E. Undrill, and G. Cameron. Designing texture filters with genetic
algorithms: An application to medical images. Signal Processing, 57:19–33, 1997.

[29] K. K. Delibasis, P. E. Undrill, and G. G. Cameron. Genetic algorithm implementation

of stack filter design for image restoration. IEE Proceedings: Vision, Image and Signal
Processing, 143(3):177–183, June 1996.

[30] K. K. Delibasis, P. E. Undrill, and G. G. Cameron. Designing Fourier descriptor-based

geometric models for object interpretation in medical images using genetic algorithms.
Computer Vision and Image Understanding: CVIU, 66(3):286–300, June 1997.

[31] A. P. Dhawan, Y. Chitre, and C. Kaiser-Bonasso. Analysis of mammographic micro-

calcifications using gray-level image structure features. IEEE Transactions on Medical
Imaging, 15(3):246–59, June 1996.

[32] W. Duch, R. Adamczak, and K. Grabczewski. A new methodology of extraction, opti-

mization and application of crisp and fuzzy logical rules. IEEE Transactions on Neural
Networks, 12(2):277–306, March 2001.

[33] R. Dybowski, P. Weller, R. Chang, and V. Gant. Prediction of outcome in critically

ill patients using artificial neural network synthesised by genetic algorithm. Lancet,
347(9009):1146–50, April 27 1996.

[34] M. O. Efe and O. Kaynak. On stabilization of gradient-based training strategies for

computationally intelligent systems. IEEE-Transactions on Fuzzy Systems, 8(5):564–
575, October 2000.

[35] M. O. Efe and O. Kaynak. A novel optimization procedure for training of fuzzy in-
ference systems by combining variable structure systems technique and levenbergŕmar-
quardt algorithm. Fuzzy Sets and Systems, 122(1):153–165, August 2001.
130 Bibliography

[36] R. Eriksson. Applying cooperative coevolution to inventory control parameter opti-

mization. Master’s thesis, University of Skövde, 1996.

[37] J. Espinosa and J. Vandewalle. Constructing fuzzy models with linguistic integrity from
numerical data-AFRELI algorithm. IEEE Transactions on Fuzzy Systems, 8(5):591–
600, October 2000.

[38] G. A. Ezzell. Genetic and geometric optimization of three-dimensional radiation ther-

apy treatment planning. Medical Physics, 23(3):293–305, March 1996.

[39] U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, editors. Advances

in Knowledge Discovery and Data Mining. AIII Press/MIT Press, March 1996.

[40] F. Fernandez, G. Galeano, and J. A. Gomez. Comparing synchronous and asynchronous

parallel and distributed GP models. In E. Lutton, J. A. Foster, J. Miller, C. Ryan, and
A. G. B. Tettamanzi, editors, Proceedings of the 4th European Conference on Genetic
Programming, EuroGP 2002, volume 2278 of LNCS, pages 327–336, Kinsale, Ireland,
3-5 April 2002. Springer-Verlag.

[41] R. A. Fisher. The use of multiple measurements in taxonomic problems. Annals of

Eugenics, 7:179–188, 1936.

[42] J. M. Fitzpatrick, J. J. Grefenstette, and D. Van Gucht. Image registration by genetic

search. In Proceedings of the IEEE SouthEast conference SOUTHEASTCON 84, pages
460–464, 1984.

[43] M. Si Fodil, P. Siarry, F. Guely, and J.-L. Tyran. A fuzzy rule base for the improved
control of a pressurized water nuclear reactor. IEEE Transactions on Fuzzy Systems,
8(1):1–10, February 2000.

[44] D. B. Fogel, editor. Evolutionary Computation: The Fossil Record. IEEE Press, Piscat-
away, NJ, 1998.

[45] D. B. Fogel. Evolutionary Computation: Toward a New Philosophy of Machine Intelli-

gence. IEEE Press, Piscataway, NJ, second edition, 1999.

[46] D. B. Fogel, E. C. Wasson III, E. M. Boughton, and V. W. Porto. Evolving artificial

neural networks for screening features from mammograms. Artificial Intelligence in
Medicine, 14(3):317–326, November 1998.

[47] L. J. Fogel. Autonomous automata. Industrial Research, 4:14–19, 1962.

[48] L. J. Fogel, A. J. Owens, and M. J. Walsh. Artificial Intelligence through Simulated

Evolution. John Wiley & Sons, New York, 1966.

[49] R. M. Friedberg. A learning machine: I. IBM Journal of Research and Development,

2:2–13, 1958.

[50] R. M. Friedberg, B. Dunham, and J. H. North. A learning machine. II. IBM Journal of
Research and Development, 3:282–287, 1959.
Bibliography 131

[51] D. E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning.

Addison-Wesley, Reading, Mass., 1989.

[52] H. F. Gray, R. J. Maxwell, I. Martinez-Perez, C. Arus, and S. Cerdan. Genetic program-

ming for classification and feature selection: Analysis of 1 H nuclear magnetic resonance
spectra from human brain tumour biopsies. NMR in Biomedicine, 11(4-5):217–224,
June-August 1998.

[53] J. J. Grefenstette and M. J. Fitzpatrick. Genetic search with approximate function evalu-
ation. In John J. Grefenstette, editor, Proceedings of the 1st International Conference on
Genetic Algorithms and their Applications, pages 112–120, Pittsburgh, PA, July 1985.
Lawrence Erlbaum Associates.

[54] H. J. Gross, B. Verwer, D. Houck, R. A. Hoffman, and D. Recktenwald. Model study

detecting breast cancer cells in peripheral blood mononuclear cells at frequencies as low
as 10(-7). Proceedings of The National Academy of Sciences of The United States of
America, 92(2):537–41, January 17 1995.

[55] M. Gudmundsson, E. A. El-Kwae, and M. R. Kabuka. Edge detection in medical images

using a genetic algorithm. IEEE Transactions on Medical Imaging, 17(3):469–74, June
1998.

[56] S. Guillaume. Designing fuzzy inference systems from data: An interpretability-

oriented review. IEEE Transactions on Fuzzy Systems, 9(3):426–443, June 2001.

[57] O. C. L. Haas, K. J. Burnham, and J. A. Mills. Optimization of beam orientation in

radiotherapy using planar geometry. Physics in Medicine and Biology, 43(8):2179–
2193, August 1998.

[58] H. Handels, Th. Roβ, J. Kreusch, H. H. Wolff, and S. J. Pöppl. Feature selection for
optimized skin tumor recognition using genetic algorithms. Artificial Intelligence In
Medicine, 16(3):283–297, 1999.

[59] F. Herrera, M. Lozano, and J. L. Verdegay. Generating fuzzy rules from examples using
genetic algorithms. In B. Bouchon-Meunier, R. R. Yager, and L. A. Zadeh, editors,
Fuzzy Logic and Soft Computing, pages 11–20. World Scientific, 1995.

[60] A. Hill, T. F. Cootes, C. J. Taylor, and K. Lindley. Medical image interpretation:

A generic approach using deformable templates. Medical Informatics, 19(1):47–59,
January-March 1994.

[61] W. Daniel Hillis. Co-evolving parasites improve simulated evolution as an optimization

procedure. Physica D, 42:228–234, 1990. Proceedings of the Ninth Annual Int. Conf.
of the Center for Nonlinear Studies, Ed. Stephanie Forrest.

[62] J. H. Holland. Outline for a logical theory of adaptive systems. Journal of the ACM,
9(3):297–314, July 1962.
132 Bibliography

[63] J. H. Holland. Processing and processors for schemata. In E. L. Jacks, editor, Associa-
tive information processing, pages 127–146. New York: American Elsevier, 1971.

[64] J. H. Holland. Adpatation in Natural and Artificial Systems. University of Michigan

Press, Ann Arbor, MI, 1975.

[65] T. P. Hong and J. B. Chen. Processing individual fuzzy attributes for fuzzy rule induc-
tion. Fuzzy Sets and Systems, 112(1):127–140, May 2000.

[66] Jörn Hopf. Cooperative coevolution of fuzzy rules. In N. Steele, editor, Proceedings
of the 2nd International ICSC Symposium on Fuzzy Logic and Applications (ISFL-97),
pages 377–381, Zurich, Switzerland, 1997. ICSC Academic Press.

[67] C. A. Hung and S. F. Lin. An incremental learning neural network for pattern clas-
sification. International Journal of Pattern Recognition and Artificial Intelligence,
13(6):913–928, September 1999.

[68] P. Husbands and F. Mill. Simulated co-evolution as the mechanism for emergent plan-
ning and scheduling. In R. Belew and K. Booker, editors, Proceedings of the 4th In-
ternational Conference on Genetic Algorithms, pages 264–270, San Diego, CA, July
1991. Morgan Kaufmann.

[69] H. Ishibushi, T. Nakashima, and T. Murata. Performance evaluation of fuzzy classifier

systems for multidimensional pattern classification problems. IEEE Transactions on
Systems, Man, and Cybernetics. Part B: Cybernetics, 29(5):601–618, October 1999.

[70] R. Jain, J. Mazumdar, and W. Moran. Application of fuzzy-classifier system to coronary

artery disease and breast cancer. Australasian Physical and Engineering Sciences in
Medicine, 21(3):141–147, September 1998.

[71] J.-S. R. Jang and C.-T. Sun. Neuro-fuzzy modeling and control. Proceedings of the
IEEE, 83(3):378–406, March 1995.

[72] M. F. Jefferson, N. Pendleton, C. P. Lucas, S. B. Lucas, and M. A. Horan. Evolution of

artificial neural network architecture: Prediction of depression after mania. Methods of
Information in Medicine, 37(3):220–225, September 1998.

[73] M. F. Jefferson, N. Pendleton, S. B. Lucas, and M. A. Horan. Comparison of a genetic

algorithm neural network with logistic regression for predicting outcome after surgery
for patients with nonsmall cell lung carcinoma. Cancer, 79(7):1338–42, April 1 1997.

[74] Jonghyeok Jeong and Se-Young Oh. Automatic rule gerenartion for fuzzy logic con-
trollers using rule-level coevolution of subpopulations. In 1999 Congress on Evolution-
ary Computation, pages 2151–2156, Piscataway, NJ, 1999. IEEE Service Center.

[75] Y. Jin. Fuzzy modeling of high-dimensional systems: complexity reduction and inter-
pretability improvement. IEEE Transactions on Fuzzy Systems, 8(2):335–344, April
2000.
Bibliography 133

[76] C.-F. Juang, J.-Y. Lin, and C.-T. Lin. Genetic reinforcement learning through symbi-
otic evolution for fuzzy controller design. IEEE Transactions on Systems, Man and
Cybernetics, Part B: Cybernetics, 30(2):290–302, April 2000.

[77] H. Juillé. Methods for Statistical Inference: Extending the Evolutionary Computation
Paradigm. PhD thesis, Brandeis University, May 1999.

[78] H.-B. Jun, C.-S. Joung, and K.-B. Sim. Co-evolution of fuzzy rules and membership
functions. In Proceedings of 3rd Asian Fuzzy Systems Symposium (AFSS’98), 1998.

[79] C. L. Karr. Genetic algorithms for fuzzy controllers. AI Expert, 6(2):26–33, February
1991.

[80] C. L. Karr, L. M. Freeman, and D. L. Meredith. Improved fuzzy process control of

spacecraft terminal rendezvous using a genetic algorithm. In G. Rodriguez, editor, Pro-
ceedings of Intelligent Control and Adaptive Systems Conference, volume 1196, pages
274–288. SPIE, February 1990.

[81] J. R. Koza. Genetic programming: A paradigm for genetically breeding populations of

computer programs to solve problems. Technical Report STAN-CS-90-1314, Dept. of
Computer Science, Stanford University, June 1990.

[82] J. R. Koza. Genetic Programming. The MIT Press, Cambridge, MA, 1992.

[83] N. Kumaravel and N. Nithiyanandam. Genetic-algorithm cancellation of sinusoidal

powerline interference in electrocardiograms. Medical and Biological Engineering and
Computing, 36(2):191–196, March 1998.

[84] N. Kumaravel, J. Rajesh, and N. Nithiyanandam. Equivalent tree representation of

electrocardiogram using genetic algorithm. Biomed Sciences Instrumentation, 33:573–
8, 1997.

[85] L. Kuncheva. Genetic algorithm for feature selection for parallel classifiers. Information
Processing Letters, 46(4):163–168, June 1993.

[86] L. Kuncheva and K. Andreeva. Dream: A shell-like software system for medical
data analysis and decision support. Computer Methods and Programs in Biomedicine,
40(2):73–81, June 1993.

[87] M. Langer, R. Brown, S. Morrill, R. Lane, and O. Lee. A generic genetic algorithm for
generating beam weights. Medical Physics, 23(6):965–71, June 1996.

[88] J. Laurikkala and M. Juhola. A genetic-based machine learning system to discover the
diagnostic rules for female urinary incontinence. Computer Methods and Programs in
Biomedicine, 55(3):217–228, March 1998.

[89] C.-T. Lin and C. S. G. Lee. Neural-network-based fuzzy logic control and decision
system. IEEE Transactions on Computers, 40(12):1320–1336, December 1991.
134 Bibliography

[90] P. Lindskog. Fuzzy identification from a grey box modeling point of view. In H. Hel-
lendoorn and D. Driankov, editors, Fuzzy Model Identification, pages 3–50. Springer-
Verlag, 1997.

[91] Y.L. Loukas. Adaptive neuro-fuzzy inference system: An instant and architecture-free
predictor for improved QSAR studies. Journal of Medicinal Chemistry, 44(17):2772–
2783, August 2001.

[92] P. J. F. Lucas. Analysis of notions of diagnosis. Artificial Intelligence, 105(1-2):295–

343, October 1998.

[93] P. J. F. Lucas. Prognostic methods in medicine. Artificial Intelligence in Medicine,

15:105–119, 1999.

[94] E. H. Mamdani. Application of fuzzy algorithms for control of a simple dynamic plant.
Proceedings of the IEE, 121(12):1585–1588, 1974.

[95] E. H. Mamdani and S. Assilian. An experiment in linguistic synthesis with a fuzzy logic
controller. International Journal of Man-Machine Studies, 7(1):1–13, 1975.

[96] O. L. Mangasarian, R. Setiono, and W.-H Goldberg. Pattern recognition via linear
programming: Theory and application to medical diagnosis. In T. F. Coleman and
Y. Li, editors, Large-Scale Numerical Optimization, pages 22–31. SIAM, 1990.

[97] O. L. Mangasarian, W. N. Street, and W. H. Wolberg. Breast cancer diagnosis and

prognosis via linear programming. Mathematical Programming Technical Report 94-
10, University of Wisconsin, 1994.

[98] A. M. Marchevsky, S. Patel, K. J. Wiley, M. A. Stephenson, M. Gondo, R. W. Brown,

E. S. Yi, W. F. Benedict, R. C. Anton, and P. T. Cagle. Artificial neural networks and
logistic regression as tools for prediction of survival in patients with stages I and II
non-small cell lung cancer. Modern Pathology, 11(7):618–625, July 1998.

[99] N. Marvin, M. Bower, and J. E. Rowe. An evolutionary approach to constructing prog-

nostic models. Artificial Intelligence in Medicine, 15(2):155–165, February 1999.

[100] G. K. Matsopoulos, N. A. Mouravliansky, K. K. Delibasis, and K. S. Nikita. Auto-

matic retinal image registration scheme using global optimization techniques. IEEE
Transactions on Information Technology in Biomedicine, 3(1):47–60, March 1999.

[101] J. M. Mendel. Fuzzy logic systems for engineering: A tutorial. Proceedings of the
IEEE, 83(3):345–377, March 1995.

[102] C. J. Merz and P. M. Murphy. UCI repository of machine learning databases, 1996.

[103] B. Mesot, C. A. Peña-Reyes E. Sanchez, and A. Perez-Uribe. SOS++: Finding smart

behaviors using learning and evolution. In Artificial Life VIII. The 8th International
Conference on the Simulation and Synthesis of Living Systems, 2002. (Submitted).
Bibliography 135

[104] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs.

Springer-Verlag, Heidelberg, third edition, 1996.

[105] D. Michie, D. J. Spiegelhalter, and C. C. Taylor, editors. Machine Learning, Neural and
Statistical Classification. Ellis Horwood, 1994.

[106] M. Mitchell. An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA, 1996.

[107] S. Mitra and Y. Hayashi. Neuro-fuzzy rule generation: Survey in soft computing frame-
work. IEEE Transactions on Neural Networks, 11(3):748–768, May 2000.

[108] S. K. Mitra and S. N. Sarbadhikari. Iterative function system and genetic algorithm-
based EEG compression. Medical Engineering and Physics, 19(7):605–617, October
1997.

[109] D. E. Moriarty. Symbiotic Evolution of Neural Networks in Sequential Decision Tasks.

PhD thesis, University of Texas, Austin, 1997.

[110] M. N. Narayanan and S. B. Lucas. A genetic algorithm to improve a neural network to

predict a patient’s response to warfarin. Methods of Information in Medicine, 32(1):55–
58, February 1993.

[111] D. Nauck and R. Kruse. A neuro-fuzzy method to learn fuzzy classification rules from
data. Fuzzy Sets and Systems, 89(3):277–288, August 1997.

[112] D. Nauck and R. Kruse. Neuro-fuzzy systems for function approximation. Fuzzy Sets
and Systems, 101(2):261–271, January 1999.

[113] B. N. Nelson. Automatic vehicle detection in infrared imagery using a fuzzy inference-
based classification system. IEEE-Transaction on Fuzzy Systems, 9:53–61, February
2001.

[114] P. S. Ngan, M. L. Wong, W. Lam, K. S. Leung, and J. C. Cheng. Medical data mining
using evolutionary computation. Artificial Intelligence in Medicine, 16(1):73–96, May
1999.

[115] A. E. Nikulin, B. Dolenko, T. Bezabeh, and R. L. Somorjal. Near-optimal region se-

lection for feature space reduction: Novel preprocessing methods for classifying MR
spectra. NMR in Biomedicine, 11(4-5):209–216, June-August 1998.

[116] T. Niwa and M. Tanaka. Analysis on the island model parallel genetic algorithms for
the genetic drifts. In B. McKay, X. Yao, C. S. Newton, J.-H. Kim, and T. Furuhashi,
editors, Proceedings of the 2nd Asia-Pacific Conference on Simulated Evolution and
Learning (SEAL-98), volume 1585 of LNAI, pages 349–356, Berlin, November 24–27
1999. Springer.

[117] S. Nolfi and D. Floreano. Co-evolving predator and prey robots: Do ’arm races’ arise
in artificial evolution? Artificial Life, 4(4):311–335, 1998.
136 Bibliography

[118] P. Nordin. Evolutionary Program Induction of Binary Machine Code and its Applica-
tion. Krehl Verlag, Munster, Germany, 1997.

[119] C.W. Omlin, L. Giles, and K.K. Thornber. Fuzzy knowledge and recurrent neural net-
works: A dynamical systems perspective. Hybrid Neural Systems. Lecture Notes in
Artificial Intelligence, 1778:123–143, 2000.

[120] S. G. Orel, N. Kay, C. Reynolds, and D. C. Sullivan. Bi-rads categorization as a predic-

tor of malignancy. Radiology, 211(3):845–880, June 1999.

[121] J. Paredis. Coevolutionary computation. Artificial Life, 2:355–375, 1995.

[122] C. S. Pattichis and C. N. Schizas. Genetics-based machine learning for the assessment
of certain neuromuscular disorders. IEEE Transactions on Neural Networks, 7(2):427–
439, March 1996.

[123] C. A. Peña-Reyes and M. Sipper. Evolving fuzzy rules for breast cancer diagnosis. In
Proceedings of 1998 International Symposium on Nonlinear Theory and Applications
(NOLTA’98), volume 2, pages 369–372. Presses Polytechniques et Universitaires Ro-
mandes, Lausanne, 1998.

[124] C. A. Peña-Reyes and M. Sipper. Designing breast cancer diagnostic systems via a hy-
brid fuzzy-genetic methodology. In 1999 IEEE International Fuzzy Systems Conference
Proceedings, volume 1, pages 135–139. IEEE Neural Network Council, 1999.

[125] C. A. Peña-Reyes and M. Sipper. A fuzzy-genetic approach to breast cancer diagnosis.

Artificial Intelligence in Medicine, 17(2):131–155, October 1999.

[126] C. A. Peña-Reyes and M. Sipper. Applying Fuzzy CoCo to breast cancer diagnosis. In
Proceedings of the 2000 Congress on Evolutionary Computation (CEC00), volume 2,
pages 1168–1175. IEEE Press, Piscataway, NJ, USA, 2000.

[127] C. A. Peña-Reyes and M. Sipper. Evolutionary computation in medicine: An overview.

Artificial Intelligence in Medicine, 19(1):1–23, May 2000.

[128] C. A. Peña-Reyes and M. Sipper. The flowering of Fuzzy CoCo: Evolving fuzzy iris
classifiers. In V. Kurková, N. C. Steele, R. Neruda, and M. Kárný, editors, Proceedings
of the 5th International Conference on Artificial Neural Networks and Genetic Algo-
rithms (ICANNGA 2001), pages 304–307. Springer-Verlag, 2001.

[129] C. A. Peña-Reyes and M. Sipper. Fuzzy CoCo: A cooperative-coevolutionary approach

to fuzzy modeling. IEEE Transactions on Fuzzy Systems, 9(5):727–737, October 2001.

[130] C.-A. Peña-Reyes and M. Sipper. Combining evolutionary and fuzzy techniques in
medical diagnosis. In M. Schmitt, H. N. Teodorescu, A. Jain, A. Jain, S. Jain, and
L. C. Jain, editors, Computational Intelligence Techniques in Medical Diagnosis and
Prognosis, volume 96 of Studies in Fuzziness and Soft Computing, chapter 14, pages
391–426. Springer-Verlag, Heidelberg, 2002.
Bibliography 137

[131] C.-A. Peña-Reyes and M. Sipper. Fuzzy CoCo: Balancing accuracy and interpretability
of fuzzy models by means of coevolution. In J. Casillas, O. Cordón, F. Herrera, and
L. Magdalena, editors, Trade-off between Accuracy and Interpretability in Fuzzy Rule-
Based Modelling, Studies in Fuzziness and Soft Computing. Physica-Verlag, 2002. (In
press).

[132] C. A. Peña-Reyes, M. Sipper, and L. Prieto. Sensitive, specific, and interpretable:

Evolving a fuzzy mammographic-interpretation assessment tool. In 2002 IEEE Interna-
tional Fuzzy Systems Conference Proceedings, pages 837–842. IEEE Neural Network
Society, 2002.

[133] W. Pedrycz, editor. Fuzzy Evolutionary Computation. Kluwer Academic Publishers,

1997.

[134] W. Pedrycz and J. Valente de Oliveira. Optimization of fuzzy models. IEEE Transac-
tions on Systems, Man and Cybernetics, Part B: Cybernetics, 26(4):627–636, August
1996.

[135] T. K. Peters, H. E. Koralewski, and E. W. Zerbst. The evolution strategy – A search

strategy used in individual optimization of electrical parameters for therapeutic carotid
sinus nerve simulation. IEEE Transactions on Biomedical Engineering, 36(7):668–675,
July 1989.

[136] V. Podgorelec and P. Kokol. Genetic algorithm based system for patient scheduling in
highly constrained situations. Journal of Medical Systems, 21(6):417–427, December
1997.

[137] R. Poli. Introduction to evolutionary computation. http://www.cs.bham.ac.uk/˜rmp-

/slide_book/, October 1996. visited: 16 March 1999.

[138] R. Poli, S. Cagnoni, and G. Valli. Genetic design of optimum linear and nonlinear QRS
detectors. IEEE Transactions on Biomedical Engineering, 42(11):1137–41, November
1995.

[139] L. Porta, R. Villa, E. Andia, and E. Valderrama. Infraclinic breast carcinoma: Ap-
plication of neural networks techniques for the indication of radioguided biopsias. In
J. Mira, R. Moreno-Díaz, and J. Cabestany, editors, Biological and Artificial Compu-
tation: From Neuroscience to Technology, volume 1240 of Lecture Notes in Computer
Science, pages 978–985. Springer, 1997.

[140] M. A. Potter. The Design and Analysis of a Computational Model of Cooperative Co-
evolution. PhD thesis, George Mason University, 1997.

[141] M. A. Potter and K. A. De Jong. A cooperative coevolutionary approach to function op-

timization. In Yuval Davidor, Hans-Paul Schwefel, and Reinhard Männer, editors, Par-
allel Problem Solving from Nature – PPSN III, pages 249–257, Berlin, 1994. Springer.
Lecture Notes in Computer Science 866.
138 Bibliography

[142] M. A. Potter and K. A. De Jong. Evolving neural networks with collaborative species.
In Proc. of the 1995 Summer Computer Simulation Conf., pages 340–345. The Society
of Computer Simulation, 1995.

[143] M. A. Potter and K. A. De Jong. Cooperative coevolution: An architecture for evolving

coadapted subcomponents. Evolutionary Computation, 8(1):1–29, spring 2000.

[144] M. A. Potter, K. A. De Jong, and J. J. Grefenstette. A coevolutionary approach to

learning sequential decision rules. In Larry J. Eshelman, editor, Proceedings of the 6th
International Conference on Genetic Algorithms (ICGA95), pages 366–372. Morgan
Kaufmann Publishers, 1995.

[145] W. K. Purves, G. H. Orians, H. C. Heller, and D. Sadava. Life. The Science of Biology,
chapter 52: Community Ecology. Sinauer, 5th edition, 1998.

[146] R. Rada, Y. Rhine, and J. Smailwood. Rule refinement. In Proceedings of the Society of
Computer Applications in Medical Care, volume 8 of Annual symposium on computer
application in medical care; proceedings, pages 62–65, 1984.

[147] I. Rechenberg. Cybernetic solution path of an experimental problem. Farborough Hants:

Royal Aircraft Establishment. Library Translation 1122, August 1965. English Trans-
lation of lecture given at the Annual Conference of the WGLR at Berlin in September,
1964.

[148] I. Rechenberg. Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien

der biologischen Evolution. Frommann-Holzboog, Stuttgart, 1973.

[149] I. Rojas, J. Gonzalez, H. Pomares, F.J. Rojas, F.J. Fernández, and A. Prieto. Multi-
dimensional and multideme genetic algorithms for the construction of fuzzy systems.
International Journal of Approximate Reasoning, 26:179–210, 2001.

[150] I. Rojas, H. Pomares, J. Ortega, and A. Prieto. Self-organized fuzzy system generation
from training examples. IEEE Transactions on Fuzzy Systems, 8(1):23–36, February
2000.

[151] C. D. Rosin and R. K. Belew. New methods for competitive coevolution. Evolutionary
Computation, 5(1):1–29, 1997.

[152] M. Russo. FuGeNeSys - A fuzzy genetic neural system for fuzzy modeling. IEEE
Transactions on Fuzzy Systems, 6(3):373–388, August 1998.

[153] B. Sahiner, H. P. Chan, N. Petrick, M. A. Helvie, and M. M. Goodsitt. Design of a

high-sensitivity classifier based on a genetic algorithm: Application to computer-aided
diagnosis. Physics in Medicine and Biology, 43(10):2853–2871, October 1998.

[154] B. Sahiner, H. P. Chan, D. Wei, N. Petrick, M. A. Helvie, D. D. Adler, and M. M.

Goodsitt. Image feature selection by a genetic algorithm: Application to classification
of mass and normal breast tissue. Medical Physics, 23(10):1671–84, October 1996.
Bibliography 139

[155] A. Lo Schiavo and A. M. Luciano. Powerful and flexible fuzzy algorithm for nonlin-
ear dynamic system identification. IEEE-Transaction on Fuzzy Systems, 9:828–835,
December 2001.

[156] P. Schroeter, J. M. Vesin, T. Langenberger, and R. Meuli. Robust parameter estimation

of intensity distributions for brain magnetic resonance images. IEEE Transactions on
Medical Imaging, 17(2):172–186, April 1998.

[157] H.-P. Schwefel. Kybernetische evolution als strategie der experimentelen forschung
in der stromungstechnik. Master’s thesis, Technical University of Berlin, Hermann
Föttinger Institute for Hydrodynamics, March 1965.

[158] H.-P. Schwefel. Evolutionsstrategie und numerische Optimierung. PhD thesis, Techni-
cal University of Berlin, Department of Process Engineering, Berlin, May 1975.

[159] H.-P. Schwefel. Evolution and Optimum Seeking. Sixth-Generation Computer Technol-
ogy Series. John Wiley & Sons., New York, 1995.

[160] R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis.
Artificial Intelligence in Medicine, 8:37–51, 1996.

[161] R. Setiono. Generating concise and accurate classification rules for breast cancer diag-
nosis. Artificial Intelligence in Medicine, 18(3):205 – 219, 2000.

[162] R. Setiono and H. Liu. Symbolic representation of neural networks. IEEE Computer,
29(3):71–77, March 1996.

[163] M. Setnes. Supervised fuzzy clustering for rule extraction. IEEE Transactions on Fuzzy
Systems, 8(4):416–424, August 2000.

[164] Y. Shi, R. Eberhart, and Y. Chen. Implementation of evolutionary fuzzy systems. IEEE
Transactions on Fuzzy Systems, 7(2):109–119, April 1999.

[165] B. Sierra and P. Larrañaga. Predicting survival in malignant skin melanoma using
bayesian networks automatically induced by genetic algorithms. An empirical compari-
son between different approaches. Artificial Intelligence in Medicine, 14(1-2):215–230,
September-October 1998.

[166] R. P. Singson, R. Alsabeh, S. A. Geller, and A. Marchevsky. Estimation of tumor stage

and lymph node status in patients with colorectal adenocarcinoma using probabilistic
neural networks and logistic regression. Modern Pathology, 12(5):479–84, May 1999.

[167] M. Sonka, S. K. Tadikonda, and S. M. Collins. Knowledge-based interpretation of MR

brain images. IEEE Transactions on Medical Imaging, 15(4):443–452, August 1996.

[168] M. Sugeno and G. T. Kang. Structure identification of fuzzy model. Fuzzy Sets and
Systems, 28(1):15–33, 1988.
140 Bibliography

[169] I. Taha and J. Ghosh. Evaluation and ordering of rules extracted from feedforward
networks. In Proceedings of the IEEE International Conference on Neural Networks,
pages 221–226, 1997.

[170] Y. Takagi and M. Sugeno. Fuzzy Identification of Systems and Its Applications to
Modeling and Control. IEEE Transactions on Systems, Man and Cybernetics, 15:116–
132, 1985.

[171] A. Tettamanzi and M. Tomassini. Evolutionary algorithms and their applications. In

D. Mange and M. Tomassini, editors, Bio-Inspired Computing Machines: Toward Novel
Computational Machines, chapter 4, pages 59–98. Presses Polytechniques et Universi-
taires Romandes, Lausanne, Switzerland, 1998.

[172] M. Tomassini. Parallel and distributed evolutionary algorithms: A review. In K. Mi-

ettinen, M. Mäkelä, P. Neittaanmäki, and J. Periaux, editors, Evolutionary Algorithms
in Engineering and Computer Science, pages 113–133. J. Wiley and Sons, Chichester,
1999.

[173] E. C. C. Tsang, X. S. Wang, and D.S. Yeung. Improving learning accuracy of fuzzy de-
cision trees by hybrid neural networks. IEEE-Transactions on Fuzzy Systems, 8(5):601–
614, October 2000.

[174] K. Uutela, M. Hamalainen, and R. Salmelin. Global optimization in the localization

of neuromagnetic sources. IEEE Transactions on Biomedical Engineering, 45(6):716–
723, June 1998.

[175] J. Valente de Oliveira. Semantic constraint for membership function optimization.

IEEE Transactions on Systems, Man, and Cybernetics. Part A: Systems and Humans,
29(1):128–138, January 1999.

[176] R. P. Velthuizen, L. O. Hall, and L. P. Clarke. Feature extraction for MRI segmentation.
Journal of Neuroimaging, 9(2):85–90, April 1999.

[177] M. D. Vose. The Simple Genetic Algorithm. MIT Press, Cambridge, MA, August 1999.

[178] P. Vuorimaa. Fuzzy self-organizing map. Fuzzy Sets and Systems, 66:223–231, 1994.

[179] T. Wanschura, D. A. Coley, W. Vennart, and S. Gandy. Automatic realignment of time-

separated MR images by genetic algorithm. Magnetic Resonance Imaging, 17(2):313–
317, February 1999.

[180] D. Whitley, S. Rana, and R. Heckendorn. Exploiting separability in search: The island
model genetic algorithm. Journal of Computing and Information Technology, 7(1):33–
47, 1999. (Special Issue on Evolutionary Computing).

[181] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Tech-
niques with Java Implementations. Data Management Systems. Morgan Kaufmann, San
Francisco, USA, 2000.
Bibliography 141

[182] A. Włoszek and P. D. Domański. Application of the coevolutionary system to the rule
generation in fuzzy systems. In Proceedings of IKAE96, pages 203–210, Murzasichle,
Poland, 1996.

[183] A. Włoszek and P. D. Domański. Application of the coevolutionary system to the fuzzy
model design. In Proceedings of the Sixth IEEE International Conference on Fuzzy
Systems, volume 1, pages 391–395, 1997.

[184] T.-P. Wu and S.-M. Chen. A new method for constructing membership functions and
fuzzy rules from training examples. IEEE Transactions on Systems, Man and Cyber-
netics, Part B: Cybernetics, 29(1):25–40, February 1999.

[185] R. R. Yager and D. P. Filev. Essentials of Fuzzy Modeling and Control. John Wiley &
Sons., New York, 1994.

[186] R. R. Yager and L. A. Zadeh. Fuzzy Sets, Neural Networks, and Soft Computing. Van
Nostrand Reinhold, New York, 1994.

[187] G. Yang, L. E. Reinstein, S. Pai, Z. Xu, and D. L. Carroll. A new genetic algo-
rithm technique in optimization of permanent 125 I prostate implants. Medical Physics,
25(12):2308–2315, December 1998.

[188] X. Yao and Y. Liu. A new evolutionary system for evolving artificial neural networks.
IEEE Transactions on Neural Networks, 8(3):694–713, May 1997.

[189] Y. Yu. Multiobjective decision theory for computational optimization in radiation ther-
apy. Medical Physics, 24(9):1445–1454, September 1997.

[190] Y. Yu and M. C. Schell. A genetic algorithm for the optimization of prostate implants.
Medical Physics, 23(12):2085–91, December 1996.

[191] Y. Yu, M. C. Schell, and J. B. Zhang. Decision theoretic steering and genetic algo-
rithm optimization: Application to stereotactic radiosurgery treatment planning. Medi-
cal Physics, 24(11):1742–50, November 1997.

[192] Y. Yu, J. B. Zhang, R. A. Brasacchio, P. G. Okunieff, D. J. Rubens, J. G. Strang,

A. Soni, and E. M. Messing. Automated treatment planning engine for prostate seed
implant brachytherapy. International Journal of Radiation Oncology, Biology, Physic,
43(3):647–652, February 1 1999.

[193] L. A. Zadeh. Fuzzy sets. Information and Control, 8:338–353, 1965.

[194] L. A. Zadeh. Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Transactions on Systems, Man and Cybernetics, SMC-3(1):28–44,
January 1973.

[195] L. A. Zadeh. The concept of a linguistic variable and its applications to approximate
reasoning. Information Science, Part I, 8:199-249; Part II, 8:301-357; Part III, 9:43-80.,
1975.
142 Bibliography

[196] S. Zahan. A fuzzy approach to computer-assisted myocardial ischemia diagnosis. Arti-

ficial Intelligence in Medicine, 21(1–3):271–275, January–March 2001.

[197] B. Zheng, Y. H. Chang, X. H. Wang, W. F. Good, and D. Gur. Feature selection for
computerized mass detection in digitized mammograms by using a genetic algorithm.
Academic Radiology, 6(6):327–332, June 1999.
Curriculum Vitæ

PERSONAL DATA

Name Carlos Andrés Peña Reyes

Birth Cali, Colombia, March 31, 1968
Nationality Colombian
Prof. Address Logic Systems Laboratory (LSL)
School of Computer and Communication Sciences - IC
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne
Switzerland
Prof. Phone (+41-21) 693 26 58
Prof. Fax (+41-21) 693 37 05
E-mail [email protected]
[email protected]
URL http://lslwww.epfl.ch/˜penha

EDUCATION

• Ph.D. in Computer Science [Expected: September 2002]

Swiss Federal Institute of Technology at Lausanne - EPFL, Lausanne, Switzerland.

• Graduate School in Computer Science

Swiss Federal Institute of Technology at Lausanne - EPFL, Lausanne, Switzerland, 1998

• Cours postgrade Informatique et biologie: du modèle à l’outil

(Graduate course: Informatics and biology: from the model to the tool)
Swiss Federal Institute of Technology at Lausanne - EPFL, Lausanne, Switzerland, 1997

• Especialización en Automatización Industrial

(Graduate cycle: Specialization on industrial automation)
Universidad del Valle, Cali, Colombia, 1997

• Ingeniería Electrónica (Electronic engineering)

Universidad Distrital Fco. José de Caldas, Bogotá, Colombia, 1992

143
144 Bibliography

ACADEMIC POSITIONS
• PhD Candidate and teaching assistant
Logic Systems Laboratory (LSL)
Swiss Federal Institute of Technology at Lausanne - EPFL
Lausanne, Switzerland, January 1999 - September 2002

• Fellowship holder
Swiss Federal Institute of Technology at Lausanne - EPFL
Lausanne, Switzerland, July 1996 - December 1998

• Assistant Instructor
Pontificia Universidad Javeriana and Corporación Universitaria Autónoma de Occidente
Cali, Colombia, January 1995 - May 1996

RESEARCH INTERESTS
• Bioinspired systems
• Computational Intelligence - Soft Computing
• Fuzzy systems
• Evolutionary computation
• Medical applications of computational intelligence techniques
• Adaptation, learning, and intelligence
• Autonomous robotics

REVIEW ACTIVITIES Referee for the following international journals:

• Fuzzy Sets and Systems, Elsevier Science

• Artificial Intelligence in Medicine, Elsevier
• Energía y Computación, Universidad del Valle, Colombia

INVITED LECTURES, TALKS, AND TUTORIALS

• Evolutionary and Coevolutionary Fuzzy Modeling, Course, Universidad Nacional,
Bogotá, Colombia, October 2002.

• Coevolutionary Fuzzy Modeling in Decision Support Systems for Cancer Diagnosis,

Invited speech, Workshop: Applications of Bioinformatics in Cancer Detection,
National Cancer Institute, Division of Cancer Prevention Bethesda, USA, August 2002.
Bibliography 145

• Evolutionary Computing, Fuzzy Logic, Fuzzy Modeling, and Evolutionary Fuzzy

Modeling, Invited lectures, Prof. Daniel Mange’s course “Systèmes et programmation
génétiques”, EPFL, Lausanne, Switzerland, January 2002 and June 2001.

• Integración de técnicas bioinspiradas. Modelado difuso evolutivo (Integration of

bioinspired techniques: Evolutionary fuzzy modeling), Universidad de Extremadura,
Centro Universitario de Mérida, Mérida, Spain, December 2000.

• Evolutionary Fuzzy Modeling, Invited lecture, Graduate School in Computer Science,

EPFL, Lausanne, Switzerland, June 2000.

• Evolutionary Fuzzy Modeling, Course, Universidad de los Andes, Bogotá and Pontificia
Universidad Javeriana, Cali, Colombia, March 2000.

• Integración de técnicas de inteligencia computacional: Lógica difusa, redes neuronales

y computación evolutiva (Integration of computational intelligence techniques: Fuzzy
logic, neural networks, and evolutionary computation), Tutorial, Universidad del Valle,
Cali, Colombia, February 1998.

• Control inteligente (Intelligent control), Tutorial, Javintec’96, Pontificia Universidad

Javeriana, Cali, Colombia, 1996.

• Introducción a los sistemas difusos (Introduction to fuzzy systems), Seminar, Tercera

Convención Nacional de Estudiantes de Ingenierías Eléctrica y Electrónica, Cali,
Colombia, 1994.

• Ingeniería de software: Técnicas de especificación para el desarrollo de proyectos (Soft-

ware engineering: Specification techniques for project development), Tutorial, IX SIDC,
Universidad Distrital, Bogotá, Colombia, 1991.

STUDENT PROJECTS SUPERVISED

• Incremental Fuzzy CoCo, Olivier Rutti, EPFL, summer 2002.

• Fuzzy Modeling by (r + 1)-Fuzzy CoCo, Michael Bruderer, EPFL, winter 2001-02.

• Test of the applicability of Fuzzy CoCo to function approximation, Akhil Gupta, Indian
Institute of Technology, Kanpur (Summer Internship Program, EPFL), summer 2001.

• Island model for the cooperative coevolutionary algorithm Fuzzy CoCo, Yves Blatter,
EPFL, winter 2000-01.
146 Bibliography

• A fuzzy logic-based tool for computer-assisted breast cancer diagnosis, Richard Greset,
EPFL, winter 2000-01.

• Real-time evolution of a neuro-fuzzy robot controller, Rodolph Perfetta, EPFL, summer

2000.

• A study of the Fuzzy CoCo algorithm, Jean-Marc Henry, EPFL, summer 2000.

• Evolution of the adaptability of a neuro-fuzzy controlled autonomous robot, Didier

Rizzotti, EPFL, winter 99-00.

• A reconfigurable neuro-fuzzy processor using on-line arithmetic, Alvaro Gutiérrez (in

cooperation with J.-L. Beuchat), EPFL, winter 99-00.

• Intelligent data analysis, Francesc Font (in cooperation with J.-L. Beuchat), EPFL,
winter 99-00

• Hybrid learning (neural-evolutionary) of a fuzzy system for breast cancer diagnosis,

Jean-Francois Philagor (in cooperation with M. Sipper), EPFL, winter 98-99 and
summer 99.

• A reconfigurable digital fuzzy coprocessor, Christophe Gogniat (in cooperation with A.

Pérez), EPFL, winter 98-99.

LANGUAGES

• Spanish: Native tongue

• French: Very good

• English: Very good

OTHER PROFESSIONAL EXPERIENCE

• General manager
Profesionales en comunicaciones PROECOM Ltda.
Cali, Colombia, March 1992 - May 1995

• Research assistant
Empresa Nacional de Telecomunicaciones TELECOM
Research Division
Bogotá, Colombia, 1991
Bibliography 147

MEMBERSHIPS
• IEEE, The Institute of Electric and Electronics Engineers
SM-85, M-93, SM-95.
• EvoNet, European Network of Excellence in Evolutionary Computing
• EUNITE, EUropean Network on Intelligent TEchnologies for Smart Adaptive Systems
• AL-IC, Latin-American Association of Computational Intelligence
Co-founder in 2002
• ACIS, Colombian Association of Researchers in Switzerland
President since 1999
• Red-CIC, Colombian Network on Computational Intelligence and Bioinspired Systems.
General coordinator since its foundation in 2000
• Gi3, Research Group on Computational Intelligence and Bio-inspired Systems
Co-founder in 1997
• Red Caldas, Colombian Network of Researchers Abroad
Member since 1992, administrator of the electronic list in 1998-99

PUBLICATIONS

Journal Papers
1. C. A. Peña-Reyes and M. Sipper, "Fuzzy Modeling by Fuzzy CoCo," International Jour-
nal of Computational Intelligence and Applications, 2002. (Submitted.)
2. C. A. Peña-Reyes and M. Sipper, "Fuzzy CoCo: A Cooperative-Coevolutionary ap-
proach to Fuzzy Modeling," IEEE Transactions on Fuzzy Systems, vol. 9, no. 5, pp.
727-737, Oct. 2001.
3. C. A. Peña-Reyes and M. Sipper, "Evolutionary Computation in Medicine: An
Overview," Artificial Intelligence in Medicine, vol. 19, no. 1, pp. 1-23, May 2000.
4. H. F. Restrepo, C. A. Peña-Reyes, and A. Perez-Uribe, "Hacia el desarrollo de nuevas
máquinas computacionales: lecciones que aprendemos de la naturaleza," Energía y
Computación, vol. IX, no. 2, pp. 40–47, 2000.
5. C. A. Peña-Reyes and M. Sipper, "A fuzzy-genetic approach to breast cancer diagnosis,"
Artificial Intelligence in Medicine, vol. 17, no. 2, pp. 131-155, Oct. 1999.

Book chapters
1. C.-A. Peña-Reyes and M. Sipper, "Fuzzy CoCo: Balancing accuracy and interpretability
of fuzzy models by means of coevolution," in Trade-off between Accuracy and Inter-
pretability in Fuzzy Rule-Based Modelling, (J. Casillas, O. Cordon, F. Herrera, and L.
Magdalena, eds.), 2002. (In press.)
148 Bibliography

2. C.-A. Peña-Reyes and M. Sipper, "Combining Evolutionary and Fuzzy Techniques in

Medical Diagnosis," in Computational Intelligence Techniques in Medical Diagnosis
and Prognosis, vol. 96, (M. Schmitt, H. N. Teodorescu, A. Jain, A. Jain, S. Jain, and L.
C. Jain, eds.), pp. 391–426, 2002.

Conferences

1. B. Mesot, C. A. Peña-Reyes E. Sanchez, and A. Perez-Uribe, "SOS++: Finding Smart

Behaviors Using Learning and Evolution," in Artificial Life VIII. The 8th International
Conference on the Simulation and Synthesis of Living Systems, 2002. (Submitted.)

2. J.L. Bermudez, C. A. Peña-Reyes, F. Rachidi, and F. Heidler, "Use of genetic algorithms

to extract primary lightning current parameters," in 5th International Symposium on
Electromagnetic Compatibility, 2002. (Accepted.)

3. C. A. Peña-Reyes, M. Sipper, and L. Prieto, "Sensitive, Specific, and Interpretable:

Evolving a Fuzzy Mammographic-Interpretation Assessment Tool," in 2002 IEEE Inter-
national Fuzzy Systems Conference Proceedings, pp. 837–842, IEEE Neural Network
Society, 2002.

4. C. A. Peña-Reyes and M. Sipper, "The Flowering of Fuzzy CoCo: Evolving Fuzzy

Iris Classifiers," in Proceedings of the 5th International Conference on Artificial Neural
Networks and Genetic Algorithms (ICANNGA 2001), (V. Kurková, N. C. Steele, R.
Neruda, and M. Kárný, eds.), pp. 304-307, 2001.

5. C. A. Peña-Reyes and M. Sipper, "Applying Fuzzy CoCo to Breast Cancer Diagnosis,"

in Proceedings of the 2000 Congress on Evolutionary Computation (CEC00), pp. 1168-
1175, 2000.

6. C. A. Peña-Reyes and M. Sipper, "Designing Breast cancer Diagnostic Systems via a

Hybrid Fuzzy-Genetic Methodology," in 1999 IEEE International Fuzzy Systems Con-
ference Proceedings, pp. 135-139, IEEE Neural Network Council, 1999.

7. C. A. Peña-Reyes and M. Sipper, "Evolving Fuzzy Rules for Breast Cancer Diagnosis,"
in Proceedings of 1998 International Symposium on Nonlinear Theory and Applications
(NOLTA’98), pp. 369-372, 1998.

8. K. Solano, R. A. Chavarriaga, R. A. Nuñez, and C. A. Peña-Reyes, "Controladores

adaptables basados en mecanismos de inferencia difusa," in Memorias del Segundo Con-
greso Asociación Colombiana de Automática, 1997.

9. J. L. López, C. E. Quintero, M. Sacanamboy, J. O. Sánchez, O.M. Agudelo, and C.

A. Peña-Reyes, "Comparación entre sistemas difusos para el control de una banda trans-
portadora dentro de un proceso de manufactura flexible," in Memorias del Segundo Con-
greso Asociación Colombiana de Automática, 1997.