Representations for Genetic and Evolutionary Algorithms
Franz Rothlauf
Representations for Genetic and Evolutionary Algorithms
ABC
Dr. Franz Rothlauf Universität Mannheim 68131 Mannheim Germany Email:
[email protected] Library of Congress Control Number: 2005936356 ISBN10 354025059X Springer Berlin Heidelberg New York ISBN13 9783540250593 Springer Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microﬁlm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable for prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springer.com c SpringerVerlag Berlin Heidelberg 2006 Printed in The Netherlands The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. AE Typesetting: by the author and TechBooks using a Springer LT X macro package
Cover design: Erich Kirchner, Heidelberg Printed on acidfree paper
SPIN: 11370550
89/TechBooks
543210
F¨ ur meine Eltern Elisabeth und Alfons Rothlauf.
Preface
Preface to the Second Edition I have been very pleased to see how well the ﬁrst edition of this book has been accepted and used by its readers. I have received fantastic feedback telling me that people use it as an inspiration for their own work, give it to colleagues or students, or use it for preparing lectures and classes about representations. I want to thank you all for using the material presented in this book and for developing more eﬃcient and powerful heuristic optimization methods. You will ﬁnd this second edition of the book completely revised and extended. The goal of the revisions and extensions was to make it easier for the reader to understand the main points and to get a more thorough knowledge of the design of highquality representations. For example, I want to draw your attention to Chap. 3 where you ﬁnd the core of the book. I have extended and improved the sections about redundancy and locality of representations adding new material and experiments and trying to draw a more comprehensive picture. In particular, the introduction of synonymity for redundant encodings in Sect. 3.1 and the integration of locality and redundancy issues in Sect. 3.3 are worth having a closer look at it. These new concepts have been used throughout the work and have made it possible to better understand a variety of diﬀerent representation issues. The chapters about tree representations have been reorganized such that they explicitly distinguish between direct and indirect representations. This distinction – including a new analysis of the edgesets, which is a direct encoding for trees – emphasizes that the developed representation framework is not only helpful for analysis and design of representations, but also for operators. The design of proper search operators is at the core of direct representations and the new sections demonstrate how to analyze the inﬂuence of such encodings on the performance of genetic and evolutionary algorithms (GEAs). Finally, the experiments presented in Chap. 8 have been completely revised considering new representations and giving a better understanding of the inﬂuence of tree representations on the performance of GEAs.
VIII
Preface
I would like to take this opportunity to thank everyone who took the time to share their thoughts on the text with me – all these comments were helpful in improving the book. Special thanks to Kati for her support in preparing this work. As with the ﬁrst edition, my purpose will be fulﬁlled if you ﬁnd this book helpful for building more eﬃcient heuristic optimization methods, if you ﬁnd it inspiring for your research, or if it is a help for you teaching students about the importance and inﬂuence of representations.
Mannheim August 2005
Franz Rothlauf
Preface to the First Edition This book is about representations for genetic and evolutionary algorithms (GEAs). In writing it, I have tried to demonstrate the important role of representations for an eﬃcient use of genetics and evolutionary optimization methods. Although, experience often shows that the choice of a proper representation is crucial for GEA’s success, there are few theoretical models that describe how representations inﬂuence GEAs behavior. This book aims to resolve this unsettled situation. It presents theoretical models describing the eﬀect of diﬀerent types of representations and applies them to binary representations of integers and tree representations. The book is designed for people who want to learn some theory about how representations inﬂuence GEA performance and for those who want to see how this theory can be applied to representations in the real world. The book is based on my dissertation with the title “Towards a Theory of Representations for Genetic and Evolutionary Algorithms: Development of Basic Concepts and their Application to Binary and Tree Representations”. To make the book easier to read for a larger audience some chapters are extended and many explanations are more detailed. During the writing of the book many people from various backgrounds (economics, computer science and engineering) had a look at the work and pushed me to present it in a way that is accessible to a diverse audience. Therefore, also people that are not familiar to GEAs should be able to get the basic ideas of the book. To understand the theoretical models describing the inﬂuence of representations on GEA performance I expect collegelevel mathematics like elementary notions of counting, probability theory and algebra. I tried to minimize the mathematical background required for understanding the core lessons of the book and to give detailed explanations on complex theoretical subjects. Furthermore, I expect the reader to have no particular knowledge of genetics and deﬁne all genetic terminology and concepts in the text. The inﬂuence of
Preface
IX
integer and tree representations on GEA performance does not necessarily require a complete understanding of the elements of representation theory but is also accessible for people who do not want to bother too much with theory. The book is split up into two large parts. The ﬁrst presents theoretical models describing the eﬀects of representations on GEA performance. The second part uses the theory for the analysis and design of representations. After the ﬁrst two introductory chapters, theoretical models are presented on how redundant representations, exponentially scaled representations and the locality/distance distortion of a representation inﬂuence GEA performance. In Chap. 4 the theory is used for formulating a timequality framework. Consequently, in Chap. 5, the theoretical models are used for analyzing the performance diﬀerences between binary representations of integers. Finally, the framework is used in Chap. 6, Chap. 7, and Chap. 8 for the analysis of existing tree representations as well as the design of new tree representations. In the appendix common test instances for the optimal communication spanning tree problems are summarized.
Acknowledgments First of all, I would like to thank my parents for always providing me with a comfortable home environment. I have learned to love the wonders of the world and what the important things in life are. Furthermore, I would like to say many thanks to my two advisors, Dr. Armin Heinzl and Dr. David E. Goldberg. They did not only help me a lot with my work, but also had a large impact on my private life. Dr. Armin Heinzl helped me to manage my life in Bayreuth and always guided me in the right direction in my research. He was a great teacher and I was able to learn many important things from him. I am grateful to him for creating an environment that allowed me to write this book. Dr. David E. Goldberg had a large inﬂuence on my research life. He taught me many things which I needed in my research and I would never have been able to write this thesis without his help and guidance. During my time here in Bayreuth, my colleagues in the department have always been a great help to overcome the troubles of daily university life. I especially want to thank Michael Zapf, Lars Brehm, Jens Dibbern, Monika Fortm¨ uhler, Torsten O. Paulussen, J¨ urgen Gerstacker, Axel P¨ urckhauer, Thomas Schoberth, Stefan Hocke, and Frederik Loos. During my time here, Wolfgang G¨ uttler and Tobias Grosche were not only work colleagues, but also good friends. I want to thank them for the good time I had and the interesting discussions. During the last three years during which I spent time at IlliGAL I met many people who have had a great impact on my life. First of all, I would like to thank David E. Goldberg and the Department of General Engineering for giving me the opportunity to stay there so often. Then, I want to say thank you
X
Preface
to the folks at IlliGAL I was able to work together with. It was always a really great pleasure. I especially want to thank Erick Cant´ uPaz, Fernando Lobo, Dimitri Knjazew, Clarissa van Hoyweghen, Martin Butz, Martin Pelikan, and Kumara Sastry. It was not only a pleasure working together with them but over time they have become really good friends. My stays at IlliGAL would not have been possible without their help. Finally, I want to thank the people who were involved in the writing of this book. First of all I want to thank Kumara Sastry and Martin Pelikan again. They helped me a lot and had a large impact on my work. The discussions with Martin were great and Kumara often impressed me with his expansive knowledge about GEAs. Then, I want to say thanks to Fernando Lobo and Torsten O. Paulussen. They gave me great feedback and helped me to clarify my thoughts. Furthermore, Katrin Appel and Kati Sternberg were a great help in writing this dissertation. Last but not least I want to thank Anna Wolf. Anna is a great proofreader and I would not have been able to write a book in readable English without her help. Finally, I want to say “thank you” to Kati. Now I will hopefully have more time for you.
Bayreuth January 2002
Franz Rothlauf
Foreword to the First Edition
It is both personally and intellectually pleasing for me to write a foreword to this work. In January 1999 I received a brief email from a PhD student at the University of Bayreuth asking if he might visit the Illinois Genetic Algorithms Laboratory (IlliGAL). I did not know this student, Franz Rothlauf, but something about the tone of his note suggested a sharp, eager mind connected to a cheerful, positive nature. I checked out Franz’s references, invited him to Illinois for a ﬁrst visit, and my early feelings were soon proven correct. Franz’s various visits to the lab brought both smiles to the faces of IlliGAL labbies and important progress to a critical area of genetic algorithm inquiry. It was great fun to work with Franz and it was exciting to watch this work take shape. In the remainder, I brieﬂy highlight the contributions of this work to our state of knowledge. In the ﬁeld of genetic and evolutionary algorithms (GEAs), much theory and empirical study has been heaped upon operators and test problems, but problem representation has often been taken as a given. In this book, Franz breaks with this tradition and seriously studies a number of critical elements of a theory of GEA representation and applies them to the careful empirical study of (a) a number of important idealized test functions and (b) problems of commercial import. Not only is Franz creative in what he has chosen to study, he also has been innovative in how he performs his work. In GEAs – as elsewhere – there appears sometimes to be a ﬁrewall separating theory and practice. This is not new, and even Bacon commented on this phenomenon with his famous metaphor of the spiders (men of dogmas), the ants (men of practice), and the bees (transformers of theory to practice). In this work, Franz is one of Bacon’s bees, taking applicable theory of representation and carrying it to practice in a manner that (1) illuminates the theory and (2) answers the questions of importance to a practitioner. This book is original in many respects, so it is diﬃcult to single out any one of its many accomplishments. I do believe ﬁve items deserve particular comment: 1. Decomposition of the representation problem 2. Analysis of redundancy
XII
Foreword to the First Edition
3. Analysis of scaling 4. Timequality framework for representation 5. Demonstration of the framework in wellchosen test problems and problems of commercial import. Franz’s decomposition of the problem of representation into issues of redundancy, scaling, and correlation is itself a contribution. Individuals have isolated each of these areas previously, but this book is the ﬁrst to suggest they are core elements of an integrated theory and to show the way toward that integration. The analyses of redundancy and scaling are examples of applicable or facetwise modeling at its best. Franz gets at key issues in run duration and population size through bounding analyses, and these permit him to draw definite conclusions in ﬁelds where so many other researchers have simply waived their arms. By themselves, these analyses would be suﬃcient, but Franz then takes the extra and unprecedented step toward an integrated qualitytime framework for representations. The importance of quality and time has been recognized previously from the standpoint of operator design, but this work is the ﬁrst to understand that codings can and should be examined from an eﬃciencyquality standpoint as well. In my view, this recognition will be understood in the future as a key turning point away from the current voodoo and black magic of GEA representation toward a scientiﬁc discussion of the appropriateness of particular representations for diﬀerent problems. Finally, Franz has carefully demonstrated his ideas in (1) carefully chosen test functions and (2) problems of commercial import. Too often in the GEA ﬁeld, researchers perform an exercise in pristine theory without relating it to practice. On the other hand, practitioners too often study the latest wrinkle in problem representation or coding without theoretical backing or support. This dissertation asserts the applicability of its theory by demonstrating its utility in understanding tree representations, both test functions and realworld communications networks. Going from theory to practice in such a sweeping manner is a rare event, and the accomplishment must be regarded as both a diﬃcult and an important one. All this would be enough for me to recommend this book to GEA aﬁcionados around the globe, but I hasten to add that the book is also remarkably well written and well organized. No doubt this rhetorical craftsmanship will help broaden the appeal of the book beyond the ken of genetic algorithmists and computational evolutionaries. In short, I recommend this important book to anyone interested in a better quantitative and qualitative understanding of the representation problem. Buy this book, read it, and use its important methodological, theoretical, and practical lessons on a daily basis. University of Illinois at UrbanaChampaign
David E. Goldberg
Contents
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 2 4
2
Representations for Genetic and Evolutionary Algorithms . 2.1 Genetic Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Genotypes and Phenotypes . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2 Decomposition of the Fitness Function . . . . . . . . . . . . . . . 2.1.3 Types of Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Genetic and Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Schema Theorem and Building Block Hypothesis . . . . . . 2.3 Problem Diﬃculty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Reasons for Problem Diﬃculty . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Measurements of Problem Diﬃculty . . . . . . . . . . . . . . . . . 2.4 Existing Recommendations for the Design of Eﬃcient Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Goldberg’s Meaningful Building Blocks and Minimal Alphabets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Radcliﬀe’s Formae and Equivalence Classes . . . . . . . . . . . 2.4.3 Palmer’s Tree Encoding Issues . . . . . . . . . . . . . . . . . . . . . . 2.4.4 Ronald’s Representational Redundancy . . . . . . . . . . . . . .
9 10 10 11 13 15 15 16 18 22 22 25
3
Three Elements of a Theory of Representations . . . . . . . . . . . . 3.1 Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1 Redundant Representations and Neutral Networks . . . . . 3.1.2 Synonymously and NonSynonymously Redundant Representations . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.3 Complexity Model for Redundant Representations . . . . .
28 28 29 31 31 33 35 35 38 45
XIV
Contents
3.1.4 Population Sizing for Synonymously Redundant Representations . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.5 Run Duration and Overall Problem Complexity for Synonymously Redundant Representations . . . . . . . . 3.1.6 Analyzing the Redundant Trivial Voting Mapping . . . . . 3.1.7 Conclusions and Further Research . . . . . . . . . . . . . . . . . . . 3.2 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Deﬁnitions and Background . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Population Sizing Model for Exponentially Scaled Representations Neglecting the Eﬀect of Genetic Drift . 3.2.3 Population Sizing Model for Exponentially Scaled Representations Considering the Eﬀect of Genetic Drift 3.2.4 Empirical Results for BinInt Problems . . . . . . . . . . . . . . . 3.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Locality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Inﬂuence of Representations on Problem Diﬃculty . . . . . 3.3.2 Metrics, Locality, and Mutation Operators . . . . . . . . . . . . 3.3.3 PhenotypeFitness Mappings and Problem Diﬃculty . . . 3.3.4 Inﬂuence of Locality on Problem Diﬃculty . . . . . . . . . . . 3.3.5 Distance Distortion and Crossover Operators . . . . . . . . . 3.3.6 Modifying BBComplexity for the OneMax Problem . . 3.3.7 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
47 49 50 57 59 59 61 65 68 72 73 74 76 78 81 84 86 89 93 95
TimeQuality Framework for a TheoryBased Analysis and Design of Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 4.1 Solution Quality and Time to Convergence . . . . . . . . . . . . . . . . . 98 4.2 Elements of the Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.2.1 Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.2.2 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.2.3 Locality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 4.3 The Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.3.1 Uniformly Scaled Representations . . . . . . . . . . . . . . . . . . . 104 4.3.2 Exponentially Scaled Representations . . . . . . . . . . . . . . . . 105 4.4 Implications for the Design of Representations . . . . . . . . . . . . . . 108 4.4.1 Uniformly Redundant Representations Are Robust . . . . 108 4.4.2 Exponentially Scaled Representations Are Fast, but Inaccurate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.4.3 Lowlocality Representations Are Diﬃcult to Predict, and No Good Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 4.5 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Contents
XV
5
Analysis of Binary Representations of Integers . . . . . . . . . . . . 117 5.1 Integer Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 5.2 Binary String Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 5.3 A Theoretical Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 5.3.1 Redundancy and the Unary Encoding . . . . . . . . . . . . . . . . 123 5.3.2 Scaling, Modiﬁcation of Problem Diﬃculty, and the Binary Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 5.3.3 Modiﬁcation of Problem Diﬃculty and the Gray Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 5.4.1 Integer OneMax Problem and Deceptive Integer OneMax Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 5.4.2 Modiﬁcations of the Integer OneMax Problem . . . . . . . . 134 5.5 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6
Analysis and Design of Representations for Trees . . . . . . . . . . 141 6.1 The Tree Design Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 6.1.1 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 6.1.2 Metrics and Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 6.1.3 Tree Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 6.1.4 Schema Analysis for Graphs . . . . . . . . . . . . . . . . . . . . . . . . 146 6.1.5 Scalable Test Problems for Graphs . . . . . . . . . . . . . . . . . . 147 6.1.6 Tree Encoding Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 6.2 Pr¨ ufer Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 6.2.1 Historical Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 6.2.2 Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 6.2.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 6.2.4 The Low Locality of the Pr¨ ufer Number Encoding . . . . . 157 6.2.5 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 169 6.3 The Characteristic Vector Encoding . . . . . . . . . . . . . . . . . . . . . . . 171 6.3.1 Encoding Trees with Characteristic Vectors . . . . . . . . . . . 171 6.3.2 Repairing Invalid Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 172 6.3.3 Bias and NonSynonymous Redundancy . . . . . . . . . . . . . . 173 6.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 6.4 The Link and Node Biased Encoding . . . . . . . . . . . . . . . . . . . . . . . 178 6.4.1 Motivation and Functionality . . . . . . . . . . . . . . . . . . . . . . . 179 6.4.2 Bias and NonUniformly Redundant Representations . . . 183 6.4.3 The NodeBiased Encoding . . . . . . . . . . . . . . . . . . . . . . . . . 184 6.4.4 A Concept for the Analysis of Redundant Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 6.4.5 Population Sizing for the LinkBiased Encoding . . . . . . . 191 6.4.6 The LinkandNodeBiased Encoding . . . . . . . . . . . . . . . . 195 6.4.7 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 6.4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 6.5 Network Random Keys (NetKeys) . . . . . . . . . . . . . . . . . . . . . . . . . 201
XVI
Contents
6.5.1 6.5.2 6.5.3 6.5.4 6.5.5
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 Uniform Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 Population Sizing and Run Duration for the OneMax Tree Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 6.5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 7
Analysis and Design of Search Operators for Trees . . . . . . . . 217 7.1 NetDir: A Direct Representation for Trees . . . . . . . . . . . . . . . . . . 218 7.1.1 Historical Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 7.1.2 Properties of Direct Representations . . . . . . . . . . . . . . . . . 219 7.1.3 Operators for NetDir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220 7.1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 7.2 The EdgeSet Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 7.2.1 Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 7.2.2 Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 7.2.3 Performance for the OCST Problem . . . . . . . . . . . . . . . . . 230 7.2.4 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 237
8
Performance of Genetic and Evolutionary Algorithms on Tree Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 8.1 GEA Performance on Scalable Test Tree Problems . . . . . . . . . . . 242 8.1.1 Analysis of Representations . . . . . . . . . . . . . . . . . . . . . . . . . 242 8.1.2 OneMax Tree Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 8.1.3 Deceptive Trap Problem for Trees . . . . . . . . . . . . . . . . . . . 251 8.2 GEA Performance on the OCST Problem . . . . . . . . . . . . . . . . . . 256 8.2.1 The Optimal Communication Spanning Tree Problem . . 257 8.2.2 Optimization Methods for the Optimal Communication Spanning Tree Problem . . . . . . . . . . . . . . 258 8.2.3 Description of Test Problems . . . . . . . . . . . . . . . . . . . . . . . 260 8.2.4 Analysis of Representations . . . . . . . . . . . . . . . . . . . . . . . . . 262 8.2.5 Theoretical Predictions on the Performance of Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 8.2.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 8.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
9
Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 9.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 9.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Contents
A
XVII
Optimal Communication Spanning Tree Test Instances . . . . 281 A.1 Palmer’s Test Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 A.2 Raidl’s Test Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 A.3 Berry’s Test Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 A.4 Real World Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 List of Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
1 Introduction
One of the major challenges for researchers in the ﬁeld of management science, information systems, business informatics, and computer science is to develop methods and tools that help organizations, such as companies or public institutions, to fulﬁll their tasks eﬃciently. However, during the last decade, the dynamics and size of tasks organizations are faced with has changed. Firstly, production and service processes must be reorganized in shorter time intervals and adapted dynamically to the varying demands of markets and customers. Although there is continuous change, organizations must ensure that the eﬃciency of their processes remains high. Therefore, optimization techniques are necessary that help organizations to reorganize themselves, to increase the performance of their processes, and to stay eﬃcient. Secondly, with increasing organization size the complexity of problems in the context of production or service processes also increases. As a result, standard, traditional, optimization techniques are often not able to solve these problems of increased complexity with justiﬁable eﬀort in an acceptable time period. Therefore, to overcome these problems, and to develop systems that solve these complex problems, researchers proposed using genetic and evolutionary algorithms (GEAs). Using these natureinspired search methods it is possible to overcome some limitations of traditional optimization methods, and to increase the number of solvable problems. The application of GEAs to many optimization problems in organizations often results in good performance and high quality solutions. For successful and eﬃcient use of GEAs, it is not enough to simply apply standard GEAs. In addition, it is necessary to ﬁnd a proper representation for the problem and to develop appropriate search operators that ﬁt well to the properties of the representation. The representation must at least be able to encode all possible solutions of an optimization problem, and genetic operators such as crossover and mutation should be applicable to it. Many optimization problems can be encoded by a variety of diﬀerent representations. In addition to traditional binary and continuous string encodings, a large number of other, often problemspeciﬁc representations have been
2
1 Introduction
proposed over the last few years. Unfortunately, practitioners often report a signiﬁcantly diﬀerent performance of GEAs by simply changing the used representation. These observations were conﬁrmed by empirical and theoretical investigations. The diﬃculty of a speciﬁc problem, and with it the performance of GEAs, can be changed dramatically by using diﬀerent types of representations. Although it is well known that representations aﬀect the performance of GEAs, no theoretical models exist which describe the eﬀect of representations on the performance of GEAs. Therefore, the design of proper representations for a speciﬁc problem mainly depends on the intuition of the GEA designer and developing new representations is often a result of repeated trial and error. As no theory of representations exists, the current design of proper representations is not based on theory, but more a result of black art. The lack of existing theory not only hinders a theoryguided design of new representations, but also results in problems when deciding which of the diﬀerent representations should be used for a speciﬁc optimization problem. Currently, comparisons between representations are based mainly on limited empirical evidence, and random or problemspeciﬁc test function selection. However, empirical investigations only allow us to judge the performance of representations for the speciﬁc test problem, but do not help us in understanding the basic principles behind it. A representation can perform well for many diﬀerent test functions, but fails for the one problem which one really wants to solve. If it is possible to develop theoretical models which describe the inﬂuence of representations on measurements of GEA performance – like time to convergence and solution quality – then representations can be used eﬃciently and in a theoryguided manner. Choosing and designing proper representations will not remain the black art of GEA research but become a well predictable engineering task.
1.1 Purpose The purpose of this work is to bring some order into the unsettled situation which exists and to investigate how representations inﬂuence the performance of genetic and evolutionary algorithms. This work develops elements of representation theory and applies them to designing, selecting, using, choosing among, and comparing representations. It is not the purpose of this work to substitute the current black art of choosing representations by developing barely applicable, abstract, theoretical models, but to formulate an applicable representation theory that can help researchers and practitioners to ﬁnd or design the proper representation for their problem. By providing an applicable theory of representations this work should bring us to a point where the inﬂuence of representations on the performance of GEAs can be judged easily and quickly in a theoryguided manner. The ﬁrst step in the development of an applicable theory is to identify which properties of representations inﬂuence the performance of GEAs and
1.1 Purpose
3
how. Therefore, this work models for diﬀerent properties of representations how solution quality and time to convergence is changed. Using this theory, it is possible to formulate a framework for eﬃcient design of representations. The framework describes how the performance of GEAs, measured by run duration and solution quality, is aﬀected by the properties of a representation. By using this framework, the inﬂuence of diﬀerent representations on the performance of GEAs can be explained. Furthermore, it allows us to compare representations in a theorybased manner, to predict the performance of GEAs using diﬀerent representations, and to analyze and design representations guided by theory. One does not have to rely on empirical studies to judge the performance of a representation for a speciﬁc problem, but can use existing theory for predicting GEA performance. By using this theory, the situation exists where empirical results are only needed to validate theoretical predictions. However, developing a general theory of how representations aﬀect GEA performance is a demanding and diﬃcult task. To simplify the problem, it must be decomposed, and the diﬀerent properties of encodings must be investigated separately. Three diﬀerent properties of representations are considered in this work: Redundancy, scaling, and locality, respectively distance distortion. For these three properties of representations models are developed that describe their inﬂuence on the performance of GEAs. Additionally, population sizing and time to convergence models are presented for redundant and nonuniformly scaled encodings. Furthermore, it is shown that lowlocality representations can change the diﬃculty of the problem. For lowlocality encodings, it can not exactly be predicted how GEA performance is changed, without having complete knowledge regarding the structure of the problem. Although the investigation is limited only to three important properties of representations, the understanding of the inﬂuence of these three properties of encodings on the performance of GEAs brings us a large step forward towards a general theory of representations. To illustrate the signiﬁcance and importance of the presented representation framework on the performance of GEAs, the framework is used for analyzing the performance of binary representations of integers and tree representations. The investigations show that the current framework considering only three representation properties gives us a good understanding of the inﬂuence of representations on GEA performance as it allows us to predict the performance of GEAs using diﬀerent types of representations. The results conﬁrm that choosing a proper representation has a large impact on the performance of GEAs, and therefore, a better theoretical understanding of representations is necessary for an eﬃcient use of genetic search. Finally, it is illustrated how the presented theory of representations can help us in designing new representations more reasonably. It is shown by example for tree representations, that the presented framework allows theoryguided design. Not black art, but a deeper understanding of representations allows us to develop representations which result in a high performance of genetic and evolutionary algorithms.
4
1 Introduction
1.2 Organization The organization of this work follows its purpose. It is divided into two large parts: After the ﬁrst two introductory chapters, the ﬁrst part (Chaps. 3 and 4) provides the theory regarding representations. The second part (Chaps. 5, 6, 7, and 8) applies the theory to the analysis and design of representations. Chapter 3 presents theory on how diﬀerent properties of representations affect GEA performance. Consequently, Chap. 4 uses the theory for formulating the timequality framework. Then, in Chap. 5, the presented theory of representations is used for analyzing the performance diﬀerences between binary representations of integers. Finally, the framework is used in Chap. 6, Chap. 7, and Chap. 8 for the analysis and design of tree representations and search operators. The following paragraphs give a more detailed overview about the contents of each chapter. Chapter 1 is the current chapter. It sets the stage for the work and describes the beneﬁts that can be gained from a deeper understanding of representations for GEAs. Chapter 2 provides the background necessary for understanding the main issues of this work about representations for GEAs. Section 2.1 introduces representations which can be described as a mapping that assigns one or more genotypes to every phenotype. The genetic operators selection, crossover, and mutation are applied on the level of alleles to the genotypes, whereas the ﬁtness of individuals is calculated from the corresponding phenotypes. Section 2.2 illustrates that selectorecombinative GEAs, where only crossover and selection operators are used, are based on the notion of schemata and building blocks. Using schemata and building blocks is an approach to explain why and how GEAs work. This is followed in Sect. 2.3 by a brief review of reasons and measurements for problem diﬃculty. Measurements of problem diﬃculty are necessary to be able to compare the inﬂuence of diﬀerent types of representations on the performance of GEAs. The chapter ends with some earlier, mostly qualitative recommendations for the design of eﬃcient representations. Chapter 3 presents three aspects of a theory of representations for GEAs. It investigates how redundant encodings, encodings with exponentially scaled alleles, and representations that modify the distances between the corresponding genotypes and phenotypes, inﬂuence GEA performance. Population sizing models and time to convergence models are presented for redundant and exponentially scaled representations. Section 3.1 illustrates that redundant encodings inﬂuence the supply of building blocks in the initial population of GEAs. Based on this observation the population sizing model from Harik et al. (1997) and the time to convergence model from Thierens and Goldberg (1993) can be extended from nonredundant to redundant representations. Because redundancy mainly aﬀects the number of copies in the initial population that are given to the optimal solution, redundant representations increase solution quality and reduce time to convergence if individuals that are similar to the optimal solution are overrepresented. Section 3.2 focuses on exponen
1.2 Organization
5
tially scaled representations. The investigation into the eﬀects of exponentially scaled encodings shows that, in contrast to uniformly scaled representations, the dynamics of genetic search are changed. By combining the results from Harik et al. (1997) and Thierens (1995) a population sizing model for exponentially scaled building blocks with and without considering genetic drift can be presented. Furthermore, the time to convergence when using exponentially scaled representations is calculated. The results show that when using nonuniformly scaled representations, the time to convergence increases. Finally, Sect. 3.3 investigates the inﬂuence of representations that modify the distances between corresponding genotypes and phenotypes on the performance of GEAs. When assigning the genotypes to the phenotypes, representations can change the distances between the individuals. This eﬀect is denoted as locality or distance distortion. Investigating its inﬂuence shows that the size and length of the building blocks, and therefore the complexity of the problem are changed if the distances between the individuals are not preserved. Therefore, to ensure that an easy problem remains easy, highlocality representations which preserve the distances between the individuals are necessary. Chapter 4 presents the framework for theoryguided analysis and design of representations. The chapter combines the three elements of representation theory from Chap. 3 – redundancy, scaling, and locality – to a timequality framework. It formally describes how the time to convergence and the solution quality of GEAs depend on these three aspects of representations. The chapter ends with implications for the design of representations which can be derived from the framework. In particular, the framework tells us that uniformly scaled representations are robust, that exponentially scaled representations are fast but inaccurate, and that lowlocality representations change the diﬃculty of the underlying optimization problem. Chapter 5 uses the framework for a theoryguided analysis of binary representations of integers. Because the potential number of schemata is higher when using binary instead of integer representations, users often favor the use of binary instead of integer representations, when applying GEAs to integer problems. By using the framework it can be shown that the redundant unary encoding results in low GEA performance if the optimal solution is underrepresented. Both, Gray and binary encoding are lowlocality representations as they change the distances between the individuals. Therefore, both representations change the complexity of optimization problems. It can be seen that the easy integer onemax problem is easier to solve when using the binary representation, and the diﬃcult integer deceptive trap is easier to solve when using the Gray encoding. Chapter 6 uses the framework for the analysis and design of tree representations. For tree representations, standard crossover and mutation operators are applied to treespeciﬁc genotypes. However, ﬁnding or deﬁning treespeciﬁc genotypes and genotypephenotype mappings is a diﬃcult task because there are no intuitive genotypes for trees. Therefore, researchers have proposed a variety of diﬀerent, more or less tricky representations which can be used in
6
1 Introduction
combination with standard crossover and mutation operators. A closer look at the Pr¨ ufer number representation in Sect. 6.2 reveals that the encoding in general is a lowlocality representation and modiﬁes the distances between corresponding genotypes and phenotypes. As a result, problem complexity is modiﬁed, and many easy problems become too diﬃcult to be properly solved using GEAs. Section 6.3 presents an investigation into the characteristic vector representation. Because invalid solutions are possible when using characteristic vectors, an additional repair mechanism is necessary which makes the representation redundant. Characteristic vectors are uniformly redundant and GEA performance is independent of the structure of the optimal solution. However, the repair mechanism results in nonsynonymous redundancy. Therefore, GEA performance is reduced and the time to convergence increases. With increasing problem size, the repair process generates more and more links randomly and oﬀspring trees have not much in common with their parents. Therefore, for larger problems guided search is no longer possible and GEAs behave like random search. In Sect. 6.4, the investigation into the redundant link and node biased representation reveals that the representation overrepresents trees that are either starlike or minimum spanning treelike. Therefore, GEAs using this type of representation perform very well if the optimal solution is similar to stars or to the minimum spanning tree, whereas they fail when searching for optimal solutions that do not have much in common with stars or the minimum spanning tree. Finally, Sect. 6.5 presents network random keys (NetKeys) as an example for the theoryguided design of a tree representation. To construct a robust and predictable tree representation, it should be non or uniformly redundant, uniformly scaled, and have highlocality. When combining the concepts of the characteristic vector representation with weighted representations like the link and node biased representation, the NetKey representation can be created. In analogy to random keys, the links of a tree are represented as ﬂoating numbers, and a construction algorithm constructs the corresponding tree from the keys. The NetKey representation allows us to distinguish between important and unimportant links, is uniformly redundant, uniformly scaled, and has high locality. Chapter 7 uses the insights into representation theory for the analysis and design of search operators for trees. In contrast to Chap. 6 where standard search operators are applied to treespeciﬁc genotypes, now treespeciﬁc search operators are directly applied to tree structures. Such types of representations are also known as direct representations as there is no additional genotypephenotype mapping. Section 7.1 presents a direct representation for trees (NetDir) as an example for the design of direct tree representations. Search operators are directly applied to trees and problemspeciﬁc crossover and mutation operators are developed. The search operators for the NetDir representation are developed based on the notion of schemata. Section 7.2 analyzes the edgeset encoding which encodes trees directly by listing their edges. Search operators for edgesets may be heuristic, considering the weights of edges they include in oﬀspring, or naive, including edges without
1.2 Organization
7
regard to their weights. Analyzing the properties of the heuristic variants of the search operators shows that solutions similar to the minimum spanning tree are favored. In contrast, the naive variants are unbiased which means that genetic search is independent of the structure of the optimal solution. Although no explicit genotypephenotype mapping exists for edgesets and the framework for the design of representations cannot be directly applied, the framework is useful for structuring the analysis of edgesets. Similarly to nonuniformly redundant representations, edgesets overrepresent some speciﬁc types of tree and GEA performance increases if optimal solutions are similar to the MST. Analyzing and developing direct representations nicely illustrates the tradeoﬀ between designing either problemspeciﬁc representations or problemspeciﬁc operators. For eﬃcient GEAs, it is necessary either to design problemspeciﬁc representations and to use standard operators like onepoint or uniform crossover, or to develop problemspeciﬁc operators and to use direct representations. Chapter 8 veriﬁes theoretical predictions concerning GEA performance by empirical veriﬁcation. It compares the performance of GEAs using different types of representations for the onemax tree problem, the deceptive tree problem, and various instances of the optimal communication spanning tree problem. The instances of the optimal communication spanning trees are presented in the literature (Palmer 1994; Berry et al. 1997; Raidl 2001; Rothlauf et al. 2002). The results show that with the help of the framework the performance of GEAs using diﬀerent types of representations can be well predicted. Chapter 9 summarizes the major contributions of this work, describes how the knowledge about representations has changed, and gives some suggestions for future research.
2 Representations for Genetic and Evolutionary Algorithms
In this second chapter, we present an introduction into the ﬁeld of representations for genetic and evolutionary algorithms. The chapter provides the basis and deﬁnitions which are essential for understanding the content of this work. Genetic and evolutionary algorithms (GEAs) are natureinspired optimization methods that can be advantageously used for many optimization problems. GEAs imitate basic principles of life and apply genetic operators like mutation, crossover, or selection to a sequence of alleles. The sequence of alleles is the equivalent of a chromosome in nature and is constructed by a representation which assigns a string of symbols to every possible solution of the optimization problem. Earlier work (Goldberg 1989c; Liepins and Vose 1990) has shown that the behavior and performance of GEAs is strongly inﬂuenced by the representation used. As a result, many recommendations for a proper design of representations were made over the last few years (Goldberg 1989c; Radcliﬀe 1991a; Radcliﬀe 1991b; Palmer 1994; Ronald 1997). However, most of these design rules are of a qualitative nature and are not particularly helpful for estimating exactly how diﬀerent types of representations inﬂuence problem diﬃculty. Consequently, we are in need of a theory of representations which allows us to theoretically predict how diﬀerent types of representations inﬂuence GEA performance. This chapter provides some of the utilities that are necessary for reaching this goal. The chapter starts with an introduction into genetic representations. We describe the notion of genotypes and phenotypes and illustrate how the ﬁtness function can be decomposed into a genotypephenotype, and a phenotypeﬁtness mapping. The section ends with a brief characterization of widely used representations. In Sect. 2.2, we provide the basis for genetic and evolutionary algorithms. After a brief description of the principles of a simple genetic algorithm (GA), we present the underlying theory which explains why and how selectorecombinative GAs using crossover as a main search operator work. The schema theorem tells us that GAs process schemata and the building block hypothesis assumes that many realworld problems are decomposable (or at least quasidecomposable). Therefore, GAs perform well for these types
10
2 Representations for Genetic and Evolutionary Algorithms
of problems. Section 2.3 addresses the diﬃculty of problems. After illustrating that the reasons for problem diﬃculty depend on the used optimization method, we describe some common measurements of problem complexity. Finally, in Sect. 2.4 we review some former recommendations for the design of eﬃcient representations.
2.1 Genetic Representations This section introduces representations for genetic and evolutionary algorithms. When using GEAs for optimization purposes, representations are required for encoding potential solutions. Without representations, no use of GEAs is possible. In Sect 2.1.1, we introduce the notion of genotype and phenotype. We brieﬂy describe how nature creates a phenotype from the corresponding genotype by the use of representations. This more biologybased approach to representations is followed in Sect. 2.1.2 by a more formal description of representations. Every ﬁtness function f which assigns a ﬁtness value to a genotype xg can be decomposed into the genotypephenotype mapping fg , and the phenotypeﬁtness mapping fp . Finally, in Sect. 2.1.3 we brieﬂy review the most important types of representations. 2.1.1 Genotypes and Phenotypes In 1866, Mendel recognized that nature stores the complete genetic information for an individual in pairwise alleles (Mendel 1866). The genetic information that determines the properties, appearance, and shape of an individual is stored by a number of strings. Later, it was discovered that the genetic information is formed by a double string of four nucleotides, called DNA. Mendel realized that nature distinguishes between the genetic code of an individual and its outward appearance. The genotype represents all the information stored in the chromosomes and allows us to describe an individual on the level of genes. The phenotype describes the outward appearance of an individual. A transformation exists – a genotypephenotype mapping or a representation – that uses the genotypic information to construct the phenotype. To represent the large number of possible phenotypes with only four nucleotides, the genotypic information is not stored in the alleles itself, but in the sequence of alleles. By interpreting the sequence of alleles, nature can encode a large number of diﬀerent phenotypic expressions using only a few diﬀerent types of alleles. In Fig. 2.1, we illustrate the diﬀerences between chromosome, gene, and allele. A chromosome describes a string of certain length where all the genetic information of an individual is stored. Although nature often uses more than one chromosome, most GEA applications only use one chromosome for encoding the genotypic information. Each chromosome consist of many alleles.
2.1 Genetic Representations
11
1 0 1 1 0 1 0 1 1 1 1 1 allele
gene
chromosome
Figure 2.1. Alleles, genes, and chromosomes
Alleles are the smallest information units in a chromosome. In nature, alleles exist pairwise, whereas in most GEA implementations an allele is represented by only one symbol. If for example, we use a binary representation, an allele can have either the value 0 or 1. If a phenotypic property of an individual, like its hair color or eye size is determined by one or more alleles, then these alleles together are denoted to be a gene. A gene is a region on a chromosome that must be interpreted together and which is responsible for a speciﬁc phenotypic property. When talking about individuals in a population, we must carefully distinguish between genotypes and phenotypes. The phenotypic appearance of an individual determines its success in life. Therefore, when comparing the abilities of diﬀerent individuals we must judge them on the level of the phenotype. However, when it comes to reproduction we must view individuals on the level of the genotype. During sexual reproduction, the oﬀspring does not inherit the phenotypic properties of its parents, but only the genotypic information regarding the phenotypic properties. The oﬀspring inherits genetic material from both parents. Therefore, genetic operators work on the level of the genotype, whereas the evaluation of the individuals is performed on the level of the phenotype. 2.1.2 Decomposition of the Fitness Function The following subsection provides some basic deﬁnitions for our discussion of representations for genetic and evolutionary algorithms. We show how every optimization problem that should be solved by using GEAs can be decomposed into a genotypephenotype fg , and a phenotypeﬁtness mapping fp . We deﬁne Φg as the genotypic search space where the genetic operators such as recombination or mutation are applied to. An optimization problem on Φg could be formulated as follows: The search space Φg is either discrete or continuous, and the function f (x) : Φg → R assigns an element in R to every element in the genotype space Φg . The optimization problem is deﬁned by ﬁnding the optimal solution x ˆ = max f (x), x∈Φg
where x is a vector of decision variables (or alleles), and f (x) is the objective or ﬁtness function. The vector x ˆ is the global maximum. We have chosen to
12
2 Representations for Genetic and Evolutionary Algorithms
illustrate a maximization problem, but without loss of generality, we could also model a minimization problem. To be able to apply GEAs to a problem, the inverse function f −1 does not need to exist. In general, the cardinality of Φg can be greater than two, but we want to focus for the most part in our investigation on binary search spaces with cardinality two. Thus, GEAs search in the binary space Φg = {0, 1}l , with the length of the string x equal to l and the size of the search space Φg  = 2l . The introduction of an explicit representation is necessary if the phenotype of a problem can not be depicted as a string or in another way that is accessible for GEAs. Furthermore, the introduction of a representation could be useful if there are constraints or restrictions on the search space that can be advantageously modeled by a speciﬁc encoding. Finally, using the same genotypes for diﬀerent types of problems, and only interpreting them diﬀerently by using a diﬀerent genotypephenotype mapping, allows us to use standard genetic operators with known properties. Once we have gained some knowledge about speciﬁc kinds of genotypes, we can easily reuse that knowledge, and it is not necessary to develop any new operators. When using a representation for genetic and evolutionary algorithms we have to introduce – in analogy to nature – phenotypes and genotypes (Lewontin 1974; Liepins and Vose 1990). The ﬁtness function f is decomposed into two parts. The ﬁrst maps the genotypic space Φg to the phenotypic space Φp , and the second maps Φp to the ﬁtness space R. Using the phenotypic space Φp we get: fg (xg ) : Φg → Φp , fp (xp ) : Φp → R, where f = fp ◦ fg = fp (fg (xg )). The genotypephenotype mapping fg is determined by the type of representation used. fp represents the ﬁtness function and assigns a ﬁtness value fp (xp ) to every individual xp ∈ Φp . The genetic operators are applied to the individuals in Φg (Bagley 1967; Vose 1993). If the genetic operators are applied directly to the phenotype it is not necessary to specify a representation and the phenotype is the same as the genotype: fg (xg ) : Φg → Φg , fp (xp ) : Φg → R. In this case, fg is the identity function fg (xg ) = xg . All genotypic properties are transformed to the phenotypic space. The genotypic space is the same as the phenotypic space and we have a direct representation. Because there is no longer an additional mapping between Φg and Φp , a direct representation
2.1 Genetic Representations
13
does not change any aspect of the phenotypic problem such as complexity, distances between the individuals, or location of the optimal solution. However, when using direct representations, we could not in general use standard genetic operators, but have to deﬁne problemspeciﬁc operators (see Sect. 7 regarding the analysis of direct representations for trees). Therefore, the key factor for the success of a GEA using a direct representation is not in ﬁnding a “good” representation for a speciﬁc problem, but in developing proper search operators. We have seen how every optimization problem we want to solve with GEAs can be decomposed into a genotypephenotype fg , and a phenotypeﬁtness mapping fp . The genetic operators are applied to the genotypes xg ∈ Φg , and the ﬁtness of the individuals is calculated from the phenotypes xp ∈ Φp . 2.1.3 Types of Representations In this subsection, we describe some of the most important and widely used representations, and summarize some of their major characteristics. In this work, we do not provide an overview of all representations which appear in the literature because every time a GEA is used, some kind of representation is necessary. This means within the scope of this research it is not possible to review all representations which have once been presented. For a more detailed overview about diﬀerent types of representations see B¨ack et al. (1997, Sect. C1). Binary Representations Binary representations are the most common representations for selectorecombinative GEAs. Selectorecombinative GEAs process schemata and use crossover as the main search operator. Using these types of GEAs, mutation only serves as background noise. The search space Φg is denoted by Φg = {0, 1}l , where l is the length of a binary vector xg = (xg1 , . . . xgl ) ∈ {0, 1}l (Goldberg 1989c). When using binary representations the genotypephenotype mapping fg depends on the speciﬁc optimization problem that should be solved. For many combinatorial optimization problems the binary representation allows a direct and very natural encoding. When encoding integer problems by using binary representations, speciﬁc genotypephenotype mappings are necessary. Diﬀerent types of binary representations for integers assign the integers xp ∈ Φp (phenotypes) in a diﬀerent way to the binary vectors xg ∈ Φg (genotypes). The most common representations are the binary, Gray, and unary encoding. For a more detailed description of these three types of encodings see Sect. 5.2. When encoding continuous variables by using binary vectors the accuracy of the representation depends on the number of bits that represent a phenotypic continuous variable. By increasing the number of bits that are used for
14
2 Representations for Genetic and Evolutionary Algorithms
representing one continuous variable the accuracy of the representation can be increased. When encoding a continuous phenotypic variable xp ∈ [0, 1] by using a binary vector of length l the maximal accuracy is 1/2l+1 . Integer Representations Instead of using binary strings with cardinality χ = 2, where χ ∈ {N+ \{0, 1}}, higher χary alphabets can also be used for the genotypes. Then, instead of a binary alphabet a χary alphabet is used for the string of length l. Instead of encoding 2l diﬀerent individuals with a binary alphabet, we are able to encode χl diﬀerent possibilities. The size of the search space increases from Φg  = 2l to Φg  = χl . For many integer problems, users often prefer to use binary instead of integer representations because schema processing is maximum with binary alphabets when using standard recombination operators (Goldberg 1990b). Realvalued Representations When using realvalued representations, the search space Φg is deﬁned as Φg = Rl , where l is the length of the realvalued chromosome. When using realvalued representations, researchers often favor mutationbased GEAs like evolution strategies or evolutionary programming. These types of optimization methods mainly use mutation and search through the search space by adding a multivariate zeromean Gaussian random variable to each variable. In contrast, when using crossoverbased GEAs realvalued problems are often represented by using binary representations (see previous paragraph about binary representations). Realvalued representations can not exclusively be used for encoding realvalued problems, but also for other permutation and combinatorial problems. Trees, schedules, tours, or other combinatorial problems can easily be represented by using realvalued vectors and special genotypephenotype mappings (see also Sect. 6.4 (LNB encoding) and Sect. 6.5 (NetKeys)). Messy Representations In all the previously presented representations, the position of each allele is ﬁxed along the chromosome and only the corresponding value is speciﬁed. The ﬁrst geneindependent representation was proposed by Holland (1975). He proposed the inversion operator which changes the relative order of the alleles in the string. The position of an allele and the corresponding value are coded together as a tuple in a string. This type of representation can be used for binary, integer, and realvalued representations and allows an encoding which is independent of the position of the alleles in the chromosome. Later, Goldberg et al. (1989) used this positionindependent representation for the messy genetic algorithm.
2.2 Genetic and Evolutionary Algorithms
15
Direct Representations In Sect. 2.1.2, we have seen that a representation is direct if fg (xg ) = xg . Then, a phenotype is the same as the corresponding genotype and the problemspeciﬁc genetic operators are applied directly to the phenotype. As long as xp = xg is either a binary, an integer, or a realvalued string, standard recombination and mutation operators can be used. Then, it is often easy to predict GEA performance by using existing theory. The situation is diﬀerent if direct representations are used for problems whose phenotypes are not binary, integer, or realvalued. Then, standard recombination and mutation operators can not be used any more. Specialized operators are necessary that allow oﬀspring to inherit important properties from their parents (Radcliﬀe 1991a; Radcliﬀe 1991b; Kargupta et al. 1992; Radcliﬀe 1993a). In general, these operators depend on the speciﬁc structure of the phenotypes xp and must be developed for every optimization problem separately. For more information about direct representations we refer to Chap. 7.
2.2 Genetic and Evolutionary Algorithms In this section, we introduce genetic and evolutionary algorithms. We illustrate basic principles and outline the basic functionality of GEAs. The schema theorem stated by Holland (1975) explains the performance of selectorecombinative GAs and leads us to the building block hypothesis. The building block hypothesis tells us that short, loworder and highly ﬁt schemata can be recombined to form higherorder schemata and complete strings with high ﬁtness. 2.2.1 Principles Genetic and evolutionary algorithms were introduced by Holland (1975) and Rechenberg (1973). By imitating basic principles of nature they created optimization algorithms which have successfully been applied to a wide variety of problems. The basic principles of GEAs are derived from the principles of life which were ﬁrst described by Darwin (1859): “Owing to this struggle for life, variations, however slight and from whatever cause proceeding, if they be in any degree proﬁtable to the individuals of a species, in their inﬁnitely complex relations to other organic beings and to their physical conditions of life, will tend to the preservation of such individuals, and will generally be inherited by the oﬀspring. The oﬀspring, also, will thus have a better chance of surviving, for, of the many individuals of any species which are periodically born, but a small number can survive. I have called this principle, by which each slight variation, if useful, is preserved, by the term Natural Selection.”
16
2 Representations for Genetic and Evolutionary Algorithms
Darwin’s ideas about the principles of life can be summarized by the following three basic principles: • There is a population of individuals with diﬀerent properties and abilities. An upper limit for the number of individuals in a population exists. • Nature creates new individuals with similar properties to the existing individuals. • Promising individuals are selected more often for reproduction by natural selection. In the following section, we brieﬂy illustrate these three principles. We have seen in Sect. 2.1 that the properties and abilities of an individual which are characterized by its’ phenotype are encoded in the genotype. Therefore, based on diﬀerent genotypes, individuals with diﬀerent properties exist (Mendel 1866). Because resources are ﬁnite, the number of individuals that form a population is limited. If the number of individuals exceeds the existing upper limit, some of the individuals are removed from the population. The individuals in the population do not remain the same, but change over the generations. New oﬀspring are created which inherit some properties of their parents. These new oﬀspring are not chosen randomly but are somehow similar to their parents. To create the oﬀspring, genetic operators like mutation and recombination are used. Mutation operators change the genotype of an individual slightly, whereas recombination operators combine the genetic information of the parents to create new oﬀspring. When creating oﬀspring, natural selection more often selects promising individuals for reproduction than lowquality solutions. Highly ﬁt individuals are allowed to create more oﬀspring than inferior individuals. Therefore, inferior individuals are removed from the population after a few generations and have no chance of creating oﬀspring with similar properties. As a result, the average ﬁtness of a population increases over the generations. In the following paragraphs, we want to describe how the principles of nature were used for the design of genetic and evolutionary algorithms. 2.2.2 Functionality Genetic and evolutionary algorithms imitate the principles of life outlined in the previous subsection and use it for optimization purposes. Researchers have proposed many diﬀerent variants of GEAs in the literature. For illustrating the basic functionality of GEAs we want to use the traditional standard simple genetic algorithm (GA) illustrated by Goldberg (1989c). This type of GEAs uses crossover as the main operator and mutation serves only as background noise. GAs are widely known and well understood. GAs use a constant population of size N , the individuals consist of binary strings with length l, and genetic operators like uniform or npoint crossover are directly applied to the genotypes. The basic functionality of a traditional simple GA is very simple. After randomly creating and evaluating
2.2 Genetic and Evolutionary Algorithms
17
an initial population, the algorithm iteratively creates new generations. New generations are created by recombining the selected highly ﬁt individuals and applying mutation to the oﬀspring. • initialize population – create initial population – evaluate individuals in initial population • create new populations – select ﬁt individuals for reproduction – generate oﬀspring with genetic operator crossover – mutate oﬀspring – evaluate oﬀspring One speciﬁc type of genetic algorithms are selectorecombinative GAs. These types of GAs use only selection and recombination (crossover). No mutation is used. Using selectorecombinative GAs gives us the advantage of being able to investigate the eﬀects of diﬀerent representations on crossover alone and to eliminate the eﬀects on mutation. This is useful if we use GAs in a way such that they propagate schemata (compare Sect. 2.2.3), and where mutation is only used as additional background noise. When focusing on GEAs where mutation functions as the main search operator, the reader is referred to other work (Rechenberg 1973; Schwefel 1975; Schwefel 1981; Schwefel 1995; B¨ack and Schwefel 1995). In the following paragraphs, we brieﬂy explain the basic elements of a GA. For selecting highly ﬁt individuals for reproduction a large number of diﬀerent selection schemes have been developed. The most popular are proportionate (Holland 1975) and tournament selection (Goldberg et al. 1989). When using proportionate selection, the expected number of copies an individual has in the next population is proportional to its ﬁtness. The chance of an individual xi of being selected for recombination is calculated as f (xi ) , N j=1 f (xj ) where N denotes the number of individuals in a population. With increasing ﬁtness an individual is chosen more often for reproduction. When using tournament selection, a tournament between s randomly chosen diﬀerent individuals is held and the one with the highest ﬁtness is chosen for recombination and added to the mating pool. After N tournaments of size s the mating pool is ﬁlled. We have to distinguish between tournament selection with and without replacement. If we perform tournament selection with replacement we choose for every tournament s individuals from all individuals in the population. Then, the mating pool is ﬁlled after N tournaments. If we perform a tournament without replacement there are s rounds. In each round we have N/s tournaments and we choose the individuals for a tournament
18
2 Representations for Genetic and Evolutionary Algorithms
from those who have not already taken part in a tournament in this round. After all individuals have performed a tournament in one round (after N/s tournaments) the round is over and all individuals are considered again for the next round. Therefore, to completely ﬁll the mating pool s rounds are necessary. The mating pool consists of all individuals who are chosen for recombination. When using tournament selection, there are no copies of the worst individual, and either an average of s copies (with replacement), or exactly s copies (without replacement) of the best individual in the mating pool. For more information concerning diﬀerent tournament selection schemes see B¨ack et al. (1997, C2) and Sastry and Goldberg (2001). Crossover operators imitate the principle of sexual reproduction and are applied to the individuals in the mating pool. In many GA implementations, crossover produces two new oﬀspring from two parents by exchanging substrings. The most common crossover operators are onepoint (Holland 1975), and uniform crossover (Syswerda 1989). When using onepoint crossover, a crossover point c = {1, . . . , l − 1} is initially chosen randomly. Two children are then created from the two parents by swapping the subp1 p1 and strings. As a result, we get for the parents xp1 = xp1 1 , x2 , . . . , xl p2 p2 p2 p1 p1 p2 p2 , x . . . x xp2 = x1 , x2 , . . . , xl the oﬀspring xo1 = x1 , x2 , . . . , xp1 c c+1 l and p2 p2 p1 p1 o2 p2 x = x1 , x2 , . . . , xc , xc+1 . . . xl . When using uniform crossover it is decided independently for every single allele of the oﬀspring from which parent it inherits the value of the allele. In most implementations no parent is preferred and the probability of an oﬀspring to inherit the value of an allele from a speciﬁc parent is p = 1/x, where x denotes the number of parents that are considered for recombination. For example, when two possible oﬀspring are considered with same probability (p = 1/2), we could get as oﬀspring p1 p2 p1 p2 p2 p1 p2 p1 and xo2 = xp2 xo1 = xp1 1 , x2 , x3 , . . . , xl−1 , xl 1 , x2 , x3 , . . . , xl−1 , xl . We see that uniform crossover can also be interpreted as (l − 1)point crossover. Mutation operators should slightly change the genotype of an individual. Mutation operators are important for local search, or if some alleles are lost during a GEA run. By randomly modifying some alleles in the population already lost alleles can be reanimated. The probability of mutation pm must be selected to be at a low level because otherwise mutation would randomly change too many alleles and the new individual would have nothing in common with its parent. Oﬀspring would be generated almost randomly and genetic search would degrade to random search. In contrast to crossover operators, mutation operators focus more on local search because they can only modify properties of individuals but can not recombine properties from diﬀerent parents. 2.2.3 Schema Theorem and Building Block Hypothesis We review explanations for the performance of selectorecombinative genetic algorithms. We start by illustrating the notion of schemata. This is followed
2.2 Genetic and Evolutionary Algorithms
19
by a brief summary of the schema theorem and a description of the building block hypothesis. Schemata Schemata were ﬁrst proposed by Holland (1975) to model the ability of GEAs to process similarities between bitstrings. A schema h = (h1 , h2 , . . . , hl ) is deﬁned as a ternary string of length l, where hi ∈ {0, 1, ∗}. ∗ denotes the “don’t care” symbol and tells us that the allele at this position is not ﬁxed. The size or order o(h) of a schema h is deﬁned as the number of ﬁxed positions (0s or 1s) in the string. A position in a schema is ﬁxed if there is either a 0 or a 1 at this position. The deﬁning length δ(h) of a schema h is deﬁned as the distance between the two outermost ﬁxed bits. The ﬁtness of a schema is deﬁned as the average ﬁtness of all instances of this schema and can be calculated as 1 f (x), f (h) = h x∈h
where h is the number of individuals x ∈ Φg that are an instance of the schema h. The instances of a schema h are all genotypes where xg ∈ h. For example, xg = 01101 and xg = 01100 are instances of h = 0 ∗ 1 ∗ ∗. The number of individuals that are an instance of a schema h can be calculated as 2l−o(h) . For a more detailed explanation of schemata in the context of GEAs the reader is referred to Holland (1975), Goldberg (1989c), or Radcliﬀe (1997). Schema Theorem Based on the notion of schemata, Holland (1975) and De Jong (1975) formulated the schema theorem which describes how the number of instances of a schema h changes over the number of generations t: δ(h) f (h, t) − pm o(h)), (1 − pc m(h, t + 1) ≥ m(h, t) ¯ l−1 f (t) where • • • • • • • •
m(h, t) is the number of instances of schema h at generation t, f (h, t) is the ﬁtness of the schema h at generation t, f¯(t) is the average ﬁtness of the population at generation t, δ(h) is the deﬁning length of schema h, pc is the probability of crossover, pm is the probability of mutation, l is the string length, o(h) is the order of schema h.
20
2 Representations for Genetic and Evolutionary Algorithms
The schema theorem describes how the number of copies that are given to a schema h depends on selection, crossover and mutation, when using a standard GA with proportionate selection, onepoint crossover, and bitﬂipping mutation. Selection favors a schema if the ﬁtness of the schema is above the average ﬁtness of the population (f (h, t) > f¯(t)). When using crossover the deﬁning length δ(h) of a schema must be small because otherwise onepoint crossover frequently disrupts long schemata. The bitﬂipping mutation operator favors low order schemata because with increasing o(h) the number of schemata which are destroyed increases. The main contribution of the schema theorem is that schemata, which ﬁtness is above average (f (h) > f¯), which have a short deﬁning length δ(h), and which are of low order o(h), receive exponentially increasing trials in subsequent generations. The theorem describes the hurdle between selection, which preserves highly ﬁt schemata, and crossover and mutation which both destroy schemata of large order or deﬁninglength. This observation brings us to the concept of building blocks (BBs). Goldberg (1989c, p. 20 and p. 41) deﬁned buildings blocks as “highly ﬁt, shortdeﬁninglength schemata” that “are propagated generation to generation by giving exponentially increasing samples to the observed best”. The notion of building blocks is frequently used in the literature but rarely deﬁned. In general, a building block can be described as a solution to a subproblem that can be expressed as a schema. A thuslike schema has high ﬁtness and its size is smaller than the length of the string. By combining BBs of lower order, a GA can form highquality overall solutions. Using the notion of genes we can interpret BBs as genes. A gene consists of one or more alleles and can be described as a schema with high ﬁtness. The alleles in a chromosome can be separated (decomposed) into genes which do not interact with each other and which determine one speciﬁc phenotypic property of an individual like hair or eye color. We see that by using building blocks we can describe – with the help of the schema theorem – how GAs can solve an optimization problem. If the subsolutions to a problem (the BBs) are short (low δ(h)) and of low order (low o(h)), then the number of correct subsolutions increases over the generations and the problem can easily be solved by a GA. The schema theorem and the concept of building blocks have attracted a lot of critical comments from various researchers (Radcliﬀe 1991b; Vose 1991; Vose 1999). The comments are mainly based on the observation that the schema theorem does not always explain the observed behavior of GEAs as it neglects the stochastic and dynamic nature of the genetic search. Diﬀerent approaches have been presented to develop and extend the schema theorem. Altenberg (1994) related the schema theorem to Price’s theorem (Price 1970), which can be viewed as a more general formulation of the schema theorem. Radcliﬀe introduced the concept of forma (compare Sect. 2.4.2) which allows us to introduce schemata not only for binary strings, but also for general genotypes. Poli et al. developed exact schema theorems in the context of genetic programming which could also be applied to general GEAs (Stephens
2.2 Genetic and Evolutionary Algorithms
21
and Waelbroeck 1999; Poli 2001a; Poli 2001b). An in depth survey of the critical comments on the schema theorem including appropriate extensions and later developments can be found in Reeves and Rowe (2003, Sect. 3). Building Block Hypothesis Using the deﬁnition of building blocks as being highly ﬁt solutions to subproblems, the building block hypothesis can be formulated. It describes the processing of building blocks and is based on the quasidecomposability of a problem (Goldberg 1989c, page 41): “Short, low order, and highly ﬁt schemata are sampled, recombined, and resampled to form strings of potentially higher ﬁtness.” The building block hypothesis basically states that GEAs mainly work due to their ability to propagate building blocks. By combining schemata of lower order which are highly ﬁt, a GEA can construct overall good solutions. The building block hypothesis can be used for explaining the high performance of GEAs in many realworld applications. It basically says that a schemata processing GEA performs well, if the problem it is applied to is quasidecomposable, that means the overall problem can be separated into smaller subproblems. If the juxtaposition of smaller, highly ﬁt, partial solutions (building blocks) does not result in good solutions, GEAs would fail in many realworld problems. Only by decomposing the overall problem into many smaller subproblems, solving these subproblems separately, and combining the good solutions, can a GEA ﬁnd good solutions to the overall optimization problem (Goldberg 2002). This observation raises the question of why the approach of separating complex problems into smaller ones and solving the smaller problems to optimality is so successful. The answer can be found in the structure of the problems themselves. Many of the problems in the real world are somehow decomposable, because otherwise all our design and optimization methods which try to decompose complex problems could not work properly. A look in the past reveals that approaching realworld problems in the outlined way has resulted in quite interesting results. Not only do human designers or engineers use the property of many complex realworld problems to be decomposable, but nature itself. Most living organisms are not just one complex system where each part interacts with all others, but they consist of various separable subsystems for sensing, movement, reproduction, or communication. By optimizing the subsystems separately, and combining eﬃcient subsystems, nature is able to create complex organisms with surprising abilities. Therefore, if we assume that many of the problems in the real world can be solved by decomposing them into smaller subproblems, we can explain the good results obtained by using GEAs for realworld problems. GEAs perform well because they try, in analogy to human intuition, to decompose the overall problem into smaller parts (the building blocks), solve the smaller
22
2 Representations for Genetic and Evolutionary Algorithms
subproblems, and combine the good solutions. A problem can be properly decomposed by identifying the interdependencies between the diﬀerent alleles. The purpose of the genetic operators is to decompose the problem by detecting which alleles in the chromosome inﬂuence each other, to solve the smaller problems eﬃciently, and to combine the subsolutions (Harik and Goldberg 1996; Harik and Goldberg 1996).
2.3 Problem Diﬃculty Previous work has shown that representations inﬂuence the behavior and performance of GEAs (Goldberg 1989c; Liepins and Vose 1990). The results revealed that when using speciﬁc representations some problems become easier, whereas other problems become more diﬃcult to solve for GEAs. To be able to systematically investigate how representations inﬂuence GEA performance, a measurement of problem diﬃculty is necessary. With the help of a diﬃculty measurement, it can be determined how representations change the complexity and diﬃculty of a problem. However, a problem does not have the same diﬃculty for all types of optimization algorithms, but diﬃculty always depends on the optimization method used. Therefore, focusing on selectorecombinative GEAs also determines the reasons of problem diﬃculty: building blocks. Consequently, in Sect. 2.3.1 we discuss reasons of problem diﬃculty and illustrate for diﬀerent types of optimization methods that diﬀerent reasons for problem diﬃculty exist. As we focus on selectorecombinative GEAs and assume that these types of GEAs process building blocks, we decompose problem diﬃculty with respect to BBs. This is followed in Sect. 2.3.2 by an illustration of diﬀerent measurements of problem diﬃculty. The measurements of problem diﬃculty are based on the used optimization method. Because we focus in this work on schemata and BBs, we later use the schemata analysis as a measurement of problem diﬃculty. 2.3.1 Reasons for Problem Diﬃculty One of the ﬁrst approaches to the question of what makes problems diﬃcult for GEAs, was the study of deceptive problems by Goldberg (1987). His studies were mainly based on the work of Bethke (1981). These early statements about deceptive problems were the origin of a discussion about the reasons of problem diﬃculty in the context of genetic and evolutionary algorithms. Searching for reasons of problem diﬃculty means investigating what makes problems diﬃcult for GEAs. Researchers recognized that there are other possible reasons of problem diﬃculty besides deception. Based on the structure of the ﬁtness landscape (Weinberger 1990; Manderick et al. 1991), the correlation between the ﬁtness of individuals describes how diﬃcult a speciﬁc problem is to solve for GEAs. By the modality of a problem, which is more popular for mutationbased
2.3 Problem Diﬃculty
23
search methods, problems can be classiﬁed into easy unimodal problems (there is only one local optimum), and diﬃcult multimodal problems (there are many local optima). Another reason for diﬃculty is found to be epistasis, which is also known as the linear separability of a problem. This describes the interference between the alleles in a string and measures how well a problem can be decomposed into smaller subproblems (Holland 1975; Davidor 1989; Davidor 1991; Naudts et al. 1997). A ﬁnal reason for problem diﬃculty is additional noise which makes most problems more diﬃcult to solve for GEAs. Many of these approaches are not focused on schemataprocessing selectorecombinative GEAs. Goldberg (2002) presented an approach of understanding problem diﬃculty based on the schema theorem and the building block hypothesis. He viewed problem diﬃculty for selectorecombinative GEAs as a matter of building blocks and decomposed it into • diﬃculty within a building block (intraBB diﬃculty), • diﬃculty between building blocks (interBB diﬃculty), and • diﬃculty outside of building blocks (extraBB diﬃculty). This decomposition of problem diﬃculty assumes that diﬃcult problems are building block challenging. In the following paragraphs, we brieﬂy discuss these three aspects of BBcomplexity. If we count the number of schemata of order o(h) = k that have the same ﬁxed positions, there are 2k diﬀerent competing schemata. Based on their ﬁtness, the diﬀerent schemata compete against each other and GEAs should increase the number of the highquality schemata. Identifying the high quality schemata and propagating them properly is the main diﬃculty of intraBB diﬃculty. Goldberg measures intraBB diﬃculty with the deceptiveness of a problem. Deceptive problems (Goldberg 1987) are most diﬃcult to solve for GEAs because GEAs are led by the ﬁtness landscape to a deceptive attractor which has maximum distance to the optimum. To reliably solve diﬃcult, for example deceptive problems, GEAs must increase the number of copies of the best BB by giving enough copies to them. One basic assumption of the schema theorem is that a problem can be decomposed into smaller subproblems. GEAs solve these smaller subproblems in parallel and try to identify the correct BBs. In general, the contributions of diﬀerent BBs to the ﬁtness function are not uniform and there can be interdependencies between the diﬀerent BBs. Because diﬀerent BBs can have diﬀerent contributions to the ﬁtness of an individual, the loss of low salient BBs during a GEA run is one of the major problems of interBB diﬃculty (compare Sect. 3.2.3). Furthermore, a problem can often not be decomposed into completely separate and independent subproblems, but there are still interdependencies between the diﬀerent BBs which are an additional source of interBB diﬃculty. Even for selectorecombinative GEAs there is a world outside the world of schemata and BBs. Sources of extraBB diﬃculty like noise have an additional inﬂuence on the performance of GEAs because selection is based on the ﬁtness
24
2 Representations for Genetic and Evolutionary Algorithms
of the individuals. Additional, nondeterministic noise randomly modiﬁes the ﬁtness values of the individuals. Therefore, selection decisions are no longer based on the quality of the solutions (and of course the BBs) but on stochastic variance. A similar problem occurs if the evaluation of the individuals is nonstationary. Evaluating ﬁtness in a nonstationary way means that individuals have a diﬀerent ﬁtness at diﬀerent moments in time. In the remainder of the subsection, we discuss that reasons of problem diﬃculty must be seen in the context of a speciﬁc optimization method. If diﬀerent optimization methods are used for the same problem then there are diﬀerent reasons of problem diﬃculty. As a result, there is no general problem diﬃculty for all types of optimization methods but we must independently identify for each optimization method the reasons of problem diﬃculty. We illustrate how problem diﬃculty depends on the used optimization method with two small examples. When using random search, the discussion of problem complexity is obsolete. During random search new individuals are chosen randomly and no prior information about the structure of the problem or previous search steps is used. As a result, all possible types of problems have the same diﬃculty. Although measurements of problem complexity, like correlation analysis or the analysis of intra, inter, or extraBB diﬃculty, lead us to believe that some problems are easier to solve than others, there are no easy or diﬃcult problems. Independently of the complexity of a problem, random search always needs on average the same number x of ﬁtness evaluations for ﬁnding the optimal solution. All problems have the same diﬃculty regarding random search, and to search for reasons of problem diﬃculty makes no sense. When comparing crossover and mutationbased evolutionary search methods, diﬀerent reasons of problem complexity exist. From the schema theorem we know that selectorecombinative GEAs propagate schemata and BBs. Therefore, BBs are the main source of complexity for these types of GEAs. Problems are easy for selectorecombinative GEAs if the problem can be properly decomposed into smaller subsolutions (the building blocks) and the intra, and interBB diﬃculty is low. However, when using mutationbased approaches like evolution strategies (Rechenberg 1973; Schwefel 1975), these reasons for problem diﬃculty are not relevant any more. Evolution strategies perform well if the structure of the solution space guides the population to the optimum (compare the good performance of evolution strategies on unimodal optimization problems). Problem complexity is not based on BBs, but more on the structure of the ﬁtness landscape. To use the notion of BBs for mutationbased optimization methods makes no sense because they propagate no schemata. We have illustrated how the diﬃculty of a problem for selectorecombinative GEAs can be decomposed into intraBB, interBB, and extra BBdiﬃculty. The decomposition is based on the assumption that GEAs decompose problems and work with schemata and BBs. The proposed BBbased reasons of problem diﬃculty can be used for selectorecombinative GEAs but can not
2.3 Problem Diﬃculty
25
be applied to other optimization methods like evolution strategies or random search. 2.3.2 Measurements of Problem Diﬃculty In the previous section, we discussed the reasons for problem diﬃculty. In the following paragraphs, we describe some measurements of problem diﬃculty. To investigate how diﬀerent representations inﬂuence the performance of GEAs, a measurement of problem diﬃculty is necessary. Based on the diﬀerent reasons for problem diﬃculty which exist for diﬀerent types of optimization methods, we discuss some common measurements of problem diﬃculty: • • • •
Correlation analysis, polynomial decomposition, Walsh coeﬃcients, and schemata analysis.
These four measurements of problem diﬃculty are widely used in the GEA literature for measuring diﬀerent types of problem diﬃculty (Goldberg 1989b; Goldberg 1992; Radcliﬀe 1993a; Horn 1995; Jones and Forrest 1995). For an overview see B¨ack et al. (1997, Chap. B2.7) or Reeves and Rowe (2003). The speciﬁc properties of the four measurements are brieﬂy discussed in the following paragraphs. Correlation analysis is based on the assumption that the high and low quality solutions are grouped together and that GEAs can use information about individuals whose genotypes are very similar for generating new oﬀspring. Therefore, problems are easy if the structure of the search space guides the search to the high quality solutions. Consequently, correlation analysis is a proper measurement for the diﬃculty of a problem when using mutationbased search approaches. Correlation analysis exploits the ﬁtness between neighboring individuals of the search space as well as the correlation of the ﬁtness between parents and their oﬀspring (For a summary see Deb et al. (1997)). The most common measurements for distance correlation are the autocorrelation function of the ﬁtness landscape (Weinberger 1990), the ﬁtness correlation coeﬃcients of genetic operators (Manderick et al. 1991), and the ﬁtnessdistance correlation (Jones 1995; Jones and Forrest 1995; Altenberg 1997). The linearity of an optimization problem can be measured by the polynomial decomposition of the problem. Each function f deﬁned on Φg = {0, 1}l can be decomposed in the form αN eTn x, f (x) = N ⊂{1,...,l}
n∈N
where the vector en contains 1 in the nth column and 0 elsewhere, T denotes transpose, and the αN are the coeﬃcients (Liepins and Vose 1991). Regarding
26
2 Representations for Genetic and Evolutionary Algorithms
the vector x having components x1 , . . . , xl , we may view f as a polynomial in the variables x1 , . . . , xl . The coeﬃcients αi describe the nonlinearity of the problem. If there are high order αN in the decomposition of the problem, the function is highly nonlinear. If the decomposition of a problem only has order 1 coeﬃcients, then the problem is linear and easy for GEAs. It is possible to determine the maximum nonlinearity of f (x) by its highest polynomial coeﬃcients. The higher the order of the αi , the more nonlinear the problem is. There is some correlation between the nonlinearity of a problem and the diﬃculty of a problem for selectorecombinative GEAs (Mason 1995), but the order of nonlinearity can only give an upper limit of the problem diﬃculty. As illustrated in the following example there could be high order αi although the problem still remains easy for a GA. The function ⎧ 1 if x = 00, ⎪ ⎪ ⎪ ⎨2 if x = 01, f (x) = (2.1) ⎪ 4 if x = 10, ⎪ ⎪ ⎩ 10 if x = 11, could be decomposed in f (x) = α0 + α1 x0 + α2 x1 + α3 x0 x1 = 1 + x0 + 3x1 + 5x0 x1 . The problem is easy for selectorecombinative GEAs (all BBs are of order k = 1), but nonlinear. Instead of decomposing a problem into its polynomial coeﬃcients, it can also be decomposed into the corresponding Walsh coeﬃcients. The Walsh transformation is analogous to the discrete Fourier transformation but for functions whose domain is a bitstring. Every realvalued function f : Φg → R over an bitstring of length l, can be expressed as a weighted sum of a set of 2l orthogonal functions called Walsh functions: f (x) =
l 2 −1
wj ψj (x),
j=0
where the Walsh functions are denoted ψj : Φg → {−1, 1}. The weights wj ∈ R are called Walsh coeﬃcients. The indices of both Walsh functions and coefﬁcients are the numerical equivalent of the binary string j. The jth Walsh function is deﬁned as: ψj (x) = (−1)bc(j∧x) , with x, j are binary strings and elements of Φg , ∧ denotes the bitwise logical and, and bc(x) is the number of one bits in x. For a more detailed explanation the reader is referred elsewhere (Goldberg 1989a; Goldberg 1989b; Vose and Wright 1998a; Vose and Wright 1998b). The Walsh coeﬃcients can be computed by the Walsh transformation: 2 −1 1 f (i)ψj (i). 2l i=0 l
wj =
2.3 Problem Diﬃculty
27
The coeﬃcients wj measure the contribution to the ﬁtness function by the interaction of the bits indicated by the positions of the 1’s in j. The larger the value of j, the higher the order of the interactions between the bits in j is. For example, w001 measures the linear contribution to f associated with bit position 2. w111 measures the nonlinear interaction between all three bits. Any function f over an l bit space can be represented as a weighted sum of all possible 2l bit interaction functions ψj . Walsh coeﬃcients are an important feature in measuring the problem diﬃculty for GEAs (Goldberg 1989a; Oei 1992; Goldberg 1992; Reeves and Wright 1994). It was shown that problems are easy for GEAs if the Walsh coeﬃcients are of order 1. Furthermore, diﬃcult problems tend to have high order Walsh coeﬃcients, but nevertheless the Walsh coeﬃcients do not give us an exact measurement of problem complexity. The highest order of the coeﬃcient wi can only give an upper limit of the problem complexity. Therefore, Walsh coeﬃcients show the same behavior as polynomials. This behavior is expected as it has already been shown that Walsh functions are polynomials (Goldberg 1989a; Goldberg 1989b; Liepins and Vose 1991). The insuﬃcient measurement of problem diﬃculty for selectorecombinative GEAs can be illustrated with the earlier example (2.1). The Walsh coeﬃcients for the former example are w = {4.25, −1.75, −2.75, 1.25}. Although the problem is quite simple, there are high order Walsh terms. If we assume that selectorecombinative GEAs process schemata and BBs, then the most natural and direct way to measure problem complexity is to analyze the size and length of the building blocks in the problem. If we assume that GEAs process building blocks, the intraBB diﬃculty of a problem can be measured by the maximum length δ(h) and size k = o(h) of the BBs h (Goldberg 1989c). A problem is denoted to be deceptive of order kmax if for k < kmax all schemata that contain parts of the best solution have lower ﬁtness than their competitors (Deb and Goldberg 1994). Schemata are competitors if they have the same ﬁxed positions. An example for competing schemata of size k = 2 for a bitstring of length l = 4 are h = 0 ∗ 0∗, h = 0 ∗ 1∗, h = 1 ∗ 0∗, and h = 1 ∗ 1∗. Therefore, the highest order kmax of the schemata that are not misleading determines the complexity of a problem for selectorecombinative GEAs. The higher the maximum order kmax of the schemata, the more diﬃcult the problem is to solve for GEAs. The average ﬁtness of the schemata for the brief example illustrated in (2.1) is shown in Table 2.1. All schemata that contain a part of the optimal order 2 schema 11 10 ﬁtness schema 01 10 00 ﬁtness 2 4 1
1 0 1* *1 ** 7 6 4.25 0* *0 1.5 2.5
Table 2.1. Average schema ﬁtness for example described by (2.1)
28
2 Representations for Genetic and Evolutionary Algorithms
solution are above average and better than their competitors. Calculating the deceptiveness of the problem based on the ﬁtness of the schemata correctly classiﬁes this problem to be very easy. Since we analyze the inﬂuence of representations on selectorecombinative GEAs, and we assume that these types of GEAs process schemata, schema ﬁtness averages are used to measure problem diﬃculty in the remainder of this work. Problems of length l are deﬁned to be fully easy if kmax = 1, and to be fully diﬃcult (compare Goldberg (1992)) if kmax = l. Therefore, when using selectorecombinative GEAs fully easy problems are the most easy problems, whereas fully diﬃcult problems are the most diﬃcult problems.
2.4 Existing Recommendations for the Design of Eﬃcient Representations Although the application of GEAs to optimization problems is not possible without using representations, mainly intuitive knowledge exists about how to choose proper representations. Up till now, there is no proven theory regarding the inﬂuence of representations on the performance of GEAs. To help users with the diﬃcult task of ﬁnding good representations, some researchers have made recommendations for the design of eﬃcient representations over the last few years. In this section, we review some recommendations which are important from the authors point of view. For a more detailed overview, the reader is referred to Ronald (1997) and Radcliﬀe (1991b). We start in Sect. 2.4.1 with the principle of meaningful building blocks and minimal alphabets which were proposed by Goldberg (1989c). About the same time, Radcliﬀe developed the concept of forma which is presented in Sect. 2.4.2. Some years later, Palmer (1994) presented more speciﬁc guidelines about proper design of representations (compare Sect. 2.4.3). Finally, we illustrate in Sect. 2.4.4 the recommendations made by Ronald (1997). 2.4.1 Goldberg’s Meaningful Building Blocks and Minimal Alphabets Some of the ﬁrst recommendations for the construction of representations were made by Goldberg (1989c). He proposed the principle of minimal alphabets and of meaningful building blocks. It is known that the design of an encoding has a strong impact on the performance of a genetic algorithm and should be chosen carefully (compare Coli and Palazzari (1995a), Ronald (1997), and Albuquerque et al. (2000)). Goldberg (1989c, p. 80) proposed two basic design principles for encodings: • Principle of meaningful building blocks: The schemata should be short, of low order, and relatively unrelated to schemata over other ﬁxed positions.
2.4. Existing Recommendations for the Design of Eﬃcient Representations
29
• Principle of minimal alphabets: The alphabet of the encoding should be as small as possible while still allowing a natural representation of solutions. The principle of meaningful building blocks is directly motivated by the schema theorem (see Sect. 2.2.3). If schemata are highly ﬁt, short, and of low order, then their numbers exponentially increase over the generations. If the highquality schemata are long or of high order, they are disrupted by crossover and mutation and they can not be propagated properly by GEAs. Consequently, representations should modify the complexity of a problem in such a way that phenotypically long or high order BBs become genotypically short, and of low order. Then, the problem becomes easier for selectorecombinative GEAs. The principle of minimal alphabets tells us to increase the potential number of schemata by reducing the cardinality of the alphabet. When using minimal alphabets the number of possible schemata is maximal. This is the reason why Goldberg advises us to use bitstring representations, because high quality schemata are more diﬃcult to ﬁnd when using alphabets of higher cardinality (Goldberg 1989c, pp. 80ﬀ). But of course we have a tradeoﬀ between the low cardinality of an alphabet and the natural expression of the problem. Therefore, sometimes a higher cardinality of the alphabet could be helpful for GEAs (Goldberg 1991b). Goldberg’s two design principles of representations are based on the assumption that GEAs process schemata and BBs, but both principles are very abstract and general, and do not provide the user with exact and applicable guidelines. 2.4.2 Radcliﬀe’s Formae and Equivalence Classes The principles from Goldberg, which are based on schemata, are made for binary representations and selectorecombinative GAs. Therefore, they are not useful in the design of proper representations when using nonbinary genotypes, or other types of search paradigms like evolution strategies or evolutionary programming as these search methods do not process schemata. Consequently, Radcliﬀe extended the notion of schemata and introduced the more general principle of forma, which describes general subsets of the search space (Radcliﬀe 1991b; Radcliﬀe 1991a; Radcliﬀe 1992; Radcliﬀe 1993b; Radcliﬀe and Surry 1994; Radcliﬀe 1994). Formae are deﬁned as equivalence classes that are induced by a set of equivalence relations. Any possible solution of an optimization problem can be identiﬁed by specifying the equivalence class to which it belongs for each of the equivalence relations. For example, if we have a search space of faces (Surry and Radcliﬀe 1996), basic equivalence relations might be “same hair color” or “same eye color”, which would induce the formae “red hair”, “dark hair”, “green eyes”, etc. Formae of higher order like “red hair and green eyes” are then constructed by composing simple formae. The search space, which includes all possible faces, can be constructed
30
2 Representations for Genetic and Evolutionary Algorithms
with strings of alleles that represent the diﬀerent formae. For the deﬁnition of formae the structure of the phenotypes is important. For example, for binary problems possible formae would be “bit i is equal to one/zero”. When encoding tree structures, possible basic formae would be “contains link from node i to node j”. Based on these basic formae, we develop in Sect. 7.1 a direct representation for trees with appropriate search operators. An unsolved problem is to ﬁnd appropriate equivalences for the diﬀerent instances of a given problem. From the equivalences the genotypic search space Φg and the genotypephenotype mapping fg can be constructed. Usually, a solution is encoded as a string of alleles, each of which indicates that the solution satisﬁes a particular equivalence. Radcliﬀe (1991a) proposed several design principles for creating appropriate equivalences for a given problem. The most important design principle is that the generated formae should group together solutions of related ﬁtness (Radcliﬀe and Surry 1994), in order to create a ﬁtness landscape or structure of the search space that can be exploited by search operators. Search operators are constructed based on the deﬁned formae. Radcliﬀe recognized that the genotypic search space, the genotypephenotype mapping, and the search operators belong together and their design can not be separated from each other (Radcliﬀe 1992). For the development of appropriate search operators that are based on predeﬁned formae he formulated the following four design principles (Radcliﬀe 1991a; Radcliﬀe 1994): • Respect: Oﬀspring produced by recombination should be members of all formae to which both their parents belong. This means for the “face example” that oﬀspring should have red hair and green eyes if both parents have red hair and green eyes. • Transmission: An oﬀspring should be equivalent to one of its parents under each of the basic equivalence relations. This means that every gene should be set to an allele which is taken from one of the parents. If one parent has dark hair and the other red hair, then the oﬀspring has either dark or red hair. • Assortment: An oﬀspring can be formed with any compatible characteristics taken from the parents. Assortment is necessary as some combinations of equivalence relations may be infeasible. This means for example, that the oﬀspring inherits the dark hair from the ﬁrst parent and the blue eyes from the second parent only if dark hair and blue eyes are compatible. Otherwise, one of the alleles is set randomly to a feasible value. • Ergodicity: The iterative use of search operators allows us to reach any point in the search space from all possible starting solutions. The recommendations from Radcliﬀe illustrate nicely that representations and search operators depend on each other and cannot be designed independently. He developed a consistent concept on how to design eﬃcient GEAs once appropriate equivalence classes (formae) are deﬁned. However, the ﬁnding of appropriate equivalence classes, which is equivalent to deﬁning the genotypic
2.4. Existing Recommendations for the Design of Eﬃcient Representations
31
search space and the genotypephenotype mapping, is often diﬃcult and remains an unsolved problem. 2.4.3 Palmer’s Tree Encoding Issues Palmer analyzed properties of tree representations (Palmer 1994). The recommendations he gave for the design of tree representations can also be applied to other types of representations. Palmer proposed the following representation issues: • An encoding should be able to represent all possible phenotypes. • An encoding should be unbiased in the sense that all possible individuals are equally represented in the set of all possible genotypic individuals. • An encoding should encode no infeasible solutions. • The decoding of the phenotype from the genotype should be easy. • An encoding should possess locality. Small changes in the genotype should result in small changes in the phenotype. Although Palmer formulated the design issues based mainly on intuition rather than on theoretical investigation, the guidelines can advantageously be used for the design of proper representations. For a more detailed description and discussion of these representation issues in the context of tree network representations, the reader is referred to Sect. 6.1.6. We will also see later in this work that the encoding of infeasible solutions (Sect. 6.3.2), a bias of the individuals (Sect. 6.4.2), or low locality of an encoding (Sect. 8.1.3) is not always necessarily disadvantageous for the performance of GEAs. 2.4.4 Ronald’s Representational Redundancy A few years ago, Ronald presented a survey of encoding issues (Ronald 1997; Ronald 1995). Representations should be chosen according to the following guidelines: • Encodings should be adjusted to a set of genetic operators in a way that the building blocks are preserved from the parents to the oﬀspring (Fox and McMahon 1991). • Encodings should minimize epistasis (Beasley et al. 1993). • Feasible solutions should be preferred. • The problem should be represented at the correct level of abstraction. • Encodings should exploit an appropriate genotypephenotype mapping process if a simple mapping to the phenotype is not possible. • Isomorphic forms, where the phenotype of an individual is encoded with more than one genotype, should not be used.
32
2 Representations for Genetic and Evolutionary Algorithms
Many of the representation issues can be put down to the principles of representations illustrated in Sect. 2.4.1. The design issue concerning isomorphic forms will be discussed in Sect. 3.1. The results will show that by using isomorphic or redundant representations the performance of GEAs can easily be increased.
3 Three Elements of a Theory of Representations
In this chapter, we study an often ignored aspect of heuristic optimization, namely the theory of representations for genetic and evolutionary algorithms. Although the importance of choosing proper representations for the performance of genetic and evolutionary algorithms is already recognized (Caruana and Schaﬀer 1988; Goldberg 1989c; Liepins and Vose 1990; Ronald et al. 1995; Radcliﬀe 1991a; Coli and Palazzari 1995b; Ronald 1997; Albuquerque et al. 2000; Kargupta 2000a; Schnier and Yao 2000; Hinterding 2000), we are still far from a complete theory of representations. Due to the fact that developing a general theory of representations is a formidable challenge, we decompose this task into smaller parts. We start by presenting three elements of representation theory which are the basis of the timequality framework of representations which we present in Chap. 4. Namely, we focus on redundancy, scaling of alleles, and locality which describes the modiﬁcation of distances between corresponding genotypes and phenotypes. We present theoretical models for these three aspects of representation theory and show how these properties of representations aﬀect the performance of GEAs. The following paragraphs discuss these three aspects of representation theory. A representation is denoted to be redundant if the number of genotypes is higher than the number of phenotypes. Therefore, for a redundant representation, a phenotype is represented on average by more than one genotype. Investigating redundancy more closely, we have to distinguish between synonymously and nonsynonymously redundant representations. Nonsynonymously redundant representations do not allow genetic operators to work properly and therefore reduce the eﬃciency of evolutionary search. When using synonymously redundant representations, GEA performance mainly depends on the change of the initial supply. Based on this observation models can be developed that allow us to determine the necessary population size and the number of generations for solving a problem. Redundant representations are uniformly redundant if each phenotype is on average represented by the same number of genotypes. Theoretical and empirical results show that represen
34
3 Elements of a Theory of Representations
tations that are synonymously and uniformly redundant do not change the behavior of GEAs. Furthermore, the results show that representations that are synonymously and nonuniformly redundant can only be used advantageously if there exists some apriori information about the optimal solution. When assigning phenotypes to genotypes, a representation fg can change the importance of the alleles. For example, if a phenotype is a list of integers, all alleles (integers) are equally relevant for calculating the ﬁtness of a phenotype. However, when encoding the phenotypic integers using binary strings, the contributions of the genotypic bits to the construction of the phenotypes and to the calculation of the corresponding ﬁtness values are no longer equal as some genotypic bits are more relevant than others. By substituting alleles with building blocks (BBs), the order of scaling of a representation describes how diﬀerent the contributions of the genotypic BBs are to the construction of the phenotypes. It is well known that if the BBs are uniformly scaled, GEAs solve all BBs implicitly in parallel. In contrast, for nonuniformly scaled BBs, domino convergence occurs and the BBs are solved sequentially starting with the most salient BB (Thierens 1995). As a result, the convergence time increases and the search is aﬀected more strongly by genetic drift. Lower salient alleles are not opposed to selection pressure unless they are aﬀected by the solving process. Therefore, some of the lower salient alleles could loose their diversity and are randomly ﬁxed. To model more exactly the eﬀects of nonuniformly scaled representations on GEA performance, we extend the work of Thierens and Goldberg (1993) and present a more general theory of nonuniformly scaled encodings. It allows us to more accurately predict the performance of GEAs using nonuniformly scaled representations under the inﬂuence of genetic drift. In general, the used representation fg should have no inﬂuence on the ability of GEAs to solve easy problems. However, previous work (Liepins and Vose 1990) has shown that representations can easily change the diﬃculty of an optimization problem. The diﬃculty of an optimization problem is determined by the mapping fp that assigns a ﬁtness value to each phenotype. Therefore, by the use of representations, fully easy (compare Sect. 2.3) problems can become fully diﬃcult, and vice versa. Section 3.3 reveals that the locality of a representation determines whether the diﬃculty of a problem is changed by the representation. The locality of a representation describes how well the distances between individuals are preserved when mapping the genotypes on the phenotypes by the representation fg . The genotypic distances depend on the type of the used search operator and the phenotypic distances are determined by the character of the optimization problem. We illustrate that highlocality representations, where neighboring phenotypes correspond to neighboring genotypes, preserve problem diﬃculty and allow GEAs to solve easy problems more reliably. Furthermore, we are able to show both theoretically and empirically, that by using representations where the locality is not preserved, fully easy problems become more diﬃcult, whereas fully diﬃcult problems become easier. Therefore, if our aim is to reliably solve problems
3.1 Redundancy
35
of bounded diﬃculty, we demand highquality representations to have low locality as only this guarantees that the diﬃculty of the problem fp remains unchanged. Finally, we discuss why, in general, it is not possible to create representations that both, preserve complexity for easy problems, and reduce complexity for diﬃcult problems. Section 3.1 shows how the usage of redundant encodings aﬀects genetic search. Based on the Gambler’s ruin model (Harik et al. 1997), we develop a quantitative model of redundancy and verify it empirically for the trivial voting mapping, which is a synonymously redundant encoding. In Sect. 3.2, we show how the behavior of GEAs changes for exponentially scaled encodings by using the existing models of genetic drift and population sizing. We use two diﬀerent drift models and develop models for the necessary population size and the convergence time that allows us to predict the solution quality more accurately than the previous models. To verify the theoretical models, we present an empirical investigation into the performance of GEAs using exponentially scaled representations. Section 3.3 shows that only representations with perfect locality guarantee that phenotypically easy problems remain genotypically easy and can still be solved using GEAs. If the locality of a representation is not perfect, the size and length of the BBs can be different for the genotypes and phenotypes and the complexity of the problem is changed. Fully easy problems can only become more diﬃcult, and fully difﬁcult problems can only become easier to solve for GEAs. The chapter ends with concluding remarks.
3.1 Redundancy This section provides the ﬁrst of three elements of a theory of representations. It identiﬁes redundancy to be important for the design of representations, distinguishes between synonymously and nonsynonymously representations, and uses existing complexity models to characterize the eﬀect of redundancy on encodings. Furthermore, it presents theoretical models on how the population size, run duration and overall problem diﬃculty is inﬂuenced by synonymously redundant encodings. The model is used for the analysis of the trivial voting mapping, which is a synonymously redundant encoding. 3.1.1 Redundant Representations and Neutral Networks Information theory provides us with a measurement of information. The information content1 (measured in Bits) of a sequence is deﬁned as the number of bits required to represent a given number of s possibilities using an optimal encoding (Shannon 1948; Shannon and Weaver 1949). It is calculated as log2 (s). Redundant encodings are less eﬃcient codings that require more bits 1
other notations are information, selfinformation, entropy, or Shannon entropy.
36
3 Elements of a Theory of Representations
to represent the information but do not increase the amount of information represented. For encoding one Bit of information content (for example the two possibilities 0 and 1) a binary string of at least length one is necessary (one bit). However, it is also possible to encode one Bit of information content using a bitstring of length l > 1. Then, more than one bit of the bitstring encodes one Bit of information, and the representation becomes redundant. We want to emphasize that it is important to distinguish between the amount of information (Bit) that should be represented and the number of bits in a string that are used to represent the information. Redundant representations are encodings where the amount of encoded information (in Bit) is lower than the used number of bits. This means, that such encodings use a higher number of alleles for encoding phenotypic information in the genotype than is necessary for constructing the phenotype. Although the practice of redundancy has steadily increased over the last few years, there is little theory regarding the inﬂuence of redundant representations on the performance of GEAs. Natural Selection and Neutral Theory: Diﬀerent Concepts for Explaining Evolution Examining the use of redundant representations reveals that redundant representations are not solely an invention of evolutionary computation researchers, but are commonly used in nature for the encoding of genetic information. Currently, in biology diﬀerent opinions exist regarding the basic concepts that underly the process of evolution in nature and the role of representations therein. Darwinism goes back to Darwin (1859) and assumes that natural selection is the driving force of evolution (compare Mayr (1991)) and that random genetic drift is unimportant. Randomly advantageous mutations are ﬁxed due to natural selection and can then be propagated from generation to generation. Genetic changes are a result of selection combined with variation operators such as crossover and random mutations. During the process of evolution the variation operators sometimes result in ﬁtter individuals, which gradually replace lessﬁt individuals in the population. The theory of natural selection has been extended and modiﬁed by the neutral theory which was proposed by Kimura (1983). It assumes that the driving force of molecular evolution is the random ﬁxation of neutral mutations rather than the ﬁxation of advantageous mutations by natural selection. Kimura observed that in nature the number of diﬀerent genotypes which store the genetic material of an individual greatly exceeds the number of diﬀerent phenotypes which determine the outward appearance. Therefore, the representation which describes how the genotypes are assigned to the phenotypes must be redundant, and neutral mutations become possible. A mutation is neutral if its application to a genotype does not result in a change of the corresponding phenotype. Because large parts of the genotype have no actual eﬀect on the phenotype, evolution can use them as a store for genetic
3.1 Redundancy
37
information that was necessary for survival in the past, and as a playground for developing new properties of the individual that could be advantageous in the future. Neutral mutations are the tool for designing these new properties without interfering with the current phenotype. Although most of the mutations are neutral, some sometimes have an eﬀect on the phenotype and bring new genetic material which was developed by neutral mutations into life. The neutral theory was highly disputed shortly after its formulation and the relative importance of neutral mutation and selection is still unclear. However, the importance of random genetic changes has generally been accepted in population genetics during the last few years. Redundant Representations in Evolutionary Computation Research Following the work of Kimura, some biological studies (Huynen et al. 1996; Huynen 1996; Schuster 1997; Reidys and Stadler 1998) focused on the neutral theory. These studies showed that the connectivity between ﬁtness landscapes can be increased by the introduction of redundant representations and neutral mutations. Diﬀerent genotypes which are assigned to the same phenotype (neutral sets) allow a population to move through the search space more easily and to ﬁnd new advantageous areas of the search space that would not have been accessible without neutral mutations. Surprisingly, the neutral theory became even more popular in the ﬁeld of genetic and evolutionary computation (Banzhaf 1994; Dasgupta 1995). There is great interest in how redundant representations and neutral search spaces inﬂuence the behavior, and especially the evolvability of GEAs (Barnett 1997; Barnett 1998; Shipman 1999; Shipman et al. 2000; Shackleton et al. 2000; Shipman et al. 2000; Ebner et al. 2001; Smith et al. 2001; Smith et al. 2001a; Smith et al. 2001b; Barnett 2001; Yu and Miller 2001; Yu and Miller 2002; Toussaint and Igel 2002). The general idea behind most of this work is that the evolvability of a population, which is deﬁned as the ability of random variations to sometimes produce improvements, is increased by the use of redundant representations. Furthermore, because redundant representations allow a population to change the genotype without changing the phenotype, the ability of a population to adapt after changes and the performance of GEAs should increase. However, in most of this work the focus has not been on the performance of GEAs, but on characteristics of the search like reachability of phenotypes, evolvability of populations, or connectivity of search spaces. No results have been presented up till now that clearly indicate the superiority of redundant representations and neutral search on practical test problems or realworld instances. Recently, Knowles and Watson (2002) presented an investigation into the performance of neutral search for NK landscapes, HIFF, and MAXSAT problems. The results showed that using arbitrary redundant representations (random boolean network mapping) does not increase the performance of
38
3 Elements of a Theory of Representations
mutationbased search for the considered test problems. In most of the problems used, adding redundancy appeared to reduce performance. Although, at the moment, the focus in investigating the role of redundant representations is mainly on neutral mutations and their eﬀects on search characteristics, there is other work which tries to address the eﬀects of redundancy on the performance of evolutionary search. Researchers used diﬀerent types of redundant representations and sometimes observed either an increase or a decrease in the performance of GEAs. Over time, diﬀerent opinions regarding the eﬀects of redundancy on the performance of GEAs have been developed. Some work noticed that redundant representations lead to a reduction in GEA performance (Davis 1989; Eshelman and Schaﬀer 1991; Ronald et al. 1995). The low performance was argued to be either due to a loss of diversity in the population, or because diﬀerent genotypes that represent the same phenotype compete against each other. Also, the larger size of the search space was listed as a reason for lower GEA performance. In contrast, other mostly applicationoriented work reports higher performance with additional redundancy (Cohoon et al. 1988; Gerrits and Hogeweg 1991; Julstrom 1999), which some researchers ascribe to an increase in diversity that hinders premature convergence. Further work considered the computational implications of genetic codelike representations in gene expression (Kargupta 2000b; Kargupta 2001). Kargupta investigated how redundant representations inﬂuence the energy of the Fourier spectrum. The results show that using redundant representations and encoding phenotypes with higher ﬁtness by a larger number of genotypes results in a higher energy of the Fourier spectrum, reduces the diﬃculty of the optimization problem, and therefore allows a more eﬀective evolutionary search. This short literature review has shown that the inﬂuence of redundant representations on the performance of GEAs is a strongly disputed topic. It can be expected that there is no easy and general answer, and not all types of redundant representations will be useful (Harvey and Thompson 1997). To ﬁnd answers, it is necessary to characterize the diﬀerent types of redundant representations regarding their speciﬁc properties, and to develop quantitative models describing how solution quality and run duration of GEAs is inﬂuenced. This approach can help to clear up some of the disputed questions and to ﬁnd out under which circumstances which type of redundant representation can be beneﬁcial for GEAs. Consequently, in the following section we develop a classiﬁcation for diﬀerent types of redundant representations and in Sects. 3.1.4 and 3.1.5 we develop quantitative models. 3.1.2 Synonymously and NonSynonymously Redundant Representations We give some basic deﬁnitions and develop a classiﬁcation for diﬀerent types of redundant representations which is based on their synonymity.
3.1 Redundancy
39
As we have discussed in Sect. 2.1.2 we have to distinguish between genotypes and phenotypes. Φg is deﬁned as the genotypic search space, where the operators crossover and mutation are applied. Φp is the phenotypic search space. The ﬁtness of an individual depends on the properties of the phenotype xp ∈ Φp . A representation fg : Φg → Φp determines which phenotypes xp ∈ Φp are represented by which genotypes xg ∈ Φg . We want to assume that every phenotype xp is assigned to at least one genotype xg . Otherwise, if a phenotype xp is not represented by some genotype xg this solution can never be found by the used optimization algorithm. A representation fg is redundant if the size of the genotypic search space is larger than the size of the phenotypic search space, Φg  > Φp . This means, there are more diﬀerent genotypes than phenotypes. When using search spaces where not all possible phenotypes or genotypes are accessible by the used search method, a representation is redundant if the number of accessible phenotypes is smaller than the number of accessible genotypes. Therefore, in general a representation fg is redundant if on average one accessible phenotype is represented by more than one genotype. Redundant representations are less eﬃcient encodings, which use an additional number of genes, but do not increase the encoded information content. Therefore, a representation is redundant if l diﬀerent phenotypes are assigned to m diﬀerent genotypes where m > l. Although the larger number of possible genotypes would allow us to encode more individuals than there are phenotypes, some of the information that exists in the genotypes is not considered. Distinguishing between Synonymously and NonSynonymously Redundant Representations To classify diﬀerent types of redundant representations we want to measure how similar the genotypes are that are assigned to the same phenotype. A representation is deﬁned to be synonymously redundant if the genotypes that are assigned to the same phenotype are similar to each other. Consequently, we denote a representation to be nonsynonymously redundant if the genotypes that are assigned to the same phenotype are not similar to each other. Therefore, the synonymity of a representation depends on the metric that is deﬁned on Φg and Φp . The metric deﬁned on Φp depends on the properties of the considered problem and the metric deﬁned on Φg depends on the used search operator. Depending on diﬀerent operators and metrics used on Φg we get diﬀerent synonymity of the representation fg . In Fig. 3.1, we illustrate the diﬀerences between synonymous and nonsynonymous redundancy. For this illustrative example we use the Euclidean distance between the individuals for indicating how similar diﬀerent individuals are. We want to formalize the classiﬁcation into synonymously and nonsynonymously redundant representations. In general, a redundant representap tion fg assigns a phenotype xp to a set of diﬀerent genotypes xg ∈ Φxg , where p p ∀xg ∈ Φxg : fg (xg ) = xp . All genotypes xg in the genotypic set Φxg represent
40
3 Elements of a Theory of Representations
x Φp
o fg
xx x x xx x o oo o o Φg o oo o
o x ox x xxo o oo x o Φ o o x
synonymous
non−synonymous
g
Figure 3.1. Synonymous versus nonsynonymous redundancy. The diﬀerent symbols indicate diﬀerent genotypes and their corresponding phenotypes. When using synonymously redundant representations (left), genotypes that represent the same phenotype are similar to each other. When using nonsynonymously redundant representations (right), genotypes that represent the same phenotype are not similar to each other but distributed over the whole search space
the same phenotype xp . A representation is synonymously redundant if the p genotypic distances between all xg ∈ Φxg are small for all diﬀerent xp . Therefore, if for all phenotypes the sum over the distances between all genotypes that correspond to the same phenotype ⎛ ⎞⎞ ⎛ 1 ⎝ ⎝ d(xg , y g )⎠⎠ , (3.1) 2 g xp g xp xp x ∈Φg y ∈Φg
where xg = y g , is reasonably small a representation is denoted to be synonymously redundant. d(xg , y g ) depends on the mutation operator used and p p measures the distance between two genotypes xg ∈ Φxg and y g ∈ Φxg which both represent the same phenotype xp . The distance between two genotypes depends on their genotypic similarity and is small if the two genotypes are similar. A diﬀerent but equivalent approach of deﬁning the synonymity of redundant representations is to use equivalence classes (compare the approach from p Radcliﬀe outlined in Sect. 2.4.2). All genotypes xg in the genotypic set Φxg belong to the same equivalence class. If the sum of the genotypic distances between the individuals that belong to the same equivalence class is small, then the representation is synonymously redundant. Synonymity and Locality of Redundant Representations The synonymity of redundant representations is related to the locality of nonredundant representations (compare Sect. 3.3). A genotype xg is a neighbor
3.1 Redundancy
41
to some other genotype y g if the distance d(xg , y g ) = dmin , where dmin = 0 is the minimal distance between two individuals in the genotypic search space. When using binary representations, dmin = 1 and two genotypes are neighbors if they diﬀer in one allele. As discussed in Sect. 3.3, the locality of a representation describes how well neighboring genotypes correspond to neighboring phenotypes. If neighboring genotypes correspond to neighboring phenotypes, a representation has high locality and small changes in the genotype result in small changes in the corresponding phenotype. In contrast, representations have low locality if neighboring genotypes do not correspond to neighboring phenotypes. There is evidence, both analytical (for example Whitley (1999) for mutationbased search or Sect. 3.3.6 for crossoverbased search), and empirical (for example Gottlieb et al. (2001) or Rothlauf and Goldberg (2000)), which shows for easy problems that lowlocality representations result in low GEA performance. The genetic operators mutation and crossover no longer work properly as they create new oﬀspring that are not similar to their parent(s). Lowlocality representations result in low GEA performance as guided search methods like GEAs, that use knowledge gained during search for determining future search steps, can only perform better than random search if on average similar solutions have similar ﬁtness (individuals are similar if there are only a few search steps between them). In general, guided search methods assume that in the neighborhood of highquality solutions other highquality solutions can be found (Manderick et al. 1991; Horn 1995; Deb et al. 1997; Christensen and Oppacher 2001). Highquality solutions are grouped together and are not scattered over the whole search space (Radcliﬀe 1991a; Whitley 2002). Therefore, to perform well, guided search methods have to search more often in the neighborhood of already found promising highquality solutions than around lowquality solutions (compare the optimal allocation of trials for the twoarmed bandit problem as discussed in Holland (1975)). This behavior guarantees high performance of the search method if on average neighboring solutions have similar properties, which means the ﬁtness values of neighboring solutions are correlated. However, search heuristics could not use any information learned during the search for determining future search steps, and consequently show low performance if the ﬁtness values of neighboring, or similar, genotypes are not correlated (Weinberger 1990; Manderick et al. 1991; Jones and Forrest 1995). On the one hand, the ﬁtness values of neighboring genotypes can be uncorrelated if the problem itself is diﬃcult, which means the ﬁtness values of neighboring phenotypes are uncorrelated. Then, guided search methods will behave as a random search even in the presence of highlocality representations. On the other hand, the ﬁtness values of neighboring genotypes are also uncorrelated if lowlocality representations are used and neighboring genotypes do not correspond to neighboring phenotypes. The lowlocality representations destroy existing correlations between phenotypes and their corresponding ﬁtness values, and genotypic neighbors no longer have similar properties and ﬁtnesses.
42
3 Elements of a Theory of Representations
In this case, search heuristics can not use any information learned during the search for determining future search steps. As a result, it makes no sense for guided search approaches to search around already found highquality genotypes and guided mutationbased search algorithms become random search. A mutation does not result in a solution with similar properties but in a random solution. Analogously, crossover is not able to create new solutions with similar properties to their parents, but creates new, random solutions. Therefore, high locality representations are a necessity for eﬃcient evolutionary search. When using lowlocality representations, no guided search is possible and guided search methods become random search. These concepts can also be applied to redundant representations. When using nonsynonymously redundant representations, genetic operators like mutation or crossover can result in an oﬀspring that is phenotypically completely diﬀerent from its parent(s). Therefore, nonsynonymously redundant representations have the same eﬀect on GEAs as when using lowlocality representations. Using neutral search spaces where the connectivity between the phenotypes is strongly increased by the use of a redundant representation allows us to reach many diﬀerent phenotypes in one single search step. However, increasing the connectivity between the phenotypes by using nonsynonymously redundant representations results in random search and decreases the eﬃciency of GEAs. As for lowlocality representations, a search step does not result in a similar phenotype but creates a randomly chosen individual. Therefore, guided search is no longer possible and guided search methods become random search. As a result, we get reduced GEA performance on problems that are easy for guided search methods (that means the ﬁtness values of similar phenotypes are correlated) when using nonsynonymously redundant representations. Examples for nonsynonymously redundant representations are the direct binary mapping, the cellular automaton mapping, or the random boolean network mapping, which have been proposed by Shackleton et al. (2000). Although the use of these types of representations strongly increases the connectivity between phenotypes, we get low GEA performance as neighboring genotypes do not correspond to neighboring phenotypes. Initial evidence of the low performance of mutationbased search when using such nonsynonymously redundant representations was shown by Knowles and Watson (2002) for the random boolean network mapping. In contrast, when using synonymously redundant representations, the connectivity between the phenotypes is not increased. Therefore, small genotypic variations can not result in large phenotypic changes but either in the same, or a similar, phenotype. Figure 3.2 illustrates this behavior and compares it to nonsynonymously redundant representations. Examples for synonymously redundant representations are the trivial voting mapping (Shackleton et al. 2000) which is investigated more closely in Sect. 3.1.6.
3.1 Redundancy
c
c
x
l
x
l
o
Φp
o
Φp
fg c c c c c c l l l l l l o
Φg
o
fg x x
o
x x
x
x c c
x l
o
43
o o
synonymous
Φg
x l l x l o o
c
o c
x
l x
c
c
o o
o x
l
non−synonymous
Figure 3.2. The eﬀects of small mutation steps for synonymously versus nonsynonymously redundant representations. The diﬀerent symbols indicate diﬀerent genotypes and their corresponding phenotypes. The arrows indicate search steps which result in neighboring individuals. When using synonymously redundant representations, a mutation results in either the same or a similar phenotype. In contrast, when using nonsynonymously redundant representations the mutation of a genotype results in completely diﬀerent phenotypes.
Formalizing Synonymously Redundant Representations In this subsection, we introduce some quantities that can be used for characterizing the properties of synonymously redundant representations. We use the deﬁnitions from Sects. 2.1.2 and 3.1.2. To describe a redundant representation, we introduce kr , the order of redundancy. kr is deﬁned as log(Φg )/ log(Φp ) and measures the amount of redundant information in the encoding. There are kr bits and 2kr diﬀerent possibilities (individuals) to encode 1 Bit of information. When using binary genotypes and binary phenotypes, the order of redundancy can be calculated as log(2lg ) , kr = log(2lp ) where lg is the length of the binary genotype and lp is the length of the binary phenotype. When using a nonredundant representation, the number of genotypes equals the number of phenotypes and kr = 1. Furthermore, we want to characterize not only to what degree a representation is redundant, but also in what way it is redundant. We are especially interested in the overrepresentation and underrepresentation of speciﬁc solutions. Therefore, we introduce r as the number of genotypes that represent the one phenotype that has the highest ﬁtness. When using nonredundant representations, every phenotype is assigned to exactly one genotype and r = 1. However, in general, 1 ≤ r ≤ Φg  − Φp  + 1.
44
3 Elements of a Theory of Representations
In the following discussion, we want to focus on how redundant representations inﬂuence the behavior of selectorecombinative GAs. Selectorecombinative GAs use crossover as the main search operator and mutation only serves as a background operator. When focusing on selectorecombinative GAs we implicitly assume that there are building blocks (BBs) and that the GA process schemata. Consequently, we must deﬁne how kr and r depends on the properties of the BBs. In general, when looking at BBs of size k there are 2k diﬀerent phenotypic BBs which are represented by 2kkr diﬀerent genotypic BBs. Therefore, kr =
kg , kp
where kg denotes the genotypic size of a BB and kp the size of the corresponding phenotypic BB. As before, a representation is redundant if kr > 1. The size of the genotypic BBs is kr times larger than the size of the phenotypic BB. Furthermore, r is deﬁned as the number of genotypic BBs of length kkr that represent the best phenotypic BB of size k. Therefore, in general, r ∈ {1, 2, . . . , 2kkr − 2k + 1}.
(3.2)
In contrast to kr , which is determined by the representation used, r depends not only on the representation used, but also on the speciﬁc problem that should be solved. Diﬀerent instances of a problem result in diﬀerent values of r. If we assume that kr is an integer (each phenotypic allele is represented by kr genotypic alleles) the possible values of the number of genotypic BBs that represent the optimal phenotypic BB can be calculated as r = ik , with i ∈ {1, 2, . . . , 2kr − 1}.
(3.3)
A representation is uniformly redundant if all phenotypes are represented by the same number of diﬀerent genotypes. Therefore, when using a uniformly redundant representation every phenotypic BB of size k = kp is represented by (3.4) r = 2k(kr −1) diﬀerent genotypic BBs. Table 3.1 gives an example for a uniformly redundant encoding. Two bits in a phenotype xp are represented by four bits in the genotype xg . Therefore, kr = 2 and r = 4. With Φp  = 2k = 22 the size of the genotypic space is Φg  = 2kkr = 24 = 16.
00 10 00 10
00, 00, 10, 10,
00 10 01 10
xg 01, 01 01, 11 11, 00 11, 11
00, 00, 11, 10,
01 11 01 11
01 01 11 11
xp 00 10 01 11
Table 3.1. An example of a uniformly redundant representation, where kr = 2 and r = 4
3.1 Redundancy
45
By introducing redundancy the search space for a GA using binary phenotypes of string length l = lp is increased from Φp  = 2l to Φg  = 2lkr . The length of the individuals increases from l = lp in the phenotypic space to lg = kr × l in the genotypic space. To represent all phenotypes, each individual xp ∈ Φp must be represented by at least one genotype xg ∈ Φg . If Φg  = Φp , and each phenotype is represented by at least one genotype, we have a nonredundant, onetoone mapping. 3.1.3 Complexity Model for Redundant Representations For modeling the eﬀects of redundant representations on the performance of GEAs, we can use the complexity model from Goldberg (1991a) and Goldberg et al. (1992). If we want to understand the eﬀects of redundancy, we must decompose the problem into smaller subproblems and try to solve these separately. We could decompose the problem step by step and subsequently collate all the subproblems. The decomposition (Goldberg 1998) takes place as follows: • • • • • •
GAs process building blocks. Problems are tractable by BBs. GAs must ensure proper supply of BBs in the initial generation. GAs must grow the high quality BBs. GAs must mix the BBs well. GAs must decide well among competing BBs.
This decomposition gives us a framework for investigating the eﬀects of redundancy on the performance of GAs. We want to examine how redundancy aﬀects the problem decomposition point by point: Using redundant encodings does not change the principal behavior of GAs. After adding redundancy, GAs still process building blocks. A problem is still tractable by building blocks when using a redundant encoding. However, the question arises as to whether the size of building blocks is changed by redundant encodings. On one hand, redundant representations increase the number of bits that are part of a building block in the genotype from k to kkr . On the other hand, the number of building blocks in the genotype that represent the same BB in the phenotype increases from 1 to on average 2k(kr −1) . Furthermore, the number of diﬀerent ﬁtness values that can be assigned to the genotypes remains constant. Taking these eﬀects into account, we assume that problems are still tractable by BBs and the larger size of the genotypic BBs is compensated by the higher number of genotypic BBs. How do redundant encodings change the initial supply of BBs in the initial population? For uniform redundancy (every phenotype xp ∈ Φp is represented by the same number of genotypes xg ∈ Φg ), the initial supply of BBs x0 /N = 1/2k is the same as for nonredundant representations. If the number of genotypes xg that represent a phenotype xp is above average, then
46
3 Elements of a Theory of Representations
the phenotype xp , and the containing schemata hp , are overrepresented in the initial population. GEAs are pushed more towards solutions that are similar to these xp . Analogously, the performance of GEAs decreases if the proportion of less ﬁt phenotypes in the initial population is increased by redundancy. As a result, the supply of BBs could be modiﬁed by redundant encodings. For a proper growth and mixing of BBs, above average BBs must be preferred by selection, and BBs should not be disrupted by the crossover operator. As the genotypic deﬁning length of a BB δ(hg ) increases when using redundant representations, GEA operators that do not obey the linkage disrupt BBs more frequently. To overcome this problem, competent GAs (M¨ uhlenbein and Paaß 1996; Goldberg 1999; Larranaga et al. 1999; M¨ uhlenbein and Mahnig 1999; Pelikan 2002) could be used. These kind of GEAs obey the linkage, and genotypic building blocks hg are not disrupted by recombination. Thus, no reduction of performance should occur, and redundancy should have no negative eﬀect on the proper mixing of BBs. However, when using redundant representations there are diﬀerent genotypes xg and y g that represent the same phenotype xp . Recombining xg and y g could result in oﬀspring that do not represent xp . This eﬀect is also known as crosscompetition among isomorphic identical BBs. As an example for a onebit problem, the genotype {00} represents the phenotype xp = 0, and the genotypes {01, 10, 11} represent the phenotype y p = 1. If the genotypes 01 and 10 are recombined, the possibilities for the oﬀspring are 00 and 11. Although both parental genotypes represent the same phenotype y p , one of the oﬀspring represents xp . Crosscompetition among isomorphic identical BBs is a result of using nonsynonymously redundant representations, which were discussed in the previous subsection. Nonsynonymously redundancy randomizes genetic search as new BBs are introduced into the search that did not exist in the parents. When using such representations, the recombination of genotypes that represent the same phenotype can result in completely diﬀerent new genotypes and phenotypes (compare also Fig. 3.2). When using synonymously redundant representations, crosscompetition can not occur. Finally, the decision making between competing BBs is not aﬀected by using redundant representations. With redundancy there are diﬀerent genotypic BBs hg that represent the same phenotypic BB hp , but the selection process does not decide between the diﬀerent hg because they all have the same ﬁtness. The ﬁtness evaluation is based on the ﬁtness of the phenotypic BBs hp , and not their genotypic representation. After recognizing that redundancy could have a major eﬀect on the supply, and a minor eﬀect on the proper mixing of building blocks, we want to quantify its eﬀect on BBs supply in the following subsection.
3.1 Redundancy
47
3.1.4 Population Sizing for Synonymously Redundant Representations In Sect. 3.1.2, we described how nonsynonymously redundant representations result in randomized search as they increase the connectivity of the search space. Therefore, in this section we want to focus on synonymously redundant representations and develop a population sizing model that describes their inﬂuence on the performance of selectorecombinative GAs. As we focus in our investigation on selectorecombinative GAs we can use the existing theory describing the behavior of selectorecombinative GAs from Harik et al. (1997) and Thierens and Goldberg (1994). They describe for nonredundant representations how the population size and the time to convergence that is necessary to solve a speciﬁc problem depend on the characteristics of the problem. Following Harik et al. (1997) the probability that a GA with a population size N converges after tconv generations to the correct solution is Pn =
1 − (q/p)x0 , 1 − (q/p)N
where x0 is the expected number of copies of the best BB in the randomly initialized population, q = 1 − p, and p is the probability of making the right choice between a single sample of each BB
d · (3.5) p=N √ 2m σBB N is the cumulative distribution function for a normal distribution, d is the signal diﬀerence between the best BB and its strongest competitor, m = 2 is the variance m − 1 with m is the number of BBs in the problem, σBB of a BB, and q = 1 − p is the probability of making the wrong decision between two competing BBs. It has been shown in Harik et al. (1997) that this random walk or Gambler’s ruin model can be used for describing the behavior of selectorecombinative GAs propagating schemata and BBs. In the following paragraphs, this model is the basis for describing the inﬂuence of synonymously redundant representations on the behavior of GAs. For a randomly initialized population with no redundancy, x0 = N/2k . The situation changes when using redundant representations. Then, the initial supply depends on the characteristics of the representation, namely r and kr . With r the number of genotypic BBs of length kkr that represent the best phenotypic BB of length k, we get x0 = N
r , 2kkr
(3.6)
where kr is the order of redundancy. The assumption that redundant representations aﬀect the initial supply of BBs is the core idea behind the proposed
48
3 Elements of a Theory of Representations
model describing the inﬂuence of synonymously redundant representations on GA performance. We assume that other eﬀects of synonymously redundant representations on GA performance can be neglected. Consequently, when using uniformly redundant representations, r = 2k(kr −1) and x0 = N/2k . These are the same values as when using nonredundant representations. Therefore, GA performance does not change when using uniformly redundant representations. 2 and the number m of BBs is not aﬀected by the use As the variance σBB of a redundant representation, the probability of GA failure α = 1 − Pn can be calculated as 1 − (q/p)x0 · (3.7) α=1− 1 − (q/p)N If we assume that x0 is small and q < p we can assume that 1 − (q/p)N converges to 1 faster than 1 − (q/p)x0 . Using these approximations (see also Harik et al. (1997)) the equation can be simpliﬁed to
x 1−p 0 · α≈ p Therefore, for the population size we get ⎛ ⎞ kkr 2 ⎝ ln(α) ⎠ · N≈ r ln 1−p
(3.8)
p
The normal distribution in (3.5) can be approximated using the ﬁrst two terms of the power series expansion √ (see Abramowitz and Stegun (1972)) as N(x) ≈ 1/2 + x/2, where x = d/ πm σBB . Substituting p from (3.5) into (3.8) we get:
1−x 2kkr ln(α)/ ln N≈ , r 1+x Since x is a small number, ln(1−x) can be approximated with −x and ln(1+x) with x. Using these approximations we ﬁnally get for the population size N : √ σBB πm 2kr k−1 ln(α) · (3.9) N ≈− r d k r The population size N goes with O 2r when using synonymously redundant representations. With increasing r the number of individuals that are necessary to solve a problem decreases. Using a uniformly redundant representation, where r = 2k(kr −1) , does not change the population size N in comparison to nonredundant representations.
3.1 Redundancy
49
3.1.5 Run Duration and Overall Problem Complexity for Synonymously Redundant Representations To describe the performance of GAs, we must calculate not only the number of individuals that are necessary for solving a problem, but also the expected number of generations until convergence. Based on M¨ uhlenbein and SchlierkampVoosen (1993) and Thierens and Goldberg (1994), Miller and Goldberg developed a convergence model for selectorecombinative GAs (Miller and Goldberg 1996b; Miller and Goldberg 1996a). The convergence time tconv depends on the length of the phenotypes l = lp and the used selection scheme. Using the selection intensity I the convergence model is
It √ + arcsin(2p(0) − 1) , p(t) = 0.5 1 + sin l where p(0) = x0 /N is the proportion of best building blocks in the initial population. I depends only on the used selection scheme. The number of generations tconv it takes to fully converge the population can be calculated by putting p(tconv ) = 1: √ l π − arcsin (2p(0) − 1) · (3.10) tconv = I 2 If we assume k = 1 and uniform redundancy (equal proportion of 1s and 0s in the initial population) we get p(0) = 0.5. Then, the number of generations until convergence simpliﬁes to √ π l · tconv = 2 I r With redundancy the initial proportion of building blocks is p(0) = 2kk r 3 (see (3.6)). Using arcsin(x) = x + o(x ) the time until convergence could be approximated by √ π r l (3.11) 1 + − kr k−1 · tconv ≈ I 2 2
With increasing r/2kr the time to convergence tconv is reduced. Therefore, the optimal solution is found after a lower number of generations if it is overrepresented by the synonymously redundant representation. For uniform redundancy r = 2k(kr −1) , we get √
1 l π tconv ≈ 1 + − k−1 · I 2 2 The time until convergence when using uniformly redundant representations is the same as without redundancy.
50
3 Elements of a Theory of Representations
After we have calculated the number of individuals that are necessary for solving a problem (see (3.9)), and the number of generations that GAs using only crossover need to converge (see (3.11)), we can calculate the absolute number of ﬁtness calls that are necessary for solving a problem: √ √ σBB πm π r l 2kr k−1 ln(α) × 1 + − kr k−1 N × tconv ≈ − r d I 2 2 √
kkr σBB πlm 2 = ln(α) (2 + π) 1− I d 4r The overall number of ﬁtness calls goes with O(2kr /r). In comparison to nonredundant representations, the number of ﬁtness calls stays constant for synonymously redundant representations if r = 2k(kr −1) . Then x0 /N = 1/2k and the representation is uniformly redundant. 3.1.6 Analyzing the Redundant Trivial Voting Mapping In the previous subsection, we developed theoretical models describing how synonymously redundant representations inﬂuence the quality of the solution and the time that is necessary to ﬁnd the good solutions. In this subsection, we investigate whether or not the proposed models allow a good prediction of GA performance for the trivial voting (TV) mapping. The TV mapping is a synonymously redundant representation and we use it for onemax and concatenated deceptive trap problems. During our investigation we are particularly interested in whether the developed models allow us to accurately predict the expected solution quality and running time of a selectorecombinative GA. The Trivial Voting Mapping We give a short introduction into the trivial voting mapping. When using the TV mapping, a set of mostly consecutive, genotypic alleles is relevant for the value of one allele in the phenotype. Each allele in the genotype can only inﬂuence the value of one allele in the phenotype. The value of the phenotypic allele is determined by the majority of the values in the genotypic alleles. In general, the diﬀerent sets of alleles in the genotype deﬁning one phenotypic allele have the same size. The TV mapping is a synonymously redundant representation as all genotypes that represent the same phenotype are similar to each other. A mutation in a genotype results either in the same corresponding phenotype, or in one of its neighbors. The TV mapping can be easily characterized using the representation parameters deﬁned in Sect. 3.1.2. The order of redundancy kr is simply the number of genotypic alleles that determine the value of one phenotypic allele. Figure 3.3 gives an example for the TV mapping.
3.1 Redundancy
51
genotype: phenotype:
Figure 3.3. The trivial voting mapping
Shackleton et al. (2000) applied the TV mapping to binary strings in the context of the neutral theory. When used for binary strings, binary genotypes xg ∈ Blg are assigned to binary phenotypes xp ∈ Blp . The length of a genotype is larger than the length of a phenotype, lg > lp . The value of one phenotypic bit is determined by the majority of the values in the corresponding genotypic bits (majority vote). However, if kr is even then the number of ones could equal the number of zeros. Therefore, half the cases that result in a tie should encode a one in the corresponding phenotypic allele, and half the cases should represent a zero. For example, for kr = 4 the genotypic BBs 1100, 1010, and 1001 represent a 1 and the genotypic BBs 0011, 0101, 0110 represent a zero. Because the majority of the votes determines the values of the corresponding phenotypic allele, the TV mapping is a uniformly redundant representation. Each phenotypic BB is represented by the same number of genotypic BBs which is 2k(kr −1) , where k is the size of the phenotypic BB. As we are not only interested in uniformly redundant representations, but also want to know how nonuniformly redundant representations inﬂuence GA performance, we extend the TV mapping to allow the encoding to overrepresent some individuals. Therefore, we want to assume that if the number of ones in the genotypic alleles xgkr i+j , where i ∈ {0, . . . , lp −1} and j ∈ {0, . . . , kr −1}, is larger or equal than a constant u then the value of the phenotypic allele xpi is set to one. Vice versa, the phenotypic allele xpi is set to zero if less than u of the corresponding genotypic alleles are set to one. Therefore, kr −1 g 0 if xkr i+j < u p xi = kj=0 g r −1 1 if j=0 xkr i+j ≥ u, where u ∈ {1, . . . , kr }. xgi (respectively xpi ) denotes the ith allele of the genotype (respectively phenotype). u can be interpreted as the number of genotypic alleles that must be set to one to encode a one in the corresponding phenotypic allele. This representation is denoted as extended trivial voting (eTV) mapping. For u = (kr + 1)/2 (kr must be odd) we get the original TV mapping. Extending the TV mapping in the proposed way allows us to investigate how nonuniform redundancy inﬂuences the performance of GAs. When using the eTV mapping, the number r of genotypic BBs that can represent the optimal phenotypic BB depends on the number of ones in the genotypic alleles that determine the value of the corresponding phenotypic allele. Considering (3.3) we get
52
3 Elements of a Theory of Representations
⎛ r=⎝
kr kr j=u
j
⎞k ⎠ ,
(3.12)
where u ∈ {1, . . . , kr }. We assume that a BB is optimal if all phenotypic bits are set to xpi = 1. k denotes the size of the phenotypic BB. To give a short illustration, we use a redundant representation with kr = 3, k = 1 (compare Fig. 3.3). The optimal BB is xpi = 1. Because u ∈ {1, . . . , kr } there are three diﬀerent values possible for r. For u = 1 the phenotypic allele xpi is set to one if at least one of the three corresponding genotypic alleles xgikr , xgikr +1 , or xgikr +2 is set to one. Therefore, a one in the phenotype is represented by 3 kr r = = 7 diﬀerent genotypic BBs (111, 110, 101, 011, 100, 010, j=1 j and 001). For u = 2, the optimal genotypic BB xpi = 1 is represented by 3 kr = 4 diﬀerent genotypic BBs (111, 110, 101, and 011) and r = j=2 j the representation is uniformly redundant. For u = 2 we get the original TV mapping. For u = 3, the optimal phenotypic BB is represented only by one genotypic BB (111). Experiments and Empirical Results Here we present empirical results when using the binary trivial voting mapping for the onemax problem and the concatenated deceptive trap problem. OneMax Problem The ﬁrst test example for our empirical investigation is the onemax problem. This problem is very easy to solve for GEAs as the ﬁtness of an individual is simply the number of ones in the binary phenotype. To ensure that recombination results in a proper mixing of the BBs, we use uniform crossover for all experiments with the onemax problem. Furthermore, in all runs we use tournament selection without replacement and a tournament size of 2. For the onemax function the signal diﬀerence d equals 1, the size k of the building 2 = 0.25. blocks is 1, and the variance of a building block σBB xg2i xg2i+1 (with kr = 2) extended TV original TV r=1 r=3 r=2 0 00, 01, 10 00 00, 01 11 01, 10, 11 10, 11 1
xpi
Table 3.2. The trivial voting mapping for kr = 2
When using the binary TV mapping for the onemax problem each bit of a phenotype xp ∈ Φp is represented by kr bits of the genotype xg ∈ Φg . The string length of a genotype xg is lg = kr ×lp and the size of the genotypic search space is Φg  = 2kr lp . Table 3.2 illustrates for kr = 2 the two possibilities (r = 1
3.1 Redundancy
53
and r = 3) of assigning genotypic BBs {00, 01, 10, 11} to one of the phenotypic BBs {0, 1} when using the extended TV mapping described in the previous paragraphs. With denoting xpi the value of the ith bit in the phenotype, the 2ith and (2i+1)th bit of a genotype determine xpi . Because the size of the BBs k = 1, the number of genotypic BBs that represent the optimal phenotypic BB is either r = 1 or r = 3 (compare (3.12)). Furthermore, Table 3.2 also lists the case where r = 2. This case is the original uniformly redundant TV mapping. The second bit of each genotypic BB does not contribute to the construction of the phenotype.
1
1
0.9 proportion of correct BBs
proportion of correct BBs
0.9
0.8
0.7 pred. kr=1 (no redund.) kr=1 (no redund.) TVM (r=2, unif. redund.) pred. eTVM with u=2 (r=1) eTVM with u=2 (r=1) pred. eTVM with u=1 (r=3) eTVM with u=1 (r=3)
0.6
0.8 0.7 0.6 pred. kr=1 (no redund.) kr=1 (no redund.) TVM (u=2, r=4, unif. redund.) pred. eTVM with u=3 (r=1) eTVM with u=3 (r=1) pred. eTVM with u=1 (r=7) eTVM with u=1 (r=7)
0.5 0.4
0.5
0.3 10
20
30
40
50
60
70
10
population size N
20
30
40
50
60
70
population size N
(a) kr = 2
(b) kr = 3
Figure 3.4. Experimental and theoretical results of the proportion of correct BBs on a 150bit onemax problem using the trivial voting mapping for kr = 2 (left) and kr = 3 (right). The lines without line points show the theoretical predictions. When using nonuniformly redundant representations, GA performance is changed with respect to the overrepresentation or underrepresentation of the highquality BBs.
extended TV mapping u=1 u=2 u=3 r 3 1 kr = 2 x0 /N 3/4 1/4 r 7 4 1 kr = 3 x0 /N 7/8 4/8 = 1/2 1/8
original TV mapping 2 2/4 = 1/2 4 2/4 = 1/2
Table 3.3. Properties of the diﬀerent TV mappings for the onemax problem (k = 1)
In Fig. 3.4(a) (kr = 2) and Fig. 3.4(b) (kr = 3), the proportion of correct BBs at the end of a run for a 150 bit onemax problem using the TV mapping is shown. For this problem 2150 diﬀerent phenotypes are represented by either
54
3 Elements of a Theory of Representations
2300 (kr = 2) or 2450 (kr = 3) diﬀerent genotypes. If we use the eTV mapping (indicated in the plots as eTVM) we can set u either to 1 or 2 (kr = 2) or to 1, 2, or 3 (kr = 3). The corresponding values for r, which can be calculated according to (3.12), as well as x0 /N are shown in Table 3.3. x0 is the expected number of copies of the best BB in the initial population and N is the population size. Furthermore, the ﬁgures show the results when using the original, uniformly redundant TV mapping, and when using the nonredundant representation with kr = 1. The lines without line points show the theoretical predictions from (3.7), and the lines with line points show the empirical results which are averaged over 250 runs. The error bars indicate the standard deviation. The results show that for the uniformly redundant TV mapping, r = 2 (kr = 2) or r = 4 (kr = 3), we get the same performance as for using the nonredundant representation (kr = 1). As in the original model proposed by Harik et al. (1997) the theoretical model slightly underestimates GA performance. As predicted by our model which we proposed in Sect. 3.1.4, GA performance does not change when using a uniformly redundant representation. Furthermore, we can see that if the optimal BB is underrepresented (u = 2 for kr = 2 and u = 3 for kr = 3) GA performance decreases. Equation 3.7 gives us a good prediction for the expected solution quality if we consider that the nonuniform redundancy of the representation changes the initial BB supply according to (3.6). If the optimal solution is overrepresented (u = 1 for both cases, kr = 2 and kr = 3) GA performance increases. Again the theoretical models give a good prediction for the expected proportion of correct BBs. Summarizing the results, we can see that using the uniformly redundant TV mapping does not change GA performance as compared to using the nonredundant representation. Only if we overrepresent the optimal phenotypic BB, does GA performance increase; likewise, if we underrepresent the optimal BB, GA performance drops. As our derived model is able to make accurate predictions for the expected solution quality, our assumption that synonymously redundant representations inﬂuence GA performance by changing the initial supply seems to be valid. In the remaining paragraphs, we perform an empirical investigation into the eﬀect of the TV mapping on the number of generations until the population of a selectorecombinative GA converges. Again we use the onemax problem and the TV mapping from above with the same parameters except the population size is set to N = 2lp to allow reliable decision making for the onemax problem (Goldberg et al. 1992). As we use tournament selection √ without replacement of size two the selection intensity I = 1/ π. Figures 3.5(a) (kr = 2) and 3.5(b) (kr = 3) show the number of generations that are necessary until 90% of all phenotypic BBs are found over the problem size which is equal to l = lp . The lines without line points show the predictions from (3.10) and the lines with line points plot the empirical results. We can see that the run duration of a GA when using the nonredundant representa
3.1 Redundancy 50
number of generations tconv
40
number of generations tconv
pred. kr=1 (no redund.) kr=1 (no redund.) TVM (r=2, unif. redund.) pred. eTVM with u=2 (r=1) eTVM with u=2 (r=1) pred. eTVM with u=1 (r=3) eTVM with u=1 (r=3)
30
20
10
55
pred. kr=1 (no redund.) kr=1 (no redund.) TVM u=2 (r=4. unif. redund.) pred. eTVM with u=3 (r=1) eTVM with u=3 (r=1) pred. eTVM with u=1 (r=7) eTVM with u=1 (r=7) 50 40 30 20 10 0
0 20
30
40
50
60
70
80
90 100
problem size l
(a) kr = 2
20
30
40
50
60
70
80
90 100
problem size l
(b) kr = 3
Figure 3.5. Theoretical predictions and experimental results for the number of generations that are necessary until 90% of all phenotypic BBs are correctly identiﬁed. The plots are for onemax problems and trivial voting mapping with kr = 2 (left) and kr = 3 (right).
tion (kr = 1) is exactly the same as when using the uniformly redundant TV mapping with kr = 2. For kr = 3 and u = 2 (uniform redundancy) the run duration is slightly increased in comparison to the nonredundant encoding. We expect that this diﬀerence increases with larger kr . In agreement with the results from Thierens (1995), we report a small underestimation of the expected number of generations when using either nonredundant, or uniformly redundant, representations. When using nonuniformly redundant variants of the eTV mapping the underestimation is larger, but nevertheless the model gives a good approximation for the expected number of generations. We can see that increasing r increases the run duration. For example, if each phenotypic bit is represented by three genotypic bits (kr = 3) and a one is represented if at least one out of three genotypic bits is set to one (u = 1) then a GA ﬁnds the good solutions after very short time (compare eTVM with u = 1). The expected number of generations shows the predicted behavior. The necessary number of genera√ tions increases by about O( l). We see that the proposed model allows us to make good predictions for the expected run duration. Concatenated Deceptive Trap Problem Our second test example uses deceptive trap functions. Traps were ﬁrst used by Ackley (1987) and investigations into the deceptive character of these func
56
3 Elements of a Theory of Representations
tions were provided by Deb and Goldberg (1993). Figure 3.6 depicts a 3bit deceptive trap problem where the size of a BB is k = 3. The ﬁtness value of a phenotype xp depends on the number of ones u in the string of length l. The best BB is a string of l ones which has ﬁtness l. Standard GEAs are misled to the deceptive attractor which has ﬁtness l − 1. For the 3bit deceptive trap 2 = 0.75. We the signal diﬀerence d is 1, and the ﬁtness variance equals σBB construct a test problem for our investigation by concatenating m = 10 of the 3bit traps so we get a 30bit problem. The ﬁtness of an individual x is m−1 calculated as f (x) = i=0 fi (u), where fi (u) is the ﬁtness of the ith 3bit trap function from Fig. 3.6. Although this function is diﬃcult for GEAs it can be solved with proper population size N . f(u) 3 2 1
1
2
3
u
Figure 3.6. A 3bit deceptive trap problem
For deceptive traps of size k = 3 we can calculate the number r of genotypic BBs that represent the optimal phenotypic BB according to (3.12). Table 3.4 summarizes, for the binary TV mapping, how r and x0 /N depends on u, which describes how many of the genotypic alleles must be set to 1 to encode a 1 in the phenotype. x0 is the expected number of copies of the best BB in the initial population and N is the population size. We have also included the properties of the original uniformly redundant TV mapping. extended TV mapping u=1 u=2 u=3 r 33 = 27 13 = 1 kr =2 x0 /N 27/64 1/64 r 73 = 343 43 = 64 13 = 1 kr =3 x0 /N 343/512 64/512 = 1/8 1/512
original TV mapping 23 = 8 8/64 = 1/8 43 = 64 64/512 = 1/8
Table 3.4. Properties of the diﬀerent TV mappings for the trap of size k = 3
By analogy to the previous paragraphs, Figs. 3.7(a) (kr = 2) and 3.7(b) (kr = 3) show the proportion of correct BBs at the end of a run over diﬀerent population sizes for ten concatenated 3bit deceptive trap problems. In this problem, 230 diﬀerent phenotypes are represented by either 260 (kr = 2) or 290 (kr = 3) diﬀerent genotypes. As before, we use tournament selection without replacement of size 2. In contrast to the onemax problem, twopoint crossover was chosen for recombination. Uniform crossover would result in an improper
3.1 Redundancy
57
mixing of the BBs because the genotypic BBs are either of length lg = kr lp = 6 (kr = 2), or of length lg = 9 (kr = 3). Again, the lines without line points show the predictions of the proposed model for diﬀerent r. Furthermore, empirical results, which are averaged over 250 runs, are shown for various values of r. The results show that for the uniformly redundant TV mapping we get the same performance as when using the nonredundant representation (kr = 1). As in the experiments for the onemax problem the proposed model predicts the experimental results well if the eTV mapping is used and some BBs are underrepresented or overrepresented.
1
1
pred. kr=1 (no redund.) kr=1 (no redund.) TVM (r=8, unif. redund.) pred. eTVM with u=2 (r=1) eTVM with u=2 (r=1) pred. eTVM with u=1 (r=27) eTVM with u=1 (r=27)
0.6
0.8 proportion of correct BBs
proportion of correct BBs
0.8
0.4
0.2
pred. kr=1 (no redund.) kr=1 (no redund.) TVM (u=2, r=64, unif. redund.) pred. eTVM with u=3 (r=1) eTVM with u=3 (r=1) pred. eTVM with u=1 (r=343) eTVM with u=1 (r=343)
0.6
0.4
0.2
0
0 10
20
30
40
50
60
70
80
90
100
10
population size N
(a) kr = 2
20
30
40
50
60
70
80
90
100
population size N
(b) kr = 3
Figure 3.7. Experimental and theoretical results of the proportion of correct BBs for ten concatenated 3bit deceptive traps. We show results for diﬀerent variants of the TV mapping and kr = 2 (left) and kr = 3 (right). The lines without line points show the theoretical predictions. As predicted, GA performance sharply decreases if the eTV mapping underrepresents the optimal BB.
The presented results show that the eﬀects of synonymously redundant representations like the TV mapping on the performance of GEAs can be explained well by a change of the initial supply of highquality BBs. If the eTV mapping favors highquality BBs then the performance of GAs is increased. If good BBs are underrepresented the performance is reduced. If the representation is uniformly redundant, GAs show the same performance as when using the nonredundant encoding. 3.1.7 Conclusions and Further Research This section investigated how redundant representations inﬂuence the performance of GEAs. It distinguished between synonymously and nonsynonymous
58
3 Elements of a Theory of Representations
ly redundant representations and illustrated that nonsynonymous redundancy does not allow genetic operators to work properly and therefore reduces the eﬃciency of evolutionary search. When using synonymously redundant representations, GEA performance depends on the change of the initial supply. Based on this observation, models were developed that give the necessary population size for solving a problem, and the number of generations as O(2kr /r), where kr is the order of redundancy and r is the number of genotypic BBs that represent the optimal phenotypic BB. As a result, uniformly redundant representations do not change the behavior of GAs. Only by increasing r, which means overrepresenting the optimal solution, does GA performance increase. In contrast, GA performance decreases if the optimal solution is underrepresented. Therefore, nonuniformly redundant representations can only be used advantageously if information exists apriori regarding the optimal solution. The validity of the proposed theoretical concepts is illustrated for diﬀerent variants of the redundant trivial voting mapping. The results show that the developed population sizing and time to convergence models allow an accurate prediction of the expected solution quality and solution time. The proposed classiﬁcation, population sizing, and time to convergence models allow us to evaluate redundant representations in a systematic and theoryguided matter. This approach will help users and researchers to answer some of the disputed questions regarding the beneﬁts of redundant representations and to use redundant representations such that they increase the performance, reliability and eﬃciency of evolutionary computation methods. In this study, we only considered crossoverbased search and neglected the inﬂuence of redundant representations on mutationbased search. However, we believe that many of the discussed topics are also relevant when using mutation. According to Sect. 3.1.2, we believe that in analogy to the results from Knowles and Watson (2002), using nonsynonymously redundant representations reduces the performance of mutationbased search. As these representations have low locality, mutation will not work properly and the search becomes random. Furthermore, there is some theoretical evidence (Radcliﬀe 1991a; Whitley et al. 1997; Rana and Whitley 1997; Whitley 1999; Whitley 2000a; Christensen and Oppacher 2001) that mutationbased search only performs well if the connectivity of the phenotypic search space is preserved by the used representation. If the connectivity is either not preserved, such as for low locality representations, or greatly increased (which results in a reduction of the relevant connectivity) like in many nonsynonymously redundant representations, the performance of mutationbased search decreases. In contrast, we expect when using synonymously redundant representations that mutationbased search will show similar behavior and performance as when using crossoverbased search. Using synonymously redundant representations introduces many plateaus in the ﬁtness landscape but does not change the structure of the search space. Mutation can still easily ﬁnd neighboring phenotypes. When using nonuniformly redundant representations, some plateaus in the ﬁtness landscape have a larger size which increases the probability
3.2 Scaling
59
that mutation ﬁnds the solution represented by the genotypes forming this plateau. As a result, the performance of mutationbased search increases if a synonymously redundant representation overrepresents the optimal solution, and decreases otherwise.
3.2 Scaling This section provides the second of three elements of a theory of representations and addresses representations which change the importance of alleles when mapping genotypes on phenotypes. Common representations for GEAs often encode the phenotypes by using a sequence of alleles. When assigning phenotypes to genotypes, a representation can change the importance of the alleles. For example, a phenotype is a list of integers and all alleles (integers) are equally relevant for calculating the ﬁtness of a phenotype. We know from previous work that the BBs (alleles) are solved in parallel if all alleles are equally relevant (Goldberg 1989c). The situation that all alleles are equally relevant is equivalent to the situation that the BBs are uniformly scaled. However, when encoding the phenotypic integers using binary strings, the contributions of the genotypic bits to the ﬁtness function are no longer equal and some bits are more relevant than others. When using such nonuniformly scaled representations, the BBs are solved sequentially and domino convergence occurs. Therefore, the time to convergence increases and the genetic search is aﬀected by genetic drift. This means that lower salient BBs are ﬁxed before they can be reached by the search process. Based on previous work (Rudnick 1992; Thierens 1995; Thierens et al. 1998; Harik et al. 1997), we describe how the performance of GEAs is inﬂuenced by the use of representations with nonuniformly scaled BBs. We develop a populationsizing model with, and without, considering genetic drift. The theoretical models are veriﬁed with empirical results. In the following subsection, we review the eﬀects of domino convergence and genetic drift. In Sects. 3.2.2 and 3.2.3 we develop population sizing models for domino convergence with, and without, considering drift. We present empirical veriﬁcation of the proposed models in Sect. 3.2.4 and end with concluding remarks. 3.2.1 Deﬁnitions and Background Representations assign phenotypes xp consisting of diﬀerent alleles xpi , where i ∈ {0, . . . , lp }, to genotypes, which also consist of diﬀerent alleles xgi , where i ∈ {0, . . . , lg }. Furthermore, the function fp assigns to each phenotype xp a corresponding ﬁtness value. Usually, the ﬁtness of a phenotype depends on the values of xpi . Therefore, diﬀerent phenotypic alleles can have a diﬀerent contribution to the ﬁtness of an individual xp . Representations are deﬁned to be uniformly scaled if the genotypephenotype mapping fg does not change
60
3 Elements of a Theory of Representations
the contribution of the alleles to the calculation of the ﬁtness function. For example, when all phenotypic alleles xpi have an equal contribution to the ﬁtness of an individual, representations are uniformly scaled if the importance of the alleles does not change by the genotypephenotype mapping, and all genotypic alleles xgi also have an equal ﬁtness contribution. Representations can also be uniformly scaled if diﬀerent phenotypic alleles have a diﬀerent contribution to the ﬁtness function as the scaling of a representation only considers the genotypephenotype mapping and not the phenotypeﬁtness mapping. Representations are deﬁned to be nonuniformly scaled if the contributions of the alleles to the ﬁtness are diﬀerent for the genotypes than the phenotypes. The most frequently used nonuniformly scaled representations are exponentially scaled representations where the genotypic alleles have an exponentially scaled contribution to the values of the phenotypic alleles. A common example for exponentially scaled representations is the binary encoding, which encodes integer phenotypic alleles using binary genotypes. The contributions of the genotypic alleles to the construction of the phenotypic alleles are exponentially diﬀerent. We want to emphasize that the scaling of a representation only considers the genotypephenotype mapping and is independent of the phenotypeﬁtness mapping. For the discussion in this section, we assume that the phenotypic alleles are uniformly scaled. This means that all phenotypic alleles have the same contribution to the ﬁtness of an individual. The term “domino convergence” was introduced by Rudnick (1992). He deﬁned domino convergence as the sequential convergence of the alleles in a bitstring. Domino convergence occurs if the alleles are nonuniformly scaled and solved sequentially. Rudnick showed that there is a convergence window of size λc . The convergence window is a set of λc contiguous alleles that have started to converge but are not yet fully converged. More salient alleles have already converged completely, whereas lower salient alleles are not yet touched by convergence. The existence of the convergence window means that not all parts of a problem are solved at the same speed. The higher the contribution of one allele to the overall ﬁtness of an individual, the earlier this allele is solved by GEAs. high salience alleles
low salience alleles
already converged
1
not yet converged
λ
ls
string position
Figure 3.8. Domino convergence
The λ model is a formal approach for modeling domino convergence in GEAs (Thierens et al. 1998). λ ∈ {1, . . . , ls } deﬁnes the dividing line between the genotypic alleles that have already converged and those that are still not
3.2 Scaling
61
touched by the selection pressure of GEAs. ls is the genotypic length of an exponentially scaled BB. We want to assume that each phenotypic BB of size lp is encoded using an (exponentially scaled) genotypic BB of size ls . λ moves on from λ = 1, where all alleles are assumed to be in the initial random state, to λ = ls where all alleles (of the genotypic BB) are converged (see Fig. 3.8). At the beginning of the genetic search process (low λ) alleles are solved that have a high contribution to the ﬁtness of an individual, whereas at the end of the GEA search (λ ≈ ls ) alleles with a low inﬂuence on the ﬁtness are solved. For intermediate states, all lower salient alleles have never been exposed to direct selection pressure and remain in their initial state as long as there is no genetic drift. In the approach used by Thierens et al. (1998), the convergence window has size λc = 1. If the population size N is not large enough, some of the lower salient genotypic alleles are randomly ﬁxed due to genetic drift. This results in a degradation of GEAs. The existence of genetic drift is widely known and has been addressed in the ﬁeld of population genetics (Kimura 1962; Kimura 1964; Gale 1990; Nagylaki 1992; Hartl and Clark 1997), and also in the ﬁeld of genetic algorithms (Goldberg and Segrest 1987; Asoh and M¨ uhlenbein 1994; Thierens et al. 1998; Lobo et al. 2000). More information regarding the inﬂuence of genetic drift on the performance of GEAs can be found in Sect. 3.1.1. Lower salient bits drift randomly because selection does not take these bits into account when deciding between solutions. As two absorbing states exist for binary alleles (all individuals have at their ith position either a zero or a one), lower salient bits could be ﬁxed because of random genetic drift before they are directly exposed to selection pressure and solved by GEAs. Thierens (1995) developed a convergence time model for nonuniformly scaled problems. This work modeled the domino convergence for tournament selection by a wave equation, and showed that the overall convergence time complexity for an exponentially scaled ﬁtness function is approximately of order O(l). This is much slower than for uniformly scaled problems where √ ranked based selection mechanisms have a convergence time of order O( l) where l is the length of the bitstring. In the following paragraphs, we want to use some existing models for developing a population sizing model for exponentially scaled representations, with, and without considering the eﬀects of genetic drift. 3.2.2 Population Sizing Model for Exponentially Scaled Representations Neglecting the Eﬀect of Genetic Drift Representations are exponentially scaled if diﬀerent genotypic alleles have a diﬀerent, exponentially scaled, contribution to the construction of the phenotypic alleles. The binary representation is the most common representative of exponentially scaled representations. We use a genetic algorithm for solving a problem and assume domino convergence due to the exponentially scaled representation. Furthermore, there
62
3 Elements of a Theory of Representations
should be no overlapping between the solving process of the current and the next allele (λc = 1). The next lower salient bit is only solved after the current allele is completely converged. Thus, we expect the developed models to only give us a lower bound on the solution quality and run duration of GAs. Although this is a very conservative assumption, this model (Thierens et al. 1998) predicts the behavior of a GA well (Lobo et al. 2000). If we want to overcome the limitation of strict sequential solving, a convergence window of larger size λc > 1 can be introduced. This would result in an overlapping solving process for the diﬀerent alleles in the convergence window. The convergence window would move through the string from the most salient to the least salient alleles and would allow us to create a more exact model for domino convergence. Assuming strictly sequential solving of alleles (λc = 1), the ﬁtness variance of an exponentially scaled string when λ bits are converged during a GA run was calculated in Thierens et al. (1998) as 2 (λ) = σN
x0 22(ls −λ) − 1 x0 x0 22(ls −λ) x0 (1 − ) ≈ (1 − ) , N N 3 N N 3
where ls is the length of the exponentially scaled string, x0 is the average number of best solutions in the initial population, and N is the population size. For x0 = N/2 (we assume that ls genotypic bits encode a phenotypic 1 2(ls −λ) 2 (λ) ≈ 12 2 . This means BB of size k = 1) the variance simpliﬁes to σN that the ﬁtness variance is determined only by the nonconverged region. The more alleles that are solved during the run, the less noise we get. The ﬁtness variance of the genotypic allele that is currently solved is x0 x0 2 (1 − )22(ls −λ) . (λ) = σBB N N 2 = 22(ls −λ−1) . As the contribution of the alleles to For x0 = N/2 we get σBB 2 becomes the ﬁtness of an individual becomes less with lower salience, σBB smaller with increasing time. The ﬁtness distance d between the best individual and its strongest competitor could be calculated as
d(λ) = 2ls −λ . If we concatenate m exponentially scaled genotypic BBs of size ls , we get competing BBs and an overall genotypic string length of lg = l = ls m. This means, we increase the number of phenotypic BBs from one to m and each phenotypic BB of size lp is encoded by ls genotypic alleles. When solving the 2 from the λth bit of the competing λth bit of a genotypic BB there is noise σBB 2 m = m−1, other BBs, and noise σN from the yet unﬁxed bits in each BB. We know from past work that the probability of making the right choice between a single sample of each bit (compare also (3.5)) is (Miller 1997): d · p=N 2 2 ) 2(m σBB + mσN
3.2 Scaling
Using the equations from above results in ⎛ ⎞ 1 ⎠ p = N ⎝ 2 xN0 1 − xN0 ( 43 m − 1)
63
(3.13)
For x0 = N/2 (we assume that each phenotypic BB has size k = lp = 1 and each phenotypic BB is encoded using ls genotypic alleles) we ﬁnally get: 2 p=N · (3.14) 4 3m − 1 The probability of deciding well is independent of the position λ in the string as long as there is no genetic drift, and the proportion of zeros and ones remains constant for the yet unﬁxed bits. This means for m exponentially scaled BBs that the probability p of deciding well is independent of the length ls of the exponentially scaled BB and that it stays constant over all alleles. Thus, the proportion of correct bits in the string at the end of a run depends only on the number of BBs m. For m = 1 (there is only one exponentially √ scaled BB) there is no noise from competing BBs and we get p = N 6 . Using p we can calculate the proportion of incorrect bits at the end of the run according to the Gambler’s ruin model (compare also (3.15)) as (Feller 1957; Harik et al. 1997) 1 − (1/p − 1)x0 · (3.15) α=1− 1 − (1/p − 1)N 2 we get: For x0 = N/2 and z = 4 m−1 3
N/2 N/2 1 1 1 − N(z) −1 − 1 N(z) 1 α=1− N = N/2 = −N/2 · 1 1 1 1 − N(z) −1 1 + N(z) −1 1 + N(z) −1 (3.16) The probability α of GA failure for exponentially scaled problems only depends on the population size N and the number of competing BBs m as long as the population size is large enough, and no genetic drift occurs. Notice that in contrast to the proportion of correct bits 1 − α, the number of correctly found exponentially scaled BBs is (1 − α)ls (each of the m genotypic BB has ls alleles). When using the ﬁrst two terms of the power series expansion as an approximation for the normal distribution (Abramowitz and Stegun 1972) from (3.14) we get 1 1 p = + √ z, 2 2π 2 where z = . Substituting this approximation in (3.15) results in 4 m−1 3
64
3 Elements of a Theory of Representations
N = 2 ln
α 1−α
⎛ / ln ⎝
1− 1+
2 πz 2 πz
⎞ ⎠·
Since z tends to be a small number, ln(1 ± z π2 ) may be approximated as ±z π2 . Using these approximations and substituting the value z into the equation ﬁnally gives
α 4 1 m−1 · (3.17) N ≈ − ln π 2 1−α 3 This rough approximation determines more clearly the variables the population size N depends on. We see that for exponentially scaled representations, the necessary population size N grows with the square root of the size m of the problem. In contrast to the more general population sizing equation from Harik et al. (1999), N does not depend on the distance d and the variance of an allele σBB if genetic drift is neglected. As already mentioned, we want to emphasize that we consider an exponentially scaled representation (genotypephenotype) mapping and do not address the phenotypeﬁtness mapping. For the developed model we assume that the size of the phenotypic BB, kp = 1, and that the phenotypic alleles are uniformly scaled this means that all phenotypic alleles have the same contribution to the ﬁtness of an individual. Finally, we give an estimation for the convergence time tconv for exponentially scaled BBs of length ls . The time until a uniformly scaled string of length m is converged can be calculated (Thierens and Goldberg 1994) as √ π m , t= 2 I where I denotes the selection√ intensity. For tournament selection without replacement of size 2, I = 1/ π. As there are m exponentially scaled BBs, and therefore m alleles of the same salience, the GEAs solve m bits in parallel. The next m lower salient bits are solved when all m currently solved bits are fully converged (λc = 1). Thus, the solving process for exponentially scaled problems is strictly serial and goes with O(ls ) (Thierens 1995, pp. 66ﬀ). The overall time to convergence can be calculated as: √ l π π m =√ , (3.18) tconv = ls 2 I m 2I where l = √lg = ls m. In contrast to an uniformly scaled √ representation (tconv = O( ls m)), the time to convergence goes with O(ls m) when using a representation, that assigns to each of the m phenotypic BBs ls exponentially scaled alleles. We see clearly that GEAs need more time to converge if nonuniformly scaled representations are used.
3.2 Scaling
65
3.2.3 Population Sizing Model for Exponentially Scaled Representations Considering the Eﬀect of Genetic Drift In the previous subsection, we have developed a population sizing and convergence time model for exponentially scaled representations. However, the model does not consider the eﬀects of genetic drift. This subsection lifts this restriction and investigates the modiﬁcations necessary for considering genetic drift. Drift aﬀects genetic search if the drift time is lower than the time to convergence, tdrif t < tconv . Low salient bits are ﬁxed due to drift before they can be reached by the solution process. The drift time has been studied in the context of GEAs (Goldberg and Segrest 1987; Asoh and M¨ uhlenbein 1994). These studies show that the expected time for an allele to converge due to genetic drift is proportional to the population size N . For an equal proportion of ones and zeros in the start population, using tournament selection of size s = 2 , the size of BBs k = 1 (Lobo et al. 2000), and random sampling with replacement, we get for the drift time tdrif t ≈ 1.4N.
(3.19)
For tconv > tdrif t genetic drift ﬁxes some low salient bits before they can be solved by the search process. Using (3.18), we can calculate the population GEAs using tournament selection of size 2 (selection size Ndrif t for which √ intensity I = 1/ π) are aﬀected by genetic drift as 5π π l· (3.20) Ndrif t < 14 m If the population size N of GEAs using an exponentially scaled representation is lower than Ndrif t then domino convergence does not reach the lowest salient alleles and these alleles are randomly ﬁxed at one of the two absorbing states 0 or 1 (we assume a binary encoding with χ = 2). In the following paragraphs, we want to propose two approaches modeling the inﬂuence of drift on the performance of GEAs using exponentially scaled representations. At ﬁrst, we need a model that describes the drift process itself. An approximation for the probability s(t) that an allele is fully converged due to genetic drift at time (generation) t was given by Kimura (1964): s(t) ≈ 1 − 6
x0 x0 (1 − ) exp(−t/N )· N N
Using this approximation we can calculate how the probability of randomly ﬁxing an allele depends on the population size N and the number of generations t. When x0 = N/2 is the expected number of 1s in the randomly initialized population (we still assume that all phenotypic BBs have size k = 1 and each phenotypic BB is represented by ls genotypic bits) we get s(t) = 1 −
3 exp(−t/N )· 2
66
3 Elements of a Theory of Representations
As s(t) is only an approximation (s(t) < 0 for small t) we deﬁne the convergence probability 0 for t < −N ln(2/3), s (t) = 3 1 − 2 exp(−t/N ) for t > −N ln(2/3).
probability of fixation due to drift
For t < −N ln(2/3) the probability that an allele converges due to genetic drift is zero. In Fig. 3.9, we plot the probability that an allele is fully converged at generation t for N = 10. 1 0.8 0.6 0.4 0.2 s(t) s’(t) ssc(t)
0 0.2 2
4
6
8
101.4 N 12
14
16
18
20
generation t
Figure 3.9. Genetic drift models
√
We know from (3.18) that GEAs need t = π2 Im generations for solving one allele in the bitstring. When considering genetic drift, the bits in the string are either solved correctly with probability 1 − α or they are randomly ﬁxed due √ π m to genetic drift. If domino convergence reaches after t = λ 2 I generations an allele, this allele is not converged due to genetic drift with probability 1−s (t). Therefore, we want to assume that with probability 1−s (t), this allele remains in its initial state and is solved correctly with probability 1 − α. In contrast, with probability s (t), this allele is converged due to genetic drift. If the allele is converged due to genetic drift, it converges to the correct solution with probability 1/2. Therefore, using the probability of GA failure α from (3.16) we can calculate the probability of GA failure when considering genetic drift as 1 π√ π√ πm) α + s (λ πm), αdrif t (λ) = 1 − s (λ 2 2√ 2 √ when using tournament selection of size 2 (I = 1/ π). 1 − s (λ π2 πm) is the probability that the allele at the λth position is not converged due to genetic √ drift at time t = λ π2 πm. The probability of error for this allele is α. √ Furthermore, the alleles are aﬀected by genetic drift with probability s (λ π2 πm), and converge to the wrong solution with probability 1/2. Consequently, the overall average percentage α ¯ drif t of incorrect alleles in one genotypic BB of length ls can be calculated as α ¯ drif t =
ls 1 αdrif t (λ) ls λ=1
(3.21)
3.2 Scaling
67
and is simply the sum over all ls bits of one BB. We refer to this model as the approximated drift model. We assume in this model that the proportion of ones and zeros for for the unconverged alleles xgi , where i > λ, is still 0.5, and therefore α does not depend on λ. We also assume that the solving time for one allele stays constant over the whole solving process, and that it is independent of λ. In the remaining paragraphs, we want to develop a second population sizing model for exponentially scaled alleles considering the eﬀects of genetic drift. Instead of using the approximation from Kimura (1964), we use a simple drift/nodrift approach. We assume that all instances of an allele are either ﬁxed due to genetic drift, or remain in the initial state. Thus, we get a linear, staircase model for the estimation of αdrif t , if we assume that for tdrif √ √t > λ π2 πm all bits can be solved with probability 1−α, and for tdrif t < λ π2 πm the remaining low salient ls − λ bits of each BB are ﬁxed randomly at the correct solution with probability 0.5. As previously, the drift time can be calculated for GAs using tournament selection of size s = 2 as tdrif t ≈ 1.4N . Therefore, the probability ssc that an allele is fully converged due to genetic drift at time t is 0 for t < tdrif t , ssc (t) = 1 for t > tdrif t . We illustrate this in Fig. 3.9. For t < tdrif t we assume no genetic drift, and for t > tdrif t all the remaining bits are randomly ﬁxed. Therefore, the probability of GA failure can be calculated as: √ α for λ < π2.8N , πm αdrif t (λ) = 2.8N 0.5 for λ ≥ π√πm . By using (3.19) and (3.20) the average percentage of incorrect alleles is:
⎧ 2.8N √ ⎨ 1 π√πm α + ls −1 1 for N < 5π √ λ=0 ls 14 πmls ,
2 λ= π2.8N πm (λ) = α ¯ drif t √ ⎩ α for N ≥ 5π 14 πmls . (3.22) For large N , no genetic drift occurs and we get the same failure probability √
bits are as for the nondrift case (3.16). For small N the most salient π2.8N πm solved correctly with probability 1 − α and the rest of the alleles are ﬁxed randomly due to genetic drift. As the drift time has a standard deviation of approximately the same order (Gale 1990, pp. 82ﬀ.) as its mean (≈ 1.4N ), the model underestimates the solution quality for t < 1.4N , and overestimates it for t > 1.4N . The probability of converging to the correct solution has a staircased slope regarding N as long as tdrif t < tconv . Thus, we refer to this model as the staircase drift model.
68
3 Elements of a Theory of Representations
3.2.4 Empirical Results for BinInt Problems In this subsection, we illustrate that the proposed models considering genetic drift predict the behavior of GEAs well for exponentially scaled representations and small populations. We show that with decreasing number of competing exponentially scaled BBs, and increasing length of the BBs, genetic drift leads to a stronger decline of GEAs. For our empirical investigation we use the integer onemax problem from Sect. 5.1. Furthermore, we encode the phenotypes (integers) as binary strings using the binary encoding (compare Sect. 5.2). As a result we get the BinInt problem (Rudnick 1992). There are ls genotypic bits that encode the phenotype and the contribution of an allele to the construction of the phenotype xp is exponentially scaled. The ﬁtness of a phenotype is its integer value. Therefore, the overall ﬁtness of an individual can be calculated as: f (xg ) =
l s −1
xgi 2i
i=0
Thus, the optimal solution is a string with only ones. If the population size is large enough, the BinInt problem is easily solved by GEAs in a stepwise way according to the domino convergence model. According to Sect. 2.1.2, the BinInt problem can be separated into a representation fg , and a phenotypeﬁtness mapping fp . Therefore, the genotype is a binary string of length ls and the phenotype is an integer. We assume that the phenotypeﬁtness mapping is the identity function and assigns an integer value to each phenotype, fp (xp ) = xp . The representation fg is exponentially scaled and assigns ls genotypic alleles to the phenotype. The value of a phenotype is calculated as xp =
l s −1
xgi 2i .
i=0
We present no results for exponentially scaled representations and deceptive problems because the population size N that is necessary to solve these types of problems is in general large enough to ensure that no genetic drift occurs, and therefore, the available population sizing models from Sect. 3.2.2 can be used. For all experiments we use uniform crossover, no mutation, and tournament selection of size two without replacement. The initial population is generated randomly (the initial state of the population has an equal proportion of zeros and ones, x0 = N/2) and a GA run is stopped after the population is fully converged. To gain statistical evidence we performed 250 runs for each problem. We present results for three test cases:
3.2 Scaling
69
• one BinInt problem (m = 1), • 10 concatenated BinInt problems (m = 10), • 50 concatenated BinInt problems (m = 50). We want to denote the correct solution (a sequence of ls ones) for one BinInt problem as a BB. For the BinInt problem, the size of the phenotypic BBs is k = 1 and the size of the genotypic BBs is ls . The ﬁtness of an individual is the sum over all m BinInt problems and can be calculated as f (xp ) =
m−1 i=0
xpi =
m−1 s −1 l
xgls i+j .
i=0 j=0
For each of these three test cases we present results for ls = 5 and ls = 10. Therefore, the maximum ﬁtness of a BinInt problem is f (xp ) = 25 = 32 (ls = 5) or f (xp ) = 210 = 1, 024 (ls = 10). In Table 3.5, we present the overall string length lg , the probability p of making the right choice between a single sample of each BB, and the overall convergence time tconv when assuming no genetic drift. If drift occurs some lower salient genotypic alleles are randomly ﬁxed at 0 or 1 before they can be reached by the search process and tconv is an upper bound for the overall convergence time. Table 3.5. Some properties of the three test cases m=1 m = 10 m = 50 ls 5 10 5 10 5 10 lg 5 10 50 100 250 500 √ p N( 6) √N( 6/37) √ √N( 6/197) √ √ √ tconv 5 π2 π 10 π2 π 5 π2 10π 10 π2 10π 5 π2 50π 10 π2 50π
In Figs. 3.10 (m = 1), 3.11 (m = 10), and 3.12 (m = 50), we present results for the diﬀerent test cases. One phenotypic integer is represented by either ls = 5 (left) or ls = 10 (right) bits. The solid lines with line points show the empirical results. We show predictions for considering no drift (dotted line), for the staircase drift model (dashed line with line points), and for the approximated drift model (dashed line). All predictions consider domino convergence. For all three cases, genetic drift has a large impact on the GA performance especially with increasing ls , and decreasing number of competing BBs m. The number of competing BBs is equivalent to the number of concatenated BinInt problems. In contrast to the nodrift model which could not predict the behavior of the GA well, both the staircase drift model, as well as the approximated drift model, are able to accurately describe the behavior of GEAs using exponentially scaled representations. Both models consider genetic drift and predict the behavior of GEAs better than the dominoconvergence model alone.
70
3 Elements of a Theory of Representations 1 proportion of correct alleles
proportion of correct alleles
1 0.95 0.9 0.85 0.8 0.75
exp. results approx. drift model staircase drift model no drift model
0.7 5
10
15
20
25
30
0.95 0.9 0.85 0.8 0.75
exp. results approx. drift model staircase drift model no drift model
0.7
35
5
number of indidividuals N
10
15
20
25
30
35
number of indidividuals N
(a) m = 1, ls = 5
(b) m = 1, ls = 10
1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5
exp. results approx. drift model staircase drift model no drift model 5 10 15 20 25 30 35 40 45 50 number of indidividuals N
(a) m = 10, ls = 5
proportion of correct alleles
proportion of correct alleles
Figure 3.10. Experimental and theoretical results of the proportion of correct alleles for one BinInt problem of length ls = 5 (left) and ls = 10 (right). The solid lines with line points show the empirical results. All shown predictions consider domino convergence. We show predictions for considering no drift (dotted line), for the staircase drift model (dashed line with line points), and for the approximated drift model (dashed line). With increasing length ls of the genotype, genetic drift has a major impact on the GA, and the GA performance declines. 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5
exp. results approx. drfit model staircase drift model no drift model 5 10 15 20 25 30 35 40 45 50 number of indidividuals N
(b) m = 10, ls = 10
Figure 3.11. Experimental and theoretical results of the proportion of correct alleles for m = 10 concatenated BinInt problems of length ls = 5 (left) and ls = 10 (right). The overall string length is l = 50 (left) and l = 100 (right). The solid lines with line points show the empirical results. We show predictions for considering no drift (dotted line), for the staircase drift model (dashed line with line points), and for the approximated drift model (dashed line). Although the GA is aﬀected by noise from the competing m − 1 problems, which superimposes the eﬀects of drift, genetic drift has a major impact on the GA performance with increasing ls .
1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5
exp. results approx. drift model staircase drift model no drift model 10 20 30 40 50 60 70 80 90 100 number of indidividuals N
(a) m = 50, ls = 5
proportion of correct alleles
proportion of correct alleles
3.2 Scaling 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5
71
exp. results approx. drift model staircase drift model no drift model 10 20 30 40 50 60 70 80 90 100 number of indidividuals N
(b) m = 50, ls = 10
Figure 3.12. Experimental and theoretical results of the proportion of correct alleles for m = 50 BinInt problems of length ls = 5 (left) and ls = 10 (right). The overall string length is l = 250 (left) and l = 500 (right). The solid lines with line points show the empirical results. We show predictions for considering no drift (dotted line), for the staircase drift model (dashed line with line points), and for the approximated drift model (dashed line). With increasing ls , genetic drift has a larger impact on the GA, and the GA performance declines.
We see that with increasing ls , the performance of GEAs declines. This behavior is expected as we know from our theoretical investigations that tconv increases linearly with ls (see (3.18)), whereas the drift time tdrif t stays constant (see (3.19)). Therefore, with increasing ls more and more lower salient alleles are ﬁxed due to genetic drift and GEA performance declines. Our results show that the inﬂuence of genetic drift is reduced for an increasing number m of BBs (BinInt problems). We know from (3.14) that with increasing m the probability of making the right choice between a single sample of each bit is reduced. Therefore, larger populations N are necessary which reduce the inﬂuence of genetic drift with increasing m. This relationship can be seen nicely in the presented plots. For m = 1 (see Fig. 3.10) we have no competing BBs, the necessary populations are very small, and the GA is strongly aﬀected by genetic drift. Therefore, the nodrift model fails completely. For m = 50, however, there is a lot of noise from the competing BBs (BinInt problems), and therefore larger populations are necessary. The inﬂuence of genetic drift is smaller, and the nodrift model when only considering domino convergence gives an acceptable prediction of the solution quality (see. Fig. 3.12). As predicted, the staircase model underestimates the proportion of correct alleles for small N (N < 57 tdrif t ), and overestimates it for large populations (N > 57 tdrif t ). The approximated drift model predicts the slope of the empirical results well, but due to the used domino convergence with strict sequential solving of the bits, it always underestimates the proportion of correct alleles. We believe that by introducing a convergence window, and assuming some
72
3 Elements of a Theory of Representations
parallel solving of the alleles, that the approximated drift model should more accurately predict the behavior of GAs. We have seen that the lower the number m of competing BBs (BinInt problems) is, and the more bits ls each BinInt problem has, the stronger is the impact of genetic drift on the solution quality. Although we only use a strictly serial solution process and no convergence window, the developed models considering genetic drift give us a good prediction of the expected proportion of correct alleles when using exponentially scaled representations. Population sizing models that neglect the eﬀect of genetic drift are not able to accurately predict the expected proportion of correct alleles for a given population size. 3.2.5 Conclusions We have illustrated the eﬀect of nonuniformly scaled representations on the performance of genetic algorithms. When using small populations and easy problems, GAs are aﬀected by genetic drift. To be able to model the eﬀects of genetic drift more accurately, we used the population sizing model from Harik et al. (1999), the domino convergence model from Rudnick (1992), and the time complexity model from Thierens (1995) and Thierens et al. (1998) and developed two population sizing models for exponentially scaled representations considering genetic drift. The approximated genetic drift model uses an approximation (Kimura 1964) for the probability that an allele is completely converged due to genetic drift after t generations, and gives us a lower bound for the solution quality. The staircase drift model assumes that genetic drift occurs as long as the convergence time is larger than the expected drift time, and that the lower salient genes are ﬁxed at the correct solution due to genetic drift with probability x0 /N . The theoretical results reveal that genetic drift has a large impact on the probability of error and convergence time when using exponentially scaled representations. Because the alleles are solved strictly in serial, exponentially scaled representations change the dynamics of genetic search. As a result the solution quality is reduced by genetic drift, and the convergence time is increased by domino convergence. The empirical investigations show that despite the assumption that the size of the convergence window λc = 1, the proposed models considering genetic drift give us, in contrast to the nodrift model, accurate predictions for the solution quality. Except for a very large number m of competing BBs, or a very low number ls of exponentially scaled genotypic alleles, the nodrift population sizing model is not able to predict the expected solution quality. When using exponentially scaled representations, researchers should be aware of the eﬀects of genetic drift as some of the alleles are ﬁxed randomly and of the eﬀects of domino convergence which increases the time to convergence.
3.3 Locality
73
3.3 Locality During the last decades, starting with work by Bagley (1967), Rosenberg (1967) and Cavicchio (1970), researchers recognized that the concept of building blocks is helpful for understanding the principles of selectorecombinative GAs, and is a key factor in determining successful use. A well designed GA should be able to preserve high quality building blocks, and increase their number over the generations (Goldberg 1989c). When using the notion of building blocks in the context of representations, we must be aware that building blocks not only exist in the genotype, but also in the phenotype. A representation transforms the structure and complexity of the building blocks from the phenotype to the genotype, and therefore the structure and complexity of the building blocks can be diﬀerent in the genotype and phenotype. This section provides the third, and ﬁnal element, of a theory of representations for GEAs and investigates how the locality of a representation modiﬁes the complexity of building blocks and changes the diﬃculty of an optimization problem. The locality of a representation describes how well genotypic neighbors correspond to phenotypic neighbors. The locality of a representation is high if genotypic neighbors correspond to phenotypic neighbors. Previous work has indicated that high locality of a representation is necessary for eﬃcient evolutionary search (Rothlauf and Goldberg 1999; Gottlieb and Raidl 1999; Gottlieb and Raidl 2000; Gottlieb et al. 2001; Rothlauf and Goldberg 2000). However, it remains unclear as to how the locality of a representation inﬂuences problem diﬃculty, and if highlocality representations always aid evolutionary search. The results show that highlocality representations do not modify the complexity of the problems they are used for. In contrast, when using lowlocality representations the genotypic BBs can be diﬀerent from the phenotypic BBs, and the complexity of the problem can be changed. Therefore, only highlocality representations can guarantee to reliably and predictably solve problems of bounded complexity. We focus our investigation on how the locality of a representation inﬂuences problem diﬃculty and do not develop models for problem diﬃculty but use selected models presented in Sect. 2.3. In the context of evolution strategies (B¨ack and Schwefel 1995), previous work developed the concept of causality (Igel 1998; Sendhoﬀ et al. 1997b; Sendhoﬀ et al. 1997a) as a measurement of problem diﬃculty. Basically, both concepts, causality and locality, address similar aspects as they describe how well the distances between individuals are preserved when mapping the phenotypes on the genotypes. However, in this study we do not classify the diﬃculty of problems but only describe how the diﬃculty of a problem is changed by the encoding. In the following subsection, the inﬂuence of representations on problem diﬃculty is discussed. We illustrate why it is helpful to use encodings that do not modify problem diﬃculty for solving problems of bounded diﬃculty. Section 3.3.2 introduces the concept of locality and shows how locality depends on
74
3 Elements of a Theory of Representations
the metrics that are deﬁned for the genotypes and phenotypes. In Sects. 3.3.3 and 3.3.4, we review the concept of problem diﬃculty introduced by Jones and Forrest (1995), and show how the locality of a representation inﬂuences problem diﬃculty. Sect. 3.3.5 introduces the distance distortion of a representation and illustrates that high locality is necessary for an encoding to have low distance distortion. In Sect. 3.3.6, we show for the fully easy onemax problem that lowlocality representations make the problem more diﬃcult. The results of the theoretical investigation are veriﬁed and illustrated in Sect. 3.3.7 by an empirical investigation for the onemax and deceptive trap problem. The results show that for highlocality representations the onemax problem remains fully easy, and the fully deceptive trap remains fully diﬃcult. In contrast, when using lowlocality representations the fully easy onemax problem becomes more diﬃcult to solve for GEAs, whereas the deceptive trap becomes easier. The section ends with concluding remarks. 3.3.1 Inﬂuence of Representations on Problem Diﬃculty We discuss the inﬂuence of representations on problem diﬃculty and why representations should preserve the complexity of a problem when assigning the genotypes to the phenotypes. It was already recognized by Liepins and Vose (1990) that by using diﬀerent representations the complexity of a problem can be completely changed. Changing the complexity of a problem means that the diﬃculty of the problem and therefore the structure of the BBs is changed (see Sect. 2.3). However, representations that modify BBcomplexity are problematic as they do not allow us to reliably solve problems of bounded diﬃculty. If the complexity of BBs is changed by the encoding, some problems become easier to solve, whereas others become more diﬃcult. To predict which problems become easier and which do not is only possible if detailed knowledge about the optimization problem exists (compare also Sects. 3.3.4 and 4.4.3). To ensure that problems of bounded complexity, that means easy problems, can be reliably solved by GEAs, representations that preserve the complexity of the building blocks should be used. Liepins and Vose (1990) showed that for a fully deceptive problem f (x) = fp (fg (x)) there is a transformation T such that the function g(x) = f [T (x)] becomes fully easy. This means that every fully deceptive problem could be transformed into a fully easy problem by introducing a linear transformation T . In general, the diﬃculty of a problem could easily be modiﬁed by using diﬀerent linear transformations T . Liepins and Vose concluded that their results underscore the importance of selecting good representations, and that good representations are problemspeciﬁc. Therefore, in order to ﬁnd the best representation for a problem, it is necessary to know how to optimize the problem and what the optimal solution for the problem is. To ﬁnd the best representation for a speciﬁc problem, Liepins and Vose proposed using adaptive representations that “simultanously search for rep
3.3 Locality
75
resentations at a metalevel” (Liepins and Vose 1990, p. 110). These representations should autonomously adapt to the problem and encode it in a proper way (compare Rothlauf et al. (2000)). Initial steps in this direction were made by Goldberg et al. (1989) with the development of the messy GA. These kinds of GAs use an adaptive encoding that adapts the structure of the representation to the properties of the problem. This approach, however, burdens the GA not only with the search for promising solutions, but also the search for a good representation. Therefore, we go back one step and ask the question, what kind of encoding should be used if there is no a priori knowledge about the problem, and no adaptive encoding should be used. Users who simply want to solve their problems by using a GEA are confronted with this kind of problem. They have no knowledge about the problem, and they do not want to do experiments to ﬁnd out what structure the problem has, what the promising areas in the search space are, and what kind of representations make the problem most easy. They just want to be sure that the GEA can solve their problem as long as the complexity of their problem is bounded and the used GEA is able to solve it. One solution for their problems are representations that preserve BBcomplexity. Using these types of representations means that the problem has the same diﬃculty in the genotypic as in the phenotypic space. Then users can be sure that the representation does not increase the complexity of the problem, and that the GEA reliably solves their problem. Wanting representations to preserve BBcomplexity raises the question of, why are we especially interested in encodings that preserve complexity? Is it not more desirable to construct encodings that reduce problem complexity, as Liepins and Vose propose? The answer is yes and no. Of course, we are interested in encodings that reduce problem diﬃculty. However, in general it is not possible to construct a representation that reduces the complexity for all possible problem instances. One result of the nofreelunch theorem (Wolpert and Macready 1995) is that if some problem instances become easier by the use of a representation, there are other types of problem instances that necessarily become more diﬃcult. Therefore, we want to at least ensure that the representation does not make problems of bounded diﬃculty – these are the class of problems we are interested in – more diﬃcult to solve. However, as shown in Sects. 3.3.6 and 3.3.7, encodings that do not preserve problem complexity always make fully easy problems more diﬃcult. Therefore, a phenotypically easy problem could even become so diﬃcult by using a “bad” encoding that it can not be solved eﬃciently any more. Of course, representations that do not preserve problem complexity make fully diﬃcult problems more easy, but as we are interested in solving only problems of bounded diﬃculty, and not all types of problems, this is not important to us. For diﬃcult problems, like the needle in the haystack problem, or the fully deceptive trap problem, the complexity of the problem is not bounded, and therefore, we are in general not interested in solving these kinds of problems with GEAs.
76
3 Elements of a Theory of Representations
The solution for all our problems would be an “optimal” representation that preserves problem complexity for problems of bounded diﬃculty and reduces the complexity for all other problems. But reaching this aim is far beyond the scope of this work. Furthermore, we believe that the no nofreelunch theorem does not allow us to get such a free lunch for every problem. Finally, we want to come back to the results of Liepins and Vose (1990) and illustrate the problems with representations that change the diﬃculty of problems. The transformation T , which can be interpreted as a genotypephenotype mapping, can modify the complexity of a problem in such a way that a fully diﬃcult deceptive trap problem becomes a fully easy onemax problem. But, using the same transformation T for a fully easy onemax problem can result in a fully deceptive trap (compare Fig. 4.1(a)). Therefore, by using this representation, we are able to solve a deceptive trap, but not the onemax problem any more. If we want to solve both types of problems we must know a priori what the problem is and adjust the representation according to the problem. However, if we do not know the problem a priori, and if we want to make sure that we can solve at least problems of bounded diﬃculty reliably, we must use representations that do not modify problem diﬃculty. In the following subsection, we deﬁne the locality of a representation formally and show in Sect. 3.3.4 that low locality is necessary for a representation to not modify problem diﬃculty. 3.3.2 Metrics, Locality, and Mutation Operators In Sects. 2.1.1 and 2.1.2, we introduced the concept of a representation which assigns genotypes xg ∈ Φg to corresponding phenotypes xp ∈ Φp . In the following paragraphs, we introduce the concept of locality and describe how the locality of a representation is based on the metric used for Φg and Φp . When using search algorithms, a metric has to be deﬁned on the search space Φ. Based on the metric, the distance dxa ,xb between two individuals xa ∈ Φ and xb ∈ Φ describes how similar the two individuals are. The larger the distance, the more diﬀerent two individuals are. In general, diﬀerent metrics can be deﬁned for the same search space. Diﬀerent metrics result in diﬀerent distances and diﬀerent measurements of the similarity of solutions. Two individuals are neighbors if the distance between two individuals is minimal. For example, when using the Hamming metric (Hamming 1980) for binary strings the minimal distance between two diﬀerent individuals is dmin = 1. Therefore, two individuals xa and xb are neighbors if the distance dxa ,xb = 1. If we use a representation fg there are two diﬀerent search spaces, Φg and Φp . Therefore, diﬀerent metrics can be used for the phenotypic and the genotypic search space. In general, the metric used on the phenotypic search space Φp is determined by the speciﬁc problem that should be solved and describes which problem solutions are similar to each other. In contrast, the metric deﬁned on Φg is not given a priori but depends on the used genotypes.
3.3 Locality
77
As diﬀerent genotypes can be used to represent the same phenotypes, diﬀerent metrics can be deﬁned on Φg . Therefore, in general, diﬀerent metrics are used for Φp and Φg which imply a diﬀerent neighborhood structure in Φg and Φp . For example, when encoding integers using binary strings the phenotype xp = 5 has two neighbors, y p = 6 and z p = 4. When using the Hamming metric, the corresponding binary string xg = 101 has three diﬀerent neighbors, ag = 001, bg = 111, and z g = 100 (Caruana and Schaﬀer 1988). The locality of a representation describes how well neighboring genotypes correspond to neighboring phenotypes. The locality of a representation is high if all neighboring genotypes correspond to neighboring phenotypes. In contrast, the locality of a representation is low if some neighboring genotypes do not correspond to neighboring phenotypes. Therefore, the locality dm of a representation can be deﬁned as dpx,y − dpmin , (3.23) dm = g dg x,y =dmin
where dpx,y is the phenotypic distance between the phenotypes xp and y p , dgx,y is the genotypic distance between the corresponding genotypes, and dpmin , resp. dgmin is the minimum distance between two (neighboring) phenotypes, resp. genotypes. Without loss of generality we want to assume that dgmin = dpmin . For dm = 0 all genotypic neighbors correspond to phenotypic neighbors and the encoding has perfect (high) locality. We want to emphasize that the locality of a representation does not only depend on the representation fg , but also on the metric that is deﬁned on Φg and the metric that is deﬁned on Φp . fg only determines which phenotypes are represented by which genotypes and says nothing about the similarity between solutions. Before we are able to describe the locality of a representation a metric must be deﬁned on Φg and Φp . In the remaining paragraphs, we brieﬂy discuss how the mutation operator used for genetic search determines the metric and the distances that are used for the genotypic space Φg . Based on the metric deﬁned on the genotypic search space Φg , search operators like mutation or crossover can be deﬁned. In EAs, and most of the individualbased search heuristics, like simulated annealing, tabu search, and others, the search operator mutation is designed to create new solutions (oﬀspring) with similar properties as its/their parent(s) (Doran and Michie 1966). In most local search methods, mutation creates oﬀspring that have a small or sometimes even minimal distance to their parents (for example the bitﬂipping operator for binary representations). Therefore, mutation operators and metrics can not be developed independently of each other but determine each other. A metric deﬁnes the mutation operator and the used mutation operator determines the metric. As the search operators are applied to the genotypes, the metric that is used on Φg is relevant for the deﬁnition of mutation operators.
78
3 Elements of a Theory of Representations
The basic idea behind using mutationbased search approaches is that the structure of the ﬁtness landscape should guide the search heuristic to the highquality solutions (Manderick et al. 1991), and that the optimal solution can be found by performing small iterated changes. It is assumed that the highquality solutions are not isolated in the search space but grouped together (Christensen and Oppacher 2001; Whitley 2002). Therefore, better solutions can easily be found by searching around already found good solutions. The search steps must be small because too large search steps would result in a randomization of the search, and guided search around good solutions would become impossible. In contrast, when using search operators that perform large steps in the search space it would not be possible to ﬁnd better solutions by searching around already found good solutions but the search algorithm would jump randomly around the search space (compare also Sect. 3.1.2). 3.3.3 PhenotypeFitness Mappings and Problem Diﬃculty As described in Sect. 2.1.2 the diﬃculty of a problem depends on the phenotypeﬁtness mapping fp . Furthermore (compare Sect. 2.3), the diﬃculty of a problem depends on the used search method and the metric that is deﬁned on the phenotypic search space Φp . The metric deﬁned on Φp determines which individuals are similar to each other and depends on the used main search operator. Both determinants of problem diﬃculty, the phenotypeﬁtness mapping fp and the metric deﬁned on Φp , are given a priori by the character of the optimization problem that should be solved and by the used search method. In Sect. 3.3.2, we described that the mutation operator and the used metric determine each other. Diﬀerent mutation operators imply diﬀerent metrics. As problem diﬃculty depends not only on fp but also on the metric deﬁned on Φp the diﬃculty of a problem is not absolute but depends on the used metric respectively search operator. The use of diﬀerent metrics and search operators result in a diﬀerent problem diﬃculty. Consequently, the diﬃculty of a problem can only be deﬁned with respect to a search operator. It makes no sense to say a problem is either easy or diﬃcult if the used search operator is not taken into account. In Sect. 2.3.2, we gave a short review of some measurements for problem diﬃculty in the context of GEAs. In the following subsections, we want to use the classiﬁcation of problem diﬃculty from Jones and Forrest (1995), which is based on the correlation analysis, for describing how the locality of a representation inﬂuences GEA performance. The diﬃculty of an optimization problem is determined by how the ﬁtness values are assigned to the phenotypes and what metric is deﬁned on the phenotypes. Combining both aspects we can measure problem diﬃculty by the ﬁtnessdistance correlation coeﬃcient ρF DC ∈ [−1, 1] of a problem (Jones 1995; Jones and Forrest 1995). ρF DC measures the correlation between the ﬁtnesses of search points and their dis
3.3 Locality
79
tances to the global optimum. We want to distinguish between three diﬀerent classes of problem diﬃculty: 1. The ﬁtness diﬀerence to the optimal solution is positively correlated with the distance to the optimal solution. With lower distance the ﬁtness difference to the optimal solution decreases. As the structure of the search space guides local search methods to the optimal solution such problems are easy for mutationbased search. 2. There is no correlation between the ﬁtness diﬀerence and the distance to the optimal solution. The ﬁtness values of neighboring individuals are uncorrelated and the structure of the search space provides no information about which solutions should be sampled next by the search method. 3. The ﬁtness diﬀerence is negatively correlated to the distance to the optimal solution. Therefore, the structure of the search space misleads a local search method to suboptimal solutions.
f opt −f
f opt −f
f opt −f
The three diﬀerent classes of problem diﬃculty are illustrated in Fig. 3.13. We show how the ﬁtness diﬀerence fopt − f  depends on the distance d to the optimal solution. In the following paragraphs, we want to discuss these three classes in some more detail.
d pos. correlation
d uncorrelated
d neg. corelation
Figure 3.13. Diﬀerent classes of problem diﬃculty
Problems are easy for mutationbased search if there is a positive correlation between an individuals’ distance to the optimal solution and the diﬀerence between its ﬁtness and the ﬁtness of the optimal solution. Many test problems that are commonly used for EAs like the sphere and corridor models for evolution strategies or the onemax problem for genetic algorithms show this behavior. Such problems are easy for local and crossoverbased search methods as the search is guided to the optimal solution by the structure of the ﬁtness landscape. Problems become much more diﬃcult if there is no correlation between the ﬁtness diﬀerence and the distance to the optimal solution. Then, the ﬁtness landscape does not guide a mutationbased search method to the optimal solution. No search heuristics can use information about a problem which was collected in prior search steps to determine the next search step. Therefore, all reasonable search algorithms show the same performance as no useful information (information that indicates where the optimal solution can be found)
80
3 Elements of a Theory of Representations
is available in the problem. Because all search strategies are equivalent, also random search is an an appropriate search method for such problems. Random search uses no information and performs well on these types of problems. Problem diﬃculty is maximal for mutationbased search methods if the ﬁtness landscape leads the search method away from the optimal solution. Then, the distance to the optimal solution is negatively correlated to the ﬁtness diﬀerence between an individual and the optimal solution. Because mutationbased search ﬁnds the optimal solution by performing iterated small steps in the direction of better solutions, all mutationbased search approaches must fail as they are mislead. All other search methods that use information about the ﬁtness landscape also fail. The most eﬀective search methods for such problems are those that do not use information about the structure of the search space but search randomly like random search. The most prominent example for such types of problems are deceptive traps. Such problems are commonly used to perform a worstcase analysis for EAs. Although we use this problem classiﬁcation for investigating the inﬂuence of locality on problem diﬃculty, we want to emphasize that in general this problem classiﬁcation is not relevant for most of the realworld problem instances. Only problems of class one can be solved eﬃciently using EAs or local search as this problem class guides the local search methods (like mutationbased EAs) to the good solutions. In contrast, for problems of class two, mutationbased search methods perform the same as random search, and for problems of class three random search performs even better. This situation is not in contrast to the observed good performance of EAs on many realworld problem instances. EAs show a good performance as most of the realworld problems are easy problems and belong to class one (compare also Sect. 3.3.1). In general, for realworld problems the ﬁtness values of neighboring solutions are correlated, and highquality and lowquality solutions are grouped together. Therefore, ﬁtness landscapes that are uncorrelated, or even deceptive, are uncommon in real world. performance mutation− based search random search
pos. correlation (class 1)
no correlation (class 2)
neg. correlation (class 3)
problem difficulty
Figure 3.14. Performance of mutationbased EA search versus random search
This situation is illustrated in Fig. 3.14. We know from the nofreelunch theorem that all search methods show on average the same performance over all possible problem instances (Wolpert and Macready 1995; Whitley 2000a).
3.3 Locality
81
Furthermore, we know that the performance of random search remains constant over all problem instances and that mutationbased evolutionary search performs well on problems of class one. Consequently, it must show low performance on other problem instances (class 3). As the performance of mutationbased search is “biased” towards problems of class one, many realworld instances can eﬃciently be solved using mutationbased EAs. 3.3.4 Inﬂuence of Locality on Problem Diﬃculty In Sect. 3.3.1, we described how representations can change the character and diﬃculty of optimization problems. In the following paragraphs, we discuss high versus lowlocality representations and examine how the locality of a representation inﬂuences problem diﬃculty. Low versus HighLocality Representations
phenotypic search space
We have seen in Sect. 3.3.2 that the metric deﬁned on Φp can be diﬀerent from the metric deﬁned on Φg . As the locality of a representation measures how well the phenotypic metric corresponds to the genotypic metric, it is possible to distinguish between high and lowlocality representations. Representations have high locality if neighboring genotypes correspond to neighboring phenotypes. In contrast, representations have low locality if neighboring genotypes do not correspond to neighboring phenotypes. Figure 3.15 illustrates the difference between high and lowlocality representations. In this example, we assume that there are 12 diﬀerent phenotypes (al) and that there is a metric deﬁned on Φp (in this example the Euclidean distance). Each phenotype (lower case symbol) corresponds to one genotype (upper case symbol). The representation fg has perfect (high) locality if neighboring phenotypes correspond to neighboring genotypes. Then a mutation step has the same eﬀect in the phenotypic and genotypic search space. a
b
c
d
e
f
g
h
i
j
k
l
genotypic search space
fg A B C D
L
E
F
G H
D F
K H
I
J
K L
E
B G
high locality
A C I J
low locality
Figure 3.15. High versus lowlocality representations
As we assume in our considerations that fg is a onetoone mapping every phenotype is represented by exactly one genotype and there are Φg ! = Φp !
82
3 Elements of a Theory of Representations
diﬀerent representations. Φg  is the size of the genotypic search space. Each of these many diﬀerent representations assigns the genotypes to the phenotypes in a diﬀerent way. A common example are diﬀerent representations for representing integer phenotypes using binary strings. Both, binary and Gray encoding, represent integers using binary strings of the same length but they diﬀer in which phenotype is represented by which genotype. Investigating the relationship between diﬀerent representations (how the genotypes are assigned to the phenotypes) and the used mutation operator (which is based on the genotypic metric) reveals that a diﬀerent assignment of genotypes to phenotypes can be equivalent to the use of a diﬀerent metric for Φg . This eﬀect is known as the isomorphism of ﬁtness landscapes (Reeves 1999). For example, it can be shown that the use of a simple bitﬂipping operator (which induces the Hamming metric) for Gray encoded problems is equivalent to the use of the complementary crossover operator (which induces a diﬀerent “nonHamming” metric) for binary encoded problems (Reeves 2000). Both metricrepresentation combinations result in the same ﬁtness landscape and therefore in the same performance of mutationbased search. Inﬂuence on Problem Diﬃculty In the following paragraphs, we examine how the locality of a representations inﬂuences the performance of GEAs. A representation transforms the phenotypic problem fp with a given phenotypic problem diﬃculty into a genotypic problem f = fp ◦ fg with a resulting genotypic problem diﬃculty that can be diﬀerent from the phenotypic problem diﬃculty. We use the problem classiﬁcation described in Sect. 3.3.3. We have seen that the phenotypic diﬃculty of an optimization problem depends on the metric that is deﬁned on the phenotypes and the function fp which assigns a ﬁtness value to every phenotype. Based on the phenotypic metric a local or crossoverbased search operator can be deﬁned (for the phenotypes). By the use of a representation, which assigns a genotype to every phenotype, a new genotypic metric is introduced which can diﬀer from the phenotypic metric. Therefore, the character of the search operator can also be diﬀerent for genotypes and phenotypes. If the locality of a representation is high, then the search operator has the same eﬀect on the phenotypes as on the genotypes. As a result, genotypic and phenotypic problem diﬃculty is the same and the diﬃculty of a problem remains unchanged by the use of an additional representation fg . Easy phenotypic problems remain genotypically easy (compare the results presented in Fig. 3.21) and diﬃcult phenotypic problems remain genotypically diﬃcult (compare Fig. 3.22). Figure 3.16 (left) illustrates the eﬀect of mutation for highlocality representations. The search operator (mutation) has the same eﬀect on the phenotypes as on the genotypes. The situation is diﬀerent when focusing on lowlocality representations. Here, the inﬂuence of the representation on the diﬃculty of a problem depends on the considered optimization problem. If the considered problem fp
3.3 Locality
phenotypic search space
high locality
83
low locality
α
b
c
d
α
k
e
f
g
h
i
f h
b
i
j
k
l
e
j
l
c g
d
genotypic search space
fg A B C D E
F
G H
I
J
K L
Figure 3.16. The eﬀect of mutation for high versus lowlocality representations
is easy (class 1) and the structure of the search space guides the search method to the optimal solution, a lowlocality representation fg randomizes the problem and makes the overall problem f more diﬃcult. When using lowlocality representations a small change in a genotype does not correspond to a small change in the phenotype, but larger changes in the phenotype are possible (compare Fig. 3.16 (right)). Therefore, when using lowlocality representations, phenotypic easy problems of class one become on average genotypic problems of class two. Lowlocality representations lead to a more uncorrelated ﬁtness landscape and heuristics can no longer extract information about the structure of the problem. Guided search becomes more diﬃcult as many genotypic search steps do not result in a similar individual but in a random one. If the problem fp is of class two, on average a lowlocality representation does not change the problem class. Although the mutationbased search becomes more random search, the performance stays constant as random search and mutationbased search show the same performance for problems of class two. Of course, representations exist that can make a problem easier and result in an overall genotypic problem f of class one; however, there are only few of these and most of the lowlocality representations simply modify the problem and do not create a ﬁtness landscape of class one which leads the search method to the good solutions. On the other hand, there are also representations fg that construct a problem f which misleads mutationbased search and transforms a problem of class two into class three. But as for lowlocality representations that transform a problem from class two into class one, there are only few such representations. Finally, we have to consider problems of class three. On average, the use of lowlocality representations transforms such problems into problems of class two as the problems become more randomized. Then, mutationbased search is less misled by the ﬁtness landscape and the problem diﬃculty for mutationbased search is reduced. On average, lowlocality representations “destroy” the deceptiveness of class three problems and turn them into problems of class two.
84
3 Elements of a Theory of Representations
Summarizing the results, we recognize that lowlocality representations have the same eﬀect as when using random search. Therefore, on average problems of class one become more diﬃcult, and problems of class three more easy to solve. As most realworld problems belong to class one, the use of lowlocality representations makes these problems more diﬃcult. Therefore, we strongly encourage researchers to use highlocality representations for problems of practical relevance. Of course, lowlocality representations make deceptive problems easier; however, these are problems which we do not expect to meet in reality and are only of theoretical interest. 3.3.5 Distance Distortion and Crossover Operators We extend the notion of locality and introduce the distance distortion of an encoding. The concept of distance distortion is related to the concept of heritability which describes that a crossover operator should create new oﬀspring that have similar properties to their parents. Appropriate measurements for heritability describe how well oﬀspring take over advantageous features of their parents (Gottlieb and Raidl 1999). When using recombinationbased search, the locality concerning small changes dm can be extended towards locality concerning small and large changes. The distance distortion dc describes how well the phenotypic distance structure is preserved when mapping Φp on Φg : dc =
2 np (np − 1)
np np
dpx,y − dgx,y ,
i=1 j=i+1
where np is the number of diﬀerent individuals, np = Φg  = Φp , and dgmin = dpmin . For dc = 0 all phenotypic distances are preserved by the representation. We see that for Φg = Φp high locality (dm = 0) results in low distance distortion (dc = 0). If, for example, our genotypic and phenotypic search space is binary, and the locality of the genotypephenotype mapping is perfect, then all distances between the individuals are preserved. However, if we assume that Φg = Φp , then high locality is a necessary, but not suﬃcient condition for the genotypephenotype mapping to have low distance distortion. Figure 3.17 illustrates the diﬀerence between representations with high versus low distance distortion. The distance distortion dc of a representation is low if the genotypic distances correspond to the phenotypic distances. If the distances between the genotypes and the corresponding phenotypes are diﬀerent, then the distance distortion dc of the representation is high. It is of interest that the locality dm and distance distortion dc do not require the deﬁnition of genetic operators a priori. It is suﬃcient to deﬁne both based on the distance metrics used for Φg and Φp . The application of mutation to an individual should result in an oﬀspring that is similar to its parent. Therefore, in many implementations, mutation creates oﬀspring who have the lowest possible distance to the parent (for example the bitﬂipping
3.3 Locality p da,b =1
cp
phenotypes
phenotypes
p da,c =3
cp
p
p
p db,c =3
85
p da,b =1
da,c =3
db,c=3
ap
ap p
p
b
b
cg ag
g
g
g da,b =1
g
da,c =3
ag
db,c=3
g da,b =3
da,c =3 g db,c =1
cg genotypes
b
g
a) low distance distortion (dc=0)
genotypes
bg
b) high distance distortion (dc >0)
Figure 3.17. The ﬁgures illustrate the diﬀerence between representations with low versus high distance distortion. If the distance distortion dc = 0 then the distances between the corresponding genotypes and phenotypes are the same. If dc > 0 genotypic distances between individuals do not correspond to phenotypic distances between individuals.
operator for binary representations). High locality of a representation is a necessary condition for successful use of mutationbased search algorithms. Otherwise, lowlocality encodings do not allow a guided search and GEAs using lowlocality representations behave like random search. The situation is similar when using crossover operators. The application of crossover operators should result in oﬀspring where the distances between the oﬀspring and its two parents are smaller than the distance between both parents. Common standard crossover operators, like npoint or uniform crossover show this behavior. The distances between genotypic oﬀspring and parents are always lower, or equal to, the distances between both parents. However, if a representation has high distance distortion, the genotypic distances do not correspond to the phenotypic distances. Then, the phenotypic distances between oﬀspring and parents are not necessarily smaller than the phenotypic distances between both parents. The application of crossover to genotypes does not result in oﬀspring phenotypes that mainly consist of substructures of their parents’ phenotypes. Therefore, the oﬀspring is not similar to its parents and the use of crossover results in random search. We see that low distance distortion of a representation is a necessary condition for good performance of crossoverbased GAs. Examining the interdependencies between locality and distance distortion shows that high locality is a necessary condition for an encoding to have low distance distortion. When using standard crossover operators such as uniform or npoint crossover, the oﬀspring could not inherit the properties of the parents if similar genotypes result in completely diﬀerent phenotypes. If the encoding has low locality, the crossover operators would create oﬀspring
86
3 Elements of a Theory of Representations
genotypes which are similar to the genotypes of the parents, but the resulting phenotypes would not be similar to the phenotypes of the parents. Thus, low locality of a representation would also result in high distance distortion. In the following subsection, we show for the fully easy onemax problem that the problem only stays fully easy if a highlocality representation is used and dm = dc = 0. All other types of representations increase the diﬃculty of the problem. 3.3.6 Modifying BBComplexity for the OneMax Problem In this subsection, we investigate the inﬂuence of locality on GA performance. We show for the fully easy onemax problem (3.24) that the use of a highlocality representation does preserve problem diﬃculty whereas other types of representations reduce GA performance. genotypes
phenotypes
000
000
001
001
010
010
011
011
100
100
101
101
110
110
111
111 fg
fitness 0 1 2 3 fp
Figure 3.18. A representation for the bitcounting function with perfect locality and no distance distortion
For the onemax, or bitcounting problem, the function fp : {0, 1}l → R l−1 assigns to every individual xp ∈ Φp the ﬁtness value i=0 xpi . As there are only l ﬁtness values that are assigned to 2l phenotypes the ﬁtness function fp is aﬀected by redundancy (see Sect. 3.1.6). The genotypephenotype mapping fg is a nonredundant onetoone mapping, and the genotypic space Φg , and the phenotypic space Φp , have the same size Φg  = Φp  = 2l and the same properties Φg = Φp . To simplify the investigation we want to assume, without loss of generality, that the phenotype with only ones is always represented by the genotype with only ones, and therefore is always the global optimum. In Fig. 3.18, a 3bit onemax problem is illustrated. The encoding used, which can be described by the genotypephenotype mapping fg , has high locality (dm = 0) and preserves the distances between the individuals when mapping the phenotypes to the genotypes (dc = 0) as the genotypephenotype mapping is the identity mapping xp = fg (xg ) = xg . As a result, the phenotypic and genotypic problem complexity is the same. We investigate how problem diﬃculty changes if we use a representation where dc = 0. For measuring problem diﬃculty we use the ﬁtness of the
3.3 Locality
87
schemata (compare Sect. 2.3.2). The ﬁtness of a schema h of size λ is deﬁned as f (h) = f (u, λ, l). It has length l, u ones, λ − u zeros in the ﬁxed positions, and l − λ don’t care positions. For the onemax problem the schema ﬁtness in terms of the function values f (u) = u can be calculated as follows:
l−λ 1 l−λ (i + u). f (u, λ, l) = l−λ i 2 i=0 For all schemata of size λ the diﬀerence between the ﬁtness f (λ, λ, l) of the best schemata with λ ones, and the ﬁtness f (λ − 1, λ, l) of its strongest competitor with λ − 1 ones is
l−λ 1 l−λ d = l−λ > 0. i 2 i=0 Thus, all schemata h that contain the global optimum xopt (a string of only ones) are superior to their competitors, and the onemax problem is phenotypically fully easy. genotypes
phenotypes
000
000
001
001
010
010
011
011
100
100
101
101
110
110
111
111 fg
fitness 0 1 2 3 fp
Figure 3.19. A lowlocality representation for the onemax problem which does not preserve the distances between the individuals (dc = 0), and therefore modiﬁes BBcomplexity. The dis2 ∗ 12 = 3/7 = 0. tance distortion dc = 7∗8
We investigate how the problem complexity changes if the distances between the individuals are changed by using a lowlocality representation. For example, the genotypephenotype mapping of two genotypes xg and y g which correspond to the phenotypes xp and y p , is changed such that afterwords xg = 000 represents y p = 001, and y g = 001 represents xp = 000 (see Fig. 3.19). Beginning with the highlocality encoding illustrated in Fig. 3.18, there are three diﬀerent possibilities when changing the mapping of two individuals: • Both individuals xg and y g have the same number of ones (uxg = uyg ) • Both individuals have a diﬀerent number of ones, and the number of different positions dxg ,yg in the two individuals xg and y g is the same as the number of diﬀerent ones (dxg ,yg = uxg − uyg ) • Both individuals have a diﬀerent number of ones, and the number of different positions is higher than the number of diﬀerent ones (dxg ,yg > uxg − uyg )
88
3 Elements of a Theory of Representations
In the following paragraphs, we investigate how the diﬃculty (measured using the notion of schemata) of the problem is changed for these three situations. If fg is modiﬁed for two genotypes xg and y g that have the same number of ones in the string, then the corresponding ﬁtness values remain unchanged (f (xg ) = f (y g )). Therefore, the ﬁtness of the schemata, and the diﬃculty of the problem both remain constant. For example, we can change the mapping of the genotypes 1001 and 0101 for a 4 bit onemax problem. Both individuals have two ones in the string and their ﬁtness is two. The diﬃculty of the problem remains unchanged. fg could be modiﬁed for two individuals xg and y g that have a diﬀerent number of ones, and therefore diﬀerent ﬁtness values. We assume that the number of diﬀerent positions in the two individuals is the same as the number of diﬀerent ones (dxg ,yg = uxg − uyg ). Before the change, the individual xg has l ones and therefore ﬁtness l; y g has h ones and ﬁtness h. We want to assume h > l. After the change, xg has ﬁtness h although it has only l ones, whereas yg has only a ﬁtness of l but h ones. Before the modiﬁcation, all schemata that lead to the global solution are superior to their competitors. Subsequently, after the modiﬁcation of the mapping the ﬁtness of all schemata h that contain y g but not xg , is reduced by (h − l)/(2l−λ ), whereas the ﬁtness of all misleading schemata containing only xg is increased by this amount. Schemata that contain xg as well as y g are not changed. As a result, the average ﬁtness of high quality schemata is reduced, whereas the ﬁtness of misleading schemata is increased. Let us illustrate this with a small 3bit example. fg from Fig. 3.18 should be modiﬁed for the genotypes 001 and 101. Therefore, xg = 001 corresponds to xp = 101 and yg = 101 corresponds to yp = 001 Then, individual xg = 001 has ﬁtness 2, and individual yg = 101 has ﬁtness 1. The ﬁtness of the schema 1** is reduced, whereas the ﬁtness of schema 0** increases. For size two schemata, the ﬁtness of 10* and 1*1 decreases, whereas the ﬁtness of 00* and 0*1 increases. As a result, the problem becomes more diﬃcult to solve for a GA. xp
xg1
xg
y g1
yg fg1
yp f g2
Figure 3.20. A decomposition of fg
Finally, we could decompose fg into two mappings fg1 and fg2 if the number of diﬀerent positions in the two genotypes xg and y g is higher than the number of diﬀerent ones dxg ,yg > uxg − uyg (see Fig. 3.20). fg1 maps xg to xg1 , and y g to y g1 . xg1 (resp. y g1 ) should have the same number of ones as xg (resp. y g ) (uxg1 = uxg , uyg1 = uyg ), but some positions are diﬀerent in the two individuals xg1 and y g1 (dxg1 ,yg1 = uxp − uyp ). Therefore, as the number of ones stays constant, fg1 does not change the ﬁtness of the schemata (compare item 1). For xg1 and xp (resp. y g1 and y p ), the number of diﬀerent ones is the
3.3 Locality
89
same as the number of diﬀerent positions. Thus, fg2 has the same properties as discussed in the previous item and increases the ﬁtness of misleading schemata, as well as reduces the ﬁtness of the highquality schemata. We see that most modiﬁcations of a highlocality genotypephenotype mapping fg make the onemax problem more diﬃcult to solve. Only when the mapping between genotypes and phenotypes is changed that have the same number of ones in the string, is the structure of the BBs preserved, and we get the same performance as for the highlocality representation from Fig. 3.18. The above proof can be applied in the same way to a fully deceptive trap problem. Then, most of the lowlocality encodings reduce the ﬁtness of the misleading schemata, and increase the ﬁtness of the highquality schemata, which makes the problem easier. In the following subsection, we present an empirical veriﬁcation of the results. 3.3.7 Empirical Results In this subsection, we present an empirical investigation into how the problem complexity is changed for the onemax problem and the deceptive trap problem if lowlocality representations are used. We experimentally show that for highlocality representations the fully easy onemax problem remains fully easy. In contrast, most of the lowlocality representations make the onemax problem more diﬃcult to solve for GAs. The situation is vice versa for the fully diﬃcult deceptive trap where lowlocality representations always makes the problem easier to solve. For a nonredundant genotypephenotype mapping fg that is deﬁned on binary genotypes of length l, there are 2l ! diﬀerent possibilities to assign the 2l genotypes to the 2l phenotypes (Assigning the genotypes to the phenotypes can be interpreted as a permutation of 2l numbers). Any of these possibilities represents a speciﬁc mapping like for example the binary encoding or the Gray encoding. Table 3.6. 24 possibilities to assign four genotypes {ag , bg , cg , dg } to four phenotypes {ap , bp , cp , dp } xp
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
ap bp cp dp
ag bg cg dg
ag bg dg cg
ag cg bg dg
ag cg dg bg
ag dg bg cg
ag dg cg bg
bg ag cg dg
bg ag dg cg
bg cg ag dg
bg cg dg ag
bg dg ag cg
bg dg cg ag
cg ag bg dg
cg ag dg bg
cg bg ag dg
cg bg dg ag
cg dg ag bg
cg dg bg ag
dg ag bg cg
dg ag cg bg
dg bg ag cg
dg bg cg ag
dg cg ag bg
dg cg bg ag
In Table 3.6, we illustrate for l = 2 that there are 22 ! = 24 possibilities to assign the four genotypes {ag = 00, bg = 01, cg = 10, dg = 11} to the four phenotypes {ap = 00, bp = 01, cp = 10, dp = 11}. Each of the 24 genotype
90
3 Elements of a Theory of Representations
phenotype mappings represents a speciﬁc representation that assigns the ﬁtness values to the genotypes in a diﬀerent way and that results in a diﬀerent diﬃculty of the problem. In the following paragraphs, we investigate how problem diﬃculty changes for the onemax and deceptive trap problem if we use lowlocality representations. The ﬁtness function fp for the fully easy lbit onemax problem is deﬁned as l−1 fp (xp ) = xpi , (3.24) i=0
and the lbit deceptive trap function is deﬁned as: l−1 l−1 p l − 1 − i=0 xpi for xi < l, p fp (x ) = i=0 l−1 p l for i=0 xi = l.
(3.25)
If we use genotypes and phenotypes of length l the number of possible representations is 2l !. To reduce this number, we assume without loss of generality that the phenotype xp with only ones, which has ﬁtness fp = l, is always assigned to the individual xg with only ones. Then, the number of diﬀerent representations is reduced to (2l − 1)!. For example, in Figs. 3.18 and 3.19 we have 2l = 8 genotypes and 2l = 8 phenotypes. Therefore, we have 8! = 40, 340 diﬀerent representations. If we assign xg = 111 always to xp = 111 then there are only 7! = 5, 040 diﬀerent representations. Every representation represents a diﬀerent genotypephenotype mapping. Furthermore, we have seen in the previous subsection that for the used phenotypeﬁtness mapping fp (onemax and deceptive trap problem) there are some genotypephenotype mappings that do not modify the BBs and therefore do not change problem diﬃculty (both individuals have the same number of ones from item 1). These mappings have the same properties and the diﬀerent fg diﬀer only for individuals that phenotypically have the same l number of ones. There are i=1 il ! representations of that kind which we want to denote as “highlocality equivalent”. For example, in Fig. 3.18 we can change fg and assign xg = 001 to xp = 010 and xg = 010 to xp = 001. Although we use a diﬀerent representation, the assignment of the ﬁtness values to the genotypes has not changed. This eﬀect is a result of the redundancy of the used onemax and deceptive trap problem which both only consider the number of ones in the phenotype. If we use these results we can calculate how many groups of diﬀerent genotypephenotype mappings exist that result in a diﬀerent structure of the BBs. If we have (2l − 1)! diﬀerent genotypephenotype mappings and use a l lbit onemax or deceptive problem, then there are i=1 il ! mappings that do not change the structure of the BBs and are equivalent to each other. Therefore, we have (2l − 1)! l l i=1 i !
3.3 Locality
91
groups of diﬀerent genotypeﬁtness mappings. Each group consists of i=1 il ! diﬀerent genotypephenotype mappings which do not aﬀect the structure of the BBs and where only the mapping is changed between genotypes and phenotypes that have the same number of ones. When using a 3bit onemax or deceptive trap problem then there are (23 −1)! = 5040/36 = 140 groups of diﬀerent genotypeﬁtness mappings with 3!3! diﬀerent properties. Each of these 140 diﬀerent groups result in a diﬀerent structure of the genotypic and phenotypic BBs. We use for our investigation 10 concatenated onemax or deceptive trap problems of size 3. Therefore, the overall string length l = 30, and the ﬁtness of an individual is calculated as the sum over the ﬁtness of the ten onemax or deceptive trap subproblems. To illustrate how genotypephenotype mappings change the complexity of a problem we measure how many of the ten BBs (a BB is a correctly solved subproblem and consists of a sequence of l = 3 ones) a GA ﬁnds dependent on the used representation. As we use a 3 bit problem there are 5,040 diﬀerent representations and 5,040/36=140 diﬀerent representations that are equivalent to each other. Only these representations that are contained in exactly one out of the 140 equivalence groups are “highlocality equivalent” (dm = 0). An example is shown in Fig. 3.18. Due to the structure of the onemax and deceptive trap problem, there are 35 other representations which, although they have low locality, do not change the structure of the BBs and are “highlocality equivalent”. These types of encodings assign genotypes and phenotypes with the same number of ones in a diﬀerent way. Figure 3.21 presents the results of our experiments for the onemax problem. We show the distribution of the number of correctly solved subproblems at the end of a GA run when using diﬀerent representations. The plot shows results for all 5,040 diﬀerent representations. The ordinate counts the number of representations that allow a GA to correctly solve a certain number of subproblems. We used a generational GA with tournament selection without replacement of size 2, uniform crossover, no mutation and a population size of N = 15. We performed 200 runs for each representation, and each run was stopped after the population was fully converged. The average number of correctly solved subproblems measures the problem diﬃculty for the GA using one speciﬁc representation. The more subproblems which could be correctly solved, the easier the problem is for GAs. How can we interpret the data in Fig. 3.21? Every bar indicates the number of diﬀerent representations that allow a GA to correctly identify a speciﬁc number of subproblems. For example, the bar of height 95 at position 7.0 means that a GA correctly solves on average between 6.975 and 7.025 subproblems for 95 diﬀerent representations. The bar at position 4.85 means that there are 4 diﬀerent representations that allow a GA to correctly solve on average between 4.825 and 4.875 subproblems. The plot shows that by using a GA with only 15 individuals we solve independently of the used representation at least 4.2 subproblems, and we are not able to correctly solve more l
frequency (number of representations)
92
3 Elements of a Theory of Representations
160 140 120 100 80 60 40 20 0 5
6
7
8
9 frequency (number of representations)
20
number of correct subproblems
all possible genotypephenotype mappings 32 highlocality representations
15
10
5
0
8.6
8.7
8.8
8.9
9
9.1
9.2
number of correct subproblems
Figure 3.21. Experimental results of the frequency of the number of correct subproblems at the end of a run for all possible encodings of a 3bit onemax problem. We present results for 10 concatenated 3bit problems. The optimal solution is always 111 so there are (23 − 1)! = 5, 040 diﬀerent representations. We use a GA with tournament selection without replacement, uniform crossover and a population size of N = 15. We perform 200 runs for every possible encoding. Only for these 36 representations that are equivalent to the highlocality representation, the fully easy onemax problem remains fully easy. All other encodings have low locality and make the problem more diﬃcult to solve for GAs.
than 9 out of ten subproblems. Furthermore, it is surprising that we have no normal distribution over the number of correct subproblems but that there are clusters. For example, there are many representations that allow a GA to solve on average between 5.8 and 6.3 subproblems but there are only a few representations that allow a GA to correctly solve on average between 6.5 and 6.8 subproblems. The reason is, that there are only 140 diﬀerent equivalence groups of representations. Although we have 5,040 diﬀerent genotypephenotype mappings, there are only 140 diﬀerent levels of problem complexity possible. The observed clusters are probably a result of these small number of diﬀerent levels of problem complexity. It is more interesting to ask how the highlocality representation from Fig. 3.18 performs? And how the performance is of the other 35 genotype
3.3 Locality
93
phenotype mappings that, although they have low locality and do not preserve the distances, are equivalent to the highlocality representation and result in the same genotypeﬁtness mapping? The small plot in Fig. 3.21 answers these questions. The bold line shows the performance of a GA using these 36 diﬀerent “highlocalityequivalent” representations. The use of these representations results in the highest GA performance. For example, there are 7 diﬀerent representations that allow a GA to correctly solve between 8.825 and 8.875 subproblems. All 7 encodings belong to the group of 36 “highlocalityequivalent” encodings. Furthermore, we see that all representations that allow a GA to correctly solve on average more than 8.75 subproblems belong to this group. These 36 encodings result in the highest proportion of correct subproblems. For these encodings the onemax problem remains fully easy and the size of the genotypic and phenotypic BBs stays kg = kp = 1. Changing the representation, that means assigning the elements of Φg in a diﬀerent way to the elements in Φp , always results in a lowlocality representation. If the genotypeﬁtness mapping is not “highlocality equivalent” then the BBcomplexity increases and the problem becomes more diﬃcult to solve. A GA has more diﬃculties in solving the problem, and the proportion of correctly solved subproblems is lower. The plot illustrates nicely that the used representation can change the complexity of the onemax problem dramatically. However, only encodings that are equivalent to the highlocality encoding allow a GA to eﬃciently solve the fully easy onemax problem. In Fig. 3.22, we present results for 10 concatenated instances of a 3bit deceptive trap. The GA parameters chosen are the same as for the onemax problem. The plots show that, as expected, the GA performs worst for “highlocality equivalent” representations. All other representations make the problem easier to solve for GAs. We have empirically shown that only highlocality representations guarantee that fully easy problems remain fully easy. Highlocality representations preserve BBcomplexity and are a good choice if we want GAs to reliably solve problems of bounded complexity. As soon as a representation has low locality and does not preserve BBcomplexity, some of the easy problems become more diﬃcult and therefore can no longer be solved by the GA. Indeed, some of the diﬃcult problems could become solvable by using lowlocality representations, but in general we are not interested in solving these types of problems. 3.3.8 Conclusions This section presented the third and ﬁnal element of a theory of representations. We investigated how the locality of a representation inﬂuences the performance of GEAs. The locality of a representation describes how well genotypic neighbors correspond to phenotypic neighbors. It is high if genotypic neighbors correspond to phenotypic neighbors. The results show that highlocality representations preserve the diﬃculty of a problem and phenotypically easy problems also remain genotypically easy. Using lowlocality
3 Elements of a Theory of Representations
160
frequency (number of representations)
94
140 120 100 80 60 40 20 0
frequency (number of representations)
4.5 20
5
5.5
6
6.5
7
7.5
8
8.5
number of correct subproblems
all possible genotpyephenotype mappings 32 highlocality representations
15
10
5
0 4.5
4.6
4.7
4.8
4.9
5
5.1
5.2
number of correct subproblems
Figure 3.22. Experimental results of the frequency of the number of correctly solved subproblems at the end of the run for all possible encodings of a 3bit deceptivemax problem. We present results for 10 concatenated 3bit problems. The optimal solution is always 111 so there are (23 − 1)! = 5, 040 diﬀerent possible encodings. The GA uses tournament selection without replacement, uniform crossover and a population size of N = 15. We perform 200 runs for every possible encoding. Only for these 36 representations that are equivalent to the highlocality encoding does the fully diﬃcult deceptive trap remain fully diﬃcult. For all other representations, the BBcomplexity is reduced and the problem becomes easier to solve for GAs.
representations is equivalent to randomizing the search process. Therefore, lowlocality representations change problem diﬃculty and make easy problems more diﬃcult and deceptive problems more easy to solve. In general, we want GEAs to be able to solve a class of problems of bounded complexity fast and reliably. However, the results have shown that the use of lowlocality representations in general changes problem diﬃculty and can only increase problem diﬃculty for fully easy problems. Therefore, easy problems that are solvable using a highlocality representation could become unsolvable when using a lowlocality representation that modiﬁes BBcomplexity. To guarantee that a GA can reliably solve problems of bounded complexity it is designed for, we recommend the use of highlocality representations.
3.4 Summary and Conclusions
95
This section nicely illustrated that representations can dramatically change the complexity of a problem. The presented work has shown that even fully diﬃcult problems can be solved easily if a proper, lowlocality, representation is used. However, using the same representation for a fully easy onemax problem can make the problem fully diﬃcult and unsolvable. Therefore, the use of lowlocality representations could be advantageous if we know that the problem is deceptive. But in general, users do not have this information and therefore, they should not use these types of representations.
3.4 Summary and Conclusions In Sect. 3.1 we described, analyzed, and modeled the eﬀect of redundant representations on the performance of GEAs. We distinguished between synonymously and nonsynonymously redundant representations and illustrated that nonsynonymous redundancy does not allow genetic operators to work properly and therefore reduces the eﬃciency of evolutionary search. For synonymously redundant representations, GEA performance depends on the change of the initial supply. Based on this observation models were developed that give the necessary population size for solving a problem, and the number of generations as O(2kr /r), where kr is the order of redundancy and r is the number of genotypic BBs that represent the optimal phenotypic BB. As a result, uniformly redundant representations do not change the performance of GAs. Only by increasing r, which means overrepresenting the optimal solution, does GA performance increase. Therefore, nonuniformly redundant representations can only be used advantageously if there exists apriori some information about the optimal solution. This was followed in Sect. 3.2 by an investigation into how the scaling of an encoding inﬂuences the performance of GEAs. We extended previous work (Rudnick 1992; Thierens 1995; Thierens et al. 1998; Harik et al. 1997) and formulated a more exact convergence model considering genetic drift for exponentially scaled representations. Representations are exponentially scaled if the contribution of the genotypic alleles to the construction of the phenotypic alleles is exponentially diﬀerent. Using the developed population sizing model, we were able to more accurately predict the behavior of GEAs using exponentially scaled representations. Finally, we presented the third and ﬁnal element of a theory of representations, namely the inﬂuence of locality on problem complexity. In Sect. 3.3, we showed that highlocality representations, which preserve the neighborhood structure when mapping genotypes to phenotypes, do not modify the diﬃculty of a problem. When using lowlocality representations, on average problem diﬃculty changes. On average, fully easy problems become more difﬁcult, and deceptive problems easier. We have discussed why representations that keep easy problems easy and make deceptive problems easier are nice
96
3 Elements of a Theory of Representations
to have, but not possible without having an exact knowledge about the optimization problem a priori. In this chapter, we identiﬁed three important elements towards a general theory of representations. We identiﬁed redundancy, scaling, and locality/distance distortion as having a major inﬂuence on the performance of GEAs. We were able to show that synonymously redundant encodings do not modify the performance of a GEA as long as the representation is uniformly redundant. Our investigation into nonuniformly scaled representations has shown that these types of encodings prolong the search process and increase the problems of GEAs with genetic drift. Finally, we have seen that lowlocality representations do not preserve BBcomplexity in general and make phenotypically easy problems more diﬃcult. Therefore, to make sure that GAs are able to reliably solve easy problems and problems of bounded complexity, the use of highlocality representations is recommended. Even by only presenting some basic elements of a general theory of representations we are able to analyze and predict the behavior and performance of GEAs using existing representations signiﬁcantly better. The presented theory gives us a deeper understanding on how existing representations inﬂuence the performance of GEAs, as well as allows us to design new representations in a more theoryguided way. By using the presented theory, on the one hand we can develop general and robust representations that can be applied to problems of unknown complexity, and on the other hand problemspeciﬁc representations which could fail for some problems, but perform well for a speciﬁc problem. Although the provided elements of representation theory already allow a guided design and analysis of representations, further research is still necessary to develop a general representation theory. Especially, the relationship between the presented elements of theory of representations should be investigated more deeply. We believe that as we are able to easily separate the eﬀects of redundant and exponentially scaled representations that there is not much interconnection and overlapping between these two elements of theory. However, for locality and its inﬂuence on BBcomplexity, the situation is different. We have seen that the modiﬁcation of problem complexity is strongly inﬂuenced by redundancy or scaling. Therefore, further research is necessary to identify the exact relations between the presented elements of theory. Finally, we want to encourage researchers to do more basic research towards the development of a general theory of representations. We believe that we provided some important parts, but there is still a long way to go. However, the path is worth following, as a general theory of representations would allow us to unleash the full power of genetic and evolutionary search and help us to solve problems fast, accurately and reliably.
4 TimeQuality Framework for a TheoryBased Analysis and Design of Representations
Over the last decades, researchers gained more and more knowledge about the principles of genetic and evolutionary algorithms (GEAs) and were able to formulate a theory describing the behavior of GEAs more precisely (B¨ ack et al. 1997; Vose 1999; Goldberg 2002; Reeves and Rowe 2003). The existing elements of GEA theory explain quite accurately the inﬂuence of many important GEA parameters, as well as selection, recombination, or mutation methods on the performance of GEAs. By using the existing GEA theory, straight forward design and the development of new, competent GAs (Goldberg et al. 1993; Harik and Goldberg 1996; Pelikan et al. 1999; Pelikan et al. 1999) became possible. However, concerning representations for GEAs, a framework which describes the inﬂuence of representations on the performance of GEAs is still missing, although it is well known that the used representation has a strong inﬂuence on GEA performance. Such a framework could help us to develop new representations in a more theoryguided manner and would be an important step towards a general theory of representations for GEAs. The purpose of this chapter is to develop a framework for a theorybased analysis and design of representations for GEAs based on the elements of theory we presented in the previous chapter. The framework should allow us to model and predict the inﬂuence of diﬀerent types of representations on the performance of genetic and evolutionary search. It should describe how redundancy, scaling, and locality of a representation inﬂuence the time to convergence and the expected solution quality. By using the framework, we would be able to theoretically compare the eﬃciency of diﬀerent representations, as well as to design new representations in a theoryguided way. The chapter starts with a brief overview of the determinants – time and quality – of GEA performance. In Sect. 4.2, the elements of the framework, namely redundancy, scaling, and locality/distance distortion, are presented. We review their inﬂuence on representations, formulate how the three properties of representations can be measured, and describe how genetic and evolutionary search is aﬀected. In Sect. 4.3, the framework itself is described. We formulate how the probability of error α and the convergence time tconv
98
4 TimeQuality Framework
depend on the diﬀerent elements of the framework. Because we are not yet able to consider the eﬀect of scaled representations in general, the section is split up into two parts concerning uniformly and nonuniformly scaled representations. For both types of scaled representations we describe how the solution quality and the time to convergence depends on the redundancy and locality of an encoding. This is followed in Sect. 4.4 by some implications of the framework on the design of representations. We show how the use of representations with diﬀerent properties aﬀects the supply of BBs, the dynamics of genetic search, or the size of BBs. The chapter ends in Sect. 4.5 with a summary and concluding remarks.
4.1 Solution Quality and Time to Convergence The following section brieﬂy reviews determinants for GEA performance. It focuses on solution quality and time to convergence. For comparing the eﬃciency of diﬀerent GEAs using diﬀerent types of representations, a measurement of GEA performance is necessary. Widely used determinants for GEA performance are the solution quality and the time to convergence. In general, the solution quality and convergence time depend on the used genetic operators, the GEA parameters, the used representation, and the optimization problem itself. The solution quality of GEAs can be measured by the probability Pn of GEA success. GEA success means that the optimal solution is found by the GEA. When using the more common probability of GEA failure α, GEA success is deﬁned as Pn = 1 − α. Earlier work by Harik et al. (1997) has shown that when using selectorecombinative GAs, the probability of error α = 1 − Pn goes with O(e−N ). With decreasing α the population size N increases exponentially. Therefore, instead of using α, the population size N that is necessary for solving a problem can also be used for comparing GEA performance. Measuring GEA performance becomes more complicated if the best solution is not known a priori. Then, Pn cannot be calculated and the best ﬁtness at the end of a GEA run can be used. It corresponds to the probability of GEA success Pn , and is determined by the used population size N . The time to convergence tconv describes how many generations selectorecombinative GEAs need to converge completely. A population is converged if there is no genetic diversity in the population after tconv generations and all individuals in the population represent the same phenotype. It was shown (Thierens and Goldberg 1993; Miller and Goldberg 1996b; Miller and Goldberg 1996a) that the convergence time mainly depends on the length of the string l and the used selection scheme. As soon as the population size N is large enough to solve the problem reliably, the convergence time tconv does not depend on N any more. To compare the overall performance of diﬀerent GEAs the number of ﬁtness evaluations nf can be used. For a given solution quality Pn 0 the total
4.2 Elements of the Framework
99
number of ﬁtness calls can be calculated as nf = N × tconv .
4.2 Elements of the Framework We focus in this section on the properties of representations that inﬂuence GEA performance and describe the elements of representation theory that are used in the framework. We review the elements, describe how we can measure them, and illustrate their eﬀects on GEAs. In Chap. 3, we presented redundancy, scaling, and locality/distance distortion as relevant properties of representations. Although we believe that these three elements are some of the most important elements of the timequality framework, there could still be others. Finding and describing them is left to further research. The section consists of three subsections which discuss the single elements, namely redundancy, scaling, and locality. In each subsection, we brieﬂy describe what we mean, illustrate how redundancy, scaling, or locality can be measured, and ﬁnally describe how GEAs are aﬀected. 4.2.1 Redundancy Section 3.1 shows that the use of redundant encodings aﬀects the performance of GEAs. In the context of representations, redundancy means that on average one phenotype is represented by more than one genotype. Therefore, Φg  > Φp  when using redundant representations. Consequently, a representation is not redundant if Φg  = Φp . Then, the number of genotypes is the same as the number of phenotypes. Because we assume that every phenotype must be represented by at least one genotype, the size of the genotypic space can not be smaller than the size of the phenotypic space. To model the eﬀects of redundancy, we distinguish between synonymous and nonsynonymous redundancy. When focusing on synonymous redundancy, the order kr of redundancy is introduced (see Sect. 3.1.2). It measures the amount of redundant information in the encoding (in bit). There are kr bits and 2kr diﬀerent possibilities (individuals) to encode 1 Bit of information content (2 possibilities). Using no redundancy in an encoding results in kr = 1. Furthermore, r is deﬁned as the number of genotypic BBs of size kkr that represent the optimal phenotypic BB of size k. Therefore, for nonredundant encodings kr = 1 and r = 1. We know from (3.2) that for redundant encodings r ∈ {1, 2, . . . , 2kkr − 2k + 1}. We assume that we use binary strings (χ = 2). For binary genotypes, there are 2k diﬀerent phenotypes and they are represented by 2kkr diﬀerent genotypes. For nonbinary genotypes we refer to Sects. 6.4.4 and 6.5.5. Using uniformly redundant representations results in
100
4 TimeQuality Framework
runif orm = 2k(kr −1) and x0 /N = r/2kkr = 1/2k . x0 denotes the the initial supply of BBs. Therefore, a representation is nonuniformly redundant if r/2kkr = 1/2k . For r/2kkr > 1/2k the optimal solution is overrepresented, and for r/2kkr < 1/2k the optimum is underrepresented. Our investigation into the eﬀects of redundancy on GEAs in Sect. 3.1.3 has shown that the supply of BBs in the initial population is inﬂuenced by the use of nonuniformly redundant encodings. If the optimal solution is overrepresented by the used synonymously redundant representation the performance of GEAs is increased, that means lower run duration tconv and lower probability of error α. The situation is reversed if the optimal solution is underrepresented and tconv and α increases. 4.2.2 Scaling In Sect. 3.2, we discussed the eﬀects of exponentially scaled representations on the performance of GEAs. Representations are uniformly scaled, if all genotypic alleles have the same contribution to the construction of the phenotypic alleles. Therefore, GEAs using uniformly scaled representations solve all alleles implicitly in parallel. In contrast, a representation is nonuniformly scaled if some genotypic alleles have a higher contribution to the construction of the phenotypic alleles than others. As a result, domino convergence occurs and the alleles are solved sequentially according to their salience. The most salient alleles are solved ﬁrst, whereas the lowest salient alleles are solved last. To more formally describe the scaling of a representation, a measurement of how strong a representation is scaled is necessary. Therefore, we describe by the order of scaling s the diﬀerence in salience for the diﬀerent alleles. When using a binary genotype of length l and ordering the alleles according to their contribution to the ﬁtness in ascending order, we deﬁne the order of scaling s ∈ [1, ∞[ as l−1 1 xci+1 , s= l − 1 i=1 xci where xci denotes the contribution of the ith most salient allele to the phenotype, and xci+1 ≥ xci , for i ∈ {1, . . . , l − 1}. Therefore, xc1 denotes the contribution of the lowest salient allele and xcl denotes the contribution of the most salient allele. When using uniformly redundant encodings the contribution of all alleles is the same which results in xci = const, for i ∈ {1, . . . , l}. Therefore, s = 1 for uniformly scaled representations. When using exponentially scaled representations the order of scaling s > 1 is constant; for binary encoded strings, we get s = 2. The order of scaling s inﬂuences the dynamics of genetic search. With increasing s, the alleles are solved more and more sequentially. Rudnick (1992) proposed the use of a convergence window for modeling the dynamic solving
4.2 Elements of the Framework
101
process when using nonuniformly scaled representations. The convergence window is a set of contiguous alleles that are not yet fully converged but have started to converge. The size λc ∈ {1, . . . , l} of the convergence window is equal to l for uniformly scaled encodings and equal to one for s → ∞. λc = 1 results in strictly sequential solving of the alleles, whereas for λc = l all alleles are solved in parallel. With increasing order of scaling s, the size λc of the convergence window is reduced. Earlier work (Thierens et al. 1998; Lobo et al. 2000) shows in correspondence to the results of Sect. 3.2.4, that the assumption of a convergence window of size λc = 1 results for s = 2 (exponentially scaled representations) in a good approximation of the dynamics of GEA search. However, for a more general theory of scaled representations a more detailed analysis of the interdependencies between λc and s is necessary. 4.2.3 Locality In Sect. 3.3, we saw that when using a representation, the neighborhood structure can be diﬀerent for the genotypes and phenotypes. In this case, the size and length of the genotypic and phenotypic building blocks are diﬀerent. We have illustrated that highlocality representations guarantee that the complexity of a problem is preserved and easy problems remain easy. Therefore, high locality is necessary for eﬃcient mutationbased search, low distance distortion is necessary for eﬃcient crossoverbased search, and high locality is a necessary condition for low distance distortion. If a representation has low locality, some genotypic neighbors are not phenotypic neighbors, and the diﬃculty of the optimization problem is changed. Our investigation into the inﬂuence of locality on problem diﬃculty has shown that only high locality guarantees that the problem diﬃculty remains unchanged. We deﬁned in (3.23) the locality dm of a representation as dpx,y − dpmin , dm = g dg x,y =dmin
where dpx,y is the phenotypic distance between the phenotypes x and y, dgx,y is the genotypic distance between the corresponding genotypes, and dpmin , respective dgmin is the minimum distance between two (neighboring) phenotypes, and respectively genotypes. Without loss of generality we want to assume that dgmin = dpmin . For dm = 0 the genotypic neighbors correspond to the phenotypic neighbors, the encoding has perfect locality, and the complexity of the phenotypic problem is not modiﬁed. We have seen in Sect. 3.3 that by using lowlocality representations (dm = 0), the complexity of the problem which can be measured by the size k of the BBs, can be changed. Distinguishing between the size kp of BBs in the phenotypic space, and the size kg of BBs in the genotypic space allows us to model the inﬂuence of dm on the performance of GEAs more exactly. Section
102
4 TimeQuality Framework
3.3 has shown that highlocality representations preserve problem diﬃculty and the problem has the same genotypic as phenotypic complexity: kg = kp , if dm = 0. The situation becomes more complicated if dm = 0. As soon as neighboring genotypes do not correspond to neighboring phenotypes, the complexity of the BBs is changed and kg = kp . Every phenotypic problem with complexity kp can be transformed by the use of a lowlocality representation into a genotypic problem with complexity kg ∈ [1, l]. For every problem there is always a representation that results in a fully easy problem, kg = 1, as well as a representation that results in a fully diﬃcult, misleading trap with kg = l (Liepins and Vose 1990). Therefore, when using lowlocality representations (dm = 0), the genotypic size of BBs kg depends not only on the genotypephenotype mapping fg , but also on the speciﬁc structure of the phenotypic problem fp . Therefore, we get kp , for dm = 0, (4.1) kg = kg (fg , fp ) , with 1 ≤ kg ≤ l , for dm = 0, where l denotes the length of the binary string. We want to illustrate with a small example why kg does not only depend on fg and kp , but also on fp . Section 3.3.1 has discussed that every fully deceptive trap with kp = l (3.25) can be transformed into a fully easy problem with kg = 1 by a linear transformation. For this purpose we want to deﬁne the genotypephenotype mapping fg as l − up − 1 if up = l, ug = (4.2) l if up = l, where l is the length of the string, ug is the number of ones in the genotype and up is the number of ones in the phenotype. This encoding has low locality (dm = 0). By the lowlocality encoding fg the phenotypically fully deceptive trap with kp = 7 becomes fully easy (see Fig. 4.1(a)). However, using the same mapping fg for a diﬀerent phenotypic problem of the same complexity (Fig. 4.1(b)) does not signiﬁcantly reduce kg . As the size kp = kg = 7 of BBs remains unchanged, and only the position of the optimal solution is changed, the problem is still fully deceptive (kg = l) after applying fg . We see that for predicting kg the knowledge of kp is not enough when using lowlocality representations. It is necessary to know fp as well as fg to predict the inﬂuence of lowlocality representations on the diﬃculty of the problem.
4.3 The Framework This section provides the timequality framework modeling the inﬂuence of representations on the performance of genetic and evolutionary algorithms.
4.3 The Framework fp(up)
f=fp(fg(ug))
7 6 5 4 3 2 1
fg
1 2 3 4 5 6 7
103
7 6 5 4 3 2 1
up
1 2 3 4 5 6 7
ug
(a) The genotypephenotype mapping fg deﬁned in (4.2) makes the fully deceptive phenotypic trap (left) fully easy (right). fp(up)
f=fp(fg (ug))
7 6 5 4 3 2 1
fg
1 2 3 4 5 6 7
up
7 6 5 4 3 2 1 1 2 3 4 5 6 7
ug
(b) Because the optimal solution is located at a diﬀerent position (up = 0) the genotypephenotype mapping fg from above does not reduce the complexity of the BBs signiﬁcantly. Figure 4.1. We show how a representation fg (4.2) modiﬁes the complexity of diﬀerent phenotypic fully deceptive problems. If the optimal solution is located at u = 7 the problem becomes fully easy (kg = 1). However, if the optimal solution is located at u = 0 the complexity of the problem remains approximately unchanged. We see that for designing a representation fg that makes diﬃcult problems easier, the structure of the phenotypic optimization problem fp must be known. Therefore, theoryguided design of lowlocality representations that reduce BBcomplexity is diﬃcult and problemspeciﬁc.
The framework allows us to theoretically predict and compare the performance of GEAs using diﬀerent types of representations. Therefore, thorough analysis and theoryguided design of representations becomes possible. Although the presented framework is not yet complete and there are still some gaps, rough approximations, unclear interdependencies, and also more as yet unknown elements, we believe that the framework is an important step towards a more general theory of representations. The framework itself is based on the characteristics of the used encoding we introduced in Sect. 4.2. There, we have seen that the redundancy of a representation can be described by the order of redundancy kr and the number of copies r which are given to the optimal solution. Furthermore, the modiﬁcation of BBcomplexity is determined by the locality dm , which measures how well phenotypic neighbors correspond to the genotypic neighbors. Finally,
104
4 TimeQuality Framework
the scaling of a representation can be described by using the order of scaling s. Currently there is no general model available for the inﬂuence of s on the performance of GEAs. Therefore, we want to focus in this framework on uniformly scaled representations (s = 1) and exponentially scaled representations with s ≥ 2. The structure of the section follows the still missing general model of the inﬂuence of scaling. Therefore, the section is split into two parts. In Sect. 4.3.1, we present the part of the framework for uniformly scaled representations and in Sect. 4.3.2 we focus on exponentially scaled representations. 4.3.1 Uniformly Scaled Representations We present the part of the framework that describes the inﬂuence of representations on GEA performance if the representations are uniformly scaled. We describe how the probability of error α and the time to convergence tconv depend on redundancy and locality. Based on the work from Harik et al. (1997) we get for the probability of error 1 − (q/p)x0 , α=1− 1 − (q/p)N where x0 is the expected number of copies of the best BB in the randomly initialized population, q = 1−p is the probability of making the wrong decision between two competing BBs, and N is the population size. From (3.6) we know that r x0 = N kkr , 2 where k is the phenotypic size of BBs, r is the number of genotypic BBs of length kkr that represent the best phenotypic BB, and kr is the order of redundancy. After some approximations (see Sect. 3.1.4) we ﬁnally model in (3.9) the inﬂuence of redundant encodings on the population size N as √ σBB πm 2kr k−1 ln(α) , (4.3) N =− r d where m = m − 1 with m is the number of BBs, d is the signal diﬀerence, 2 is the variance of the BBs. The probability α of GA failure can be and σBB calculated as:
N dr √ (4.4) α = exp − 2kr k−1 σBB πm We have described in Sect. 4.2.3 that the problem diﬃculty measured by the size of BBs k is modiﬁed by the locality dm of the representation. In (4.1), the genotypic size kg of the BBs is calculated as kp , if dm = 0, kg = kg (fg , fp ) ,where 1 ≤ kg ≤ l , if dm = 0,
4.3 The Framework
105
where fg is the genotypephenotype mapping (the used representation), and fp is the optimization problem with the size kp of the phenotypic BBs. Substituting kg into 4.4 we get for uniformly scaled representations:
N dr √ α = exp − . (4.5) 2kr kg −1 σBB πm . We see that using reThe probability of error α goes with O exp 2k−r r kg dundant representations (kr > 1) without increasing r has the same inﬂuence on GEA performance as increasing the size of BBs kg . From Sect. 3.1.5 we get for the time to convergence for a uniformly scaled representation (3.10) √ x l π 0 − arcsin 2 − 1 , (4.6) tconv = I 2 N where l is the length of the phenotypes, and I is the selection intensity. Substituting x0 from (3.6) into (4.6) yields √ r l π − arcsin( kr k−1 − 1) . tconv = I 2 2 When considering the eﬀect of locality (4.1) we ﬁnally get for the time to convergence √ r l π − arcsin( kr kg −1 − 1) . (4.7) tconv = I 2 2 tconv increases with larger kg and decreasing r/2kr . With 0 < 2krrkg < 1 we can calculate upper and lower bounds for the expected time to convergence as √ l π 0 < tconv < I If r/2kr kg ≈ 1 most of the randomly created genotypic individuals represent the phenotypic optimum. Therefore, GEAs converge very fast and tconv → 0. If either kg is a large number or r/2kr is small then there is only a small fraction of optimal BB in the initial population and GEAs need many generations to converge. 4.3.2 Exponentially Scaled Representations We describe the inﬂuence of redundancy and locality on the performance of GEAs if the representations are exponentially scaled. In contrast to the previous subsection where the size of the convergence window λc is equal to the string length and all alleles are solved in parallel, we assume that the alleles are solved strictly in serial and λc = 1.
106
4 TimeQuality Framework
As illustrated in Sects. 3.2 and 4.2.2 we can use the domino convergence model for estimating the performance of GEAs using exponentially scaled representations. We assume that the alleles are solved strictly in serial and there are no interdependencies between the ls alleles in an exponentially scaled BB. However, it is possible to concatenate m exponentially scaled BBs of length ls . When using exponentially scaled representations the maximum size of BBs is k = 1. All schemata of order k = 1 that contain the best solution have higher ﬁtness than their competitors. Therefore, it makes no sense to consider the eﬀect of locality on GEA performance when using exponentially scaled representations. Section 4.2.3 has shown that low locality modiﬁes the size of BBs k and results in interdependencies between the alleles. However, if kg > 1, the domino convergence model can not be used any more, because we can then not assume that the alleles are still solved sequentially. Therefore, we assume in the following that k = 1 and the representation does not modify the size of BBs when mapping the phenotypes onto the genotypes. When using redundant representations we know from (3.6) that r x0 = kkr , N 2 where x0 is the expected number of copies of the best BBs in the initial population, N is the population size, k is the size of BBs, m = m − 1 with m is the number of BBs, kr is the order of redundancy, and r is the number of genotypic BBs of size kkr that represent the best phenotypic BB. As we have seen in Sect. 3.2.2, the probability p of making the right choice between a single sample of each BB remains constant for the ls bits in the exponentially scaled BB if we assume that all alleles which are not yet touched by the solving process remain in their initial state. Substituting x0 /N into (3.13), we get ⎞ ⎛ p = N ⎝
2 2k r 1 − r
1 r 2k r
( 43 m − 1)
⎠·
(4.8)
As illustrated above, k = 1 and there are m competing BBs with ls exponentially scaled alleles. Furthermore, with x0 = 2Nkrr we get from (3.15) for the probability of error x
Nr
(1/p − 1) 0 − (1/p − 1)N (1/p − 1) 2kr − (1/p − 1)N α= = · 1 − (1/p − 1)N 1 − (1/p − 1)N
(4.9)
We want to approximate (4.9) in analogy to Sect. 3.1.4. If we assume that x0 is small we get from (4.9)
x 1−p 0 α≈ · p When using the ﬁrst two terms of the power series expansion of the normal distribution for approximating (4.8) we get
4.3 The Framework
α ≈ exp x0 ln
1−x 1+x
107
,
where x = 1/ π xN0 (1 − xN0 )( 43 m − 1). Because x is a small number we can assume that ln(1 − x) ≈ −x and ln(1 + x) ≈ x. Using these approximations we get ⎞ ⎛ 2 ⎠· α ≈ exp ⎝−x0 x0 x0 4 π N (1 − N )( 3 m − 1) If we approximate
− x0 /N ) by x0 /N we get for the probability of error ⎞ ⎛ Nr ⎠· α ≈ exp ⎝− k −1 2 r π 2krr ( 43 m − 1)
x0 N (1
Simplifying this equation yields ﬁnally ⎛ α ≈ exp ⎝−
√ 2n r
2kr π( 43 m − 1)
⎞ ⎠·
(4.10)
Using this rough approximation we appreciate that α is reduced with increasing r/2kr and N . α is also reduced with a smaller number m of competing BBs. The reader should notice that α does not depend on the length ls of an exponentially scaled BB, as we assumed that the alleles remain in their initial state as long as they are not reached by the search window. We have seen in Sect. 3.2 that genetic drift reduces the performance of GEAs when using exponentially scaled representations. Genetic drift can be considered by either the approximated drift model (3.21) or the staircase drift model (3.22). By substituting the probability of error α either from (4.9) or from (4.10) into either (3.21) or (3.22) we get the average percentage of incorrect alleles α ¯ . For example, we can calculate the overall percentage α ¯ of incorrect alleles using the approximated drift model as:
ls −1 π√ 1 π√ 1 πm) α + s (λ πm) , 1 − s (λ α ¯= ls 2 2 2
(4.11)
λ=0
with
0 s (t) = 1−
3 2
for t < −N ln(2/3), exp(−t/N ) for t > −N ln(2/3).
With increasing ls , more and more of the lower salient alleles are ﬁxed randomly and α ¯ is reduced. The time to convergence for the m alleles of the same salience can be calculated by using (3.10) as
108
4 TimeQuality Framework
√ tconv =
r m π − arcsin kr −1 − 1 · I 2 2
(4.12)
As before we assume that k = 1. After m alleles of the same salience are converged the GEAs tries to solve the next m alleles with the next lower salience. Because each of the m BBs consists of ls alleles with diﬀerent salience and the solving process is strictly serial, we get for the overall time to convergence √ r m π − arcsin kr −1 − 1 , (4.13) tconv = ls I 2 2 The time to convergence increases linearly with the length of an exponentially scaled BB ls . With larger r/2kr the time to convergence is reduced.
4.4 Implications for the Design of Representations The purpose of this section is to describe some of the important implications of the framework on the behavior of GEAs. We show how the inﬂuence of diﬀerent types of representations on the performance of GEAs can be described by using the presented framework. Based on the framework, we see that representations that overrepresent a speciﬁc solution can result in high GEA performance, but are not robust concerning the location of the optimal solution. When using exponentially scaled representations, the framework tells us that there is a tradeoﬀ between the accuracy of the solution quality and convergence time. Because lowlocality representations aﬀect the size of BBs, the behavior of GEAs using lowlocality representations is diﬃcult to predict. The section starts by illustrating the eﬀects of nonuniformly redundant representations. We have seen in Sects. 3.1 and 4.2.1 that redundancy affects the supply of BBs in the initial population. Therefore, representations that overrepresent a speciﬁc solution result in high GEA performance but are not robust. Section 4.4.2 illustrates that the scaling of a representation inﬂuences the dynamics of genetic search. GEAs using exponentially scaled representations deliver rough approximations of the optimal solution after a few generations, but the overall time to convergence is increased in comparison to uniformly scaled representations. Finally, we show in Sect. 4.4.3 the eﬀects of lowlocality representations. If dm = 0, the genotypic problem complexity depends on fg and fp and easy problems fp on average become more diﬃcult to solve for GEAs. Therefore, when using lowlocality representations the performance of GEAs for a speciﬁc problem is diﬃcult to predict but on average lowlocality representations make easy problems more diﬃcult. 4.4.1 Uniformly Redundant Representations Are Robust Section 3.1 has illustrated the eﬀects of redundancy on the performance of GEAs. The results have shown that the quality of the solutions and the time
4.4 Implications for the Design of Representations
109
to ﬁnd them can be increased if we focus the genetic search on some speciﬁc areas of the search space. We described in Sect. 4.2.2 the inﬂuence of synonymous redundancy by r denoting the number of genotypic BBs that represent the optimal phenotypic BB and kr denoting the order of redundancy. Therefore, r/2kr can be used for characterizing redundancy in an encoding. Our framework in the previous section tells us how the solution quality (α and tconv ) depends on r/2kr . For uniform redundant representations (see (4.5) and (4.7))
N dr √ α = exp − , 2kr kg −1 σBB πm and √
2r l π − arcsin( kr kg − 1) . tconv = I 2 2 When neglecting the eﬀect of genetic drift we get for exponentially scaled representations (see (4.10) and (4.13)) ⎞ ⎛ √ r 2N ⎠, α ≈ exp ⎝− 2kr π( 43 m − 1) √ r m π − arcsin kr −1 − 1 I 2 2 kr We see that α goes for uniformly scaled representations with O(exp(−r/2 )) k r and for exponentially scaled representations with O(exp(− r/2 )). The time to convergence tconv is reduced for both types of representations with increasing r/2kr . Therefore, GEA performance increases with larger r/2kr . As a result designing eﬃcient representations seems to be quite an easy task. Initially it appears that we simply have to increase r/2kr and are rewarded with high performing GEAs. Therefore, we have to investigate if there are any problems associated with increasing r/2kr . When using synonymously redundant representations, the order of redundancy kr does not depend on the structure of the optimal solution. However, r depends by deﬁnition on the structure of the optimal solution. r measures how many genotypic BBs of size kkr represent the optimal phenotypic BB of size k. On average ravg = 2k(kr −1) genotypic BBs represent one of the 2k phenotypic BBs. Therefore, if r > ravg for some phenotypic individuals, there must also be some individuals with r < ravg . That means if some individuals are overrepresented by a speciﬁc representation there must be others which are underrepresented. We have learned from the framework that solution quality increases with increasing r/2kr . If we have uniform redundancy (r = 2k(kr −1) ), GEAs perform the same as without redundancy. For uniformly redundant representations, the performance of GEAs is independent of the location of the optimal and
tconv = ls
110
4 TimeQuality Framework
solution. If a speciﬁc phenotype xpr is overrepresented with r > 2k(kr −1) , GEAs searching for optimal solutions that are similar to xr perform better. However, when using this representation and searching for solutions that have a large distance to xpr , GEAs perform worse. The situation is vice versa if xpr is underrepresented. We see that by increasing r/2kr , we reduce the robustness of the representation. A representation is denoted to be robust if the performance of a GA is independent of the location of the optimal solution in the search space. We illustrate this behavior of redundant encodings in Fig. 4.2. The Figure shows how the performance of GEAs depends on the over or underrepresentation of the phenotype xpr .
performance r 2 kk r r 2 kk r r 2 kk r
xpr
Φp
1 2k
Figure 4.2. We show how the performance of GEAs using redundant representations depends on the location of a speciﬁc individual xpr in the search space Φp . r determines the number of genotypes that represent a speciﬁc phenotype xpr . The performance of GEAs is independent of r if all phenotypes are uniformly represented (r = 2k(kr −1) for all phenotypes). If xpr is overrepresented (r > 2k(kr −1) ) GEAs perform better when searching for individuals similar to xpr , and worse for individuals with a larger distance to xpr . If xpr is underrepresented the situation is reversed.
We see that when designing representations, redundancy is helpful if the number of copies r that are given to the optimal solution xpr is above average. However, to systematically increase the number of copies of the optimal solution, it is necessary to know where the optimal solution is located in the search space Φp . Otherwise, if we do not know where the optimal solution can be found, it is not possible to increase the number of copies of the best solution or solutions that are similar to the best solution in a systematic way by using redundant representations. Therefore, problemspeciﬁc knowledge is necessary to increase the value of r. If we do not have any problemspeciﬁc knowledge about the structure of the problem, either nonredundant, or uniformly redundant representations, should be used. Both types of encodings guarantee that GEAs perform robustly, that means independently of the structure of the optimal solution.
4.4 Implications for the Design of Representations
111
4.4.2 Exponentially Scaled Representations Are Fast, but Inaccurate In Sect. 4.3, we examined the eﬀect of uniformly and nonuniformly scaled representations on the performance of GEAs. We saw (see also Sect. 4.2.2) that a diﬀerent scaling of representations modiﬁes the dynamics of genetic search. When using uniformly redundant representations all alleles are solved in parallel, whereas for exponentially scaled BBs the alleles are solved strictly serially. In the following paragraphs, we want to illustrate that GEAs using exponentially scaled representations deliver fast solutions which are inaccurate. In our framework, we have presented two diﬀerent models for scaled representations. For uniformly scaled representations, we assumed that all alleles are solved in parallel, and that the size of the convergence window is the same as the string length. We get from (4.7) for l = ls m and kg = 1 √
r ls m π unif orm − arcsin( kr − 1) · tconv = I 2 2 −1 When using exponentially scaled representations, we use the domino convergence model and the size of the convergence window λc = 1. Therefore, we get from (4.13) for the overall time to convergence √ r m π texp − arcsin = l − 1 · s conv I 2 2kr −1 In each of the m exponentially scaled BBs of size ls , the alleles are solved strictly sequentially. We see that when using exponentially scaled representations, the ﬁrst alleles are converged to the correct solution after a short time and we get a ﬁrst rough approximation of the correct solution. Furthermore, the alleles are solved from the most salient to the least salient. When using exponentially scaled representations the most salient allele has the same contribution to the phenotype as all lower salient alleles together. Because the low salient alleles do not signiﬁcantly change the phenotype (and the corresponding ﬁtness), we get an acceptable approximation after a few generations. The situation is diﬀerent when examining the number of generations that are necessary until the whole string is converged. GEAs using uniformly scaled orm = representations converge faster and ﬁnd the optimal solution after tunif conv √ exp tconv / ls generations. We can compare the diﬀerent times to convergence orm texp < tunif < texp conv conv ,
where texp = texp conv /ls denotes the time after m alleles of the same salience are converged (moving the convergence window to the next lower salient allele). GEAs always need more time to completely converge when using exponentially scaled representations than when using uniformly scaled representations.
112
4 TimeQuality Framework
correct alleles
ls m
2m
exponentially scaled uniformly scaled
m texp 2texp
ls t exp
ls t exp
generations
Figure 4.3. Number of correctly identiﬁed alleles over the number of generations for GEAs using uniformly versus nonuniformly scaled representations. The number of correctly identiﬁed alleles corresponds to the accuracy of the solution. GEAs using nonuniformly scaled representations provide an inaccurate solution to the problem more rapidly, but need longer to ﬁnd the exact solution.
We want to illustrate the inﬂuence of scaling on the dynamics of genetic search in Fig. 4.3. The ﬁgure shows the number of correctly identiﬁed alleles over the number of generations using uniformly scaled versus nonuniformly scaled representations. The number of correctly identiﬁed alleles is a measurement for the accuracy of the solution we get. The plots show that GEAs using nonuniformly scaled representations steadily improve the solution quality. After a few generations GEAs already provide us with a correct, but yet inaccurate solution to the problem. GEAs using uniformly scaled representations do not give us approximations after a few generations, but allow us to ﬁnd the exact optimum faster. We see that nonuniformly scaled representations rapidly deliver correct, but inaccurate solutions, whereas uniformly scaled representations do not produce early results, but give us the optimal solution faster. If we do not want to spend much time and we are not interested in high accuracy, nonuniformly scaled representations are a considerable choice. On the other hand, if we need exact solutions, we should use uniformly scaled representations. GEAs using uniformly scaled representations can ﬁnd the exact optimum in a shorter length of time. 4.4.3 Lowlocality Representations Are Diﬃcult to Predict, and No Good Choice Section 3.3 illustrated how the complexity of an optimization problem can be modiﬁed by using lowlocality representations. The genotypic size kg of BBs is determined by the locality dm (see Sect. 4.2.3). If the locality is high (dm = 0) the genotypic problem complexity is the same as the phenotypic problem complexity. However, for dm = 0 the genotypic size of BBs kg depends not only on the used representation but also on the speciﬁc optimization problem. We have already illustrated in Sect. 4.3.2 that when looking at exponentially scaled representations it makes no sense to consider the eﬀect of dm on
4.4 Implications for the Design of Representations
113
GEA performance. The performance of GEAs is modeled using the domino convergence model, which assumes strictly serial solving of the alleles. However, if the representation modiﬁes the size of BBs and kg = 1, there are interdependencies between the alleles, and the domino convergence model can not be used any more. Therefore, we want to focus in the following on uniformly scaled representations. For uniformly scaled representations, the probability of error (4.5) can be approximated as
N dr √ , α = exp − 2kr kg −1 σBB πm and the time to convergence (4.7) √
2r l π tconv = − arcsin( kr kg − 1) , I 2 2 where
kp kg = kg (fg , fp ) ,where 1 ≤ kg ≤ l
, if dm = 0, , if dm = 0.
The probability of error goes with O(exp(1/2kg )), and the time to convergence tconv increases with increasing kg . The genotypic size kg of BBs is inﬂuenced by dm . If dm = 0, kg is the same as kp and the complexity of the problem is preserved when mapping the phenotypes onto the genotypes. The situation becomes diﬀerent if dm = 0. Then, kg depends on the used representation fg and on the optimization problem fp . We have illustrated in the small example shown in Sect. 4.2.3 (see Fig. 4.1) that predicting GEA performance using lowlocality representations is diﬃcult. It depends on where the optimal solution is located, on the representation fg , and the structure of the problem fp . We illustrate the problem with lowlocality representations in Fig. 4.4. In Fig. 4.4(a), the locality is high and the size of BBs is the same in the genotypic and the phenotypic space. If we assume that the used GEAs can solve a problem up to kg = kmax , we can be sure that all problems with kp < kmax can be solved reliably. The situation becomes diﬀerent if dm = 0 (Fig. 4.4(b)). Then, kg depends not only on kp , but also on fg , and on fp . Using a lowlocality representation can result for problems with the same kp in problems with diﬀerent kg . Therefore, we can not predict which types of problems remain solvable and which are no longer solvable. Originally easy problems could become fully diﬃcult and fully diﬃcult problems could become fully easy. To predict the performance of GEAs using lowlocality representations is not possible if we have no detailed knowledge about fp . We summarize that we can not predict the performance of GEAs using lowlocality representations if we have no exact knowledge about the optimization problem fp and the used representation fg . However, if we know that a problem is too diﬃcult to be solved by the used GEAs, a representation with dm = 0 can sometimes advantageously use this problemspeciﬁc information.
114
4 TimeQuality Framework kg
kg
l
l
k max
k max
solvable
1
k max
solvable l
kp
(a) dm = 0
1
l
kp
(b) dm = 0
Figure 4.4. We illustrate how the genotypic size kg of BBs depends on the phenotypic size kp of BBs when using lowlocality versus highlocality representations. For highlocality representations (dm = 0), easy problems up to complexity kp can be reliably solved (Fig. 4.4(a)). If dm = 0 (Fig. 4.4(b)), kg does not correspond to kp , but depends on the used representation and the optimization problem. Then, we cannot predict kg without knowing fp .
Then, the representation modiﬁes kp and there is a chance that the problem becomes so easy that it can be solved (kg < kmax ). When a problem is fully diﬃcult (kp = l), on average, lowlocality representations make the problem easier (kg ≤ l) and more likely to be solved by GEAs. If we have no information a priori about a problem, or if we know that a problem is easy to solve for GEAs, we strongly favor the use of highlocality representations. These types or representations allow GEAs to reliably solve easy problems up to some complexity bound.
4.5 Summary and Conclusions We presented in this chapter a timequality framework for a theorybased analysis and design of representations for genetic and evolutionary algorithms. The chapter started with the determinants of GEA performance. The performance of GEAs is determined by the expected quality of the solutions and the number of generations that are necessary to ﬁnd them. This was followed in Sect. 4.2 by a description of the three elements the framework consists of. We presented how redundancy, scaling, and locality of a representation are measured and how they aﬀect GEA performance. In Sect. 4.3, we presented the main part of the chapter: the framework. Based on the work outlined in Chap. 3, we showed how the probability of error α and the time to convergence tconv is inﬂuenced by diﬀerent properties of representations. Finally, we presented in Sect. 4.4 some implications of the framework on the design of representations.
4.5 Summary and Conclusions
115
Based on the three elements of representation theory outlined in Chap. 3, the presented framework theoretically describes how diﬀerent types of representations inﬂuence GEA performance. The framework provides us with some important beneﬁts. It gives us a theoretical model for a better understanding of the inﬂuence of representations on GEA performance. Furthermore, it allows us to model and predict the performance of GEAs using a speciﬁc representation for diﬀerent types of optimization problems. Therefore, a theorybased use, analysis, and design of representations becomes possible by using the outlined framework. Based on the results from Chap. 3, the framework shows that synonymously redundant representations increase GEA performance, if the optimal solution is overrepresented by the representation. However, if some speciﬁc individuals are overrepresented, others remain underrepresented, and the performance depends on the structure of the optimal solution. Only uniformly redundant representations are robust concerning the structure of the optimal solution. By modifying the scaling of a representation the dynamics of genetic search are changed. If GEAs use exponentially scaled representations the domino convergence model can be used because the alleles are solved serially according to their salience. Therefore, the most salient alleles are solved after a few generations and a rough approximation of the optimal solution is available. However, to solve all alleles, GEAs using exponentially scaled representations need a larger number of generations compared to using uniformly scaled representations. The presented framework reveals that the locality of a representation is crucial for the performance of GEAs. The locality describes how well genotypic neighbors correspond to phenotypic neighbors when mapping the phenotypes on the genotypes. The analysis shows that locality inﬂuences the genotypic size of BBs. When using highlocality representations, the complexity of the problem is preserved and easy problems remain easy. For lowlocality encodings, the size of BBs is modiﬁed, and for predicting the resulting problem complexity, exact knowledge about the optimization problem and the used representation is necessary. We believe that the representation framework developed in this chapter is an important step toward a more general theory of representations for GEAs. Although it is not yet completed and there are still many open questions and shortcomings, it provides us with a much better understanding of the principles of representations and allows a more theoryguided design of representations. We want to encourage researchers to use the presented framework as a basis for a more detailed and extensive investigation into representations.
5 Analysis of Binary Representations of Integers
In the previous chapter, we presented a framework which describes the eﬀects of representations on the performance of GEAs. We illustrated the elements of the framework for problems where the phenotypes and genotypes are both bitstrings. The question is still open as to whether the framework also holds true for problems where the genotypes and phenotypes are diﬀerent. This question can be answered by examining problems where the genotypes are still binary but the phenotypes are integers. Integer optimization problems are common in many realworld applications. Although the most natural way for representing integer problems is to use an integer representation with integer genotypes, previous work has shown that by using genotypes that have a lower cardinality of the alphabet (for example binary genotypes) the possible number of schemata can be increased in comparison to integer strings (Goldberg 1990b). Consequently, researchers have developed diﬀerent types of binary representations for integers. The most common are binary, Gray, and unary representations. Previous work has shown that these three representations have diﬀerent properties and inﬂuence GEA performance differently (Caruana and Schaﬀer 1988; Whitley 1999; Whitley et al. 1997; Whitley 2000a). The purpose of this chapter is to use the framework presented in the previous chapters to explain the performance diﬀerences of GEAs when using diﬀerent binary representations for integers. For our investigation, we use two types of integer problems: variants of the easy integer onemax problem, and a diﬃcult deceptive integer trap problem. For encoding integer phenotypes, we use binary, Gray, or unary representations. The analysis of the unary encoding using the previously presented elements of representation theory reveals that the encoding is nonsynonymously redundant and does not represent the phenotypes uniformly. Therefore, the performance of GEAs depends on the structure of the optimal solution. If the good solutions are overrepresented by the encoding, GEAs perform well, whereas, if the good solutions are underrepresented, GEAs fail.
118
5 Analysis of Binary Representations of Integers
The binary encoding uses exponentially scaled alleles to represent integer values. Therefore, the convergence behavior is aﬀected by domino convergence and genetic drift. However, the analysis shows that genetic drift only results in a reduction of GEA performance for easy problems and small populations. An investigation into the locality of the binary encoding reveals that the locality is low and increases the diﬃculty of fully easy problems and reduces the diﬃculty of fully diﬃcult problems. Although the Gray encoding was designed to overcome the problems with the Hamming cliﬀ (Schaﬀer et al. 1989), it also has low locality and changes the diﬃculty of problems. Focusing on selectorecombinative GAs, a schema analysis for the integer onemax problem reveals that using Gray encoding results in larger BBs in comparison to binary encoding. As a result, in comparison to the binary encoding, the diﬃculty of the easy integer onemax problem increases for selectorecombinative GAs. These results are not contradictory to the FreeLunch theorem from Whitley (1999) and Whitley (2000a) regarding the Gray encoding but conﬁrm the results therein. The diﬀerence can be found in the used search method. We investigate the inﬂuence of Gray encoding on recombinationbased search approaches, whereas Whitley (1999) looks at mutationbased search methods. The work basically counts the number of local optima, which is lower when using Gray than binary encoding. Therefore, the performance of mutationbased search approaches on easy problems is higher when using Gray than when using binary encodings. After a brief presentation of the integer problems in Sect. 5.1, Sect. 5.2 describes the Gray, binary and unary encodings and analyzes their properties. This is followed in Sect. 5.3 by a theoretical comparison of the three encodings using the elements of theory presented in Chap. 3. We illustrate how the unary encoding is aﬀected by redundancy, how the exponential scaling of BBs inﬂuences the performance of the binary encoding, and how Gray encoding does not preserve problem diﬃculty well. Based on the elements of representation theory, we are able to make theoretical predictions about GEA performance. In Sect. 5.4, these predictions are ﬁnally conﬁrmed by empirical results. The chapter ends with concluding remarks.
5.1 Integer Optimization Problems In this section, we present integer problems we want to use for a comparison of diﬀerent representations deﬁned on binary genotypes. To be able to make a fair comparison between diﬀerent representations, the problem must be deﬁned on the integer phenotypes independently of the used binary representation. The diﬃculty of the problem is determined by the phenotypeﬁtness mapping fp . The diﬃculty of the problem can be changed by using an additional genotypephenotype mapping fg , which assigns binary genotypes to integer phenotypes. When assuming that the ﬁtness function fp assigns a real number to every individual in the phenotypic space, we get for
5.1 Integer Optimization Problems
119
the phenotypeﬁtness mapping: fp (xp ) : N → R. We want to use integerspeciﬁc variations of the onemax and the fullydeceptive trap problem. Traditionally, these problems are deﬁned on binary strings, but we want to deﬁne them in a similar way for integers. The integer onemax problem is deﬁned as fp (xp ) = xp .
(5.1)
A more general variant of the integer onemax problem (denoted as genonemax ) can be deﬁned as: fp (xp ) = xmax − xp − a,
(5.2)
where xp ∈ N, xmax = max(xp ), and a ∈ {0, 1, . . . , xmax }. For a = xmax the genonemax problem becomes the standard integer onemax problem (compare (5.1) and Fig. 5.1(a)). The diﬃculty of diﬀerent genonemax problems is independent of the parameter a. a only changes the location of the optimal solution in the search space and evolutionary search algorithms should show the same performance for diﬀerent values of a. Two examples for the genonemax problem are given in Fig. 5.1. f(xp)
f(xp)
6 5 4 3 2 1
6 5 4 3 2 1
0
1 2 3 4 5 6
(a) a = 6
x
p
0 1 2 3 4 5 6
x
p
Figure 5.1. Two examples for the general integer onemax problem (genonemax)
(b) a = 4
Furthermore, the integer deceptive trap problem can be deﬁned as xp if xp = xmax , f (xp ) = xmax − xp − 1 else,
(5.3)
where xp ∈ N. The onemax problems for integers are fully easy problems, whereas the integer deceptive trap should be fully diﬃcult to solve for GEAs. For measuring the similarity of individuals, we need to deﬁne a metric for the genotypic search space Φg and phenotypic search space Φp (see Sect. 3.3.2). As we use binary genotypes, we use the Hamming distance (Hamming 1980) on Φg , and the distance between two genotypes xg and y g of length l is deﬁned
120
5 Analysis of Binary Representations of Integers
l−1 as dxg ,yg = i=0 xgi − yig . The distance measures the number of alleles that are diﬀerent in both genotypes. The more bits two genotypes have in common, the more similar they are. The Hamming metric is chosen with respect to the bitﬂipping operator. Using this mutation operator results in an individual that has the lowest possible genotypic distance from its parent. Following the Hamming metric for the genotypes, we measure the distance between two phenotypes xp and y p (xp , y p ∈ N) as dxp ,yp = xp − y p . The distance between two phenotypes is simply the diﬀerence between both integers.
5.2 Binary String Representations After we have deﬁned the optimization problems, we present possible binary representations fg for integers. The representation fg assigns binary genotypes xg to integer phenotypes xp . Instead of using binary strings with cardinality χ = 2 for the genotypes, higher χary alphabets could also be used. Then, a χary alphabet is used for the string of length l instead of a binary alphabet. Therefore, instead of encoding 2l diﬀerent individuals with a binary alphabet, we are able to encode χl diﬀerent possibilities. However, Goldberg (1990b) has shown that schema processing is maximum with binary alphabets. Focusing on binary representations, we have a large number of diﬀerent genotypephenotype mappings that we can use as representations. If we use a redundancyfree encoding and want to encode 2l phenotypes with 2l possible genotypes, then there are (2l )! diﬀerent possibilities for the genotypephenotype mapping fg (see Sect. 3.3.7). Nevertheless, for our comparison we want to focus on the three most widely used representations deﬁned on binary strings: • binary representation, • Gray representation, and • unary representation. In contrast to the unary encoding, the binary and Gray encoding allows us to encode information redundancyfree. For the encoding of s possibilities, both encodings use log2 (s) bits (compare Sect. 3.1.1). The unary encoding uses s − 1 bits for encoding only s diﬀerent possibilities and is a redundant representation. In the following paragraphs, we want to brieﬂy review the important properties of the three diﬀerent encodings: When using the binary encoding, each integer phenotype xp ∈ Φp = {1, . . . , xmax } is represented by a binary genotype xg of length l = log2 (xmax ). The genotypephenotype mapping fg is deﬁned as xp = fg (xg ) =
l−1 i=0
2i xgi ,
5.2 Binary String Representations
121
with xgi denoting the ith bit of xg . Using the binary encoding for the integer onemax problem (5.1) results in the BinInt problem (compare Sect. 3.2.4). Since the bits in the string are exponentially scaled, we must use the domino convergence model and GAs are aﬀected by genetic drift (see Sect. 3.2). The bits are solved sequentially, and the low salient bits can be ﬁxed randomly before they are reached by the solving process. Furthermore, the encoding has problems associated with the Hamming cliﬀ (Schaﬀer et al. 1989). The Hamming cliﬀ describes the eﬀect that some neighboring phenotypes (the phenotypes have a distance of one) are represented by completely diﬀerent genotypes (the distance between the genotypes is much larger than one). Therefore, the locality of the binary encoding is low. As a result, especially mutationbased search approaches have problems when using this encoding because they rely on a high locality of the encoding. We have seen in Sect. 3.3.5 that high locality is a necessary condition for a representation to preserve BBcomplexity. Therefore, the ability of the binary encoding to preserve problem complexity is reduced in comparison to highlocality representations. However, the encoding also has some very interesting properties: It is linear, very compact and redundancyfree. For an example of the binary encoding, the reader is referred to Table 5.1 To overcome problems with the Hamming cliﬀ and the diﬀerent scaling of the alleles in binary strings, the Gray encoding was developed (Caruana and Schaﬀer 1988; Schaﬀer et al. 1989). When using Gray encoding, the average contribution of the genotypic alleles to the construction of the phenotype is the same for each allele in the binary genotype. Therefore, the Gray encoding is uniformly scaled (compare Sect. 3.2.1). The Grayencoded string itself can be constructed in two steps. At ﬁrst, the phenotype is encoded using the binary encoding, and subsequently the binaryencoded string can be converted into the corresponding Grayencoded genobin bin type. The binary string xbin ∈ {0, 1}l = {xbin 1 , x2 , . . . , xl } is converted to Gray l the corresponding Grayencoded string x ∈ {0, 1} = {xGray , . . . , xGray } 1 l l l by the mapping γ : B → B : xbin if i = 1, Gray i = xi bin bin otherwise, xi−1 ⊕ xi where ⊕ denotes addition modulo 2. The decoding of a Grayencoded string is as follows: i xbin = yjGray , i j=1
for i = {1, . . . l}. As mentioned before, a Grayencoded string has the same length as a binaryencoded string and the encoding is redundancyfree. Furthermore, the representation overcomes the problems with the Hamming cliﬀ. All neighboring phenotypes are also neighboring genotypes. However, as the number of neighbors is diﬀerent for the genotypic and phenotypic search space,
122
5 Analysis of Binary Representations of Integers
not all genotypic neighbors can correspond to phenotypic neighbors. Each phenotype has two neighbors (except xp = 0 and max(xp )), whereas each genotype has l = log2 (max(xp )) neighbors. Therefore, there are more genotypic than phenotypic neighbors and the locality of the encoding is low (dm = 0). However, when using the Gray encoding, for all genotypes xg ∈ Φg there exists a neighboring genotype y g ∈ Φg who corresponds to the neighboring phenotype, {y g (dyg ,yg = 1) ∧ (dyp ,xp = 1)}. This property gives Gray encoding an advantage over binary encoding when using mutationbased operators like the bitﬂipping operator (Whitley 1999) as there is always one genotypic mutation that results in a neighboring phenotype. In contrast, for the binary encoding there exist genotypes (for example xg = 1000 which corresponds to xp = 8) that do not have a genotypic neighbor which corresponds to the phenotypic neighbor (for example xg = 0111 which corresponds to xp = 7). For the binary encoding, mutation can not directly move from xg to y g although dxp ,yp = 1. For Gray encoding, there is always one direct move from xg to y g if dxp ,yp = 1. As before, Table 5.1 shows an example for the Gray encoding. Finally, the unary encoding can be used for encoding integers. A phenotype xp is encoded by the number u of ones in the corresponding genotype xg . With the length l = max(xp ) of the string and xgi as the ith bit of xg we get xp = fg (xg ) =
l−1
xgi .
i=0
number of genotypes that represent x
p
In contrast to the binary and Gray encoding, a string of length l = s − 1 is necessary for representing s diﬀerent phenotypes xp . Therefore, the genotypephenotype mapping is no longer a onetoone mapping but redundant. When encoding the phenotypic space Φp = {0, 1, . . . , l} using a unary string of length l, each of the l + 1 phenotypes xp ∈ Φp is represented by xlp diﬀerent genotypes xg . The number of genotypes that represent xp is illustrated for l = 7 in Fig. 5.2. Some phenotypes are represented by only one genotype (xp = 0 and xp = 7), whereas xp = 3 and xp = 4 are represented by 35 genotypes.
40 35 30 25 20 15 10 5 0 0
1
2
3
4 x
p
5
6
7
Figure 5.2. Redundancy of the unary encoding for l = 7
5.3 A Theoretical Comparison
123
The unary encoding is uniformly scaled and phenotypic neighbors correspond to genotypic neighbors. However, it has low locality as there are more genotypic than phenotypic neighbors (compare the discussion of the Gray encoding from above), and some genotypic neighbors do not correspond to phenotypic neighbors. We can illustrate this with an example. xp = 2 and y p = 3 are neighboring phenotypes. The corresponding genotypes xg = 0011 and y g = 0111 are also neighbors. However, other corresponding genotypes such as xg = 0011 and y g = 1110 are not neighbors. As the locality is low, the encoding also has low distance distortion. As not all genotypes that correspond to the same phenotype are similar to each other, the unary encoding is a nonsynonymously redundant encoding (compare Sect. 3.1.2). Finally, we want to give a brief example for the three diﬀerent types of encodings. Table 5.1 illustrates how the phenotypes xp ∈ {0, . . . , 7} can be represented by the binary, Gray, and unary encoding.
xp 0 1 2 3 4 5 6 7
xg binary 000 001 010 011 100 101 110 111
Gray 000 001 011 010 110 111 101 100
0000001, 0000011, 0000111, 0001111, 0011111, 0111111,
unary 0000000 0000010, . . . ,0100000, 0000101, . . . ,1010000, 0001011, . . . ,1101000, 0010111, . . . ,1110100, 0101111, . . . ,1111010, 1011111, . . . ,1111101, 1111111
1000000 1100000 1110000 1111000 1111100 1111110
Table 5.1. An example for using binary, Gray, and unary encodings
5.3 A Theoretical Comparison This section uses the framework of representations we presented in Chap. 4 to theoretically compare the performance of GEAs using binary, Gray and unary representations. The framework allows us to make predictions about the performance of GEAs which will be empirically veriﬁed in Sect. 5.4. In particular, we illustrate the eﬀects of the nonuniform redundancy of the unary encoding, how the genetic search process is prolonged by the eﬀect of exponentially scaled BBs for the binary encoding, and how the complexity of the problem is not well preserved by the Gray and binary encoding. 5.3.1 Redundancy and the Unary Encoding We know from Sect. 3.1 that redundancy reduces GEA performance if the encoding underrepresents good solutions. Furthermore, we know from the previous section (see Fig. 5.2) that the unary representation is a nonsynonymously
124
5 Analysis of Binary Representations of Integers
redundant encoding as one phenotype is encoded on average by more than one genotype but not all genotypes that encode one phenotype are similar to each other. Due to its nonsynonymous redundancy, the use of the unary encoding randomizes genetic search and the genotypic problem diﬃculty is diﬀerent from the phenotypic problem diﬃculty. In general, when using nonsynonymous redundant encodings, phenotypically easy problems become more diﬃcult whereas phenotypically diﬃcult problems become more easy. However, when using the unary encoding for the integer onemax and deceptive trap problem, the integer onemax problem remains genotypically fully easy for recombinationbased search (Goldberg 1989b; Deb and Goldberg 1993; Deb and Goldberg 1994) as all schemata containing the global optimum are still superior to their competitors (compare also Sects. 3.3.6 and 3.3.7). It also remains easy for mutation as it is a unimodal problem for mutationbased search and the structure of the ﬁtness landscape guides mutation towards the global optimum. Analogously, the fully deceptive trap remains fully diﬃcult. The only real handicap of the unary encoding seems to be the nonuniform redundancy. Therefore, we want to neglect the nonsynonymous redundancy and focus on the overrepresentation and underrepresentation of the optimal solution. We want to predict the performance of GEAs using the unary encoding for the integer onemax and integer deceptive trap problem from Sect. 5.1. We assume that Φp  = s. Therefore, for both problems, the integer onemax and the integer deceptive trap problem, the length of the unary encoded string is l = s − 1. Thus, 2s−1 diﬀerent genotypes only encode s diﬀerent phenotypes. log2 (s) Bits of information content (see Sect. 3.1.2) are encoded by s − 1 bits. Therefore, we get for the order of redundancy (see Sect. 3.1.2) kr =
s−1 , for s > 1. log2 (s)
On average, kr bits of a unary encoded bitstring are necessary for encoding one Bit of information content. This means, on average 2kr diﬀerent genotypes represent only two diﬀerent phenotypes. When using the unary encoding for the integer onemax or deceptive trap deﬁned in Sect. 5.1, the optimal phenotype (xp,opt = l) is represented by only one genotype (a string of only ones). Therefore, the number of genotypic BBs that represent the best phenotypic BB is r = 1. k r From (3.9), we get for the population size N = O 2r when using redundant encodings. Therefore, the necessary population size N when using unary encoding for the integer onemax and deceptive trap problem is increased in comparison to an encoding with uniform or no redundancy as s−1
Nunary = Nredundancyf ree × 2 log2 (s)
−1
.
The equation shows that with increasing string length l = s − 1, the necessary population size when using unary encoding increases exponentially. This eﬀect
5.3 A Theoretical Comparison
125
Nunary/Nredundancyfree
30 25 20 15 10 5 0 5
10
15
20
25
30
Figure 5.3. Necessary population size Nunary when using unary encoding
s
35
20
15
overrepresented
30 25
underrepresented
40
underrepresented
number of genotypes that represent x
p
is illustrated in Fig. 5.3. For even small problems the necessary population size Nunary is unreasonably high. Obviously the use of the unary encoding results for the proposed integer onemax and deceptive trap problem in a low GA performance. However, we know from Sect. 3.1 that the performance of nonuniformly redundant encodings depends on the speciﬁc problem they are used for. GEAs using synonymously redundant encodings only show low performance if the good solutions are underrepresented. Therefore, we want to investigate in the remaining paragraphs for which problems the unary encoding performs well. For both problems, the integer onemax and deceptive trap problem, the optimal solution xp,opt = l is strongly underrepresented by only one genotype and GEAs using the unary encoding show low performance. When changing the position of the optimal solution xp,opt (like in the genonemax problem) and the optimal solution is not a string with only ones (or only zeros), GEAs can perform better when using the unary encoding. This means, if we use a diﬀerent integer ﬁtness function fp and the optimal solution would be, for example, xp,opt = l/2 (resulting in a genonemax problem with a = l/2), then the optimal solution would be strongly overrepresented and GEAs using the unary encoding would be able to solve this problem very eﬀectively.
10 5
0
0
1
2
3
4
x average
p
5
6
7
Figure 5.4. Areas of over and underrepresented phenotypes when using unary encoding
126
5 Analysis of Binary Representations of Integers
We want to illustrate the problem of the unary encoding with underrepresented solutions in Fig. 5.4 more clearly. On average γavg
s−1 1 s−1 = s xp =0 xp
< γavg the optimal genotypes represent one phenotype. Therefore, if xs−1 p,opt is underrepresented and the performance of the GA is reduced. solution xp,opt > γavg the optimal solution is overrepresented and redunHowever, if xs−1 p,opt dancy helps the GA in ﬁnding the optimal solution xp,opt . We see that the performance of GEAs using the unary encoding depends on the structure of the optimization problem we want to solve. The unary encoding can be a good choice for the integer onemax problem or the integer deceptive trap problem if the optimal solution is not strongly underrepresented. 5.3.2 Scaling, Modiﬁcation of Problem Diﬃculty, and the Binary Encoding In Sect. 3.2, we illustrated how the search process of GEAs is prolonged by using nonuniformly scaled encodings. If integers are encoded using the binary representation, domino convergence occurs and GEAs are aﬀected by genetic drift. As a result, the probability α of GEA failure increases for small population sizes N . However, scaling only aﬀects GEAs for the easy integer onemax problem because the optimal solution can be found even with small populations, but not for the more diﬃcult deceptive trap problem. Here, the necessary population size is large enough that no drift occurs. For further details regarding the eﬀects of exponentially scaled encodings, such as the binary encoding on GEA performance, we refer to Sect. 3.2 as well as to results presented in the literature (Thierens 1995; Thierens et al. 1998; Lobo et al. 2000). The performance of GEAs using binary encoding is not only aﬀected by the exponential scaling of the encoding, but also by problems associated with locality and the Hamming cliﬀ (Caruana and Schaﬀer 1988; Caruana et al. 1989; Schaﬀer et al. 1989). Binary encodings have low locality (dm = 0) as not all neighboring genotypes correspond to neighboring phenotypes. As an example we can chose the genotypes xg = 011 and xg = 111, which are neighbors and have distance dxg ,yg = 1. However, the corresponding phenotypes xp = 3 and y p = 7 have distance dxp ,yp = 4. As we know from Sect. 3.3.5 that lowlocality encoding change problem diﬃculty, we expect that the nonredundant binary encoding changes the structure and complexity of the BBs. How exactly the structure of the BBs is changed is exemplarily measured in the following paragraphs (compare Table 5.3).
5.3 A Theoretical Comparison
127
5.3.3 Modiﬁcation of Problem Diﬃculty and the Gray Encoding The nonredundant Gray encoding has low locality (dm = 0), as the genotypes have a larger number of neighbors than the phenotypes. Therefore, the complexity of BBs is modiﬁed and the problem diﬃculty for GEAs is changed (compare Sect. 3.3) when mapping phenotypic integers on genotypic bitstrings. As a result, fully easy integer problems remain not fully easy. Mutationbased search approaches using Gray encoding perform better than using the binary encoding, as there is always one genotypic mutation that allows the search method to reach a neighboring phenotype in one search step. This performance advantage of Gray encoding in comparison to binary encoding has already been described in other work (Whitley et al. 1997; Rana and Whitley 1997; Whitley and Rana 1997; Rana and Whitley 1998; Whitley 1999; Whitley 2000a; Whitley 2000b). This work formulated a FreeLunch theorem for the use of Gray encoding and mutationbased search approaches. GEAs using mutation as the main search operator perform better on easy problems (these are the problems which we are interested in) when using Gray encoding than when using binary encoding. It was shown that the number of local optima introduced by Gray encoding is smaller than by binary encoding. We know that using Gray encoding changes the problem diﬃculty for GEAs. When using crossoverbased search, the schema analysis is an appropriate method to measure problem diﬃculty (compare Sect. 2.3.2). For both, binary and Gray encoding, dc = 0 and genotypic and phenotypic problem diﬃculty is diﬀerent when using crossoverbased search approaches. The distance distortion dc is high, as the structure of the genotypic and phenotypic search space is diﬀerent and a phenotype has a lower number of neighbors than a genotype. The following analysis of the schemata ﬁtness reveals for the integer onemax and deceptive trap problem that in comparison to the binary encoding, Gray encoding does not preserve the complexity of BBs as well. This leads to a lower GA performance. To investigate how well BBcomplexity is preserved, we analyze the ﬁtness of the schemata for a 3bit problem (s = 23 = 8) using Gray versus binary encoding. In Table 5.2, we present the binary and Grayencoded genotypes, and the resulting ﬁtness values for the integer onemax and deceptive trap problem. In Table 5.3, we present the average ﬁtness of the schemata for the two problems. Reviewing problem complexity, the problem is fully deceptive if all schemata of lower order containing the global optimum are inferior to their competitors (Deb and Goldberg 1994). Analogously, the problem is fully easy if all schemata containing the global optimum are superior to their competitors (compare Sect. 2.3.2). The analysis shows that for the fully easy integer onemax problem with binary encoding, all schemata containing the global optimum xg,opt = 111 are superior to their competitors. Although the binary encoding has low locality, the fully easy integer onemax problem remains fully easy, and the binary encoding preserves the diﬃculty of the problem well. The schema analysis for
128
5 Analysis of Binary Representations of Integers
Table 5.2. Using binary and Gray encoding for an integer onemax and deceptive trap problem (s = 8). The resulting length of the genotypes l = 3. genotype xg phenotype xp
binary Gray binary Gray
integer deceptive trap problem (s = 8)
integer onemax problem (s = 8)
ﬁtness
order schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness schema ﬁtness
binary Gray integer fone−max (xp ) fdeceptive (xp )
3 2 111 11* 1*1 7 6.5 6 01* 0*1 2.5 2 10* 1*0 4.5 5 00* 0*0 0.5 1 100 10* 1*0 7 6.5 5.5 11* 1*1 4.5 5.5 01* 0*1 2.5 1.5 00* 0*0 0.5 1.5 111 11* 1*1 7 3.5 4 01* 0*1 3.5 4 10* 1*0 1.5 1 00* 0*0 5.5 5 100 10* 1*0 7 3.5 4.5 11* 1*1 1.5 0.5 01* 0*1 3.5 4.5 00* 0*0 5.5 4.5
000 000 0 0 6
*11 5 *01 3 *10 4 *00 2 *00 3.5 *11 3.5 *01 3.5 *00 3.5 *11 5 *01 3 *10 2 *00 4 *00 6.5 *11 2.5 *01 2.5 *00 2.5
001 001 1 1 5
010 011 2 2 4
**1 11 **0 3
1 *1* 4.5 *0* 2.5
0 1** *** 5.5 3.5 0** 1.5
1** 5.5 0** 1.5
*0* 3.5 *1* 3.5
**0 *** 3.5 3.5 **1 3.5
**1 4 **0 3
*1* 2.5 *0* 3.5
1** *** 2.5 3.5 0** 4.5
1** 2.5 0** 4.5
*0* 4.5 *1* 2.5
**0 *** 4.5 3.5 **1 2.5
011 010 3 3 3
100 110 4 4 2
101 111 5 5 1
110 101 6 6 0
111 100 7 7 7
Table 5.3. Schemata ﬁtness for the integer onemax and deceptive trap problem using binary versus Gray encoding. The integer onemax problem remains fully easy when using the binary representation. Using Gray encoding makes the problem more diﬃcult as some of the high quality schemata have the same ﬁtness as the misleading schemata. The situation for the deceptive trap is the opposite one. The fully diﬃcult deceptive trap becomes easier to solve when using Gray encoding.
5.4 Experimental Results
129
the Gray encoding reveals that the schemata containing the global optimum xg,opt = 100 are not always superior to their competitors. Therefore, the problem is not fully easy anymore, and the Gray encoding changes problem diﬃculty and does not preserve the easiness of the integer onemax problem. The schemata analysis of the integer trap problem reveals that the problem remains not fully deceptive when using the binary encoding. Some of the schemata containing the global optimum xg,opt = 111 are superior to their competitors (*11 and **1). However, when using Gray encoding even more schemata containing the global optimum are not inferior to their competitors (1*0, *00, *0*, **0). The phenotypically fully diﬃcult problem is not fully diﬃcult anymore.
5.4 Experimental Results In this section, we present an experimental veriﬁcation of the performance diﬀerences between the three diﬀerent representations we discussed in the previous section. 5.4.1 Integer OneMax Problem and Deceptive Integer OneMax Problem We compare the performance of GAs using binary, Gray, and unary encoding for the integer onemax and deceptive trap problems as deﬁned in Sect. 5.1. We performed 250 runs for each problem instance and each run was stopped after the population was fully converged (all individuals in the population are the same). For the integer onemax problem we used uniform crossover, and for the integer deceptive trap we used twopoint crossover. As selection method we used tournament selection without replacement of size two. We used no mutation as we want to focus on selectorecombinative GEAs. The Figs. 5.5, 5.6, 5.7, and 5.8 present results for the integer onemax problem, and the Figs. 5.9 and 5.10 for the integer deceptive trap problem. The plots show for diﬀerent representations the proportion of correctly solved subproblems at the end of the run (left) and the run duration tconv (right) with respect to the population size N . For the onemax problem, we concatenated 20 subproblems of order 2 (s = 22 = 4, see Fig. 5.5), 3 (s = 8, Fig. 5.6), 4 (s = 16, Fig. 5.7), and 5 (s = 32, Fig. 5.8)1 . The ﬁtness of an individual is calculated as the sum of the ﬁtness of the 20 concatenated subproblems. Because large integer deceptive traps are not solvable by GAs in a reasonable time, we only present results for the deceptive trap problem of order 2 (s = 4, Fig. 5.9), and 3 (s = 8, Fig. 5.10). Using binary or Gray encoding results for the order 2 problems in a string length l = 40, for order 3 in l = 60, for order 1 The order r of a problem is deﬁned as r = log2 s and describes the length of the corresponding binary or Grayencoded string.
130
5 Analysis of Binary Representations of Integers
1
26 25
0.98
avg. run duration tconv
proportion of correct subproblems
4 in l = 80, and for order 5 in l = 100. When using unary encoding we need 20 × 3 = 60 bits for order 2, 20 × 7 = 140 bits for order 3, 20 × 15 = 300 bits for order 4, and 20 × 31 = 620 bits for order 5 problems.
0.96 0.94 0.92 binary Gray unary
0.9 0.88
24 23 22 21 20
binary Gray unary
19 18
20
40
60
80
population size N
100
20
40
60
80
100
population size N
Figure 5.5. Integer onemax problem of order 2. We concatenated m = 20 subproblems and the size of the search space Φp  = 22 = 4. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). Due to the low complexity of the problems all three representations perform about the same. However, binary encoding is much faster in ﬁnding the good solutions.
Due to the problems of the unary encoding with redundancy, which result in an underrepresentation of the optimal solution, GAs using unary encoding show decreasing performance with increasing problem size. Therefore, for onemax problems of order more than three the GA performance is signiﬁcantly worse than when using Gray or binary encoding. Although the onemax problem remains fully easy, GEA performance is reduced because the optimal solution is strongly underrepresented. Only for the almost trivial onemax problem of order 2 or 3 has the unary encoding a comparable performance. The plots nicely illustrate that only for small onemax problems the beneﬁts from the preservation of BBcomplexity can compensate the performance reduction caused by the underrepresentation of the optimal solution. For deceptive traps of order more than 2, unary encoding fails completely because the problem remains fully diﬃcult and the optimal solution is underrepresented. Furthermore, the plots show that due to the preservation of BBcomplexity, a GA using unary encoding performs in comparison to Gray or binary encoding relatively better for the easy onemax problem than for the deceptive trap. The failure of the encoding for the deceptive trap can be better understood if we recognize that an order 3 problem results in a fully deceptive BB of length l = 7. This problem is only solvable with much larger population sizes.
1
131
40 38
0.95
avg. run duration tconv
proportion of correct subproblems
5.4 Experimental Results
0.9 0.85 0.8 binary Gray unary
0.75 0.7
36 34 32 30 28 26
binary Gray unary
24 22
20
40
60
80
100
20
population size N
40
60
80
100
population size N
1
60 55
0.8
avg. run duration tconv
proportion of correct subproblems
Figure 5.6. Integer onemax problem of order 3. We concatenated m = 20 subproblems and the size of the search space Φp  = 23 = 8. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). Binary encoding performs the best. Because the optimal solutions are underrepresented, GAs using the unary encodings perform worse than when using Gray encoding for small population sizes.
0.6 0.4 binary Gray unary
0.2
40
60
population size N
45 40 35 binary Gray unary
30 25
0 20
50
80
100
20
40
60
80
100
population size N
Figure 5.7. Integer onemax problem of order 4. We concatenated m = 20 subproblems and the size of the search space Φp  = 24 = 16. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). Because the binary encoding preserves BBcomplexity better than Gray encoding, a GA using binary representations performs best. Because of problems with redundancy, the unary encoding performs worst and needs the most ﬁtness evaluations. The error bars indicate the standard deviation of some results.
5 Analysis of Binary Representations of Integers 1
90 80
0.8
avg. run duration tconv
proportion of correct subproblems
132
0.6 0.4 binary Gray unary
0.2 0
70 60 50 40 binary Gray unary
30 20
20
40
60
80
100
20
population size N
40
60
80
100
population size N
1 35
0.9
avg. run duration tconv
proportion of correct subproblems
Figure 5.8. Integer onemax problem of order 5. We concatenated m = 20 subproblems and the size of the search space Φp  = 25 = 32. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). As before, binary encoding performs best. It becomes obvious that with increasing problem size GEAs using unary encoding have increasing diﬃculty in ﬁnding the good solutions. Furthermore, the performance diﬀerences between the binary and Gray encoding become larger with increasing problem size. The error bars indicate the standard deviation of some results.
0.8 0.7 binary Gray unary
0.6
30 25 binary Gray unary
20
0.5 20
40
60
80
population size N
100
20
40
60
80
100
population size N
Figure 5.9. Integer deceptive trap problem of order 2. We concatenated m = 20 subproblems and the size of the search space Φp  = 22 = 4. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). Gray encoding performs slightly better than binary encoding as it preserves the structure of the subproblems worse. Because unary encoding strongly underrepresents the optimal solution, it performs the worst.
133
1 50 0.8
avg. run duration tconv
proportion of correct subproblems
5.4 Experimental Results
0.6 binary Gray unary
0.4 0.2 0
45 40 35 30 binary Gray unary
25 20
20
40
60
population size N
80
100
20
40
60
80
100
population size N
Figure 5.10. Integer deceptive trap problem of order 3. We concatenated m = 20 subproblems and the size of the search space Φp  = 23 = 8. We show the average proportion of correct subproblems at the end of run (left) and the average running time (right). Gray encoding performs signiﬁcantly better than binary encoding as it makes the fully diﬃcult deceptive trap problem easier to solve. Unary encoding fails as it has problems with redundancy. The error bars indicate the standard deviation of some results.
As expected, Gray encoding performs worse than binary encoding for the onemax problem, and better for the deceptive trap problem. Because for the integer onemax and deceptive trap problem Gray encoding preserves BBcomplexity less than binary encoding (compare Table 5.3), the fully easy integer onemax problem becomes more diﬃcult to solve, whereas the fully diﬃcult deceptive trap is easier to solve for a GA using Gray encoding. Finally, the inﬂuence of exponentiallyscaled representations like the integer encoding on the performance of GEAs can be seen for the onemax problem. For small population sizes N , genetic drift has a larger impact on the solution quality. Therefore, for the easy integer onemax problem and small population sizes N , GAs using binary encoding perform only slightly better than Gray encoding. For larger population sizes, however, the eﬀect of genetic drift is reduced and GAs using binary representation perform relatively better. We see that the empirical results nicely verify the theoretical predictions from the previous section. Fig. 5.11 summarizes some of the results for the integer onemax problem and shows the proportion of correct subproblems at the end of the run over the order of the problem. Due to the underrepresentation of the optimal solutions the performance of a GA using unary encoding strongly decreases with increasing problem size. A GA using binary encoding performs best in comparison to Gray and unary encoding, as the exponential scaling of the representation aﬀects a GA only for small populations, and the encoding preserves BBcomplexity better than the Gray encoding. As a result,
5 Analysis of Binary Representations of Integers 1
binary Gray unary
0.8 0.6 0.4 0.2 0 2
3
4
5
6
8
order of problem
10
proportion of correct subproblems
proportion of correct subproblems
134
1
binary Gray unary
0.8 0.6 0.4 0.2 0 2
3
4
5
6
8
10
order of problem
Figure 5.11. Proportion of correct subproblems at the end of the run over the order of the problem for a population size of 20 (left) and 40 (right) for binary, Gray, and unary encoding. The ﬁgures are plotted for the integer onemax problem. It can be seen that with increasing order of the problem the performance of the unary representation strongly decreases. When using Gray or binary encoding the performance of the GA declines much less.
the easy onemax problem remains easier with the binary encoding than with the Gray encoding. Although Gray encoding preserves BBcomplexity worst, it still signiﬁcantly outperforms unary encoding which fails for the onemax and deceptive trap due to the underrepresentation of the optimal solution. 5.4.2 Modiﬁcations of the Integer OneMax Problem This section investigates how the performance of Gray and binary encoding depends on the properties of the optimal solution for mutationbased and crossoverbased search. For the experiments we use the genonemax problem as deﬁned in (5.2). MutationBased Search Using Simulated Annealing We investigate how the locality of an encoding inﬂuences the performance of mutationbased search approaches. In our investigations, we assume that the genonemax problem deﬁned in (5.2) is easy for mutationbased search independently of the position of the optimal solution a. We want to use simulated annealing (SA) as a representative for a mutationbased search approach because it uses only mutation, and can in contrast to for example an (1+1) evolution strategy, solve diﬃcult, and multimodal, problems more easily. Simulated annealing can be modeled as a GA with population size N = 1 and Boltzmann selection (Mahfoud and Goldberg 1995). In each generation, a genotypic oﬀspring xg,o is created by applying mutation to the parent xg,p . Therefore, if we use bitﬂippingmutation, xg,o always has genotypic distance 1 to its parent xg,p . If xg,o has higher ﬁtness
5.4 Experimental Results
135
10 8 6 4 2
binary Gray 0
500
1000
number of fitness calls
1500
number of correct subproblems
number of correct subproblems
than xg,p , it replaces xg,p . If it has lower ﬁtness, it replaces xg,p with prob g,o g,p (x ) ability P (T ) = exp − f (x )−f . By lowering the temperature T , the T probability of accepting worse solutions decreases. For further information the reader is referred to van Laarhoven and Aarts (1988). For our investigation we concatenate 10 integer genonemax problems of diﬀerent length l. When using Gray or binary encoding, each of the 10 phenotypic integers xp ∈ {0, . . . , 2l − 1} corresponds to l bits in the genotype. Therefore, for l = 5, the overall length of a genotype is 50. The ﬁtness of an individual is calculated as the sum over the ﬁtness of the 10 subproblems. The ﬁtness of one subproblem is calculated according to (5.2). Figure 5.12 presents results using SA for two instances (a = 15 and a = 31) of the genonemax problem with l = 5. We show the number of correctly solved subproblems over the number of ﬁtness evaluations. The start temperature Tstart = 50 is reduced in every step by the factor 0.995. Therefore, Tt+1 = 0.995×Tt . Mutation is deﬁned to randomly change one bit in the genotype. We performed 100 runs and each run was stopped after 2000 mutation steps. The results show that mutationbased search approaches using Gray encoding always solve all 10 subproblems. In contrast, for a = 15, mutationbased search using binary encoding gets stuck in local optima because the optimal solution lies in areas with problems with the Hamming cliﬀ.
10 8 6 4 2
binary Gray 0
500
1000
1500
number of fitness calls
Figure 5.12. We use SA and show the number of correctly solved subproblems over the number of ﬁtness calls (search steps) for a = 31 (left) and a = 15 (right).
To generalize our investigation and to determine how the performance of mutationbased search depends on the structure of the integer optimization problem, Fig. 5.13 illustrates how SA performance depends on the value of the optimal solution a. We show results for l = 3 (left) and l = 5 (right). We only change a and use the same parameter settings as before. The plots show that using Gray encoding allows SA to reliably ﬁnd the optimal solution independently of the location of the best solution. Using binary encoding often results in lower SA performance and the SA gets stuck in local optima.
5 Analysis of Binary Representations of Integers
10 9 8 binary Gray 7 0
1
2
3
4
5
6
7
number of correct subproblems
number of correct subproblems
136
10 9 8 7 6 binary Gray
5 4 0
a (value of optimal solution)
5
10
15
20
25
30
a (value of optimal solution)
Figure 5.13. We use SA and show the number of correctly solved subproblems at the end of a run over the location of the optimal solution a for l = 3 (left) and l = 5 (right).
This performance diﬀerences between Gray and binary encoding has already been observed in other work (Schaﬀer et al. 1989; Whitley et al. 1997; Whitley 1999). The binary encoding has problems associated with the Hamming cliﬀ and has low locality. Due to its low locality, small changes of the genotype do not always result in small changes of the corresponding phenotype. In addition, due to the Hamming cliﬀ, neighboring phenotypes are not always neighboring genotypes (for example xp = 15 and xp = 16). Therefore, mutationbased search can never move directly from xp = 15 to xp = 16. Only if the optimal solution consists of almost only ones or zeros does binary encoding show similar performance to Gray encoding. We can explain the performance diﬀerences between Gray and binary encoding by focusing on the Hamming distances between neighboring individuals. Although both encodings, Gray and binary encoding, have low locality, and not all genotypic neighbors correspond to phenotypic neighbors, there are some diﬀerences in the neighborhood structure. Table 5.4 presents the properties of neighboring individuals for Gray versus binary encoding for l = 3. We show the l = 32 = 8 possible phenotypes xp and their phenotypic neighbors y p , {y p dxp ,yp = 1}. Furthermore, we show for Gray and binary encoding the corresponding genotypes xg , the genotypes y g that correspond to the phenotypic neighbors y p , and the average genotypic Hamming distances avg(dgxg ,yg ) between xg and y g , where avg(dgxg ,yg ) =
1 np 
dxg ,yg .
{y g dxp ,yp =1}
np  denotes the number of phenotypic neighbors y p . As already discussed in Sect. 5.2, for Grayencoded genotypes there is always a neighboring genotype that corresponds to a phenotypic neighbor. Therefore, only one genotypic mutation step is necessary to reach all possible phenotypic neighbors and the average distances between xg and the genotypes y g that represent the phenotypic neighbors y p is one (avg(dgxg ,yg ) = 1).
5.4 Experimental Results
137
Table 5.4. Properties of neighboring individuals for Gray and binary encoding xp 0 1 2 3 4 5 6 7 p {y dxp ,yp = 1} 1 0, 2 1, 3 2, 4 3, 5 4, 6 5, 7 6 xg 000 001 011 010 110 111 101 100 g p Gray {y dx ,yp = 1} 001 000,011 001,010 011,110 010,111 110,101 111,100 101 avg(dgxg ,yg ) 1 1 1 1 1 1 1 1 xg 000 001 010 011 100 101 110 111 binary {y g dxp ,yp = 1} 001 000,010 001,011 010,100 011,101 100,110 101,111 110 avg(dgxg ,yg ) 1 1.5 1.5 2 2 1.5 1.5 1
For binary encoding, avg(dgxg ,yg ) ≥ 1. The average genotypic distance between the genotypes xg and the genotypes y g that represent the phenotypic neighbors of xg increases for xp → (xpmax /2). For xpmax and xpmin it is minimal (avg(dgxg ,yg ) = 1), and for xp = xpmax /2 it is maximal (avg(dgxg ,yg ) = (1 + l)/2, where l = s − 1 is the length of the genotype). As a result, the genonemax problem is, independently of the position of the optimal solution, easy for mutationbased search using Gray encoding as there is always a neighboring genotype which corresponds to a neighboring phenotype. In contrast, for the binary encoding, there is not always a neighboring genotype that corresponds to the phenotypic neighbors. As avg(dgxg ,yg ) increases for a → (xpmax /2), the diﬃculty of genonemax problems for mutationbased search increases with a → (xpmax /2). This behavior of mutationbased search using the binary encoding can be nicely observed in Fig. 5.13. CrossoverBased Search Using Genetic Algorithms We investigate for the genonemax problem how Gray and binary encoding inﬂuence the performance of crossoverbased search. We have seen discussed in Sect. 5.2 that both representations have low locality; however, Gray encoding better preserves the phenotypic neighborhood structure (compare results from the previous paragraphs). Due to the low locality of the encodings, oﬀspring produced by standard crossover mechanisms could have nothing in common with their parents. For example, if we use binary encoding and uniform crossover we can get from the parents xp = 4 (xg = 100) and y p = 3 (y g = 011) the oﬀspring z p = 7 (z g = 111). The oﬀspring has phenotypically nothing in common with its parents and the phenotypic distances between the oﬀspring and its parents are much larger than the distances between both parents (compare Sect. 3.3.5). Therefore, both encodings change the diﬃculty of the easy integer problem. As before, we concatenate 10 integer genonemax problems with l = 5 (xp ∈ {0, . . . 31}). For our investigation we use a selectorecombinative standard GA (Goldberg 1989c) using only uniform crossover and no mutation. The population size is set to N = 20 and we use tournament selection without replacement of size 2. We performed 100 runs, and each run was stopped after the population was fully converged.
5 Analysis of Binary Representations of Integers 10 8 6 4 2
binary Gray
0 20
40
60
80
100
number of correct subproblems
number of correct subproblems
138
10 8 6 4 2
binary Gray
0 20
number of generations
40
60
80
100
number of generations
Figure 5.14. We use a selectorecombinative GA and show the number of correctly solved subproblems over the number of generations for a=31 (left) and a=15 (right).
10 9 8 binary Gray
7 0
1
2
3
4
5
6
a (value of optimal solution)
7
number of correct subproblems
number of correct subproblems
In Fig. 5.14, we show the number of correctly solved subproblems over the number of generations for a = 31 (left) and a = 15 (right). For a = 31, the genonemax problem is equivalent to the onemax problem and we get the same results as presented in Sect. 5.4.1. The results show that selectorecombinative GAs using binary encoding outperform Gray encoding for a = 31. For a = 15, the situation changes and GAs using Gray encoding perform signiﬁcantly better than binary encoding. As before, we want to generalize our investigation and show in Fig. 5.15 how the average number of correctly solved subproblems at the end of the run depends on the value of the optimal solution a. We show results for l = 3 (left) and l = 5 (right). It can be seen that GAs using binary encoding perform better than Gray encoding if a is either small (a → xpmin ) or large (a → xpmax ). Otherwise, GAs using Gray encoding perform better.
7 6 5 4 binary Gray
3 0
5
10
15
20
25
30
a (value of optimal solution)
Figure 5.15. We use a selectorecombinative GA and show the number of correctly solved subproblems at the end of a run over the location of the optimal solution a for l = 3 (left) and l = 5 (right).
5.5 Summary and Conclusions
139
When using crossoverbased search, the diﬃculty of the original optimization problem fp only remains unchanged if the genotypephenotype mapping fg does not change the distances between the individuals (dc = 0). However, both encodings, Gray and binary, have high distance distortion (and low locality) and change the distances between corresponding genotypes and phenotypes. Therefore, both encodings change the diﬃculty of the optimization problem. As a result, GA performance strongly varies for diﬀerent a, although the diﬃculty of fp remains constant and is independent of a. GA performance can not accurately be predicted due to the low locality of the encodings. General statements regarding the performance (for example measured by using the schemata analysis as illustrated in Sect. 5.3.3) of binary or Gray encoding for crossoverbased search are not possible as GA performance depends on the speciﬁc problem that should be solved.
5.5 Summary and Conclusions Section 5.1 started this chapter by presenting two types of integer problems. Integer onemax problems are fully easy problems, whereas the integer fully deceptive trap is an example of a fully diﬃcult problem. In Sect. 5.2, we presented and examined the binary, Gray and unary encoding, which are common representations for integer phenotypes. This is followed in Sect. 5.3 by a theoretical comparison of the expected performance of GEAs using the three diﬀerent representations. We showed that using the nonsynonymously redundant unary encoding reduces GEA performance if the optimal solution is underrepresented. Therefore, the necessary population size for solving the integer onemax and the integer deceptive trap problem is increased. Using binary encoding results in a more compact, redundancyfree representation, but the alleles are exponentially scaled. Therefore, genetic drift occurs for small and easy problems and larger population sizes are necessary. Due to the problems of the binary encoding with the Hamming cliﬀ and its low locality, the performance of GEAs is reduced for fully easy problems, and increased for fully diﬃcult problems. Although Gray encoding was developed to overcome the problems of the binary encoding with the Hamming cliﬀ, an analysis of the average ﬁtness of the schemata for the integer onemax problem shows that the Gray encoding also has low locality and preserves problem diﬃculty for selectorecombinative GAs less than binary encoding. Thus, the performance of selectorecombinative GAs decreases for the easy integer onemax problem and increases for the integer deceptive trap problem. To verify the theoretical prediction, we performed an empirical investigation into the performance of mutationbased and crossoverbased GEAs using the diﬀerent encodings in Sect. 5.4. The results conﬁrmed the theoretical predictions. Throughout this entire chapter, we have used the representation framework from Chap. 4 for the analysis of binary representations for integers. The analysis has shown that the pieces of representation theory can eﬀectively be
140
5 Analysis of Binary Representations of Integers
used for predicting the performance of GEAs. We were able to explain the diﬀerences in performance of the binary, Gray, and unary representations by using the outlined theory about redundant, exponentially scaled, and lowlocality representations. In particular, we gained the following insights: We have seen that the binary encoding is exponentially scaled. However, the inﬂuence of the exponential scaling of the alleles on GEA performance can be neglected as it only aﬀects the performance of GEAs for easy problems and small √ population sizes. Although the time to convergence is increased from O( l) to O(l), we can easily overcome the negative eﬀects of the exponential scaling on GEA performance by using larger population sizes. When using the nonuniformly and nonsynonymously redundant unary encoding, redundancy aspects become important. GEAs using the unary encoding fail for the integer onemax and deceptive trap as the optimal solution is strongly underrepresented for these two types of problems. Therefore, GEAs using the unary encoding perform signiﬁcantly worse in comparison to GEAs using the nonredundant binary or Gray encoding. Finally, the investigation of locality reveals that both, Gray and binary encoding, have low locality. Not all neighboring genotypes correspond to neighboring phenotypes. Therefore, the inﬂuence of these representations on GEA performance is diﬃcult to predict and depends on the speciﬁc problem that is solved. The locality of the binary encoding is worse than the Gray encoding, as for Grayencoded genotypes there is always a neighboring genotype that corresponds to a neighboring phenotype. When using the binary encoding, it is not possible for all genotypes to reach a neighboring phenotype with one genotypic mutation. To give a ﬁnal recommendation for selectorecombinative GEAs is diﬃcult. Both encodings, the binary and the Gray encoding, change the distances between the individuals and therefore change the complexity of the optimization problem. Thus, the resulting problem diﬃculty depends not only on the used representation but also on the considered optimization problem (compare Sect. 4.4.3). We have seen that some easy problems like the integer onemax problem become easier when using the binary encoding than when using the Gray encoding. However, there are other easy problems that become more diﬃcult when using the binary encoding than when using the Gray encoding (compare the results presented in Sect. 5.4.2). When using mutationbased GEAs instead of crossoverbased GAs, the Gray encoding is the best choice (Whitley 1999). For this type of search process, Gray encoding allows a more eﬃcient search as it better preserves the neighborhood structure (there is always a genotypic neighbor that corresponds to a phenotypic neighbor). Although the locality of Gray encoding also is not perfect (the number of genotypic neighbors is higher than the number of phenotypic neighbors), the performance of mutationbased search approaches on easy problems, and problems of bounded complexity, is higher when using Gray rather than binary encodings.
6 Analysis and Design of Representations for Trees
In the previous chapter, we illustrated that our framework modeling the inﬂuence of representations on the performance of GEAs not only works for binary phenotypes, but also for problems where the phenotypes are integers. However, it is possible to go one step further and to look at problems where the phenotypes and genotypes are completely diﬀerent. One example for these types of problems are tree optimization problems. Trees are special types of graphs. Representations for trees must incorporate the additional restriction of a graph to be a tree. Therefore, if the genotypes are strings, there is a large semantic gap between tree structures (phenotypes) and strings (genotypes). In contrast to general network problems, where a representation simply has to indicate which links are used for the graph, no natural or intuitive “good” tree representations exist which are accessible for GEAs. As a result, researchers have proposed a variety of diﬀerent tree representations with diﬀerent properties. However, up till now no theorybased analysis exists about how GEA performance is inﬂuenced by the diﬀerent types of tree representations. The purpose of this chapter is to ﬁll this gap and to analyze, based on the timequality framework from Chap. 4, the inﬂuence of some of the most widely used tree representations on GEA performance. We use the existing theory about redundant, exponentially scaled, and lowlocality representations to predict GEA behavior. Furthermore, the framework is used to design a new representation, the network random key (NetKey) representation. The analysis and design of direct representations, where both, genotypes and phenotypes, are trees is presented in Chap. 7. Because analyzing all known tree representations is beyond the scope of this work, we focus on some of the most widely used tree representations that assign trees to diﬀerent types of strings: Pr¨ ufer numbers (Pr¨ ufer 1918), the characteristic vector encoding (Celli et al. 1995; Berry et al. 1997; Ko et al. 1997; Dengiz et al. 1997c; Dengiz et al. 1997b; Dengiz et al. 1997a; Berry et al. 1999; Premkumar et al. 2001), and the link and node biased encoding (Palmer 1994). Analyzing these representations shows that Pr¨ ufer numbers have low locality, that the redundant
142
6 Analysis and Design of Representations for Trees
characteristic vector encoding is aﬀected by stealth mutation, and that the link and node biased encoding is not uniformly redundant. This chapter is structured as follows. In the ﬁrst section, we introduce the tree design problem and develop some basic requisites for graph problems. This is followed in Sect. 6.2 by an investigation into the Pr¨ ufer number encoding. It focuses on the Pr¨ ufer numbers’ missing high locality which is necessary for GEAs to perform well on easy problems and problems of bounded diﬃculty. Section 6.3 presents the characteristic vector encoding, which is a redundant encoding for trees. The encoding is uniformly redundant and the performance of GEAs is independent on the structure of the optimal solution. However, GEA performance is reduced as the encoding is nonsynonymously redundant. In Sect. 6.4, we show that the redundant link and node biased encoding is biased towards stars if a nodespeciﬁc bias is used and biased towards the minimum spanning tree if both, link bias and node bias, are small. Finally, Sect. 6.5 presents the new network random key encoding. In analogy to random keys, the links of a tree are represented as ﬂoating numbers, and a construction algorithm constructs the corresponding tree from the keys. The NetKey representation allows us to distinguish between important and unimportant links, is uniformly redundant, uniformly scaled, and has high locality. Due to its uniform redundancy, the performance of GEAs is independent of the structure of the optimal solution. The chapter ends with concluding remarks.
6.1 The Tree Design Problem This section provides the background for analyzing how tree representations aﬀect GEA performance. After a brief deﬁnition of the network design problem, Sect. 6.1.2 focuses on metrics and distances for graphs. This is followed by an illustration of diﬀerent tree structures like stars or lists. To be able to measure the phenotypic diﬃculty of a tree problem, we introduce in Sect. 6.1.4 a schema analysis for graphs. Based on the schema analysis, we present in Sect. 6.1.5 scalable test problems for graphs. The onemax tree problem, which is similar to the well known onemax problem, is a fully easy problem, whereas the deceptive trap for trees is fully diﬃcult. Finally, the section ends with a review of former design criteria for tree encodings as provided by Palmer (1994).
6.1.1 Deﬁnitions This subsection provides the necessary deﬁnitions for analyzing tree problems. We deﬁne a network as a graph G with n nodes and a maximum of n(n−1) links connecting the nodes. If the network is fully connected it has at least n − 1 links. We assume that all links are undirected (they can be used in both directions) and that a network is always fully connected. Therefore, the maximum number of possible links is n(n − 1)/2. The position of the nodes
6.1 The Tree Design Problem
143
in the graph is given a priori and the distances between two diﬀerent nodes a and b are deﬁned by using the Euclidean distance metric as (6.1) da,b = (xa − xb )2 + (ya − yb )2 , where x denotes the abscissa and y the ordinate of a node in a Cartesian coordinate system.
14
3
Figure 6.1. A 15 nodes tree with the path connecting nodes 3 and 14 emphasized.
The basic purpose of the network is to transport objects, for example goods or information, from some nodes in the network to other nodes. Therefore, a rule is necessary for how to transport the objects through the network. The rule for how to route the traﬃc through the network is based on the used routing algorithm. If the number of links in a fully connected network is larger than n − 1, the routing of the traﬃc through the network can be dynamically changed dependent on the current traﬃc load, the delay, the failure of nodes or links, or other criteria. In contrast, if the number of links in a fully connected network is equal to n − 1 there is only one unique path from every node to every other node and no dynamic routing is necessary. A tree T is deﬁned as an undirected and connected graph with no cycles. For a tree T with n nodes there are exactly n−1 links. It was found by Cayley (1889) that for a graph with n nodes, there are exactly nn−2 possible trees. A tree structure has some remarkable beneﬁts: It represents the network structure with the lowest number of possible links to still obtain a connected graph. Furthermore, no dynamic routing is necessary as there is only one possible path for the traﬃc between any two nodes (compare Fig. 6.1). Finally, the size of the search space Φtree  = nn−2 is much smaller than for general networks Φ 2n(n−1)/2 .1 However, the use of trees also has some drawbacks: Trees are very vulnerable to link or node failures. If one link or one node fails, the tree divides up into two unconnected subtrees which can not communicate with each other. However, despite this fact, trees are widely used for communication networks (Minoux 1987; Abuali et al. 1995; Elbaum and Sidi 1996; G¨ uls 1996; Tang et al. 1997; Brittain et al. 1997; Streng 1997; Gargano et al. 1998; Gerstacker 1999; Brittain 1999; Chu et al. 1999; Chu and Premkumar 1999; Knowles et al. 1999; Grasser 2000; Gaube 2000; Edelson and Gargano 1
We assume that there is only one possible capacity for a link. For diﬀerent types of lines with k diﬀerent capacities the number of possible network structure increases to Φ kn(n−1)/2 .
144
6 Analysis and Design of Representations for Trees
2000; Edelson and Gargano 2001; Premkumar et al. 2001; Chou et al. 2001). The network design problem itself is deﬁned as follows: Based on the • • • • •
number of network nodes n, locations of the n nodes, traﬃc demands between all n nodes, available capacities for the links, cost of the links dependent on the capacity and length,
we determine the • topology (structure) of the network, • capacity of the links, • routing of the traﬃc through the network. The general aim of the design process is to minimize the overall cost of the network with the constraint that all traﬃc demands between the nodes must be satisﬁed. If we focus on tree structures, the capacity of the links as well as the routing of the traﬃc is determined by the topology. This means for trees that the optimization problem simpliﬁes down to ﬁnding the optimal structure of the tree. 6.1.2 Metrics and Distances As illustrated in Sect. 3.3.2, a metric is necessary for the genotypic and phenotypic space Φg and Φp to deﬁne genetic operators like mutation or recombination. The application of the mutation operator to a genotype should result in the smallest possible change in the individual, and should generate an oﬀspring with distance 1 for the genotypes and the phenotypes. The recombination operator should ensure that the oﬀspring inherit substructures from the parents. In terms of metric, the distance between an oﬀspring and its parents should be lower than the distance between the two parents (compare Sect. 3.3.5). In accordance with Chap. 3, the Hamming metric (Hamming 1980) is used for the genotypes. Thus, the Hamming distance between two binary genotypes xg ∈ {0, 1}l and y g ∈ {0, 1}l of length l is deﬁned as dxg ,yg =
l−1
xgi − yig .
i=0
The distance d measures the number of alleles that are diﬀerent in both individuals. Similarly, the distance between two diﬀerent phenotypes (trees) is measured by using the Hamming distance dh for trees. The Hamming distance between two trees measures the number of diﬀerent links in the two trees. Therefore, the minimum Hamming distance between two diﬀerent trees is dh = 2.
6.1 The Tree Design Problem
G
G
i
j
145
Figure 6.2. Two graphs Gi and Gj with di,j = 1. The Hamming distance between the two graphs is 2.
As illustrated in Fig. 6.2 the minimal Hamming distance between two trees is two, although they have n − 2 links of all n − 1 links in common. To simplify the metric, we deﬁne the distance di,j ∈ {0, 1, 2, . . . n − 1} between two trees Gi and Gj by half of the number of diﬀerent links (di,j = 12 dhi,j ). It can be calculated as n−1 a−1 1 i j lab − lab , dpGi ,Gj = di,j = 2 a=1 b=0
i lab
where is 1 if the link from node a to node b exists in tree Gi and 0 if it does not exist in Gi . Then, the number of links that the two trees Gi and Gj have in common can easily be calculated as n − 1 − di,j . A mutation of a tree should result in the exchange of one link, and the distance between parent and child is dparent,child = 1. 6.1.3 Tree Structures When focusing on trees, diﬀerent basic topological structures can be identiﬁed. In general, we can distinguish between • stars, • lists, and • arbitrary trees. Figure 6.3 illustrates the diﬀerent tree types. The degree of a node is deﬁned as the number of links which are connected to the node.
(a) star
(b) list
(c) arbitrary tree
Figure 6.3. Diﬀerent tree structures
A star (Fig. 6.3(a)) has one center and all other nodes are connected to the center. Therefore, the center of the network has degree n − 1 and all other nodes have degree 1. For a network with n nodes there are n diﬀerent stars.
146
6 Analysis and Design of Representations for Trees
A failure of a link or a node (except the center) disconnects only the aﬀected node. However, if the center node fails, no further communication over the network is possible. For a list (Fig. 6.3(b)) two nodes have degree one (leaf nodes), and all other nodes have degree 2. There are many more possible lists than stars as the number of possible lists is 12 n!. A link or a node failure results in two separate sublists. Finally, there are arbitrary trees (Fig. 6.3(c)) which have no special structure except that they are trees. The degree of a node can vary from 1 to n − 1. As for stars and nlists, the sum over the degrees deg(i) of all n nodes can be calculated as i=1 deg(i) = 2(n − 1). 6.1.4 Schema Analysis for Graphs In this subsection, we deﬁne schemata for graphs in analogy to schemata deﬁned on bitstrings (compare Sect. 2.2.3). Schema analysis is helpful in determining whether graph problems are easy or diﬃcult to solve for selectorecombinative GAs. When assuming that GEAs process schemata, the analysis of schema ﬁtness is the appropriate method to measure problem diﬃculty (see Sect. 2.3.2). The BB hypothesis (see Sect. 2.2.3) deﬁnes building blocks to be highly ﬁt schemata of short deﬁning length and low order. Consequently, problems are fully easy if all schemata of order one that contain the optimum have higher ﬁtness than their competitors. Problems are diﬃcult if all lower order schemata containing the global optimum are inferior to some of their competitors. The onemax problem is an example of a fully easy problem, whereas the fully deceptive trap of order k is an example for a fully diﬃcult problem. B
2
1
6
5 A
C
4
3 D
Figure 6.4. Labeling of links for n = 4
We can measure the diﬃculty of network problems for selectorecombinative GEAs by introducing schema analysis for graphs. To formally deﬁne schemata, we have to label the possible links in a graph with numbers {1, 2, . . . , n(n − 1)/2}. Figure 6.4 illustrates an example of labeling the links in a graph with n = 4 nodes. Then, a schema is a string of length l = n(n − 1)/2 and the symbol at the ith position describes the existence of a link. 1 indicates that the link is established, 0 indicates no link, and * indicates don’t care (dashed line). Don’t care means that the link is either established or not. Figure 6.5 illustrates some possible schemata for a 4 node network using the labeling from Fig. 6.4.
6.1 The Tree Design Problem
11****
11**00
001100
1*1*1*
11111*
147
******
Figure 6.5. Some schemata for graphs
When using schemata for trees, there is the additional restriction that each tree has exactly n − 1 links and it must be connected. Therefore, there must be n − 1 ones in each solution string, and the string must encode a connected tree. This means that the average ﬁtness of a schema must be calculated only from the trees that are represented by the schema. Other nontrees that are represented by the schema do not aﬀect the schema ﬁtness. This implies, for example, that schemata with more than n − 1 ones, or more than 1 2 n(n − 1) − (n − 1) zeros, do not exist because they do not encode a valid tree. Using schema analysis, the diﬃculty of a network problem can easily be measured by the maximum order of the building blocks k. If a problem is fully easy then all lower order schemata that contain the global optimum are superior to their competitors. All building blocks have order one (k = 1). A problem is fully deceptive if all lower order schemata that contain the optimum are inferior to their competitors. To ﬁnd the optimum, GEAs must be able to ﬁnd BBs of order k = n(n − 1)/2. In general, the order of the largest BB determines the complexity of a problem. In contrast to binary strings, the length of the schemata has no meaning in the context of graphs as the labeling of the nodes does not aﬀect problem diﬃculty. We have seen in Sect. 3.3 that lowlocality representations modify problem diﬃculty. With the analysis of graph schemata, we can compare the diﬃculty of graph problems deﬁned on the phenotypes to the diﬃculty of the corresponding genotypic problems deﬁned on strings. This allows us to more easily recognize whether a tree encoding modiﬁes the diﬃculty of a problem. In the following paragraphs, we consequently deﬁne a fully easy and a fully diﬃcult scalable test problem based on the schema analysis for graphs. 6.1.5 Scalable Test Problems for Graphs To examine the performance of optimization algorithms for the topological design of trees, standard test problems should be used. Motivated by the previous subsection, we deﬁne a fully easy and a fully diﬃcult scalable tree problem. The onemax tree problem is based on the integer onemax problem (compare Sect. 5.1) (Ackley 1987). An optimal solution Topt is chosen either randomly or by hand. The structure of this optimal solution Topt can be determined: It can be a star, a list, or a random tree with n nodes.
148
6 Analysis and Design of Representations for Trees
fitness
fitness
For the calculation of the ﬁtness fi of a solution Ti , the distance di,j between two trees Ti and Tj is used (compare Sect. 6.1.2). Using this metric, the ﬁtness fi of a solution Ti depends on the distance di,opt between Ti and the optimal solution Topt . We can distinguish two types of problems: maximization and minimization problems. When deﬁning a minimization problem the ﬁtness fimin of an individual Ti is deﬁned as the distance di,opt to the optimal solution Topt . Therefore, fimin = di,opt , where fimin ∈ {0, 1, . . . , n − 1}. An individual has ﬁtness of n − 2 if it only has one link in common with the best solution. If the two individuals do not diﬀer (Ti = Topt ), the ﬁtness of Ti is fimin = 0. If our example tree from Fig. 6.7 is chosen as the optimal solution and we have a minimization problem, the star with center D would have ﬁtness (cost) of 1, because the two trees diﬀer at one edge2 (di,opt = 1).
n1 n2 n3
n1 n2 n3
2 1
2 1
d i,opt
0
1 2
n2n1
(a) onemax problem
tree
0
d i,opt 1 2
n2n1
(b) deceptive tree problem
Figure 6.6. Scalable maximization test problems for trees
trap
When deﬁning a maximization problem, the ﬁtness fimax of an individual Ti is deﬁned as the number of edges it has in common with the best solution Topt (compare Fig. 6.6(a)). Therefore, fimax = n − 1 − di,opt . If we have a maximization problem, and our example network from Fig. 6.7 is chosen as the optimal solution, the star with center D would have ﬁtness f max = 3 because the two networks have three links in common, and the distance between the two trees is 1. Because both test problems (minimization and maximization problem) are similar to the standard onemax problem, they are easy to solve for mutationbased GEAs, but somewhat harder for recombinationbased GAs (Goldberg et al. 1993). The existing knowledge about solving the standard integer onemax problem can be used for the onemax tree problem. Using the schemata for trees introduced in the previous subsection shows that all building blocks of the onemax tree problem have order 1. All schemata that contain the global optimum Topt are superior to their competitors. Therefore, the onemax tree problem is fully easy. For a network with 4 nodes, the schemata are already of length l = 6, and there are 36 = 729 diﬀerent schemata. Due to the limited space, we want to leave the explicit calculation 2
AC respectively AD
6.1 The Tree Design Problem A
C
B
D
E
schema schema ﬁtness 11**0* 2.33 11**1* 2 01**0* 1.67
10**0* 1.67
01**1* 1.33
10**1* 1.33 00**0* 1 00**1* 0.67
149
Figure 6.7. A ﬁve node tree represented trees 111000 110100 110001 110010 011100 011001 010101 101100 101001 100101 011010 010110 010011 101010 100110 100011 001101 001110 001011 000111
ﬁtness of trees 3 2 2 2 2 2 1 2 2 1 2 1 1 2 1 1 1 1 1 0
Table 6.1. An example of calculating the average schema ﬁtness for a 4 node onemax tree maximization problem where 111000 is the optimal solution. All schemata (in our example 11**0*) which contain the global optimum have higher ﬁtness than their competitors. The problem is fully easy.
of all schema ﬁtnesses to the reader. We only illustrate in Table 6.1 the ﬁtness calculation for schemata where the ﬁrst, second, and ﬁfth position are ﬁxed for a 4 node onemax tree problem. The optimal tree is deﬁned as xgopt = 111000. Obviously, the schema 11**0* which contains the global optimum is superior to all its competitors. In analogy to this fully easy onemax tree problem, we deﬁne a fully diﬃcult deceptive trap problem for trees. As before, we choose an optimal solution Topt with ﬁtness n − 1 (assuming a maximization problem) either by hand or randomly. Then, the ﬁtness of all other individuals Ti = Topt is deﬁned as fi = dopt,i − 1. The ﬁtness function is illustrated in Fig. 6.6(b). This problem is fully diﬃcult as all schemata with k < n(n − 1)/2 containing the global optimum are inferior to their misleading competitors. Both, mutationbased and crossoverbased search approaches have great problems in ﬁnding the global optimum. Using again the example network from Fig. 6.7 as the optimal solution, a star with center D has ﬁtness 0 because the distance to the optimal solution is 1. The optimal solution itself has ﬁtness 4.
150
6 Analysis and Design of Representations for Trees
6.1.6 Tree Encoding Issues We review the tree encoding issues as described by Palmer (1994) and Palmer and Kershenbaum (1994b) and relate them to the insights into the basic elements of representation theory we gained in Chap. 3. According to Palmer (compare also Sect. 2.4.3), tree representations should possess the following properties: • • • •
A representation should be able to represent all possible trees. It should be unbiased in the sense that all trees are equally represented. A representation should be capable of representing only trees. The construction of the phenotype from the genotype and vice versa should be easy. • A representation should possess locality concerning small changes. • The schemata should encourage short, low order schemata. We discuss these issues and relate them to the theory of representations outlined in Chap. 3 and 4. The issue that a representation should be able to represent all possible trees is almost trivial. As long as we have no special knowledge about the problem we want to solve, it makes no sense to use a representation that might not represent some of the possible solutions. Otherwise, it could happen that GEAs search for the optimal solution, but the optimal solution can never be reached because it can not be encoded. If we have knowledge about the optimization problem, we can weaken this issue and demand representations to at least encode all the solutions we are interested in, and those which could be the optimal solution. A representation is unbiased if all tress are represented by the same number of genotypes. Problems with biased encodings can be explained by the concept of redundant encodings illustrated in Sect. 3.1. If some phenotypes are overor underrepresented, the encoding is biased, and the performance of GEAs is changed. The inﬂuence on the performance of GEAs by biased encodings can be modeled by using the Gambler’s ruin model (Harik et al. 1999). Section 3.1 has shown that as long as the highquality solutions are overrepresented, a bias increases performance. If the highquality solutions are underrepresented, a decline of GEA performance is unavoidable. Therefore, with respect to a robust encoding which can be used for problems of unknown complexity, it is desirable to use unbiased encodings. Some tree representations can also represent nontrees. These kind of representations are aﬀected by two problems: Firstly, it could be diﬃcult to generate valid initial populations. Secondly, the application of genetic operators can result in invalid solutions. The question arises of how to handle invalid solutions, and what to do with nontrees. In general, there are two possibili
6.2 Pr¨ ufer Numbers
151
ties3 : Invalid solutions can either be repaired, or they can be left unchanged in the population and hopefully they will disappear by the end of the run. Repairing invalid solutions means that some of the trees are represented not only by valid individuals but also by some invalid solutions. Therefore, the representation is redundant. Phenotypes are not uniformly represented by the genotypes if the repair process is somehow shifted. Only a completely unbiased repair process which does not favor some tree structures guarantees an unbiased population and uniform redundancy. To keep invalid solutions in the population could sometimes be helpful for GEAs (Orvosh and Davis 1993). Nevertheless, it must be ensured that the optimal solution at the end of the run is valid. Otherwise, the application of GEAs to tree design problems is useless as it does not result in valid solutions. To drive GEAs towards valid solutions, researchers often use penalties for invalid solutions. However, additional penalties change the ﬁtness function and with it the behavior of GEAs. Therefore, they should be used very carefully. An easy construction of the phenotype from the genotype, and vice versa, is necessary for an eﬃcient implementation of GEAs. However, it depends on the complexity of the ﬁtness function whether the computational eﬀort for the genotypephenotype mapping signiﬁcantly aﬀects the run duration of the computer experiments. In contrast to costly ﬁtness evaluations, a slightly more complicated genotypephenotype mapping could often be neglected. The problem of locality is part of the larger question of how well the encoding preserves the complexity of a problem. As illustrated in Sect. 3.3.4, high locality guarantees that the problem diﬃculty remains unchanged for mutationbased search. If the locality of an encoding is low, it becomes more diﬃcult to solve easy problems and problems of bounded diﬃculty. Finally, Palmer listed Goldberg’s basic design principle of meaningful building blocks (compare Sect. 2.4.1) and demanded encodings to encourage short, low order schemata. Otherwise, “long schemata cause genetic algorithms to drift” (Palmer 1994, p. 40). However, as illustrated in Sect. 3.2, drift is caused by nonuniformly scaled alleles and domino convergence, and not by the length and the size of the building blocks. The size and length of the building blocks determine the complexity of a problem for selectorecombinative GAs. We recognize that the tree design issues from Palmer can be well understood by using the framework presented in Chap. 4.
6.2 Pr¨ ufer Numbers Pr¨ ufer numbers are a widely used representation for trees. The purpose of this section is to use the framework from Chap. 4 for an investigation into the 3 Of course, there is a third possibility: To remove the individual from the population. However, we do not consider this case.
152
6 Analysis and Design of Representations for Trees
properties of Pr¨ ufer numbers. The analysis focuses on the low locality of the encoding and shows how GEA performance is aﬀected. The section starts with an historical review of the use of the Pr¨ ufer number encoding in the context of genetic and evolutionary search. The review shows a strong increase in interest into the encoding over the last 5 to 10 years. This is followed by the construction and deconstruction process of Pr¨ ufer numbers. Section 6.2.3 illustrates the beneﬁts and drawbacks of the encoding. The use of the Pr¨ ufer number encoding is very charming due to its advantageous properties, although the low locality of the encoding has already been identiﬁed by Palmer (1994) to be its main drawback. Subsequently, in Sect. 6.2.4 we present a deeper investigation into how exactly the low locality damages the performance of GEAs. In analogy to Sect. 3.3, we illustrate why high locality is necessary for an encoding to preserve problem diﬃculty and perform random walks through the search space. After an analysis of the neighborhood structure of Pr¨ ufer numbers, we ﬁnally present empirical results for diﬀerent tree structures using mutation and recombinationbased evolutionary search methods. The section ends with concluding remarks. 6.2.1 Historical Review We give a brief historical review of the development and use of the Pr¨ ufer number encoding in the context of GEAs. Cayley (1889) identiﬁed the number of distinct spanning trees on a complete graph with n nodes as nn−2 (Even 1973, pp. 103104). Later, this theorem was very elegantly proven by Pr¨ ufer (1918) by the introduction of a onetoone correspondence between spanning trees and a string of length n−2 over an alphabet of n symbols. This string is denoted as Pr¨ ufer number, and the genotypephenotype mapping is the Pr¨ ufer number encoding. It is possible to derive a unique tree with n nodes from the Pr¨ ufer number of length n − 2 and vice versa (Even 1973, pp. 104106). Of course there are other onetoone mappings from strings of n − 2 labels onto spanning trees on the n labeled links. One example is the Blob Code which was developed and proposed by Picciotto (1999). Julstrom (2001) compared this encoding to Pr¨ ufer numbers and found for easy problems a higher performance of GEAs using the Blob Code than Pr¨ ufer numbers. Later, in the context of GEAs, several researchers used the Pr¨ ufer number encoding for the representation of trees. Palmer used the encoding in his doctoral thesis at the beginning of the nineties (Palmer 1994; Palmer and Kershenbaum 1994a; Palmer and Kershenbaum 1994b), and compared the performance of Pr¨ ufer numbers with some other representations for the optimal communication spanning tree problem. However, he noticed that the Pr¨ ufer number encoding has low locality and therefore is not a good choice for encoding trees. The low performance of the encoding was conﬁrmed by Julstrom (1993) who used Pr¨ ufer numbers for the rectilinear Steiner problem, and also observed low GEA performance using this encoding.
6.2 Pr¨ ufer Numbers
153
About the same time, Abuali et al. (1994) used Pr¨ ufer numbers for the optimization of probabilistic minimum spanning trees (PMST) with GEAs. The investigation focused more on the inﬂuence of diﬀerent operators than on the performance of Pr¨ ufer numbers. However, at the end of the work, the conclusion was drawn that in contrast to Palmer and Julstrom, Pr¨ ufer numbers “lead to a natural GEA encoding of the PMST problem” (Abuali et al. 1994, p. 245). Some years later, similar results were reported by Zhou and Gen (1997) who successfully used the Pr¨ ufer encoding for a degree constraint minimum spanning tree problem. The degree constraint was considered by repairing invalid solutions that violate the degree constraints. Furthermore, Pr¨ ufer numbers were used for spanning tree problems (Gen et al. 1998; Gen et al. 1998), the timedependent minimum spanning tree problem (Gargano et al. 1998), the ﬁxedcharge transportation problem (Li et al. 1998) and a bicriteria version of it (Gen and Li 1999), and a multiobjective network design problem (Kim and Gen 1999). Most of this work reported good results when using Pr¨ ufer numbers, and labeled the encoding to be (very) suitable for encoding spanning trees. As an example of positive results we want to cite Kim and Gen (1999), who wrote: “The Pr¨ ufer number is very suitable for encoding a spanning tree, especially in some research ﬁelds, such as transportation problems, minimum spanning problems, and so on.”4 However, other relevant work by Krishnamoorthy et al. (1999), who used Pr¨ ufer numbers for the degree constraint spanning tree problem, from Julstrom (2000) who compared a list of edges encoding with Pr¨ ufer numbers, or from Gottlieb and Eckert (2000) who used Pr¨ ufer numbers for the ﬁxed charge transportation problem showed that Pr¨ ufer numbers result in a low GEA performance. A summarizing study by Gottlieb et al. (2001) compared the performance of Pr¨ ufer numbers for four diﬀerent network problems and concluded that Pr¨ ufer numbers always perform worse than other encodings, and are not suitable for encoding trees when using GEAs. To explain the diﬀerences between the good and bad results obtained by GEAs using Pr¨ ufer numbers, Rothlauf and Goldberg (1999) investigated the locality of the encoding more closely. It was shown that Pr¨ ufer numbers only have high locality if they encode stars. For all other tree types the locality is low which leads to a degradation of GEAs (see also Rothlauf and Goldberg (2000) and Rothlauf et al. (2001)). Therefore, the diﬀerences in performance could be well explained if one assumes that the performance of GEAs depends on the structure of the optimal solution. Obviously, researchers who report good solutions when using Pr¨ ufer numbers used problems where the optimal solution is more starlike and therefore easy to ﬁnd for GEAs. However, when using Pr¨ ufer numbers for more general, nonstar like problems, a strong decrease in GEA performance is inescapable. The results from Roth4
Special thanks to Bryant A. Julstrom for his help with ﬁnding this statement.
154
6 Analysis and Design of Representations for Trees
lauf and Goldberg (1999) were conﬁrmed by Gottlieb and Raidl (2000) who investigated the eﬀects of locality on the dynamics of evolutionary search. We have seen that the performance of GEAs using Pr¨ ufer numbers is a strongly discussed topic. Some researchers report good results and favor the use of Pr¨ ufer numbers. Other researchers, however, point to the low locality of the encoding, report worse results and advise us not to use Pr¨ ufer numbers. A closer investigation into how locality depends on the structure of the tree could solve these contradictory results. As the work from Rothlauf and Goldberg (2000) indicates that the locality of Pr¨ ufer numbers strongly depends on the structure of the tree, GEAs show good results if the good solutions are starlike, and worse results for all other types. In Sect. 6.2.4, we review the main results from Rothlauf and Goldberg (1999) and Rothlauf and Goldberg (2000) and extend it with additional work. 6.2.2 Construction We review the construction rule for the Pr¨ ufer number encoding. We present both sides of the story: How a Pr¨ ufer number can be constructed from a tree, and how a tree can be constructed from a Pr¨ ufer number. The Construction of the Pr¨ ufer Number from a Tree The degree deg(i) of a node i denotes the number of links that are connected to the node. Thus, as a fully connected tree has exactly n − 1 links, the degree deg(i) of a node i lies between 1 and n − 1. A node has degree one if it is a leaf node. It has degree n − 1 if it is the center of a star. There are always at least two nodes which have degree 1. The Pr¨ ufer number itself encodes an nnode tree with a string of length n − 2, and each element of the string is of base n. As the mapping is onetoone, a Pr¨ ufer number is a unique encoding of a tree, and there are nn−2 diﬀerent possible Pr¨ ufer numbers (Cayley 1889; Pr¨ ufer 1918). For the construction of the Pr¨ ufer number from a tree, we label all nodes with numbers from 1 to n. Then, the Pr¨ ufer number can be constructed from a tree by the following algorithm: 1. Let i be the lowest numbered node of degree 1 in the tree (deg(i) = 1). 2. Let j be the one node which is connected to i (there is exactly one). The number of the jth node is the furthest right digit of the Pr¨ ufer number. 3. Remove node i and the link (i, j) from the tree and from further consideration. 4. Go to 1 until only two nodes (that means one link) are left. After termination of the construction rule, we have a Pr¨ ufer number with n−2 digits which represents the tree. An eﬃcient implementation of this algorithm uses a priority queue implemented in a heap to hold the nodes of degree 1. The algorithm’s time complexity is then O(n log n).
6.2 Pr¨ ufer Numbers 1
2
3
4
5
6
155
Figure 6.8. A tree and the corresponding Pr¨ ufer number P = 2565
Let us demonstrate the construction of the Pr¨ ufer number with a brief example. The network in Fig. 6.8 has 6 nodes. Therefore, the Pr¨ ufer number consists of 4 digits. The lowest numbered node with degree 1 is node 1. This node is connected to node 2 so the Pr¨ ufer number starts with a 2. We remove node 1 from further consideration and search for the lowest numbered node with degree 1. We identify node 2 which is connected to node 5. The Pr¨ ufer number becomes 25. After removing node 2, node 3 is the lowest numbered node which is eligible (it has degree 1). Node 3 is connected to 6 so we get 256. The node 4 is the lowest eligible node and we add 5 to the Pr¨ ufer number. Finally, only two nodes remain in the tree. The algorithm stops and the resulting Pr¨ ufer number is 2565. The Construction of the Tree from the Pr¨ ufer Number The construction of the tree from the Pr¨ ufer number follows the construction of the Pr¨ ufer number from the tree. It goes as follows: 1. Let P be a Pr¨ ufer number with n − 2 digits. All node numbers which are not in P can be used for the construction of the tree (are eligible). 2. Let i the lowest numbered eligible node. Let j be the leftmost digit of P . 3. Add the link (i, j) to the tree. 4. Designate i as no longer eligible and remove the leftmost digit j from the Pr¨ ufer number. 5. If j does not occur anywhere else in the remaining Pr¨ ufer number, designate j as eligible. 6. Go to 2 until no digits remain in the Pr¨ ufer number. If no digits are left, then there are exactly two numbers, r and s, which are eligible. Finally, add the link (r, s) to the tree. We also illustrate this construction rule with a brief example. We want to construct the tree from the Pr¨ ufer number P = 2565. Eligible nodes are 1, 3 and 4. As 1 is the lowest eligible node, and 2 is the leftmost digit of the Pr¨ ufer number, we add the link (1, 2) to the tree. 1 is then no longer eligible, and 2 does not occur anywhere else in the string. Therefore, the nodes 2, 3 and 4 are eligible and P becomes 565. Now, 2 is the lowest eligible node and we add the link (2, 5) to the tree. As 5 occurs somewhere else in the string, we do not designate 5 as eligible. Thus, we only remove 2 from our pool of eligible numbers, and then we can add the link (3, 6) to the tree. Now, only the nodes 4 and 6 are eligible and P = 5. We continue with adding (4, 5). Finally, all
156
6 Analysis and Design of Representations for Trees
digits are removed from P and the numbers 5 and 6 remain eligible. The link (5, 6) is added to the tree and the algorithm terminates. We have constructed the tree illustrated in Fig. 6.8. 6.2.3 Properties We analyze the properties of the Pr¨ ufer number encoding by using the design issues from Palmer and Kershenbaum (1994a) and Palmer (1994) (see Sect. 6.1.6). Furthermore, we relate these properties to the more general properties of representations we developed in Chap. 3. Beneﬁts The Pr¨ ufer number encoding is a very elegant and interesting encoding with some remarkable beneﬁts: • • • •
Every tree can be represented by a Pr¨ ufer number. Only trees are represented by Pr¨ ufer numbers. Every Pr¨ ufer number represents exactly one tree. All trees are represented uniformly (unbiased).
A look at the construction rule of the Pr¨ ufer number shows that a Pr¨ ufer number is able to represent all possible trees. Because every tree has at least two nodes with degree 1, the construction rule can be applied to every tree. The user should notice that the original intent of the Pr¨ ufer number was to prove Cayley’s theorem (Cayley 1889) by introducing Pr¨ ufer numbers. It was also shown by Pr¨ ufer (1918) that Pr¨ ufer numbers only represent trees. Therefore, a Pr¨ ufer number can be randomly created and it always represents a tree. In contrast to many other representations, no repairing of a randomly chosen individual is necessary. Furthermore, it is also not necessary to repair individuals that are generated by genetic operators in each generation. The ﬁrst three beneﬁts of the Pr¨ ufer numbers can be summarized by denoting the Pr¨ ufer number encoding as a onetoone mapping. The mapping is not only surjective, but also bijective. One consequence of a onetoone mapping is that all trees are uniformly represented as each tree is represented by exactly one speciﬁc Pr¨ ufer number. The number of diﬀerent trees for a graph with n nodes is nn−2 , and there ufer numbers for an n node tree. Therefore, are also exactly nn−2 diﬀerent Pr¨ GEAs using Pr¨ ufer numbers have no problems with redundancy. GEAs using Pr¨ ufer numbers can not be aﬀected by the over or underrepresentation of some individuals. These advantages make Pr¨ ufer numbers an interesting encoding for trees. However, the use of Pr¨ ufer numbers is connected to some serious drawbacks.
6.2 Pr¨ ufer Numbers
157
Drawbacks The Pr¨ ufer number has the disadvantages of • complex calculation and • low locality. In comparison to some other representations, the construction of the Pr¨ ufer number is more complex and not straightforward. But, it can be done using the help of a heap in O(n log n). This seems to be acceptable for most problems. The most important disadvantage of the Pr¨ ufer number is the low locality of the representation. Small changes in the Pr¨ ufer number string can lead to large changes in the represented network. This means, the mapping from the phenotype to the genotype is not homogeneous. Therefore, the basic mutation operator that searches the local solution space around an individual does not generate oﬀspring that are similar to their parents. A descendant does not inherit the important properties of its parents. Thus, mutation works not as a local search, but more as a random search over the solution space (compare Sect. 3.3.4).
2
Prüfer number: 2 2 3 3
2 2 3 1
1
4
5
(a) A tree and its Pr¨ ufer number P = 2233
2
Prüfer number:
3 6
1
3 4
5
6
(b) A tree and its Pr¨ ufer number P = 2231
Figure 6.9. The low locality of the Pr¨ ufer number encoding. A change of one digit changes three links in the corresponding tree.
A small example illustrates the low locality of the encoding. Changing the last digit in the Pr¨ ufer number of Fig. 6.9(a) from 3 to 1 yields 2231, which decodes to the links (2,4), (2,5), (3,2), (1,3), and (1,6) (compare Fig. 6.9(b)). Only two of the original tree’s ﬁve links exist in the oﬀspring. 6.2.4 The Low Locality of the Pr¨ ufer Number Encoding As illustrated in the previous subsection, the Pr¨ ufer number encoding is affected by low locality. The purpose of this subsection is to investigate the locality of the encoding more closely. The locality of Pr¨ ufer numbers is examined by performing two diﬀerent investigations: Firstly, we perform random walks through the search space and examine the distances between parents and oﬀspring. Secondly, we investigate the neighborhood of the genotypes and phenotypes. We examine the locality of the neighboring individuals, and determine their number. Finally, we present an empirical veriﬁcation of the theoretical predictions for mutation and crossoverbased evolutionary search.
158
6 Analysis and Design of Representations for Trees
Random Walks We present experiments for evaluating the locality of the Pr¨ ufer number encoding. For this purpose we perform random walks through the genotypic and phenotypic search space and analyze the resulting changes in the corresponding phenotypes/genotypes. ufer number Figure 6.10 shows the encoding of a tree xp ∈ Φp as a Pr¨ g1 ufer number as a bitstring xg ∈ Φg . The x ∈ Φg1 and the encoding of the Pr¨ ufer number itself genetic operators are applied to the genotypes xg . The Pr¨ is a sequence of integers and is represented as a bitstring using the binary encoding (compare Sect. 5.2). Therefore, the mapping from the bitstring to the Pr¨ ufer number fg : Φg → Φg1 is aﬀected by scaling and has the properties discussed in Sect. 5.3.3. The mapping from the Pr¨ ufer numbers to the trees fg1 : Φg1 → Φp is described in Sect. 6.2.2. Notice that for a tree with n nodes, the Pr¨ ufer number has n − 2 digits, and the bitstring (n − 2)log2 (n) bits. tree
1
2
3
4
5
6
xp
Prüfer number 2 5 6 5
xg1
bitstring
xg
010 101 110 101
Figure 6.10. A tree, its Pr¨ ufer number and the corresponding bitstring
The locality of the Pr¨ ufer number encoding can be measured by performing a random walk through one of the solution spaces Φg , Φg1 , or Φp , and measuring the distances between parent xp and oﬀspring xo in the other two solution spaces. A random walk through a search space Φ is deﬁned by performing iteratively small changes. Therefore, the distance between parent xp and oﬀspring xo in the solution space we are performing our random walk in is dxp ,xo = 1. According to Fig. 6.10 there are three diﬀerent possibilities: 1. A random walk through Φg (a step changes one bit of the bitstring, dgxo ,xp = 1). ufer num2. A random walk through Φg1 (a step changes one digit of the Pr¨ = 1). ber, dg1 xo ,xp 3. A random walk through Φp (a step changes one edge in the tree, dpxo ,xp =1). A random walk through Φg means randomly changing one bit of xg and examining how many links change in the corresponding tree xp . Furthermore, the change of one bit in the bitstring xg results in the change of exactly one digit of the Pr¨ ufer number xg1 . A random walk through Φg1 means randomly changing one digit of the Pr¨ ufer number xg1 and measuring how many links are diﬀerent in the resulting xp . Notice that the change of one digit in xg1 results in up to log2 (n) diﬀerent bits in xg . A random walk through Φp means
1
cumulative frequency
cumulative frequency
6.2 Pr¨ ufer Numbers
0.8 0.6 0.4 0.2 0 0
2
4
6
8
1 0.8 0.6 0.4 0.2 0
10
0 2 4 6 8 10 12 14 16
p
d
159
d
xp,xo
(a) n = 16
p xp,xo
(b) n = 32
1
cumulative frequency
cumulative frequency
Figure 6.11. Distribution of phenotypic distances dpxp ,xo for neighboring bitstrings xp , xo ∈ Φg on 16 and 32 nodes. We perform a random walk through Φg (dgxp ,xo = 1), and show how many links are diﬀerent if one bit of the bitstring is changed.
0.8 0.6 0.4 0.2 0 0
2
4
6 p
d
xp,xo
(a) n = 16
8
10
1 0.8 0.6 0.4 0.2 0 0 2 4 6 8 10 12 14 16 18 d
p xp,xo
(b) n = 32
ufer Figure 6.12. Distribution of phenotypic distances dpxp ,xo for neighboring Pr¨ numbers xp , xo ∈ Φg1 on 16 and 32 nodes. We perform a random walk through Φg1 (dg1 ufer number xp ,xo = 1), and show how many links are diﬀerent if one digit of the Pr¨ is changed.
that one link of the tree xp is replaced by a randomly chosen link, and the diﬀerence of bits/digits in the bitstring/Pr¨ ufer number is examined. Figures 6.11, 6.12, 6.13, and 6.14 present the results of the random walks. We show the cumulative frequency over the distances dxp ,xo between the parent xp and the oﬀspring xo . The phenotypic distance between the parent xp and oﬀspring xo is denoted as dpxp ,xo (for trees xp , xo ∈ Φp ), as dg1 xp ,xo (for Pr¨ ufer numbers xp , xo ∈ Φg1 ), or as dgxp ,xo (for bitstrings xp , xo ∈ Φg ). In all our experiments, the start individual for the bitstring, the Pr¨ ufer number, or the tree is chosen randomly. To gain statistically signiﬁcant results independently of the start individual, 400 steps (mutations) were carried out in each of the 20 runs. Thus, we performed overall 8,000 steps in the search space.
6 Analysis and Design of Representations for Trees
1
cumulative frequency
cumulative frequency
160
0.8 0.6 0.4 0.2 0 0
1 0.8 0.6 0.4 0.2 0
5 10 15 20 25 30 35
0
20
g
d
40 d
xp,xo
(a) n = 16
60
80
100
g xp,xo
(b) n = 32
1
cumulative frequency
cumulative frequency
Figure 6.13. Distribution of genotypic distances dgxp ,xo for neighboring trees xp , xo ∈ Φp on 16 and 32 nodes. We perform a random walk through Φp (dpxp ,xo = 1), and show how many bits are diﬀerent if one link of the tree is changed.
0.8 0.6 0.4 0.2 0 0
2
4
6
8 10 12 14
g1 d x ,x p o
(a) n = 16
1 0.8 0.6 0.4 0.2 0 0
5
10 15 20 25 30 d
g1 xp,xo
(b) n = 32
Figure 6.14. Distribution of genotypic distances dg1 xp ,xo for neighboring trees xp , xo ∈ Φp on 16 and 32 nodes. We perform a random walk through Φp (dpxp ,xo = 1), and show how many digits of the Pr¨ ufer number are diﬀerent if one link of the tree is changed.
The results for a random walk through Φg (Fig. 6.11) and Φg1 (Fig. 6.12) show that only about 40% of the one bit/digit changes lead to a change of one link in the tree (dpxp ,xo = 1). More than 35% (16 nodes) or 50% (32 nodes) of all one bit/digit changes result in trees xpo with at least four diﬀerent links (dpxp ,xo ≥ 4). Therefore, the locality of the mapping from the genotype (bitstring as well as Pr¨ ufer number) to the phenotype is low. Low genotypic distances (dgxp ,xo = 1 or dg1 xp ,xo = 1) do not correspond to low phenotypic distances dpxp ,xo . When walking through the phenotypic solution space (trees) the plots in Fig. 6.13 show that only about 50% of all one link changes result in a change of less than eight bits (16 nodes), respectively 20 bits (32 nodes) in the bitstring.
6.2 Pr¨ ufer Numbers
161
For the Pr¨ ufer number (Fig. 6.14) about 75% of the neighboring trees are diﬀerent in more than one digit. Therefore, the locality of the phenotypegenotype mapping is also low. The random walks through Φg and Φp have shown that the locality of the Pr¨ ufer number representation is low. Most of the small steps in the phenotypic and genotypic search space result in unacceptably high changes in the corresponding genotypic and phenotypic search space. The following paragraphs investigate whether the locality of the Pr¨ ufer numbers encoding is uniformly low everywhere in the search space, or if there are some areas of high locality.
Analysis of the Neighborhood Performing random walks through the diﬀerent search spaces has revealed that the locality of the Pr¨ ufer number encoding is low. Therefore, we investigate whether the locality of the encoding is uniformly low, or if there are diﬀerences in locality for diﬀerent areas of the search space. The search space can be separated into diﬀerent areas by making assumptions about the structure of the represented tree such as being a star or a list. To investigate whether the locality of the Pr¨ ufer number encoding is different for diﬀerent areas of the search space, we choose an individual x with speciﬁc properties and examine its locality. We examine all individuals y with distance dx,y = 1 and measure the resulting genotypic or phenotypic distance. As an individual is deﬁned to be a neighbor to another individual if the distance between the two individuals is 1, our examination is nothing more than an examination of the neighborhood of speciﬁc individuals. In analogy to the random walks, we investigate the neighborhood of individuals in all three search spaces Φg , Φg1 and Φp : • Neighborhood of an individual xg ∈ Φg (all neighbors y g ∈ Φg that are diﬀerent in one bit from the examined individual, dgxg ,yg = 1). • Neighborhood of a Pr¨ ufer number xg1 ∈ Φg1 (all neighbors y g1 ∈ Φg1 that g1 diﬀer in one digit, dxg1 ,yg1 = 1). • Neighborhood of a tree xp ∈ Φp (all neighbors y p ∈ Φp that have distance dpxp ,yp = 1). We examine the complete neighborhood of an individual either in Φg , Φg1 , or Φp and measure the corresponding distances in the two others. This investigation can be performed for four diﬀerent types of networks: (i) Star: One node is of degree n − 1 and the rest of the nodes have degree 1. (ii) Random list: Two nodes are of degree 1 (the ﬁrst and the last node of the list) and all other nodes have degree 2. The numbering of the nodes is random.
162
6 Analysis and Design of Representations for Trees
(iii) Ordered list: Like random list, but the nodes in the list are connected in ascending order. Node k is connected to k + 1, node k + 1 is connected to k + 2 and so on. If the highest numbered node n is not a leaf node, then it is connected to node 1. (iv) Tree: An arbitrary tree. We distinguish between ordered and random lists because the locality of the Pr¨ ufer number encoding is slightly diﬀerent for ordered and random lists. Figures 6.15, 6.16, 6.17, and 6.18 examine the neighborhood of stars, random and ordered lists, and arbitrary trees with 16 and 32 nodes. A bitstring representing a tree has either length l = 56 (16 nodes) or l = 150 (32 nodes); a Pr¨ ufer number encoding a tree has either 14 (16 nodes) or 30 (32 nodes) digits. For every problem instance, the complete neighborhood of 1000 randomly chosen individuals x is examined. Figure 6.15 shows distributions of phenotypic distances dpxp ,yp between a randomly chosen bitstring xg and all neighboring y g , where dgxg ,yg = 1. Figure 6.16 shows distributions of phenoufer number xg1 and all typic distances dpxp ,yp between a randomly chosen Pr¨ g1 g1 neighboring y , where dxg1 ,yg1 = 1. Figures 6.17 and 6.18 show distributions of genotypic distances (dgxg ,yg and dg1 xg1 ,y g1 ) between a randomly chosen tree xp and all neighboring graphs y p , where dpxp ,yp = 1; that is, for spanning trees that diﬀer in one link. Figures 6.15 and 6.16 reveal that the neighborhood of a bitstringencoded tree, as well as a Pr¨ ufer number representing a spanning tree, depends on the structure of the encoded tree. If the bitstring/Pr¨ ufer number encodes a star, all genotypic neighbors also have a phenotypic distance of one. This means that the locality of the bitstring/Pr¨ ufer number is perfect for stars. If the bitstring/Pr¨ ufer number encodes an ordered list, the genotypic neighbors have a maximum phenotypic distance of 4, independently of the number of nodes. This means, a change of one bit/digit of xg or xg1 that encode an ordered list results in a maximum phenotypic distance dpxp ,yp = 4. In contrast, bitstrings and Pr¨ ufer numbers that encode random lists or random trees, show low locality, and most of the genotypic neighbors are phenotypically completely diﬀerent. The neighborhoods of trees are examined in Figs. 6.17 and 6.18. Similarly to the genotypic neighborhood, all neighbors of a star have a genotypic distance in the Pr¨ ufer number space of one (dg1 xg1 ,y g1 = 1 illustrated in Fig. 6.18). g ufer number y g1 using a binary encodHowever as a bitstring y encodes a Pr¨ ing, up to log2 (n) bits are changed in the bitstring y g that represents the ufer number Pr¨ ufer number y g1 (Fig. 6.17). The change of one digit in a Pr¨ can result in a change of up to log2 (n) bits in the bitstring. Therefore, the locality of the Pr¨ ufer number encoding is high for stars. For arbitrary trees and lists, the change of one link often results in a completely diﬀerent bitstring/Pr¨ ufer number. Therefore, the locality of the Pr¨ ufer number encoding is low for trees and lists.
1
cumulative frequency
cumulative frequency
6.2 Pr¨ ufer Numbers
0.8 0.6 star tree ordered list random list
0.4 0.2 0 0
2
4
6
8
163
1 0.8 0.6 0.4 0.2
star tree ordered list random list
0
10
0 2 4 6 8 10 12 14 16
dpxp,yp
dpxp,yp
(a) 16 nodes
(b) 32 nodes
1
cumulative frequency
cumulative frequency
Figure 6.15. Distribution of phenotypic distances dpxp ,yp , where {y g dgxg ,yg = 1}. The graphs illustrate how many links are diﬀerent in the tree y p when examining the complete neighborhood of a randomly chosen genotypic bitstring xg .
0.8 0.6 star tree ordered list random list
0.4 0.2 0 0
2
4
6 p p p x ,y
d
(a) 16 nodes
8
10
1 0.8 0.6 0.4 0.2
star tree ordered list random list
0 0 2 4 6 8 10 12 14 16 18 d
p p p x ,y
(b) 32 nodes
= 1}. Figure 6.16. Distribution of phenotypic distances dpxp ,yp , where {y g1 dg1 xg1 ,y g1 The graphs illustrate how many links are diﬀerent in the tree y p when examining the complete neighborhood of a randomly chosen genotypic Pr¨ ufer number xg1 .
The results show that the locality of the Pr¨ ufer number encoding is highly irregular and depends on the phenotypic structure of the encoded tree. If a Pr¨ ufer number encodes a list or arbitrary tree, the locality of the encoding is low. Then, most of the genotypic neighbors of a Pr¨ ufer numberencoded genotype are phenotypically completely diﬀerent. However, if Pr¨ ufer numbers encode stars, the locality of the encoding is perfect. All genotypic neighbors of a star are also phenotypic neighbors. These results raise two new questions: Why do Pr¨ ufer numbers which encode stars have high locality? How large are the areas of high locality? We answer these questions in the following paragraphs.
1 0.8 0.6 star tree ordered list random list
0.4 0.2 0 0
cumulative frequency
6 Analysis and Design of Representations for Trees
cumulative frequency
164
1 0.8 0.6 star tree ordered list random list
0.4 0.2 0
5 10 15 20 25 30 35
0
20
40
60
dgxg,yg
dgxg,yg
(a) 16 nodes
(b) 32 nodes
80
100
1 0.8 0.6 star tree ordered list random list
0.4 0.2 0 0
2
4
6 d
8 10 12 14
g1 g1 g1 x ,y
(a) 16 nodes
cumulative frequency
cumulative frequency
Figure 6.17. Distribution of genotypic distances dgxg ,yg , where {y p dpxp ,yp = 1}. The graphs illustrate how many bits are diﬀerent in y g when examining the complete neighborhood of a randomly chosen tree xp .
1 0.8 0.6 star tree orderd list random list
0.4 0.2 0 0
5
10 15 20 25 30 d
g1 g1 g1 x ,y
(b) 32 nodes
, where {y p dpxp ,yp = 1}. Figure 6.18. Distribution of genotypic distances dg1 xg1 ,y g1 The graphs illustrate how many digits are diﬀerent in the Pr¨ ufer number yg1 when examining the complete neighborhood of a randomly chosen tree xp .
Number of Neighbors We have seen that Pr¨ ufer numbers only have high locality if they encode stars. Therefore, we focus on two issues: Firstly, we want to reveal why Pr¨ ufer numbers have high locality when encoding stars. And secondly, we want to know how large the areas are of high locality. Finding answers for these questions helps us to more accurately predict the behavior of GEAs for diﬀerent tree optimization problems. The investigation will show that the number of neighbors has a major impact on the answers to both questions. The previous investigations have shown that some Pr¨ ufer numbers have high locality. A Pr¨ ufer number representing a star has perfect locality (dm = 0) because all phenotypic neighbors of a star are also genotypic neighbors.
6.2 Pr¨ ufer Numbers
165
number of neighbors
To shed light on the question of why exactly stars have perfect locality, we calculate the number of neighbors for both, Pr¨ ufer numbers and trees. A Pr¨ ufer number xg1 uses n − 2 digits of base n to encode a tree xp with n nodes. Because we can change each of the n−2 digits to n−1 diﬀerent integers, each Pr¨ ufer number has exactly (n − 1) × (n − 2) neighbors. Furthermore, each Pr¨ ufer number xg1 with n − 2 digits is encoded as a bitstring xg of length (n − 2) × log2 (n). So each bitstring xg has (n − 2) × log2 (n) neighbors. A change of one digit in the Pr¨ ufer number xg1 can result in up to log2 (n) g ufer number encoding, diﬀerent bits in x . As we are mainly interested in the Pr¨ we want to focus in the following paragraphs only on Pr¨ ufer strings (integers) and neglect the encoding of Pr¨ ufer strings as bitstrings. The reader should notice that the number of neighbors of a Pr¨ ufer number is independent of the structure of the encoded tree. A star on n nodes has (n − 1)(n − 2) neighbors obtained by replacing one of its links with another feasible link. Therefore, for stars the number of neighbors is the same for the phenotypes and for the genotypes. Furthermore, a star’s neighbors are represented by the neighbors of its Pr¨ ufer number obtained by changing one of the number’s symbols; as already mentioned, these neighbors number also (n − 1)(n − 2). For stars, the genotypic and phenotypic neighborhoods coincide, and therefore locality is maximal. This seems auspicious, but tree localities vary with the shape of the tree. A list is a spanning tree with two leaves and n − 2 nodes of degree 2. In a list’s Pr¨ ufer number, all the symbols are distinct, and each Pr¨ ufer number has, as already mentioned, (n − 1)(n − 2) neighbors. However, a list on n n−1 n−1 n−1 nodes has i=1 i(n − i) − 1 = n i=1 i − i=1 i2 − (n − 1) = 12 n2 (n − 1) − 1 1 6 n(n − 1)(2n − 1) − n + 1 = 6 n(n − 1)(n + 1) − n + 1 neighbors (Gerstacker 1999). Therefore, for lists the number of phenotypic neighbors is much higher than the number of genotypic neighbors. Stars and lists have the smallest and largest phenotypic neighborhoods, respectively. All other spanning trees fall between these extremes which Fig. 6.19 plots as a function of the number n of nodes. 300 250 200 150 100 50 0
tree
star list 6
8
10
12
14 n
16
18
20
Figure 6.19. Phenotypic neighborhood sizes for lists and stars, as functions of the number of nodes. The values for all other trees lie between these curves.
We see that the number of neighbors of a phenotype increases when modifying a star towards a list. However, the number of neighbors of a Pr¨ ufer number remains constant and is independent of the structure of the encoded
166
6 Analysis and Design of Representations for Trees
tree. So there is a mismatch between the number of neighbors of nonstar trees and of Pr¨ ufer numbers. Therefore, the locality of all Pr¨ ufer numbers not encoding a star could not be perfect as phenotypes always have a higher number of neighbors than genotypes. The results concerning the number of neighbors are summarized and illustrated for some example networks in Table 6.2. The number of neighbors for arbitrary trees is between the number for stars and lists and must be separately calculated for each tree. It depends on the degrees of the nodes in the tree. Table 6.2. Number of neighbors for graphs, Pr¨ ufer numbers, and bitstrings graph xp
prop. of highlocality individuals
nodes
(n − 1)(n − 2) − 1)(n + 1) − n + 1
Pr¨ ufer number xg1
bitstring xg
(n − 1) (n − 2)
(n − 2) × log2 (n)
n
star list
8
star list tree
42 77 62.45 (avg.)
42
18
16
star list tree
210 665 447.2 (avg.)
210
56
32
star list tree
930 5425 2595 (avg.)
930
150
1 n(n 6
1 0.01 0.0001 1e06 1e08 1e10 1e12 1e14 10
12
14
16 n
18
20
22
Figure 6.20. The proportion of spanning trees on n nodes whose Pr¨ ufer numbers have high locality, deﬁned as diﬀering from the Pr¨ ufer number of a star in no more than imax = 5 digits.
After we have explained why Pr¨ ufer numbers can only have high locality when encoding stars, we focus on the question of how large the areas of
6.2 Pr¨ ufer Numbers
167
high locality are. GEAs only search eﬀectively in regions near stars. In these areas, the locality is high and the encoding allows a guided search. To approximate the number of individuals with high locality, we extend the deﬁnition of neighbors to include trees whose Pr¨ ufer numbers xg1 diﬀer in at the most imax digits (imax n). This means, we assume an individual xg1 to have high locality if its distance dg1 xg1 ,star towards a star is equal or lower than imax . distance imax toThe number of individuals xg1 which have the maximum imax n−2 (n − 1)i ; this wards a star (dg1 x,star ≤ imax ) can be calculated as i=0 i 2imax ). However, the number of spanning trees on n nodes, and value is O(n thus the size of the search space, is nn−2 . Therefore, as Fig. 6.20 illustrates, the proportion of these highlocality individuals is small even for moderate n and diminishes exponentially as n grows. The areas of high locality grow more slowly with increasing problem size n than the overall search space. As a result, we expect GEAs using Pr¨ ufer numbers and searching for stars to perform worse with increasing problem size n. We were able to explain the high locality of Pr¨ ufer numbers representing stars by calculating the number of neighbors for diﬀerent tree structures. The number of neighbors for Pr¨ ufer numbers stays constant, whereas for phenotypes the number of neighbors is diﬀerent for diﬀerent tree types. For stars, there is a onetoone correspondence and the number of neighbors is the same for genotypes and phenotypes. All other types of trees like lists or arbitrary trees have a higher number of neighbors than stars (or Pr¨ ufer numbers). Furthermore, the areas of high locality are very tiny as they grow with O(n2imax ), where imax n, whereas the search space grows with O(nn−2 ). Thus, the locality is in general low and GEAs searching for optimal trees that are diﬀerent from stars must fail. Performance We verify experimentally that GEAs using the Pr¨ ufer number encoding do not perform well when searching for good solutions in areas where the locality is low. We present results for GEAs only using onepoint crossover and for simulated annealing using only mutation. Both search algorithms are applied to the fully easy onemax tree problem from Sect. 6.1.5. Simulated annealing (SA) can be modeled as a GEA with population size N = 1 and Boltzmann selection (Goldberg 1990a; Mahfoud and Goldberg 1995). In each generation, an oﬀspring is created by applying one mutation step to the parent. Therefore, the new individual has distance 1 to its parent. If the oﬀspring has higher ﬁtness than its parent it replaces the parent. If it has lower ﬁtness it replaces the parent with the metropolis probability f −f P (T ) = exp(− of f springT parent ), where f denotes the ﬁtness of an individual. The acceptance probability P depends on the actual temperature T which is reduced during the run according to a cooling schedule. With lowering T , the probability of accepting worse solutions decreases. Because SA uses only mutation, and can in contrast to for example a (1 + 1) evolution strategy
168
6 Analysis and Design of Representations for Trees
solve diﬃcult multimodal problems, we use it as a representative of mutationbased evolutionary search algorithms. For further information about simulated annealing we refer to Cavicchio (1970) and Davis (1987). In Fig. 6.21, we present results for GEAs with µ + λ selection using onepoint crossover and no mutation on 16 and 32 node onemax tree problems. µ + λ selection means that we generate λ oﬀspring from µ parents and that we choose the best µ individuals from all µ + λ individuals as parents for the next generation. This selection scheme assures that a once found best individual is preserved during a GA run and not lost again. The structure of the optimal solution is determined to be either a star, list, or an arbitrary tree. For the 16 node problems, we chose µ = λ = 400, and for the 32 node problems µ = λ = 1500. We performed 250 runs and each run was stopped after the population was fully converged. Figure 6.22 presents results for using simulated annealing. The start temperature Tstart = 100 is reduced in every step by the factor 0.99. Therefore, Tt+1 = 0.99 × Tt . Mutation is deﬁned to randomly change one digit of the Pr¨ ufer number. We performed 250 runs and each run was stopped after 5000 iterations. The results in Figs. 6.21 and 6.22 show that if the optimal solution is a randomly chosen star, both search algorithms, the recombinationbased GA and the mutationbased SA are able to ﬁnd the optimal star easily. A search near stars is really a guided search and both algorithms are able to ﬁnd their way to the optimum. However, if the optimal solution is a list or an arbitrary tree, GEAs can never ﬁnd the optimal solution and are completely misled. Exploring the neighborhood around an individual in an area of low locality results in a blind and random search. Individuals that are created by mutating one individual, or by recombining two individuals, have nothing in common with their parent(s). The results show that good solutions can not be found if they lie in areas of low locality. A degradation of the evolutionary search process is unavoidable. Evolutionary search using the Pr¨ ufer number encoding could only work properly if the good solutions are stars. Near stars the locality is high and a guided search is possible. Furthermore, the empirical results conﬁrm the theoretical predictions from Sect. 3.3.4 that high locality is a necessary condition for mutation and recombinationbased GEAs. If the locality of an encoding is low, the diﬃculty of a problem is changed by the representation, and easy problems like the onemax tree problem become diﬃcult and can not be solved any more. The presented empirical results also shed some light on the contradicting statements about the performance of GEAs using Pr¨ ufer numbers. Researchers who investigate problems in which good solution are starlike see acceptable results and favor the use of Pr¨ ufer numbers. Other researchers with nonstarlike optimal solutions, however, observe low performance and advise not to use the encoding. Furthermore, we have seen that the Pr¨ ufer number encoding has low locality. Therefore, Pr¨ ufer numbers change the diﬃculty of the problem. Fully easy problems like the onemax tree problem become more
6.2 Pr¨ ufer Numbers 16
30 best fitness
14 best fitness
169
12 10 star arbitrary tree list
8 6 0
10
20
30
25 20 15
star arbitrary tree list
10 40
50
0
generation
20
40
60
80
100
generation
16 14 12 10 8 6 4 2 0
30 25 20 fitness
fitness
Figure 6.21. The performance of a GA for a 16 (left) and 32 (right) node onemax tree problem. The plots show the ﬁtness of the best individual over the run. The structure of the best solutions has a large inﬂuence on the performance of GAs. If the best solution is a star, GAs perform well. If GAs have to ﬁnd a best solution that is a list or a tree, they degrade and cannot solve the easy onemax problem.
star arbitrary tree list
15 10
star arbitrary tree list
5 0
0
500
1000
1500
2000
generation
0
1000 2000 3000 4000 5000 generation
Figure 6.22. The performance of simulated annealing for a 16 (left) and 32 (right) node onemax tree problem. The plots show the ﬁtness of the best individual over the run. As for recombinationbased approaches, the mutationbased simulated annealing fails if the best solution is not a star.
diﬃcult, whereas fully diﬃcult problems become more easy. Results about the performance of Pr¨ ufer numbers on fully diﬃcult problems are presented later in Sect. 8.1.3. 6.2.5 Summary and Conclusions This section presented an investigation into the locality of Pr¨ ufer numbers and its eﬀect on the performance of GEAs. We started with a historical review on the use of Pr¨ ufer numbers. In Sect. 6.2.2, we presented the construction rules for Pr¨ ufer numbers. This was followed by a brief overview over the beneﬁts and drawbacks of the encoding. Although the encoding has some remarkable
170
6 Analysis and Design of Representations for Trees
advantages, it is aﬀected by low locality. Consequently, we focused in the main part of our investigation (Sect. 6.2.4) on the low locality of the Pr¨ ufer number encoding. We examined the locality of Pr¨ ufer numbers more closely by performing random walks through the search space. This was followed by an analysis of the neighborhood of the genotypes and phenotypes. The analysis showed diﬀerences in locality. To explain the diﬀerences, we calculated the number of neighbors for Pr¨ ufer numbers and trees dependent on the structure of the network. Finally, we empirically veriﬁed the theoretical predictions by using recombinationbased and mutationbased search algorithms for solving the onemax tree problem. The historical review showed that there has been a great increase in interest in the Pr¨ ufer number encoding over the last few years. However, the suitability of Pr¨ ufer numbers for encoding trees is strongly disputed as some researchers report good results whereas others report failure. By performing random walks through the search space, the low locality of the encoding can be nicely illustrated. A small modiﬁcation of a genotype mostly results in a completely diﬀerent phenotype. The analysis of the neighborhood of individuals answers the question of whether the locality is low everywhere in the search space, and gives an explanation for the contradicting results from diﬀerent researchers. The results show that the locality of Pr¨ ufer numbers representing stars is high. However, all other types of networks like lists or arbitrary trees lack locality and the genotypic neighbors of a Pr¨ ufer number representing a list or an arbitrary tree have on average not much in common with each other. Therefore, the low locality of Pr¨ ufer numbers does not reduce GEAs performance in all areas of the solution space to the same extent. This can explain the diﬀerent results using Pr¨ ufer numbers existing in the literature. To answer the questions of why exactly Pr¨ ufer numbers encoding stars have high locality, and how large the areas of high locality are, we investigated the number of neighbors a Pr¨ ufer number has. The analysis shows that for Pr¨ ufer numbers the number of neighbors remains constant. For phenotypes, however, the number of neighbors varies with the structure of the tree. Stars have as many neighbors as the corresponding Pr¨ ufer numbers and therefore, the locality around stars is high. When modifying stars towards lists, the number of phenotypic neighbors increases, which makes it impossible to obtain high locality for problems other than stars. Furthermore, the areas of high locality are only of order O(nconst ), whereas the whole search space grows with O(nn−2 ). Thus, the regions of high locality become very small with increasing problem size n, which reduces the performance of GEAs on larger problems. The results show that Pr¨ ufer numbers have low locality and change the difﬁculty of problems. Researchers should be careful when using Pr¨ ufer numbers on problems of unknown complexity because easy problems become more difﬁcult when using the Pr¨ ufer number encoding. As we assume that most of the realworld problems are relatively easy and can be solved using GEAs, GEAs using Pr¨ ufer numbers are likely to fail when used on realworld problems.
6.3 The Characteristic Vector Encoding
171
6.3 The Characteristic Vector Encoding The characteristic vector (CV) encoding is a common approach for encoding graphs (Davis et al. 1993; Berry et al. 1997; Ko et al. 1997; Dengiz et al. 1997c; Dengiz et al. 1997b; Dengiz et al. 1997a; Berry et al. 1999; Premkumar et al. 2001). Representative examples for the use of the CV encoding for trees can be found in Tang et al. (1997) and Sinclair (1995). The purpose of this section is to use the framework from Sect. 4 for the analysis of the CV encoding. The investigation shows that CVs are able to represent invalid solutions. Therefore, the encoding is redundant and a repair mechanism is necessary that constructs valid trees from invalid genotypes. Examining the redundancy of CVs shows that they are uniformly redundant but aﬀected by nonsynonymous redundancy. Nonsynonymous redundancy is a result of the repair process, which brings already extinguished schemata back into the population. As repairing infeasible solutions works like additional mutation and randomizes and prolongs the search process we denote the eﬀect of nonsynonymous redundancy as stealth mutation. In the following subsection, we describe the functionality of the CV encoding. This is followed in Sect. 6.3.2 by a discussion about how to deal with representations that are able to represent invalid solutions. We illustrate that the CV encoding can represent invalid solutions and we propose a repair mechanism for the encoding. In Sect. 6.3.3, we investigate the eﬀects of the repair mechanism. We show that CVs are uniformly redundant because the proposed repair mechanism is unbiased. However, repairing invalid solutions results in nonsynonymous redundancy which increases the run duration tconv . The section ends with a brief summary. 6.3.1 Encoding Trees with Characteristic Vectors We brieﬂy describe the CV encoding and review some of its important properties. The CV encoding can be used for the encoding of trees. Further information and examples for its use can be found in Berry et al. (1994) and Palmer (1994). A
C
B
D
E
Figure 6.23. A ﬁve node tree
A CV is a binary vector that indicates if a link is used or not in a graph. For an nnode graph there exist n(n − 1)/2 possible links, and a CV of length l = n(n − 1)/2 is necessary for encoding an nnode graph. All possible links must be numbered, and each link must be assigned to a position in the vector.
172
6 Analysis and Design of Representations for Trees
In Table 6.3, we give an example of a CV for a 5 node tree. The nodes are labeled from A to E. The link from node A to B is assigned to the ﬁrst position in the string, the link from A to C is assigned to the second position, and so on. To indicate if the ith link is established, the value at position i is set to one. If no link is established, the value of the allele is set to zero. The tree that is represented by Table 6.3 is shown in Fig. 6.23. 0 1 0 0 0 1 0 1 0 1 AB AC AD AE BC BD BE CD CE DE
Table 6.3. The CV for the tree in Fig. 6.23
The CV encoding has interesting properties. Firstly, it is able to represent all possible trees. The encoding can also represent nontrees, and we will discuss this problem in the next subsection. Furthermore, all alleles of a CVencoded genotype have the same contribution to the construction of the phenotype. Therefore, we expect no problems due to nonuniformly scaled alleles, domino convergence, or genetic drift. Finally, the locality dm of the encoding is high for feasible genotypes. Feasible genotypes that are neighbors correspond to neighboring phenotypes. Two neighboring phenotypes xp and y p that have the phenotypic distance dpxp ,yp = 1 diﬀer in exactly two positions in the feasible genotype, dgxg ,yg = 2. In general, we can calculate the phenotypic distance dp between two individuals x and y as dpx,y = 0.5dgx,y , where dg denotes the genotypic Hamming distance and dp denotes the distance between trees as deﬁned in Sect. 6.1.2. Therefore, the locality of the CV is high if only feasible genotypes are considered. However, as already mentioned, the CV encoding can also represent nontrees. When considering also infeasible solutions, the encoding has low locality and problems with nonsynonymous redundancy as the phenotypic neighbors do not correspond to genotypic neighbors any more. We discuss this problem in Sect. 6.3.3. 6.3.2 Repairing Invalid Solutions We describe how to deal with invalid solutions (nontrees) which can be represented by the CV encoding. The most common approach is to repair invalid genotypes. Every CV that represents a tree must have exactly n − 1 ones, the represented graph must be connected, and there are no cycles allowed. This makes the construction of trees from randomly chosen CV demanding as most of the randomly generated CVs are invalid, and not trees. For an nnode network, ufer 1918). The there are 2n(n−1)/2 possible CVs, but only nn−2 valid trees (Pr¨ ln(n) nn−2 probability of randomly getting a tree is 2n(n−1)/2 < 2ln(2)n < 3 ln(n)/n. Therefore, the chance of randomly creating a CV that represents a tree decreases exponentially with the problem size n (Palmer 1994).
6.3 The Characteristic Vector Encoding
173
Randomly chosen CVs which should represent a tree can be invalid in two diﬀerent ways: • There are cycles in the represented graph. • The graph is not connected. We get a cycle for the example in Table 6.3 if we set the ﬁrst allele (AB) to one. Then, the genotype does not represent a tree any more because there is a cycle of ACDBA. Furthermore, if we alter any of the ones in the CV to zero we get two disconnected trees. For example, we could set AC to zero and get two disconnected subtrees. Subtree one consists of the node A and subtree two consists of a star with center D. If we do not want to remove invalid solutions from the population, there are two diﬀerent possibilities to handle invalid solutions: Firstly, we can ignore the invalid solutions and leave them in the population, or secondly, we repair them. Some GEA approaches report to some extent good results when accepting invalid solutions (Orvosh and Davis 1993; Davis et al. 1993). However, when leaving invalid solutions in the population we must ensure that a valid individual is created at the end of the run. Furthermore, we must ﬁnd a way to evaluate invalid solutions. This can be diﬃcult as for tree problems a ﬁtness function exists for trees, but not for nontrees. Finally, the largest problem is that for tree problems, the probability of generating a valid individual drops exponentially, O(exp(−n)), and therefore, only a very small fraction of the individuals are valid at all. Due to these problems, most of the traditional GEAs choose the second possibility and repair all infeasible solutions. In general, the repair process for invalid solutions consists of two steps (Berry et al. 1994): 1. Remove links from the tree that cause cycles. 2. Add links to obtain a connected tree. When repairing a CV that should represent a tree, in a ﬁrst step, the cycles in the graph must be identiﬁed. If we randomly choose xg = 1100010100 (encodes a graph with n = 5 nodes), we have a cycle ACDBA in xp . Consequently, we remove one randomly chosen link (AC, CD, DB or BA). When we choose the link AC we get y g = 1000010100. As there are no more cycles in y g , we can stop removing links and continue checking whether the graph is fully connected. As there are only three ones in y g and we have a tree with 5 nodes, the tree can not be connected and we have to add one link. As the node E is separated from the rest of the tree, a link from E to a randomly chosen node A, B, C or D, has to be added. The link CE is chosen and we ﬁnally get z g = 1000010110. A closer look at the repair mechanism shows that the order of the repair steps does not matter. 6.3.3 Bias and NonSynonymous Redundancy The CV encoding is redundant as there are infeasible genotypes that have to be repaired to represent a feasible phenotype. Therefore, each phenotype
174
6 Analysis and Design of Representations for Trees
is encoded by exactly one feasible genotype and a large number of infeasible genotypes. Consequently, we have to investigate whether the representation is uniformly redundant. Otherwise, GEA performance depends on the structure of the optimal solution. Furthermore, we have to investigate whether the encoding is synonymously redundant as synonymous redundancy is necessary for guided search. A Redundant and Unbiased Encoding We illustrate that the CV encoding with repair mechanism is a redundant encoding. Furthermore, we show that when using the repair mechanism from the previous subsection the encoding is unbiased that means uniformly redundant. The genotypephenotype mapping fg constructs a valid tree xp ∈ Φp with the help of the repair mechanism from every possible feasible or infeasible xg ∈ Φg . This means that nn−2 phenotypes are represented by 2n(n−1)/2 genotypes. Therefore, the CV encoding is a redundant encoding independently of whether the infeasible individuals are repaired or remain unrepaired in the population. If they remain unrepaired in the population they must be evaluated using the ﬁtness function deﬁned on the phenotypes (that means on trees). Therefore, the infeasible and unrepaired individuals must be assigned in some way to the feasible trees and we have the same situation as when the infeasible genotypes are repaired. Recognizing that the CV is a redundant encoding, we can use the insights into the eﬀects of redundant encodings from Sect. 3.1. Therefore, we are interested as to whether the CV encoding is biased that means nonuniformly redundant. A closer look at the repair mechanism from Sect. 6.3.2 shows that the removed, respective added links are chosen randomly. We know that an encoding is unbiased if every phenotype is represented on average by the same number of genotypes. The random repair process shows exactly this behavior as it does not favor any particular genotype. Therefore, the CV encoding is unbiased, that means uniformly redundant.
n 8 16 32
min(dprnd,star ) dprnd,M ST unbiased CV unbiased CV µ (σ) µ (σ) µ (σ) µ (σ) 3.67 (0.64) 3.66 (0.64) 5.16 (0.99) 5.19 (0.99) 10.91 (0.78) 10.91 (0.79) 13.08 (1.07) 13.08 (1.09) 26.25 (0.82) 26.23 (0.83) 29.08 (1.31) 29.05 (1.31)
Table 6.4. Minimum distance min(dprnd,star ) to stars and distance dprnd,M ST to the MST
To examine whether the CV encoding is unbiased, we randomly create CVencoded solutions xgrnd , repair them, and measure for the repaired solutions xprnd the average minimum distance min(dprnd,star ) towards a star and the average distance dprnd,M ST to the minimum spanning tree (MST). The MST is a spanning tree with minimum costs, where the costs of the tree are the sum
6.3 The Characteristic Vector Encoding
175
of the distance weights dij of the used edges (compare (8.3)). In Table 6.4, we present the mean µ and the standard deviation σ of the two distances. We show the average minimum distance to one of the n stars and the average distance to the MST. We randomly created 10,000 CVencoded individuals and compare the CV encoding to an unbiased representation like the Pr¨ ufer number encoding. The nodes are placed randomly on a twodimensional grid of size 1000×1000 and the distance weights dij are the Euclidean distances between the nodes i and j. The numbers indicate that the CV is about unbiased this means uniformly redundant. Because the CV encoding is uniformly redundant, GEA performance should be independent of the structure of the optimal solution. We present in Fig. 6.24 results for the performance of GEAs for a 16 node onemax tree problem (compare Sect. 6.1.5) using only uniform crossover, no mutation, and tournament selection of size 3. The plots show the probability of success Psucc = 1−α (ﬁnding the optimal solution) and the ﬁtness of the best individual at the end of the run (population is completely converged). We performed 250 runs for Topt is either a star, a list or an arbitrary tree. Before evaluating an infeasible individual in each generation, we repair it according to the algorithm outlined in Sect. 6.3.2, and the repaired CV replaces the infeasible solution. The plots show that GEA performance is about independent of the structure of the optimal solution.
best fitness at end of the run
1
Psucc
0.8 0.6 0.4 star arbitrary tree list
0.2 0 10
15
20
25
30
N
(a) probability of success
35
14 12 10 star arbitrary tree list
8 6 10
15
20
25
30
35
N
(b) ﬁtness at end of the run
Figure 6.24. Performance of GEAs using the CV encoding for optimizing a 16 node onemax tree problem, where Topt is either a star, a list, or an arbitrary tree. Invalid solutions are repaired. GEA performance is about independent of Topt . Therefore, the redundant CV encoding using the repair mechanism from Sect. 6.3.2 is unbiased.
Finally, we brieﬂy discuss the possibility of using speciﬁc mutation and recombination operators that always create only valid solutions. Then, no repair
176
6 Analysis and Design of Representations for Trees
steps are necessary and we do not have to worry any more about a bias. Every individual that is created during the GEA’s run would be valid. However, one problem is the creation of the initial, feasible solutions as the fraction of feasible solutions is tiny. Furthermore, the creation of thuslike “intelligent” crossover and mutation operators leads to a direct encoding (compare Chap. 7). Then, the genetic operators are, in contrast to the used standard crossover or bitﬂipping mutation, not based on the Hamming distance between individuals any more (compare Sect. 3.3.5). Instead, the operators are problemspeciﬁc and there exists no explicit representation (Φg = Φp ). We have illustrated that the CV encoding is uniformly redundant as phenotypes are represented on average by the same number of infeasible genotypes. Therefore, GEA performance is about independent of the structure of the optimal solution. NonSynonymous Redundancy and Stealth Mutation When recombining two feasible, CVencoded, parents, the oﬀspring are often underspeciﬁed. They do not represent a fully connected tree and the repair mechanism we presented in Sect. 6.3.2 has to insert links randomly to construct a valid solution. Therefore, links which do not exist in both parents could be used for the construction of the oﬀspring. It could also happen that a link that does not not exist in any of the individuals in the population can ﬁnd its way back into the population by the repair mechanism. This eﬀect caused by the repair process results in crosscompetition and should be denoted as stealth mutation. Stealth mutation is a result of the nonsynonymous redundancy (compare Sect. 3.1.2) of the CV encoding. The CV encoding is nonsynonymously redundant as not all infeasible genotypes that encode the same phenotype are similar to each other. We present a brief example. We assume a genotype consisting of only ones. This infeasible genotype is repaired so that all possible phenotypes are constructed with the same probability. Although, on average, all phenotypes are represented by about the same number of infeasible genotypes, it is not possible to determine which infeasible solution encodes which feasible solution and the encoding is nonsynonymously redundant. This property of the CV encoding, that the genotypephenotype mapping of infeasible solutions is random (due to the repair process), results in the situation that not all infeasible genotypes that are decoded to the same phenotype are similar to each other. When using the notion of BBs, stealth mutation results in a continuous supply of new BBs during a run. Due to crosscompetition (compare the eﬀects of nonsynonymous redundancy described in Sect. 3.1.2), new BBs are created randomly in the population during the GEArun, even if they are not present in the start population, or not properly mixed and lost. As the CV encoding is nonsynonymously redundant, the diﬃculty of a problem is changed and easy problems become more diﬃcult to solve as the search process is randomized. GEAs show low performance and the running time of GEAs increases.
6.3 The Characteristic Vector Encoding
177
In M¨ uhlenbein and SchlierkampVoosen (1993) and Thierens and Goldberg (1994), the √ time until convergence was found for the onemax problem as tconv = π l/2I with the selection intensity I and the string length l. I depends only on the selection of size 2 we √ used selection scheme and for tournament get I = 1/ π. With l = n(n − 1)/2, we get tconv ≈ π π2 n. n denotes the size of the problem (number of nodes). 200 empirical results prediction
180 run duration tconv
160 140 120 100 80 60 40 20 8
10 12 14 16 18 20 22 24 problem size n
Figure 6.25. Run duration tconv over problem size n for the onemax tree problem using tournament selection of size 2 and uniform crossover. For selectorecombinative GAs, tconv should grow linearly with the problem size n. However, if GEAs use the CV encoding tconv grows due to problems with nonsynonymous redundancy approximately exponentially with n.
To illustrate the eﬀects of stealth mutation and nonsynonymous redundancy on the performance of GEAs we compare in Fig. 6.25 for the onemax tree problem the theoretical prediction for the run duration tconv with the empirical results for the CV encoding. For the empirical analysis, we use a simple GA with no mutation, tournament selection of size s = 2, and uniform crossover. The population size N is large enough to reliably ﬁnd the optimal solution. The results show a nonlinear dependency of the run duration tconv over the problem size n. The plots indicate that GEAs using CVs struggle because of more repair operations and stealth mutation, which are a result of nonsynonymous redundancy. The search for good solutions depends more on the random eﬀects of mutation than on recombination. Due to the nonsynonymous redundancy, with higher problem size the probability of randomly ﬁnding the correct link decreases, and the run duration of GEAs using CVs increases. 6.3.4 Summary This section examined important properties of the CV encoding. We started with a description of the encoding and brieﬂy reviewed its important properties. In Sect. 6.3.2, we illustrated the problem of infeasible solutions and how infeasible solutions can be repaired. This was followed in Sect. 6.3.3 by an investigation into the properties of the encoding. We recognized that the CV encoding is redundant and unbiased. Therefore, GEA performance is independent of the structure of the optimal solution. Furthermore, it was discussed
178
6 Analysis and Design of Representations for Trees
that the repair process results in stealth mutation and nonsynonymous redundancy which increases run duration and reduces GEA performance. We have seen in this section that an encoding which can represent not only valid, but also invalid, solutions encodes a valid phenotype by more than one genotype. Therefore, such an encoding is redundant and the results about redundant encodings from Sect. 3.1 can be used. Redundancy is independent of whether the invalid solutions are repaired, or if they remain untouched in the population. In both situations, it is necessary to evaluate the invalid genotypes and to assign a ﬁtness value to every invalid solution. Furthermore, we have seen that an encoding which can represent invalid solutions is unbiased if the construction of a valid phenotype from an invalid genotype is unbiased and does not favor some phenotypes. Our investigation in the CV encoding has shown that the encoding is uniformly redundant. Only nn−2 valid trees are encoded by nn−2 valid and 2n(n−1)/2 − nn−2 invalid solutions. To repair invalid solutions, we presented a randomized and unbiased repair mechanism. The CV encoding is unbiased, and GEAs perform independently of the structure of the optimal solution. However, repairing invalid solutions results in nonsynonymous redundancy. We denoted this eﬀect as stealth mutation and showed that it increases the run duration tconv .
6.4 The Link and Node Biased Encoding The link and node biased (LNB) encoding belongs to the class of weighted encodings and was developed by Palmer (1994). It encodes a tree using weights for the nodes and links of a tree. Additional encoding parameters such as the linkspeciﬁc and the nodespeciﬁc bias are necessary to control the importance of link and node weights. The encoding was proposed to overcome the problems of characteristic vectors, predecessor representations (compare also Raidl and Drexel (2000)) and Pr¨ ufer numbers. Later, Abuali et al. (1995) compared diﬀerent representations for probabilistic minimum spanning tree (PMST) problems and in some cases found the best solutions by using the LNB encoding. Raidl and Julstrom (2000) observed solutions superior to those of several other optimization methods using a representation similar to the LNB encoding for the degreeconstrained minimum spanning tree (dMST) problem. This section examines the LNB encoding using the framework from Chap. 4. It investigates whether the encoding is uniformly redundant, and how the encoding parameters (linkspeciﬁc and nodespeciﬁc bias) inﬂuence redundancy. The investigation reveals that all versions of the LNB encoding are synonymously redundant. Furthermore, the commonly used simpler version of the encoding, which uses only a nodespeciﬁc bias, is biased towards stars, and the initial population is dominated by only a few starlike individuals. Focusing on the linkbiased (LB) encoding, which only uses a linkspeciﬁc
6.4 The Link and Node Biased Encoding
179
bias, reveals that solutions similar to the minimum spanning tree (MST) are overrepresented. The general LNB encoding uses both biases, a nodespeciﬁc bias and a linkspeciﬁc bias. Similarly to the nodebiased encoding, the LNB encoding is biased towards stars if a nodespeciﬁc bias is used. When using a small linkbias, the encoding is, in analogy to the LB encoding, biased towards the MST. However, by increasing the linkspeciﬁc bias, the bias of the LNB encoding is reduced and it becomes uniformly redundant for a large linkspeciﬁc bias. The LNB encoding has great problems if the two biases are too small, because then only MSTlike phenotypes can be represented. At the extreme, if both biases are zero, only the MST can be represented independently of the LNBencoded genotype. In Sect. 3.1, we have seen that the performance of GEAs using a synonymously redundant encoding increases if the good solutions are overrepresented, and decreases if they are underrepresented. Therefore, the performance of GEAs using the LNB encoding is high if the optimal solution is similar to stars or MSTs. For all other problems, however, a reduction of GEA performance is unavoidable unless the linkspeciﬁc bias is high. In the following subsection, we give a brief description of the LNB encoding. In Sect. 6.4.2, we discuss that the bias of an encoding is equivalent to nonuniform redundancy (compare Sect. 3.1). This is followed by an investigation into whether the NB encoding is uniformly redundant. Section 6.4.4 focuses on the LB encoding and Sect. 6.4.5 presents a population sizing model for the onemax tree problem based on the population model from Sects. 3.1.4 and 6.5.5. In Sect. 6.4.6, we investigate for the LNB encoding the eﬀect of both encoding parameters, the nodespeciﬁc bias and the linkspeciﬁc bias. Finally, we present in Sect. 6.4.7 some empirical GEA performance results for the onemax tree problem. The section ends with concluding remarks. 6.4.1 Motivation and Functionality We review the motivation and the resulting properties of the LNB encoding as described in Palmer (1994), Palmer and Kershenbaum (1994a), and Palmer and Kershenbaum (1994b). As the costs of a communication or transportation network strongly depend on the length of the links, network structures that prefer short distance links often tend to have higher ﬁtness. Furthermore, it is useful to run more traﬃc over the nodes near the gravity center of an area than over nodes at the edge of this area (Kershenbaum 1993; Cahn 1998). Thus, it is desirable to be able to characterize nodes as either interior (some traﬃc only transits), or leaf nodes (all traﬃc terminates). As a result, the more important a link is, and the more transit traﬃc that crosses the node, the higher is, on average, the degree of the node. Nodes near the gravity center tend to have a higher degree than nodes at the edge of the network. Hence, the basic idea of the LNB encoding is to encode the importance of a node or link. The more important the node or link is, the more traﬃc that should transit over it.
180
6 Analysis and Design of Representations for Trees
When applying this idea to tree encodings, the distance weight matrix, which deﬁnes the distance weights between any two nodes, is biased according to the importance of the nodes or links. If a node or link is not important, the modiﬁed distance weight matrix should increase the distance weight of all links that are connected to this node. Doing this will result with high probability in a leaf node if the encoded phenotype is constructed from the distance weights by the help of Prim’s algorithm. The NodeBiased Encoding When using the nodebiased (NB) encoding, the genotype b holds weights for each node, and has length n for an n node tree. There is a distance weight matrix which assigns a nonnegative distance weight dij to the n(n − 1)/2 diﬀerent edges (ij) of a tree. The values dij of the distance weight matrix are modiﬁed according to b using the weighting function dij = dij + P2 (bi + bj )dmax ,
(6.2)
where bi ∈ [0, 1], dmax = max(dij ), and i, j ∈ {0, . . . , n − 1}. The node bias P2 controls the inﬂuence of the node weights bi on the construction of xp . When using the NB encoding for trees, we get the phenotype xp from the genotype b by calculating the minimum spanning tree (MST) for the modiﬁed distance weight matrix D . Prim’s algorithm (Prim 1957) was used in the original work. By running Prim’s MST algorithm, nodes i that correspond to a low bi will probably be interior nodes of high degree in the tree. Nodes j that correspond to high bj will probably be leaf nodes. Thus, the higher the weights bi of a node i, the higher is the probability that node i will be a leaf node. To get the tree’s ﬁtness, the tree xp is evaluated by using the original distance weight matrix D. We illustrate the functionality of the NB encoding with a small example. The NBencoded genotype b = {0.7, 0.5, 0.2, 0.8, 0, 1} holds the node weights. A distance weight matrix for the 5node problem is deﬁned as ⎛ ⎞ − 2 1 3 4 ⎜2 − 5 6 3 ⎟ ⎜ ⎟ ⎟ (6.3) D=⎜ ⎜ 1 5 − 4 3 ⎟· ⎝ 3 6 4 − 10 ⎠ 4 3 3 10 − For the construction of the tree xp from b, we ﬁrst have to calculate all values dij of the modiﬁed distance weight matrix. Using P2 = 1, we get for example d0,1 = 2+(0.7+0.5)×10 = 14. When calculating the remaining di,j according (6.2) we get for the modiﬁed distance matrix:
6.4 The Link and Node Biased Encoding
⎛
− ⎜ 14 ⎜ D = ⎜ ⎜ 10 ⎝ 18 12
14 10 18 − 12 19 12 − 14 19 14 − 11 6 19
181
⎞
12 11 ⎟ ⎟ 6⎟ ⎟· 19 ⎠ −
Using Prim’s algorithm for the modiﬁed distance matrix D , we ﬁnally get the tree illustrated in Fig. 6.26. The represented tree is calculated as the MST using the distance matrix D . For example, d0,2 = 10 < d0,i , where i ∈ {1, 3, 4}. Because the link between node 0 and node 2 has the lowest distance weight d , it is used for the construction of the represented tree.
0
2
4
1
Figure 6.26. An example tree for the nodebiased encoding
3
Palmer noticed in his original work that each bi modiﬁes a whole row and a whole column in the distance weight matrix. Thus, not all possible solution candidates can be encoded by the NB encoding (Palmer 1994, pp. 6667). The LinkBiased Encoding When using the linkbiased (LB) encoding, each genotype holds weights bij for the n(n−1)/2 diﬀerent edges of a tree. As for the LB encoding, the weights bij are ﬂoating values between zero and one. The original distance weights dij are modiﬁed by the link weights bij as dij = dij + P1 bij dmax ,
(6.4)
where dij are the modiﬁed distance weights, dmax = max(dij ), and P1 is the linkspeciﬁc bias. The parameter P1 controls the inﬂuence of the linkspeciﬁc weights bij and has a large impact on the structure of the tree. For P1 = 0, the linkspeciﬁc weights have no inﬂuence and only the MST calculated based on dij can be represented. 2
A
D 1
4 5
0
B
3
C
Figure 6.27. An example tree for the LB encoding (The numbers indicate the number of a link)
We illustrate the construction of a tree from the genotype b with a brief example. We use the LB encoding and for representing a tree with n = 4 nodes
182
6 Analysis and Design of Representations for Trees
the genotype is of length l = n(n − 1)/2 = 6. For the example we want to use the LBencoded genotype bij = {0.1, 0.6, 0.2, 0.1, 0.9, 0.3}. With P1 = 1 and using the distance weights dij = {10, 30, 20, 40, 10, 20} we can calculate the modiﬁed distance weights according to (6.4) as dij = {14, 54, 28, 44, 56, 32}. Notice that dmax = 40. The represented tree, which is calculated as the MST tree using the modiﬁed distance weights d , is shown in Fig. 6.27. The six possible edges are labeled from 0 to 5 and the tree consists of the edges between A and B (dAB = 14), A and D (dAD = 28), and C and D (dCD = 32). The LinkandNodeBiased Encoding As an extension of the NB encoding, Palmer (1994) introduced a second, extended version of the LB encoding with an additional link bias. For the linkandnodebiased (LNB) encoding, the genotype holds weights not only for the n nodes but also for all possible n(n−1)/2 links, and has l = n(n+1)/2 elements (n node weights bi and n(n − 1)/2 link weights bij ). Therefore, the weighting function is dij = dij + P1 bij dmax + P2 (bi + bj )dmax ,
(6.5)
with the node weights bi , the link weights bij , the linkspeciﬁc bias P1 , and the nodespeciﬁc bias P2 . With proper parameter setting, the LNB encoding can represent all possible trees. However, in comparison to the LB encoding, the length of a genotype increases from n to n(n + 1)/2. We present a brief example for the LNB encoding. The example genotype holds the node weights bi = {0.7, 0.5, 0.2, 0.8, 0, 1} and the link weights ⎛ ⎞ − 0.1 0.6 0.2 0.8 ⎜ 0.1 − 0.1 0.9 0.5 ⎟ ⎜ ⎟ ⎟ bij = ⎜ ⎜ 0.6 0.1 − 0.3 0.2 ⎟ · ⎝ 0.2 0.9 0.3 − 0.4 ⎠ 0.8 0.5 0.2 0.4 − With P1 = 1, P2 = 1, and using the distance weight matrix from (6.3) we can calculate d0,1 = 2 + 0.1 × 10 + (0.7 + 0.5) × 10 = 15. Consequently, we get for the modiﬁed distance weight matrix ⎛ ⎞ − 15 16 20 20 ⎜ 15 − 13 28 16 ⎟ ⎜ ⎟ ⎟ D = ⎜ ⎜ 16 13 − 17 8 ⎟ · ⎝ 20 28 17 − 23 ⎠ 20 16 8 23 − Finally, we calculate the MST using the modiﬁed distance matrix D and get the tree shown in Fig. 6.28. Using the diﬀerent variants of the LNB encoding makes it necessary to determine the value of one or two additional encoding parameters, P1 and
6.4 The Link and Node Biased Encoding 0
2
1
183
4
3
Figure 6.28. An example tree for the LNB encoding
P2 . In the original work from Palmer, only results for the NB encoding with P2 = 1 are presented. For the diﬀerent variants of the LNB encoding, the construction of the phenotype can be implemented with a Fibonacci heap and goes with O(n2 ). The structure of the represented tree depends not only on the node and link weights, but also on the given distance weights of the links. The same LB, NB, or LNBencoded genotype can represent diﬀerent phenotypes if diﬀerent distance weights are used. Therefore, we assume in our experiments that the distance weights remain constant and do not change during the run of a GA. 6.4.2 Bias and NonUniformly Redundant Representations Representations are redundant, if they assign a discrete, noninﬁnite number of diﬀerent phenotypes to genotypes that consist of real values. Each phenotype can be represented by an inﬁnite number of diﬀerent genotypes. Consequently, the diﬀerent variants of the LNB encoding are redundant. Furthermore, the representations are synonymously redundant. Genotypes that represent the same phenotype are next to each other in the mutational space. Small mutations of the node weights bi or link weights bij often do not change the represented phenotype, or only slightly by one edge. Even large mutations that completely change a node or link weight only result in a change of up to two edges. As a result of the synonymous redundancy of the diﬀerent variants of the LNB encoding, the models from Sect. 3.1 can be used to predict the inﬂuence of the redundant LNB encodings on the performance of GEAs. In Sect. 3.1, we have seen that uniform redundancy does not change the performance of GEAs in comparison to nonredundant representations. An encoding is uniformly redundant if all phenotypes are represented by the same number of genotypes. However, if some individuals are overrepresented by the encoding, the performance of GEAs is inﬂuenced. If the optimal solution is similar to the overrepresented individuals, GEA performance increases. If the optimum is similar to underrepresented individuals, a degradation of GEA performance is unavoidable. As a result, if the encoding is not uniformly redundant, GEA performance depends on the structure of the optimal solution. We illustrate that a bias of a representation is equivalent to nonuniform redundancy. Therefore, the results about redundant representations from Sect. 3.1 can be also used for biased representations. Palmer deﬁned a biased encoding in his thesis (Palmer 1994, pp. 39) as: “It (a representation) should be unbiased in the sense that all trees are equally represented; i.e., all trees should be represented by the
184
6 Analysis and Design of Representations for Trees
same number of encodings. This property allows us to eﬀectively select an unbiased starting population for the GA and gives the GA a fair chance of reaching all parts of the solution space.” When comparing this deﬁnition of bias to the deﬁnition of redundant encodings (compare Sect. 3.1.2), we see that both deﬁnitions are essentially the same. An encoding is biased if some individuals are over, or underrepresented. Furthermore, Palmer correctly recognized, in agreement with the results about redundant encodings, that a widely usable, robust encoding should be unbiased. However, in contrast to Palmer’s statement that only unbiased encodings allow an eﬀective search, we have seen that biased (nonuniformly redundant) encodings can be helpful if the encoding is synonymously redundant and overrepresents solutions similar to the optimal solution. The reader should be careful not to confuse the bias of a representation with the ability to encode all possible phenotypes. Palmer and Kershenbaum (1994a) have already shown that the simple NB encoding is not able to represent all phenotypes. The bias of an encoding describes whether the phenotypes that can be represented are uniformly represented but does not consider the number of represented phenotypes. As the diﬀerent variants of the LNB encoding are redundant, the performance of GEAs goes with O(r/2kr ). The question arises regarding whether the encodings are nonuniformly redundant, or not. Palmer developed the LNB encoding with the intent to create a uniformly redundant encoding. Therefore, to be able to judge the performance of GEAs using the LNB encoding, we investigate in the following subsections whether the synonymously redundant variants of the LNB encoding are uniformly redundant. 6.4.3 The NodeBiased Encoding The NB encoding is not capable of representing all possible trees (Palmer 1994). The purpose of this subsection is to investigate whether the LB encoding uniformly encodes the phenotypes that can be encoded. The subsection extends prior work (Gaube 2000; Gaube and Rothlauf 2001) by new results. We start with a distance weight matrix where all dij have the same value. This is followed by an investigation where the position of the nodes are chosen randomly and the distance weights dij are the Euclidean distances between the nodes. All Distance Weights are Equal We assume that all distance weights dij are equal, dij = dkl , for i, j, k, l ∈ {0, . . . , n − 1} and i = j, k = l. Thus, diﬀerent values of dij are a result of diﬀerent node weights bi (6.2). We denote by bl = minn−1 i=0 bi the lowest weight of b. It is the weight for the lth node. Then, dil < dij for bl = min{b0 , . . . bn−1 },
6.4 The Link and Node Biased Encoding
185
where i, j, l ∈ {0, . . . , n − 1}, i = l, i = j and l = j. As Prim’s algorithm chooses these n − 1 links with the lowest modiﬁed distance weights that do not create a cycle, the only phenotype that could be represented by the NB encoding is a star with center l. Therefore, the LB encoding is uniformly redundant and encodes only stars if all distance weights dij are the same. For a tree with n nodes, the number of possible stars is n, whereas the number of all possible trees is nn−2 . Thus, only a small fraction of trees could be represented by the nodebiased encoding.
nonstar 0%
star with center l l=0 l=1 l=2 l=3 25.01% 24.97% 24.92% 25.10%
Table 6.5. Average percentage of represented phenotypes (n = 4)
Although an empirical proof of a theoretical prediction is redundant, Table 6.5 presents an empirical veriﬁcation of these results for a small 4 node problem. There are 16 possible trees, and 4 of them are stars with center l, where l ∈ {0, 1, 2, 3}. For the experiments, we created 1,000 random LBencoded solutions for 1,000 diﬀerent distance weights dij . We see that it is not possible to create nonstars, and that the stars are represented uniformly. As a result, for equal distance weights, the NB representation is uniformly redundant (it represents the n diﬀerent stars unbiased) but it can only represent a small portion of the solution space (only stars). Random Distance Weights We randomly placed the nodes on a twodimensional quadratic plane of size 1000×1000 and calculated the distance weights dij using the Euclidean distance between the nodes i and j (6.1). For the experiments, we randomly created 500 NBencoded genotypes xgrnd . Figure 6.29 shows the average minimum phenotypic distance min(dprnd,star ) between a randomly created nodebiased individual xprnd and a star. The distance dprnd,star measures how similar the phenotype of a randomly created NB vector is to one of the n stars. If dprnd,star is low, xprnd has many edges in common with one of the n stars. We performed experiments for 8, 16, and 32 node problems and positioned the nodes 250 times randomly on the 1000×1000 square. The plots show that in comparison to an unbiased encoding (like Pr¨ ufer numbers), where the average distance towards a star stays constant, with increasing nodespeciﬁc bias P2 , the minimum phenotypic distance min(dprnd,star ) decreases. Therefore, a randomly created NBencoded individual becomes more and more starlike and the NB encoding is nonuniformly redundant as it overrepresents solutions similar to stars. This result is not surprising when we take a closer look at (6.2). The original distance weights dij are modiﬁed by an additional bias. With increasing
6 Analysis and Design of Representations for Trees 12 nodebiased unbiased
3
30
10 min(dprnd,star)
min(dprnd,star)
4
2 1 0
25
nodebiased unbiased
8
min(dprnd,star)
186
6 4 2 0
0
2
4
6 P2
(a) 8 node
8
10
nodebiased unbiased
20 15 10 5 0
0
2
4
6 P2
(b) 16 node
8
10
0
2
4
6
8
10
P2
(c) 32 node
Figure 6.29. Average minimum phenotypic distance min(dprnd,star ) of a randomly generated NBencoded genotype to a star. With increasing node bias P2 , the NB encoding is strongly biased towards stars. For large P2 , only one of the n stars can be encoded.
node bias P2 , the inﬂuence of dij relatively decreases and dij only depends on the nodespeciﬁc weights bi and bj , and no longer on the distance weights dij . Therefore, with large P2 , dij can be completely neglected and the NB encoding can only encode stars. We have the same situation as when all distance weights are the same. Thus, the results described in the previous paragraphs hold true. We see that with increasing P2 , every randomly created NBencoded individual will decode to a star. However, not only for very large, but even for reasonable values of P2 , the NB encoding is strongly biased. To investigate how often diﬀerent phenotypes are represented by a randomly created NBencoded solution, we ordered the represented phenotypes according to their frequency. In Fig. 6.30, we plot how the cumulative frequencies of the ordered number of copies of a speciﬁc phenotype depend on P2 for a tree with four nodes (there are 42 = 16 diﬀerent phenotypes). The frequencies are ordered in ascending order. This means that rank 1. corresponds to the phenotype that is most often encoded (encoded with the highest probability), and rank 16. to the phenotype that is encoded with the lowest probability. We performed 1,000 experiments with diﬀerent node locations and randomly generated 1,000 NBencoded genotypes for each experiment. The presented values are averaged over these 106 diﬀerent genotypes. If the encoding is unbiased (uniformly redundant), all individuals are created with the same probability and the cumulative frequency linearly increases with the rank. All 16 possible phenotypes are represented uniformly with probability 1/16=0.0625. However, for the NB encoding, some phenotypes (stars) are created more often. For example, when using the NB encoding with P2 = 0.5, a randomly generated LBencoded genotype represents with a probability of about 0.5 the same phenotype (rank 1.). The next most frequent phenotype is encoded by a randomly chosen genotype with probability of about 0.25. The line shows that for P2 = 0.5 about 90% of all randomly
6.4 The Link and Node Biased Encoding
187
1
cumulative frequency
0.8
0.6
0.4
0.2
P2=0 P2=0.05 P2=0.5 P2=1 P2=10 P2=∞ unbiased encoding
0 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. rank
Figure 6.30. Distribution of diﬀerent phenotypes (n = 4)
generated NBencoded genotypes only represent three diﬀerent phenotypes. Furthermore, we see that for P2 large enough (P2 > c), the encoding can only represent a maximum of n = 4 diﬀerent phenotypes which are stars. These 4 stars are encoded unbiased (each of the four individuals is created with 25% probability), similarly to the situation where all distance weights are the same. The results show that for medium values for P2 some phenotypes are strongly overrepresented, whereas some others are not represented at all. For lower values of P2 , the node weights bi have a lower inﬂuence on dij and the represented phenotypes are MSTlike. Figure 6.30 shows that with decreasing P2 , fewer and fewer diﬀerent phenotypes can be represented. For P2 = 0 only one phenotype, the MST, can be encoded. Figure 6.31 illustrates the frequency of the ﬁrst, second, and ﬁfth most frequent phenotype over P2 . The plots are based on Fig. 6.30. We see that with increasing P2 , the diversity decreases, and for P2 10 only the four most frequent phenotypes (they are the four stars) are uniformly represented with about 25% frequency. With lower P2 , the phenotypes are biased towards the MST, and for P2 → 0, only one phenotype (the MST) can be represented. The results show, that the NB encoding is nonuniformly redundant and overrepresents either stars or the MST. For P2 → ∞, the encoding can only represent stars. When using medium values for P2 , a few starlike individuals dominate a randomly created population, and it is impossible to create some phenotypes. For small values of P2 , the represented phenotypes become more and more MSTlike, and for P2 → 0 only the MST can be encoded. 6.4.4 A Concept for the Analysis of Redundant Representations In this section, we want to illustrate how the theoretical insights from Sect. 3.1 can be used for predicting the inﬂuence of the redundant LB encoding on the performance of GEAs.
188
6 Analysis and Design of Representations for Trees 1
1. rank 2. rank 5. rank
frequency
0.8
0.6 minimum spanning tree like tree structures
star−like tree structure
0.4
0.2
0 0.001
0.01
0.1
1
10
100
P2
1000
Figure 6.31. Frequency of how often the ﬁrst, second, and 5th most frequent phenotype is encoded by a randomly chosen LBencoded genotype (n = 4)
Section 6.4.1 illustrated for the LB encoding that the mapping from the genotypes to the phenotypes depends on the linkspeciﬁc bias P1 . Therefore, to be able to predict the expected GEA performance when using the LB encoding, it must be investigated how P1 inﬂuences the characteristics of the encoding. The following investigation is an example performed for the LB encoding, but the investigation approach is general and can be transferred to any other representation. Factors needing to be examined are: • • • •
Size of the search space. Synonymy of the redundant representation. Order of redundancy. Over and underrepresentation.
These aspects are discussed in the following paragraphs. When using a (redundant) representation it is important that all possible phenotypes can be represented. A representation should assign at least one genotype to all phenotypes of interest. Otherwise, if no genotype is assigned to some phenotypes, the search space is reduced by the encoding and some possible solutions can never be found by the used search method. The inﬂuence of this eﬀect on the performance of a GEA is twofold. If the number of accessible solutions is reduced but the optimal solution is still accessible, GEA performance increases. On the other hand, if the optimal solution is no longer accessible, all search methods must fail. Therefore, a reduction of the phenotypic search space should be avoided if no problemspeciﬁc knowledge regarding the optimal solution exists. When using the LB encoding, the number of accessible solutions depends on P1 (Gaube and Rothlauf 2001). If P1 is very large, all possible phenotypes can be represented using this encoding. At the other extreme, for P1 very small (P1 → 0), only the MST calculated from the distance weights dij can be represented. As long as P1 1, every possible phenotype can be encoded as the additional overall bias P1 bij dmax (compare (6.4)) can
6.4 The Link and Node Biased Encoding
189
always be larger than dij . If P1 1 some of the possible trees can not be encoded using the LB encoding. In our proposed model (Sect. 3.1.4) describing the inﬂuence of redundant representations on GEA performance, we assume that nonuniform redundancy changes the initial supply. If we want to use this model for predicting GEA performance we must ensure that the considered representation is synonymously redundant. If a representation is not synonymously redundant, the standard search operators no longer work properly and GEAs fail (compare Section 3.1.2). The LB encoding is synonymously redundant independently of the parameter P1 . Even if the number of accessible solutions decreases with lower values of P1 , a mutation operator always results in the same, or a slightly diﬀerent, phenotype. When comparing diﬀerent types of redundant representations, an important measure is the order kr of redundancy. In Section 3.1.4, we saw that the population size N goes with O(2kr ) for synonymously redundant representations. Therefore, kr has a strong inﬂuence on GEA performance. For the realvalued LB encoding we can assume that kr remains independent of P1 . Finally, when using a redundant representation it must be investigated whether some phenotypes are over or underrepresented. Section 3.1.4 has shown that the necessary population size N goes with O(1/r). In general, the parameter r is problemspeciﬁc and depends on the speciﬁc instance of a problem. GEA performance remains unchanged if a synonymously redundant representation is uniformly redundant. If a representation is nonuniformly redundant, some instances of a problem will become easier for the search method (those where the optimal solution is overrepresented) and some instances will become more diﬃcult (those where the optimal solution is underrepresented). For the LB encoding, solutions that are similar to the MST are increasingly overrepresented with decreasing P1 . For very small P1 only a tiny fraction of genotypes represent a solution diﬀerent from the MST. Only for large values of P1 → ∞ is the LB encoding uniformly redundant. As a result, there is a continuum between uniform redundancy (P1 → ∞) and complete nonuniform redundancy (P1 = 0), which can be controlled by the parameter P1 . In the remaining paragraphs of this subsection, we investigate this continuum and examine how the overrepresentation of speciﬁc edges can be controlled by the representationspeciﬁc parameter P1 . We want to start with an investigation into how similar a randomly created individual is compared to the MST. The similarity between two trees (the MST and the randomly created individual) is measured by calculating the distance between both trees. The distance between two trees measures the number of diﬀerent edges (Sect. 6.1.2). In Fig. 6.32, we show the phenotypic distance of a randomly created linkbiased individual to the MST for n = 16 and n = 28. The error bars show the standard deviations. The dotted lines indicate the distance of a randomly created individual towards the MST when using a nonredundant representation (for example Pr¨ ufer numbers). The results show that for large values of P1 a randomly created LB individual has about the same distance
190
6 Analysis and Design of Representations for Trees
distance to MST
25 20 15 10 5 0 0.01
16 nodes 28 nodes 0.1
1
10
100
1000
linkspecific bias P1
Figure 6.32. Phenotypic distance of a randomly generated LB individual to the minimum spanning tree for trees with n = 16 and n = 26 nodes. The distance between two individuals indicates the number of links that are diﬀerent.
towards the MST as a nonredundant encoding. Therefore, it can be assumed that with a large enough P1 , the LB encoding is uniformly redundant. With decreasing values of P1 the represented trees become more and more MSTlike and the LB encoding becomes more and more nonuniformly redundant. We see that the overrepresentation of a speciﬁc solution (the MST) strongly depends on the linkspeciﬁc bias P1 . To be able to calculate the overrepresentation of speciﬁc edges (we need this for the population sizing model we derive in Sect. 6.4.5) we want to examine how Pr , which is the probability that an edge contained in a randomly created LB individual is also contained in the MST, depends on P1 . For nonredundant or uniformly redundant representations the probability Pru can be calculated as Pru =
n
Pru
8 12 16 20 24 28
0.25 0.1667 0.125 0.1 0.0834 0.0714
Pr for P1 =1,000,000 µ(σ) 0.2497 (0.013) 0.1665 (0.017) 0.1250 (0.015) 0.0998 (0.010) 0.0832 (0.012) 0.0713 (0.012)
2n · n(n − 1)
(6.6)
Table 6.6. A comparison between Pru and Pr for P1 = 1, 000, 000.
Table 6.6 compares for diﬀerent problem sizes n, the probability Pru for nonredundant representations to empirical results for Pr (mean µ and standard deviation σ) when using a large linkspeciﬁc bias (P1 = 1, 000, 000). It can be seen that for large values of P1 the probability Pr (that an edge contained in a randomly created LB individual is also contained in the MST) equals the probability Pru (that a randomly chosen edge is part of the MST). Therefore, for large values of P1 the encoding becomes uniformly redundant.
6.4 The Link and Node Biased Encoding
191
1 8 nodes 16 nodes 24 nodes 28 nodes
Pr
0.5
0.2
0.1
0.05 0.01 0.1
1
10
100 1000
linkspecific bias P1
1e+006
Figure 6.33. Probability that a link of a randomly generated LBencoded tree is part of the MST over the linkspeciﬁc bias P1
Consequently, Fig. 6.33 plots, for the LB encoding, how Pr depends on the linkspeciﬁc bias P1 . The results show the mean and the standard deviation for 8, 16, 24, and 28 node trees. For large values of P1 (P1 > 100), Pr equals Pru and we get the values shown in Table 6.6. With decreasing P1 , the edges contained in a randomly created individual are also more and more frequently contained in the MST. For small values of P1 , all edges of a randomly created individual are with high probability Pr and also part of the MST. After discussing how the redundancy of the LB encoding and the overrepresentation of speciﬁc edges depends on the linkspeciﬁc bias P1 , we want to formulate a model based on the results from Sect. 3.1.4, which describes how GEA performance depends on P1 . 6.4.5 Population Sizing for the LinkBiased Encoding The previous section has shown that with decreasing linkspeciﬁc bias P1 the LB encoding overrepresents solutions similar to the MST. This section shows that we are able to give good predictions on how GA performance depends on the linkspeciﬁc parameter P1 by combining the population sizing model (6.9) for the network random keys (compare Sect. 6.5.5), which only holds for the uniformly redundant LB encoding (P1 must be large), with the population sizing model from Sect. 3.1.4, which explains the inﬂuence of nonuniformly redundant representations on GEA performance. We formulate the population sizing model for the LB encoding and present experimental results. In Sect. 6.5.5, we present a population sizing model for the onemax tree problem which was derived for the network random key representation. The network random key representation is almost identical to the LB encoding using large values of P1 . Both encodings are synonymously and uniformly redundant representations deﬁned on realvalued strings of the same length. Only the construction of the tree from the genotypic weights is diﬀerent. Network random keys use Kruskal’s algorithm and do not consider the original distance weights dij , whereas the LB encoding uses Prim’s algorithm for the
192
6 Analysis and Design of Representations for Trees
construction of the phenotypes and considers dij . Therefore, the population sizing model for the network random keys is valid for the LB encoding if a large linkspeciﬁc bias P1 is used. A large value of P1 is necessary to ensure uniform redundancy. The model is formulated as (compare (6.9)) √ √ π π ln(α) n(n − 1)(n − 2) ≈ − ln(α)n1.5 , N =− 4 4 where α is the probability of failure, and n is the number of nodes. It can be seen that the necessary population size N increases with O(n1.5 ). For further information regarding the model, the reader is referred to Sect. 6.5.5. In the following experiments, the optimal solution for the onemax tree problem is always the MST. We want to calculate for a GA using the LB encoding the population size N that is necessary for ﬁnding the optimal solution (the MST) with some probability Popt . The optimal solution is correctly found by a GA if all of the n − 1 links of the optimal solution are correctly identiﬁed. Therefore, Popt = (1 − α)n−1 , where α is the probability of error for one link. We get for
log(Popt ) α = 1 − exp . n−1 Substituting α into (6.9) results in
√ log(Popt ) π ln 1 − exp n(n − 1)(n − 2). N =− 4 n−1
(6.7)
This population sizing model should give us good predictions for the expected minimal population size using the LB encoding with a large linkspeciﬁc bias P1 . The large linkbias ensures that the encoding is uniformly biased. Figure 6.34 shows the theoretical prediction from (6.7) and the experimental results for the LB encoding with P1 = 1, 000, 000. The plots show the necessary population size N over the problem size n for Popt = 0.95. We performed 500 runs for each population size and the resolution for N is 1. We used a standard generational GA with uniform crossover and no mutation. In all runs, we use tournament selection without replacement of size 3 and each run is stopped after the population is fully converged. Because the encoded phenotype depends on the distance weights dij , for every run we randomly placed the nodes on a 1000x1000 square. dij is calculated as the Euclidean distance between the two nodes i and j. Although the population sizing model from (6.7) slightly overestimates the necessary population size N , it still allows a good approximation of the experimental results. As we are mainly interested in investigating the inﬂuence of P1 on the solution quality, and not on the development of a highly accurate population sizing model, we are satisﬁed with the accuracy of this population sizing model. It can be seen that the necessary population size N increases approximately with O(n1.5 ).
6.4 The Link and Node Biased Encoding
193
400
N
200 100 50
experimental results prediction 8
12
16
20 24 28
n
Figure 6.34. Necessary population size N over the problem size n for the onemax tree problem. The optimal solution is the MST and P1 = 1, 000, 000 to ensure uniform redundancy. The results show that the used population sizing model gives an acceptable approximation of the expected GA performance.
In the following paragraphs, we want to consider that the LB encoding becomes nonuniformly redundant with decreasing P1 . With lower P1 , the links that are contained in the MST are overrepresented by the encoding. Therefore, GEA performance increases and the population size that is necessary to ﬁnd the optimal solution (the MST) decreases. We know from Sect. 3.1.4 that the necessary population size N goes with O(2kr /r). r is the number of genotypic BBs that represent the optimal phenotypic BB. For the onemax tree problem we can assume that the size of the BBs k equals one and that each possible link is one phenotypic BB. We have to determine how the diﬀerent phenotypic BBs (the possible edges in the tree) are overrepresented by the LB encoding. In Sect. 6.4.4, we have introduced the probability Pr that a link contained in a randomly created individual is also part of the optimal solution. We can assume that the probability Pr is proportional to r (Pr = const × r). Doubling the probability Pr means that a speciﬁc link of a randomly created individual is twice as often also contained in the optimal solution (the MST). Therefore, doubling Pr has the same eﬀect as doubling r. Furthermore, we can assume that the character of the LB encoding does not change for diﬀerent values of P1 and that kr remains constant. Therefore, the population size N when using the LB encoding goes with O(1/Pr ). From (6.7) we ﬁnally get
N =−
√ log(Popt ) Pru π ln 1 − exp n(n − 1)(n − 2), Pr 4 n−1
(6.8)
where Pru indicates Pr for P1 → ∞ (compare (6.6)). The values of Pr depend on the linkspeciﬁc bias P1 and are shown in Fig. 6.33 for diﬀerent problem sizes n.
194
6 Analysis and Design of Representations for Trees
The theoretical predictions and the empirical results for diﬀerent problem sizes are shown in Figs. 6.35(a) (8 and 20 node onemax tree problem), 6.35(b) (12 and 24 node onemax tree problem), 6.35(c) (16 and 28 node tree onemax tree). The results are split into three plots due to illustrative purposes. The plots show how the necessary population size N depends on the linkspeciﬁc bias P1 . The probability of ﬁnding the optimal solution (the MST) is Popt = 0.95. For determining the relationship between P1 and Pr , which we discussed in Section 6.4.4, we used the results plotted in Fig. 6.33. The lines show the theoretical predictions from (6.8) and the points show the experimental results. In all runs, the optimal solution was the MST and we used the same parameters as for the uniformly redundant LB encoding, whose details are described above.
10
100
N
N
100
10
n=8 (exp.) n=8 (prediction) n=20 (exp.) n=20 (prediction)
1
n=12 (exp.) n=12 (prediction) n=24 (exp.) n=24 (prediction)
1 0.01 0.1 1 10 100
1e+006
0.01 0.1 1 10 100
P1
1e+006
P1
(a) 8 and 20 node onemax tree
(b) 12 and 24 node onemax tree
N
100
10
n=16 (exp.) n=16 (prediction) n=28 (exp.) n=28 (prediction) 0.01 0.1 1 10 100
1e+006
P1
(c) 16 and 28 node onemax tree Figure 6.35. We show how the population size N which is necessary for ﬁnding the optimal solution with probability Popt = 0.95 depends on the linkspeciﬁc bias P1 . In all runs, the optimal solution was the MST. The results show that the proposed population sizing model gives good predictions for the expected solution quality. For small values of P1 the populations size N strongly decreases as the size of the search space collapses and only the optimal solution (the MST) can be represented.
6.4 The Link and Node Biased Encoding
195
The results show that the proposed population sizing model (6.8) gives us a good prediction on how the performance of a GA depends on the linkspeciﬁc bias P1 . There is only a small diﬀerence between the predicted value for N and the actual experimental results. As expected, the population size N declines with decreasing P1 and the problem becomes easier to solve for a GA. Furthermore, we can see that for small values of P1 < 1 the necessary population size N strongly declines and the experimental population size drops much faster than predicted. This is because for P1 < 1 (compare Sect. 6.4.4) the LB encoding does not allow us to encode all possible trees and the search space collapses. Only trees that are similar to the MST can be encoded. Small values of P1 result in high values of Pr (compare Fig. 6.33) which means that most of the links of a randomly created individual are also part of the optimal solution (the MST). In the extreme cases, for P1 → 0 (Pr → 1), the LB encoding can only encode the optimal solution (the MST) and the necessary population size N → 0. This subsection has illustrated that the proposed theoretical concepts describing the inﬂuence of synonymously redundant representations on the performance of GAs can be used for realvalued representations like the LB encoding. The presented results have shown that the proposed theory from Sect. 3.1 predicts the expected GA behavior well. 6.4.6 The LinkandNodeBiased Encoding Palmer (1994) proposed the LNB encoding to overcome some of the problems with the nodebiased (NB) encoding. In this subsection, we investigate how the nonuniform redundancy which we have noticed for the NB and LB encoding is inﬂuenced by the interplay between the two parameters P1 and P2 . In analogy to Sect. 6.4.3, we investigate the bias of the LNB encoding by randomly creating LNBencoded genotypes xgrnd and measuring their minimal phenotypic distance min(dprnd,star ) towards one of the n stars. The more links an individual has in common with one of the stars, the more starlike it is and the lower is the distance. In Fig. 6.36, we present results for randomly created LNBencoded genotypes xgrnd with 8, 16, and 32 nodes. The average minimum distance min(dprnd,star ) towards one of the n stars is plotted over P1 and P2 , and compared to an unbiased encoding (Pr¨ ufer numbers). The parameters P1 and P2 vary between 0 and 1, and we generated 1,000 LNBencoded genotypes xgrnd . The distance weights dij are the Euclidean distances between the nodes i and j. We present results averaged for randomly positioning the nodes 250 times on a twodimensional 1000×1000 grid. The results show for diﬀerent n that the nonuniform redundancy of the LNB encoding depends on the nodespeciﬁc bias P2 . With increasing P2 the individuals are biased towards stars. With P2 dominating P1 , we notice the same behavior as for the NB encoding. With increasing P2 , the encoding can only represent starlike structures, and for P2 large enough, the distance of an individual towards one of the n stars would become zero.
196
6 Analysis and Design of Representations for Trees
p min(d rnd,star)
min(dprnd,star)
4
11 10 9 8
3.5 3 2.5 0
0 P2
0.5
0.5 P1
11
0
0
0.5
0.5
11
P2
linkandnodebiased unbiased
linkandnodebiased unbiased
(a) 8 node
(b) 16 node
P1
p min(d rnd,star)
27 25 23 21
0
0
0.5
0.5 P2
11
P1
linkandnodebiased unbiased
(c) 32 node Figure 6.36. Average minimum phenotypic distance of a randomly generated LNBencoded individual with 8, 16, or 32 nodes to a star. By increasing the nodespeciﬁc bias P2 , an individual is strongly biased towards a star. Higher values for the linkspeciﬁc bias P1 result in a lower bias. Small values of P1 and P2 result in a bias towards the MST.
As discussed in the previous section, with increasing linkspeciﬁc bias P1 the LNB encoding becomes uniformly redundant. To more closely investigate the dependency of the bias on P1 , Fig. 6.37 shows how min(dprnd,star ) depends on P1 for P2 = 0. In accordance with the results from Sect. 6.4.4, with increasing P1 , min(dprnd,star ) stays about constant, and the encoding becomes approximately uniformly redundant. In comparison to nonuniform representations, there is still a small bias. However, as it is very small, we can ignore it. Finally, we want to emphasize that small values of P1 and P2 reduce the performance of the encoding, as the link and node weights have no inﬂuence on the construction of the phenotype. The encoding can only represent trees similar to the MST. The results show that with increasing P2 , a randomly created LNBencoded tree is strongly biased towards stars. By increasing the
6.4 The Link and Node Biased Encoding
3.8 3.7 3.6 0
2
4
6
8
10
P1
(a) 8 node
0
2
4
28 rnd,star)
rnd,star) p
min(d
min(d
p
rnd,star)
4 3.9
12 11.8 linkbiased unbiased 11.6 11.4 11.2 11 10.8 10.6
27.5
linkbiased unbiased
27
p
linkbiased unbiased
4.1
min(d
4.2
197
26.5 26
6
P1
(b) 16 node
8
10
0
2
4
6
8
10
P1
(c) 32 node
Figure 6.37. Average minimum phenotypic distance of a randomly generated LNBencoded individual to a star over the linkspeciﬁc bias P1 . The nodespeciﬁc bias P2 = 0. For large P1 , the encoding becomes unbiased. For P1 → 0, only the MST can be encoded.
linkspeciﬁc bias P1 the LNB encoding becomes less biased. For P1 large enough, the encoding is approximately uniformly redundant if the nodebias is small. 6.4.7 Experimental Results The previous subsections have shown that the NB encoding as well as the LNB encoding are biased towards stars. With increasing nodebias P2 , starlike structures are strongly overrepresented in a randomly generated population. We know from Sect. 3.1 that redundancy favors genetic search if the optimal solutions are overrepresented by the encoding, and hurts genetic search if the optimal solutions are underrepresented. Therefore, we expect high GEA performance if the optimum is a star, and low GEA performance if the optimum is a nonstar such as a random list. Furthermore, we have seen that with a large linkbias P1 , the encoding becomes uniformly redundant. This means, the performance of GEAs should be independent of the structure of the optimal solution with P1 large enough. Finally, we know that the LNB encoding can not work for low values of P1 and P2 because then the node and link weights have no inﬂuence on the structure of the phenotypes and the encoding can only represent MSTlike trees. At the extreme, if P1 = P2 = 0, the genotype has no inﬂuence at all, and only the MST can be represented. To investigate how the performance of GEAs using the LNB encoding depends on P1 and P2 , we use the onemax tree problem from Sect. 6.1.5. We deﬁne the best solution to be either a star or a random list and present the performance of GEAs in Figs. 6.38, 6.39, 6.40, and 6.41. We use a simple GA on a n = 16 node problem with only onepoint crossover, no mutation, tournament selection of size 3, a population size of N = 300, and terminate the run after the population is fully converged. For each parameter setting, we
198
6 Analysis and Design of Representations for Trees
14
avg. fitness of best individual
avg. fitness of best individual
perform 100 runs with diﬀerent randomly chosen positions of the 16 nodes (on a twodimensional 1000×1000 grid). The distance weights dij are calculated according the Euclidean metric (compare (6.1)).
12 10 8
P2=0.0 P2=0.2 P2=0.4 P2=0.6 P2=0.8 P2=1.0 P2=2.0
6 4 2 0 0
10
20
30
40
generation
50
P2=0.0 P2=0.2 P2=0.4 P2=0.6 P2=0.8 P2=1.0 P2=2.0
14 12 10 8 6 4 2 0 0
10
20
30
40
50
generation
(a) Topt is a random star
(b) Topt is a random list
fitness of best individual after 20 gen.
Figure 6.38. Average ﬁtness of the best individual over the number of generations for diﬀerent values of the nodespeciﬁc bias P2 (P1 = 0). The best solution Topt for the 16 node onemax tree problem is either a star or a list. The results reveal that GAs perform better with increasing P2 if Topt is a star. If Topt is a list, GAs fail. We see that the performance of GAs using only the LB encoding strongly depends on the structure of the optimal solution.
16 14 12 10 8 6 4 star list
2 0 0
0.5
1 P2
1.5
2
Figure 6.39. We compare for a 16 node onemax tree problem how the ﬁtness of the best individual after 20 generations depends on P2 for diﬀerent optimal solutions Topt (P1 = 0). If Topt is a star, the ﬁtness increases with larger P2 . If Topt is a list, GAs fail and the ﬁtness at the end of the run is independent of P2 .
14
avg. fitness of best individual
avg. fitness of best individual
6.4 The Link and Node Biased Encoding
12 10 8
P1=0.0 P1=0.5 P1=1.0 P1=2.0 P1=4.0 P1=8.0 P1=16
6 4 2 0 0
10
20
30
40
generation
(a) Topt is a random star
50
199
14 12 10 8
P1=0.0 P1=0.5 P1=1.0 P1=2.0 P1=4.0 P1=8.0 P1=16
6 4 2 0 0
10
20
30
40
50
generation
(b) Topt is a random list
fitness of best individual after 20 gen.
Figure 6.40. Average ﬁtness of the best individual over the number of generations for diﬀerent values of the linkspeciﬁc bias P1 (P2 = 0). The best solution Topt for the 16 node onemax tree problem is either a star or a list. The results reveal that with P1 large enough, the LB encoding is unbiased, and the performance of GAs is independent of the structure of Topt . 16 14 12 10 8 6 4 star list
2 0 0
2
4
6
8 10 12 14 P1
Figure 6.41. We compare for a 16 node onemax tree problem how the ﬁtness of the best individual after 20 generations depends on P1 for diﬀerent optimal solutions (P2 = 0). We see that the performance of GAs is independent of the structure of the optimal solution. Therefore, if the linkbias P1 is large enough, the LB encoding is unbiased.
Figures 6.38(a) and 6.38(b) present the average ﬁtness of the best individual over the run dependent on diﬀerent values of the nodespeciﬁc bias P2 (P1 = 0) if the optimal solution is either a star or a list. We see that with increasing P2 , GEAs ﬁnd the optimal star much faster, whereas GEAs fail completely when searching for the optimal list. The reader should also notice that with increasing P2 the initial population becomes more and more starlike, and the average ﬁtness of the best individual in the initial population becomes higher if the optimum is a star. If the nodebias is very small (P2 → 0) only the MST can be encoded and the only individual a GA can ﬁnd is the MST. As a result, GAs fail for small values of P2 . The problem with
200
6 Analysis and Design of Representations for Trees
P2 becomes more obvious when looking at how the best solution at the end of the run depends on P2 as illustrated in Fig. 6.39. If the optimum is a star, GEAs perform better with increasing P2 . However, if the optimum solution is a list, GEAs are not able to ﬁnd the optimal solution. The situation is diﬀerent when investigating the inﬂuence of the linkspeciﬁc bias P1 (P2 = 0) on the performance of GEAs as illustrated in Figs. 6.40 and 6.41. For both problems, GEAs work properly with P1 large enough. GEAs searching for a star perform as well as when searching for a list. If P1 is large enough, the encoding is (almost) unbiased and GA performance is independent of the structure of the optimal solution. As we have seen before, the encoding is slightly biased towards lists and GAs perform slightly better when searching for optimal lists. However, as the eﬀect is very small we want to neglect it. The experimental results conﬁrm the theoretical predictions from the previous subsections. With a large nodespeciﬁc bias P2 , the LNB encoding is strongly biased towards stars, and GEAs fail if the optimal solution is not a star. With a large linkspeciﬁc bias P1 , the encoding becomes uniformly redundant and GEAs well perform independently of the structure of the optimal solution. If both biases are very small, only the MST can be represented and GEAs not searching for the MST fail. 6.4.8 Conclusions Analyzing the notion of biased encodings as given by Palmer (1994) we recognize that a bias is equivalent to nonuniform redundancy. Therefore, we can use the results from Sect. 3.1 for analyzing the performance of GEAs using the LNB encoding. The performance of GEAs using redundant representations goes with O(r/2kr ), where r denotes the number of copies that are given to the best phenotypic BB and kr denotes the order of redundancy. Palmer, who introduced the LNB encoding, drew the conclusion that the “... new Link and Node Bias (LNB) encoding was shown to have all the desirable properties ...” (Palmer 1994, pp. 90) illustrated in Sect. 6.1.6 including those to be unbiased that means uniformly redundant. However, we have seen that this claim is not true. With increasing nodespeciﬁc bias P2 , the LNB encoding becomes more and more biased towards stars and for P2 → ∞, the LNB encoding is only able to represent stars. Therefore, GEAs using a large nodespeciﬁc bias can not work properly. Fortunately, the encoding becomes uniformly redundant with increasing linkspeciﬁc bias P1 . For P1 → ∞, all phenotypes are about uniformly represented, and GEAs perform independently of the structure of the optimal solution. Finally, the LNB encoding has problems if both biases are small because then the encoding can only represent phenotypes that are similar to the min
6.5 Network Random Keys (NetKeys)
201
imum spanning tree. At the extreme, for P1 → 0 and P2 → 0, the genotype has no inﬂuence on the phenotype and only the MST can be encoded. Because optimal solutions for the optimal communication spanning tree problem (see Sect. 8.2) often tend to be star or MSTlike, the LNB encoding could be a good choice for this problem. In general, however, the encoding has some serious problems, especially when using the simpliﬁed NB encoding. Researchers should therefore be careful when using this encoding for other problems because some trees are not encoded at all, and a randomly generated LNBencoded genotype can be biased towards stars or MSTs. As a result, we strongly encourage users to use, as long as they have no idea about the structure of the optimal solution, high values for the linkspeciﬁc bias, and to discard the nodespeciﬁc bias. Otherwise, GEAs are likely to have large problems in ﬁnding optimal nonstar or nonMST trees, and a reduction of GEA performance is unavoidable.
6.5 Network Random Keys (NetKeys) The representation framework we developed in Chaps. 3 and 4 tells us that highquality representations should be robust, allow genetic operators to work properly, and have high locality. Redundant representations should be synonymously and nonuniformly redundant if no knowledge regarding the optimal solutions exists. However, Pr¨ ufer numbers, the CV encoding, and the LNB encoding have shown to have problems with some of these issues. Pr¨ ufer numbers have low locality and make easy problems more diﬃcult, the CV encoding is nonsynonymously redundant, and the LNB encoding is nonuniformly redundant. Therefore, Pr¨ ufer numbers and the CV encoding are not suitable encodings for easy problems as they make these problems more diﬃcult. Only the LNB encoding can result in good GEA performance if the encoding parameters P1 and P2 are properly set. The purpose of this section is to design a new representation by combining the advantageous properties of the CV and the LNB encoding. As a result, we get the synonymously and uniformly redundant network random key encoding (NetKeys). The NetKey encoding belongs to the class of weighted encodings. In contrast to other representations such as the CV encoding which can only indicate whether a link is established or not, weighted encodings use weights for the genotype and can thus encode the importance of links. Consequently, an additional construction algorithm is necessary which constructs a valid tree from the genotypic weights (the random key sequence of length l = n(n−1)/2). The section starts by illustrating how the NetKey encoding can be designed by combining the CV encoding with some elements of the LNB encoding. This is followed in Sect. 6.5.2 by the functionality of the NetKeys. We illustrate the random keys which store the importance of the links as weighted vectors, and the construction algorithm which constructs a valid tree from a random key sequence. Section 6.5.3 summarizes the properties of the NetKey encoding.
202
6 Analysis and Design of Representations for Trees
This is followed by an investigation into whether the synonymously redundant NetKey encoding is uniformly redundant, or not. For randomly created genotypes, we measure the distance towards stars and MST, and provide empirical veriﬁcation that GEAs using NetKeys perform independently of the structure of the optimal solution. Before closing the section with concluding remarks, Sect. 6.5.5 presents a model for the population sizing and run duration of GEAs using NetKeys for the onemax tree problem. 6.5.1 Motivation We have seen in Sect. 6.3.3 that the CV encoding is uniformly redundant. A repair mechanism is necessary that assigns all infeasible genotypes to feasible solutions. Because an allele only indicates if a link is established or not, the repair mechanism must rely on random link insertion or deletion. Therefore, the CV encoding is nonsynonymously redundant, as not all infeasible genotypes that are assigned to one phenotype are similar to each other. We want to use the functionality of the CV encoding as the basis for NetKeys. When trying to improve the CV encoding, we have to overcome the problem of nonsynonymous redundancy. Furthermore, we must ensure that the new encoding remains uniformly redundant. Therefore, we replace the binary alleles, which only indicate if a link is established or not, by continuous alleles which encode the importance of a link by a weighted value (a randomly chosen number). We have seen in Sect. 6.4 that the linkbiased (LB) encoding which uses genotypic weights is synonymously redundant. When combining the link weights of the LB encoding with the principles of the CV encoding, we get the NetKey encoding which is synonymously redundant (all genotypes that represent one phenotype are similar to each other) and nonuniformly redundant (no phenotypes are overrepresented). By using a weighted instead of a binary encoding, the NetKey encoding inherits most of the properties (for example the synonymous redundancy) from the LB encoding. In contrast to the LB encoding, the alleles of the NetKeys directly encode the importance of a link and not a bias of the distance matrix. Furthermore, we do not use Prim’s algorithm (Prim 1957) which constructs the MST from the modiﬁed distance matrix, but Kruskal’s algorithm (Kruskal 1956). The construction algorithm is unbiased and uses only the genotypic weights and no distance weights for the construction of the phenotype. 6.5.2 Functionality For describing the functionality of the NetKey encoding, we have to separate the representation into two parts: Firstly, the genotype which is a sequence of random keys. It stores the importance of the links as a weighted vector of length l = n(n − 1)/2. Secondly, the construction algorithm (genotypephenotype mapping) which constructs a tree (phenotype) from a random key sequence (genotype).
6.5 Network Random Keys (NetKeys)
203
Random Keys By substituting the zeros and ones in the CV encoding by continuous values that can describe the importance of the links, the ﬁrst part of NetKey functionality is deﬁned. However, the idea to use a weight for describing the importance of an allele is not new, and has already been presented in a different context as the so called random key (RK) encoding. For other work about weighted encodings in the context of tree representations the reader is referred to Palmer (1994) or Raidl and Julstrom (2000). The RK representation for encoding permutations was ﬁrst presented by Bean (1992). Later, the encoding was also proposed for single and multiple machine scheduling, vehicle routing, resource allocation, quadratic assignment the, and traveling salesperson problems (Bean 1994). Norman and Bean (1994) reﬁned this approach (Norman and Bean 2000) and applied it to multiple machine scheduling problems (Norman and Bean 1997). An overview of using RKs for scheduling problems can be found in Norman (1995). In Norman and Smith (1997) and Norman et al. (1998), RKs were used for facility layout problems. In Knjazew (2000) and Knjazew and Goldberg (2000), a representative of the class of competent GAs (fast messy GA (Goldberg et al. 1993)) was used for solving ordering problems with RKs. The RK representation uses random numbers for the encoding of a solution. A key sequence of length l is a sequence of l distinct real numbers (keys). The values are initially chosen at random, are ﬂoating numbers between zero and one, and are only subsequently modiﬁed by mutation and crossover. An example for a key sequence is r = (0.07, 0.75, 0.56, 0.67). Of importance for the interpretation of the key sequence is the position and value of the keys in the sequence. If we assume that Zl = {0, . . . , l − 1} then a permutation σ can be deﬁned as a surjective function σ : Zl → Zl . For any key sequence r = r0 , . . . , rl−1 , the permutation σr of r is deﬁned as the sequence with elements (σr)i = rσ(i) . The permutation rs corresponding to a key sequence r of length l is the permutation σ such that σr is decreasing (for example, i < j ⇒ (σr)i > (σr)j ). The ordering corresponding to a key sequence r of length l is the sequence σ(0), . . . , σ(l − 1), where σ (also denoted as rs ) is the permutation corresponding to r. This deﬁnitions say that the positions of the keys in the key sequence r are ordered according to the values of the keys in descending order. In our example, we have to identify the position of the highest value in the key sequence r (r1 = 0.75). The next highest value is r3 = 0.67. We continue ordering the complete sequence and get the permutation rs = 2, 4, 3, 1. In the context of scheduling problems, this permutation can be interpreted as a list of jobs that are executed on one machine (We start with job 2, then continue with job 4, job 3, and job 1). From a key sequence of length l, we can always construct a permutation of l numbers. Every number between 0 and l − 1 appears in the permutation only once as the position of each key is unique. Here are some properties of the encoding.
204
6 Analysis and Design of Representations for Trees
• A valid permutation rs of l numbers can be created from all possible key sequences r as long as there are no two keys ri that have the same value (ri = rj for i = j and i, j ∈ {0, l − 1}). Therefore, every random key sequence r can be interpreted as a permutation rs . • There are many possibilities for the construction of a key sequence r from a permutation rs . All elements ri can be scaled up by some factor and r still represents exactly the same permutation rs . As long as the relative ordering of the keys in r remains the same, diﬀerent key sequences always represent the same permutation. It is necessary that rs is a permutation of l numbers, otherwise no key sequence r can be constructed from rs . • RKs encode both, the relative position of a number in the permutation rs (encoded by the value of the key at position i in comparison to all other keys) and the absolute position of i in rs . The relative position of a number i in the permutation rs is determined by the numbers that precede and follow i. It is determined directly by the weights ri . All numbers j in the sequence rs that follow i correspond to lowervalued keys (rj < ri ), whereas all numbers j that precede i correspond to highervalued keys (rj > ri ). In the context of scheduling problems, all jobs where the corresponding key has a higher value than the ith key are executed before job i, and all jobs with a corresponding key with lower value are executed after i. In contrast, the absolute position of a number i in the permutation rs cannot be encoded directly, but is only indirectly determined by the value of the ith key. The absolute position describes at which position in the permutation rs a number i appears. A large value at the ith position results in a position at the beginning of the permutation rs , and a low value results in a position at the end of rs . • The distinction between relative and absolute position of a number in the permutation rs is important for the synonymity of RKs. The synonymity of a redundant encoding, which is based on the locality of a nonredundant encoding, describes how similar the genotypes are that represent the same phenotype. A representation is synonymously redundant if mutating a genotype changes the corresponding phenotype only slightly. A look at RKs shows that they are synonymously redundant when used for ordering problems. A small change in the genotype (the key sequence r) leads to a small change in the phenotype (the permutation rs ). The change of one key changes the relative position of exactly one number. However, one must be careful with the deﬁnition of the phenotypic neighborhood. If the absolute position of the numbers in rs is relevant for the phenotype, a change of one key is disastrous and the representation is nonsynonymously redundant. If the value of the key ri with the highest value is modiﬁed, only the number i changes its relative position in the permutation rs , but up to l numbers change their absolute position in the permutation. However, as we use RKs to represent a permutation of numbers, only the relative, and not the absolute positions of the numbers in the permutation must be
6.5 Network Random Keys (NetKeys)
205
considered. And for problems where the relative positions of numbers are important, RKs are synonymously redundant. • When using GEAs with RKs, standard crossover and mutation operators can be used and are expected to work well. No repair mechanism, or problemspeciﬁc operators, are necessary when using this encoding for ordering problems. The standard one or multipoint crossover schemes work well (Bean 1994) because the relative ordering of the positions in the parents is preserved and transferred to the oﬀspring (Fox and McMahon 1991). Due to the synonymous redundancy of the encoding we expect standard mutation operators to work well and to construct oﬀspring that are similar to their parents. We have seen that RKs have interesting properties. When using them for the encoding of trees, we still have to deﬁne exactly how a tree can be constructed from them. Constructing Trees from Random Keys After we have presented RKs as the basis for the NetKey encoding, we still have to deﬁne a construction algorithm which creates a valid tree from a RK sequence. Both elements, the RKs and the construction algorithm are necessary for the new NetKey encoding. To get a synonymously and uniformly redundant encoding, we demand the construction algorithm to preserve the synonymity of the RK encoding and not to favor some phenotypes but to work uniformly. We have seen that we are able to give priority to the objects in the permutation when using RKs. As NetKeys use continuous variables that could be interpreted as the importance of the link, it is possible to distinguish between more and less important links. The higher the value of the allele, the higher the probability that the link is used for the tree. When constructing the tree, the positions of the keys in the key sequence r are interpreted in the same way as for the CV. The positions are labeled and each position represents one possible link in the tree. From a key sequence r of length l = n(n − 1)/2, a permutation rs of l numbers can be constructed as illustrated above. Then the tree is constructed from the permutation rs as follows: 1. Let i = 0, T be an empty tree with n nodes, and rs the permutation of length l = n(n − 1)/2 that can be constructed from the key sequence r. All possible links of T are numbered from 1 to l. 2. Let j be the number at the ith position of the permutation rs . 3. If the insertion of the link with number j in T would not create a cycle, then insert the link with number j in T . 4. Stop, if there are n − 1 links in T . 5. Increment i and continue with step 2.
206
6 Analysis and Design of Representations for Trees
The construction rule is based on Kruskal’s algorithm (Kruskal 1956) and only considers the weights of the RK vector for building the tree. With this rule, we can construct a unique, valid tree from every possible RK sequence. Thus, the NetKey encoding is now completely described: The new encoding uses RKs which allows us to give priority to some links, and the construction rule uses this information and gradually builds a valid tree. A
C
B
D
E
Figure 6.42. A ﬁve node tree
We want to illustrate the functionality of the NetKey encoding with an example. We use the key sequence from Table 6.7. The permutation rs = 10, 8, 6, 9, 2, 7, 1, 5, 4, 3 can be constructed from the random key sequence r. We start constructing the tree T by adding the link DE (position 10) to the tree. This is followed by adding CD (position 8) and BD (position 6). If we add the link CE (position 9) to the tree, the cycle CEDC would be created, so we skip CE and continue by adding AC (position 2). Now we have a tree with four edges and terminate the construction algorithm. We have constructed the tree shown in Fig. 6.42. position 1 2 3 4 5 6 7 8 9 10 value 0.55 0.73 0.09 0.23 0.40 0.82 0.65 0.85 0.75 0.90 link AB AC AD AE BC BD BE CD CE DE
Table 6.7. A key sequence r
The computational eﬀort for constructing the phenotype from the genotype is similar for the NetKey and the LB representation. The calculation of the permutation from the key sequence r can be done in O(l log(l)) (sorting an array of l numbers). The process of constructing the graph from the permutation rs is comparable to repairing an invalid graph that is constructed from a CV and its eﬀort depends on the speciﬁc structure of the phenotype. In analogy to the LB encoding with a large linkspeciﬁc bias P1 → ∞, NetKeys are synonymously redundant. A mutation (changing the value of one key) results either in no change of the corresponding phenotype if the relative ordering is not changed, or the change of two edges if the relative position is changed. Therefore, the maximum phenotypic distance dpxp ,yp between two neighboring genotypes xg and y g is one (compare Sect. 6.1.2 about the deﬁnition of distance). The reader should observe that a mutation of one key of the genotype often dramatically changes the absolute positions of the numbers in the permutation rs . However, the construction rule we deﬁned is
6.5 Network Random Keys (NetKeys)
207
only based on the relative ordering of rs . Therefore, we do not have to worry about the change of the absolute positions. 6.5.3 Properties We summarize the properties of the NetKey encoding. The use of the NetKey encoding has some remarkable advantages: • The encoding is synonymously redundant and standard crossover and mutation operators work properly. • The encoding allows a distinction between important and unimportant links. • There are no unfeasible solutions. In this section, we brieﬂy discuss these properties. In Sect. 3.3.2, we have stipulated that mutation operators must create an oﬀspring which is genotypically and phenotypically similar to its parent. Therefore, a small genotypic distance between two individuals should correspond to a small phenotypic distance. Then, the encoding has high locality. Based on locality, Sect. 3.1.2 introduced the synonymous redundancy of a redundant representation which is equivalent to the high locality of a nonredundant representation. When using synonymously redundant representations, all genotypes that correspond to a phenotype are similar to each other (have small distances). A glance at the NetKey encoding shows that the mutation of one key results either in the same, or in a neighboring tree, which makes it synonymously redundant. Furthermore, in Sect. 3.3.5, we have determined that recombination operators create an oﬀspring which inherits the properties of its parents. In terms of metric, the distance of an individual to its parents should be smaller than the distance between both parents. In terms of links, an oﬀspring should inherit the links from its parents. Standard recombination operators, like npoint or uniform crossover, show this behavior when used for NetKeys: if a link exists in a parent, the value of the corresponding key is high in comparison to the other keys. After recombination, the corresponding key in the oﬀspring has the same high value and is therefore also used with high probability for the construction of the oﬀspring. As a result, both types of operators, mutation and recombination, work well when used for the NetKey encoding. GEAs using NetKeys are able to distinguish between important and unimportant links. In contrast to the CV encoding, which only stores information about whether a link is established or not, GEAs using NetKeys are able to identify the important links in the tree. As the CV encoding can not store information regarding the importance of a link, the repair process must delete or insert links randomly. High quality links can be accidentally removed, or low quality links can ﬁnd their way back into the population. Finally, NetKeys always encode valid trees. No over or underspeciﬁcation of a tree is possible. The construction process which builds a tree from a RK
208
6 Analysis and Design of Representations for Trees
ensures that NetKeys only encode valid solutions. Thus, we do not need an additional repair mechanism. We see that the NetKey encoding has some remarkable beneﬁts. However, we have not yet investigated whether the encoding is uniformly redundant. We want to do this in the following subsection. 6.5.4 Uniform Redundancy The NetKey encoding is a synonymously redundant encoding. To ensure that GEAs perform independently of the structure of the optimal solution, NetKeys should be uniformly redundant, i.e. unbiased. This section examines the bias of the NetKey encoding. We measure, in analogy to Sect. 6.3.3, for randomly created NetKey genotypes xgrnd , the minimum phenotypic distance min(dprnd,star ) towards stars, and the average phenotypic distance dprnd,M ST towards the MST. This is followed by empirical evidence of the uniform redundancy of the encoding. NetKeys are redundant because they encode a ﬁnite number of phenotypes using continuous genotypes. We know from Sect. 4.4.1 that GEA performance depends on the location of the optimal solution in the search space if an encoding is nonuniformly redundant. GEAs searching for the optimal solution only perform well if the encoding is not biased towards the lowquality solutions. Table 6.8. Mean and standard deviation of min(dprnd,star ) and dprnd,M ST for randomly created NetKey genotypes xgrnd n 8 16 32
min(dprnd,star ) unbiased NetKey µ σ µ σ 3.67 0.643 3.75 0.602 10.91 0.783 11.00 0.759 26.25 0.818 26.34 0.800
dprnd,M ST unbiased NetKey µ σ µ σ 5.16 0.993 5.24 0.961 13.08 1.072 13.13 1.041 29.08 1.311 29.07 1.319
In Table 6.8, we present for randomly created NetKey genotypes xgrnd the average minimum distance min(dprnd,star ) to a star, and the average distance dprnd,M ST to the MST. The MST (compare (8.3)) is calculated based on the distance weights dij which are the Euclidean distances between the nodes i and j. The nodes are randomly placed on a twodimensional 1000×1000 grid. We randomly create 10,000 solutions for each problem instance and show the mean µ and the standard deviation σ of the distances. The numbers indicate that although there is a small bias, we can view the NetKey encoding as uniformly redundant and assume that all phenotypes are represented uniformly. The small bias is expected to be a result of the construction process. For examining how GA performance depends on the structure of the optimal solution, we use the onemax tree problem from Sect. 6.1.5. Our GA
1 0.8 0.6 0.4
random star arbitrary tree random list
0.2 0 10
20
30
40
50
fitness at end of the run
probability of sucess
6.5 Network Random Keys (NetKeys) 7 6.5 6 5.5
random star arbitrary tree random list
5 4.5 10
1 0.8 0.6 random star arbitrary tree random list
0.2 0 20
40
60
80 100 120
star arbitrary tree random list
0.6 0.4 0.2 0 50
100
150
200
population size n
40
50
60
12 10
random star arbitrary tree random list
8 20
40
60
80
100 120
population size n
250
fitness at end of the run
probability of sucess
0.8
30
14
population size N 1
20
population size N fitness at end of the run
probability of sucess
population size N
0.4
209
30 25 20 star arbitrary tree random list
15 10 50
100
150
200
250
population size n
Figure 6.43. Performance of GAs using NetKeys for 8 (top), 16 (middle) and 32 (bottom) onemax tree problems. The plots show either the probability of success (left) or the ﬁtness at the end of a run (right). The GAs search for the optimal star, list, or arbitrary tree. The plots indicate that GA performance is independent of the structure of the optimal solution. Therefore, NetKeys are uniformly redundant.
only uses uniform crossover, no mutation, tournament selection without replacement of size 3, and stops after the population is fully converged. On the left of Fig. 6.43, we show the probability of ﬁnding the optimal solution (a randomly chosen star, list, or arbitrary tree) over the population size N . The right ﬁgures show the ﬁtness of the best individual at the end of the run over the population size N . Both plots conﬁrm that GEAs perform almost independently of the structure of the optimal solution. The reader should observe that the ﬁgures indicate a slightly better performance for stars. We
210
6 Analysis and Design of Representations for Trees
believe that omitting invalid links during the construction process results in this small inﬂuence on GA performance. We ignore this small bias and assume that NetKeys are uniformly redundant. Our results indicate that the NetKey encoding is approximately uniformly redundant and GEAs using NetKeys perform nearly independently of the structure of the optimal solution. 6.5.5 Population Sizing and Run Duration for the OneMax Tree Problem We examine the necessary population size and run duration of GAs using NetKeys. We present theoretical models for the onemax tree problem (compare Sect. 6.1.5) and provide empirical veriﬁcation. Population Sizing When extending the population sizing equation of Harik et al. (1997) from a binary alphabet to a χary alphabet, we get: Nmin = −
σf √ χk ln(α) π, 2 d
where χ is the cardinality of the alphabet, α is the probability of failure, σf is the overall variance of the function, and d is the signal diﬀerence between the best and second best BB. For calculating σf we have to investigate how to decide among the competing BBs. For the onemax tree problem we have to ﬁnd these n − 1 links the optimal solution is constructed from. The key sequence r that represents a tree with n nodes consists of l = n(n − 1)/2 diﬀerent keys. For the construction of the tree, these n − 1 keys ri are used that have the highest value. Therefore, we can split the n(n − 1)/2 diﬀerent keys ri in n/2 diﬀerent groups of size (n − 1). Finding the optimal solution means that all keys ri with the links i contained in the optimal solution can be found in one group, which is considered for the construction of the tree and thus contains the keys with the n − 1 highest values of ri . A good decision among competing BBs means deciding between the n/2 diﬀerent groups of size n − 1 and identifying the correct one. A key ri can belong either to the one group that is considered for the construction of the tree (the key has a high value), or to one of the n − 2 groups that are not considered. This is similar to the needle in a haystack model and the standard deviation for such a case is (Goldberg et al. 1992) 2l(n − 2) (n − 1)(n − 2) √ = ≈ n. σf = n n As we have n/2 diﬀerent partitions, the cardinality χ of the alphabet is n/2. Using these results and with k = 1 (the onemax tree problem is fully easy,
6.5 Network Random Keys (NetKeys)
211
and there are no interdependencies between the alleles), and d = 1, we get an approximation for the population size Nmin : √ √ π π ln(α) n(n − 1)(n − 2) ≈ − ln(α)n1.5 (6.9) Nmin = − 4 4
minimum population size Nmin
The necessary population size Nmin goes with O(n1.5 ).
200
100
50 experimental results prediction 25 8
10 12 14 16
20
problem size n
26
Figure 6.44. Minimum pop size Nmin for NetKeys over the problem size n for the onemax tree problem. The probability of ﬁnding the optimal solution is Pn = 0.95. The population size goes with O(n1.5 ).
In Fig. 6.44, the minimum necessary population size Nmin that is necessary for solving the onemax tree problem with probability Pn = 1 − α = 0.95 is shown over the problem size n. We use a simple GA with tournament selection without replacement of size 3, uniform crossover and the NetKey encoding. We perform 500 runs for each population size and for N > Nmin the GA is able to ﬁnd the optimum with probability p = 0.95 (α = 0.05). Although, we have to make some assumptions in our derivation, and the exact inﬂuence of the construction algorithm of the phenotype from the genotypic weights is diﬃcult to describe theoretically, the population sizing model gives us a good approximation of the expected population size N which goes with O(n1.5 ). Run Duration In M¨ uhlenbein and SchlierkampVoosen (1993) and Thierens√ and Goldberg (1994), the time until convergence is deﬁned as tconv = π l/2I with the selection intensity I and the string √ length l. I depends only on the used ack et al. selection scheme and is I = 3/(2 π) for a tournament size of 3 (B¨ 1997, C 2.3). With l = n(n − 1)/2, we get tconv ≈ const × n. Therefore, the run duration tconv of a GA should go linearly with the problem size n of the onemax tree problem. In Fig. 6.45, we show the run duration tconv over the problem size n for tournament selection without replacement of size 3 and uniform crossover. tconv measures the number of generations until the population is completely
212
6 Analysis and Design of Representations for Trees
run duration tconv
60 50 40 30 20 8 10 12 14 16 18 20 22 24 26 problem size n
Figure 6.45. Run duration tconv over problem size n for the onemax tree problem using tournament selection without replacement and uniform crossover
converged. The population size was chosen as N = 2 × Nmin and Nmin is from Fig. 6.44. The population size N is large enough to ensure that the optimal solution was found in all 500 runs which we performed for every problem instance. The results show that tconv grows, as predicted, linearly with increasing n. 6.5.6 Conclusions This section presented the NetKey encoding. We started by illustrating how we can combine characteristic vectors with some elements of the linkbiased encoding to get the NetKey encoding. This was followed in Sect. 6.5.2 by the functionality of NetKeys. We explained the principles of random keys and illustrated how we can construct a valid tree from a random key sequence. After all components of the NetKey encoding were deﬁned, we summarized in Sect. 6.5.3 important properties of the new encoding. Because NetKeys are a redundant encoding, Sect. 6.5.4 presented an investigation into the bias of the encoding. Finally, based on existing theory, we developed in Sect. 6.5.5 a population sizing and run duration model for GAs using NetKeys and solving the onemax tree problem. The section demonstrates that using the framework outlined in Chap. 4 allows theoryguided design of highquality representations. Based on the insights into the principles of representations, we were able to develop the new NetKey encoding. NetKeys are based on the CV encoding, but use continuous weights for encoding information about the represented tree. Therefore, NetKeys are similar to the linkbiased encoding with a large linkspeciﬁc bias P1 → ∞, but use a diﬀerent construction rule for the phenotypes. The investigation into the properties of the NetKeys revealed that the encoding is synonymously and uniformly redundant, that standard crossover and mutation operators work properly, and that the representation does not change problem diﬃculty. Based on the presented results, we encourage further study of NetKeys for encoding both trees and other networks. The use of existing theory for
6.6 Conclusions
213
formulating a population sizing model as well as a time to convergence model illustrated the beneﬁts we can get from using existing theory. We encourage users to use existing theory for predicting GEA behavior more frequently. Finally, even though more work is needed, we believe that the properties presented are suﬃciently compelling to immediately recommend increased application of the NetKey encoding.
6.6 Conclusions In this chapter, we used the framework from Chap. 4 for the analysis and design of tree representations. By doing this, we were able to illustrate the relevance of the basic design principles more clearly and to understand the inﬂuence of common tree representations on the performance of GEAs. We started in Sect. 6.1 by providing the necessities for analyzing tree representations. We deﬁned the network design problem and presented the used metric for graphs. Based on the schema analysis for graphs we presented in Sect. 6.1.5 scalable test problems for trees (onemax tree problem and deceptive trap tree problem). The section ended with a review of design criteria for trees as given by Palmer (1994). Section 6.2 presented an investigation into the properties of the Pr¨ ufer number encoding. After an historical review, the construction rule, and known properties of the encoding, we focused in Sect. 6.2.4 on the low locality of the encoding. We performed random walks through the search spaces and showed that the locality of the representation is low. This was followed by an investigation into the locality of neighboring individuals. The section ended with empirical evidence of the theoretical predictions of GEA performance. Section 6.3 focused on the characteristic vector (CV) encoding as an example for a uniformly redundant representation. We described how trees can be represented by the CV encoding, and how invalid solutions can be handled by repairing them. Examining the repair process for CVs revealed that the encoding is nonsynonymously redundant as genotypes that represent the same phenotype are not similar to each other. Therefore, GEA performance is reduced and the time to convergence tconv increases. In Sect. 6.4, we examined the link and node biased (LNB) encoding. We started by illustrating the motivation for developing the encoding and described its diﬀerent variants. This was followed by illustrating that the LNB encoding is synonymously and nonuniformly redundant. A closer examination of the nonuniform redundancy revealed that the LNB encoding overrepresents stars if a nodespeciﬁc bias P2 is used and it overrepresents the MST if a linkspeciﬁc bias P1 is used. Only for P1 → ∞ does the LNB encoding become uniformly redundant. As a result, GEA performance depends on the structure of the optimal solution. Finally, we veriﬁed the theoretical predictions concerning GEA performance by empirical results.
214
6 Analysis and Design of Representations for Trees
In the last section, we designed the NetKey representation using the insights into the inﬂuence of representations on GEA performance. NetKeys are synonymously and uniformly redundant and encode a tree using a weighted vector. We discussed the motivation for designing the new encoding, reviewed its functionality, and summarized important properties of NetKeys. Section 6.5.4 analyzed the bias of the encoding and showed that it is nonuniformly redundant. Furthermore, in Sect. 6.5.5 we used the concepts from Sect. 3.1.4 and developed a population sizing model for NetKeys and the onemax tree problem. In this section, we used the framework about the inﬂuence of representations on GEA performance for the analysis of existing tree representations and for the design of a new representation. The framework allowed us to predict how the performance of GEAs, measured by run duration and solution quality, is aﬀected by the used representation. We were able to compare representations in a theorybased manner, to predict the performance of GEAs using diﬀerent representations, and to analyze representations guided by theory. The analysis showed that the proposed elements of the framework – redundancy, scaling, and locality – can be used for analyzing representations. We want to brieﬂy summarize the insights our analysis revealed: Our investigation into the locality of the Pr¨ ufer number encoding has shown that the locality is diﬀerent in diﬀerent areas of the search space. For trees that are similar to stars, the encoding has high locality, and BBcomplexity is the same for the genotypes and phenotypes. However, for nonstars, the encoding has low locality and easy phenotypic problems, where the optimal solution is a nonstar, become more diﬃcult to solve when using Pr¨ ufer numbers. These insights explain the inconsistent statements about Pr¨ ufer number’s performance in the literature. If the optimal solution was accidentally starlike, the encoding shows an acceptable performance; if it was nonstarlike, GEAs fail. We presented the CV encoding as an example of a uniformly redundant representation that is nonsynonymously redundant. We recognized that an encoding that allows the representation of invalid solutions, like the CV encoding, is redundant. Such an encoding is uniformly redundant if the repair process is unbiased, that means it does not overrepresent some phenotypes. However for the CV encoding, the repair process results in nonsynonymous redundancy which has the same eﬀect as low locality for nonredundant representations. Nonsynonymous redundancy reduces solution quality and increases the run duration tconv . The investigation into the synonymously redundant variants of the LNB encoding illustrated the inﬂuence of nonuniform redundancy on GEA performance. The LNB encoding overrepresents either stars or the MST. Only for the linkspeciﬁc bias P1 → ∞ does the encoding become uniformly redundant. In general, GEAs using the LNB encoding have large problems in ﬁnding optimal solutions if these are not a star or the MST. The analysis of the redundant linkbiased encoding, which was shown in Sect. 6.4.4, can be generalized and
6.6 Conclusions
215
is helpful for the analysis of redundant representations. By examining the size of the search space, the synonymity of the representation, the order of redundancy, and the over and underrepresentation of solutions, the inﬂuence of a redundant representation on GEA performance can be analyzed. The section about the NetKey encoding illustrated how the representation framework can be used for the design of highquality representations. The NetKey encoding is synonymously and uniformly redundant and allows eﬃcient GEA search. Last but not least, in Sect. 6.1, we presented a schema analysis for graph problems. Using it, we were able to measure the phenotypic problem complexity of a graph problem and to classify problems to be easy or diﬃcult. Furthermore, it can help us to judge if encodings preserve problem diﬃculty because we can measure if the problem complexity remains constant when mapping the phenotypes on the genotypes. Based on the schema analysis for graphs, we provided a fully diﬃcult deceptive trap and a fully easy onemax tree optimization problem. Both scalable test problems are helpful for comparing the performance of diﬀerent tree encodings. Furthermore, the test problems allow users to easily examine if GEA performance depends on the structure of the optimal solution. This is important for investigating eﬀects that can be caused by nonuniformly redundant encodings. This chapter has applied the principles of representations from Chap. 3 to common tree encodings. By identifying Pr¨ ufer numbers to have low locality, the LNB encoding to be redundant but biased, and the CV encoding to be uniformly and nonsynonymously redundant, we were able to predict the behavior and performance of GEAs using these representations. In general, by applying the presented theory of representations to other not mentioned, or new representations, the behavior of GEAs using these encodings can be much better predicted. Therefore, we want to encourage researchers to use the presented theory for representations from Chaps. 3 and 4 for analyzing other representations.
7 Analysis and Design of Search Operators for Trees
When using GEAs for tree problems it is necessary to encode a solution (tree) such that evolutionary search operators like crossover or mutation can be applied. There are two diﬀerent possibilities for doing this: indirect representations usually encode a tree (phenotype) as a list of strings (genotypes) and apply standard search operators to the genotypes. The phenotype is constructed by an appropriate genotypephenotype mapping (representation). As seen in the previous chapter, there are many indirect representations for trees such as NetKeys, the LNB encoding, the CV encoding, or Pr¨ ufer numbers. In contrast, direct representations encode a tree as a set of edges and apply search operators directly to the set of edges. Therefore, no representation is necessary. Instead, treespeciﬁc search operators must be developed as standard search operators can no longer be used. This chapter uses the insights into representation theory for the analysis and design of search operators for trees. In contrast to Chap. 6, where standard search operators are applied to treespeciﬁc genotypes, here treespeciﬁc search operators are directly applied to the phenotypes as there is no additional genotypephenotype mapping. Section 7.1 presents a direct representation for trees (NetDir) as an example for the design of direct tree representations. Search operators are directly applied to trees and problemspeciﬁc crossover and mutation operators are developed. The search operators for the NetDir representation are developed based on the notion of schemata from Sect. 6.1.4. Section 7.2 analyzes the edgeset encoding (Raidl and Julstrom 2003) which encodes trees directly by listing their edges. Search operators for edgesets are either heuristic considering the weights of edges they include in oﬀspring, or naive, including edges without regard to their weights. Analyzing the properties of the heuristic variants of the search operators shows that solutions similar to the minimum spanning tree are favored. In contrast, the naive variants are unbiased which means that genetic search is independent of the structure of the optimal solution. Although no explicit genotypephenotype mapping exists for edgesets and the framework for the design of representations can not be directly applied, it is useful for structuring the analysis of
218
7 Analysis and Design of Search Operators for Trees
edgesets. The results of the analysis show that similarly to nonuniformly redundant representations, edgesets overrepresent some speciﬁc types of trees, and GEA performance increases if optimal solutions are similar to the MST. Analyzing and developing direct representations nicely illustrates the tradeoﬀ between designing either problemspeciﬁc representations or problemspeciﬁc operators. For eﬃcient GEAs, it is necessary to design either problemspeciﬁc representations and to use standard operators such as onepoint or uniform crossover, or to develop problemspeciﬁc operators and use direct representations.
7.1 NetDir: A Direct Representation for Trees The purpose of this section is to develop a direct representation for trees (NetDir) and to illustrate that when using direct representations, the design task of ﬁnding proper representations is substituted by the search for good crossover and mutation operators. When using GEAs based on the notion of schemata, these problemspeciﬁc operators must obey the linkage in the phenotypes and process BBs properly. Therefore, by using direct representations it is not possible to get rid of the diﬃculties in designing eﬃcient optimization methods. The section starts with a brief historical review of direct representations for trees. In Sect. 7.1.2, we discuss the properties of direct representations. We demonstrate the beneﬁts and drawbacks of using direct representations for GEAs. Because NetDir directly represents trees as graph structures and not as a list of alleles, standard genetic operators can not be used any more. Therefore, problemspeciﬁc operators are necessary. Consequently, in Sect. 7.1.3 we develop mutation and crossover operators for the NetDir representation. The section ends with a short summary. 7.1.1 Historical Review One of the ﬁrst approaches to direct representations for trees was presented by Piggott and Suraweera (1993). Oﬀspring individuals are created by randomly copying n − 1 edges from both parents to the oﬀspring. However, the creation of an oﬀspring does not ensure that the oﬀspring represents a fully connected tree. Therefore, a penalty for invalid solutions is necessary. Li and Bouchebaba (1999) overcame the problem of invalid solutions and designed more advanced operators such as path crossover and mutation. These operators always generate feasible new solutions. Although Li and Bouchebaba did not compare their new representation with other representations, the results presented were promising. Raidl (2000) introduced edge crossover and edge insertion mutation for a degree constrained tree problem. New oﬀspring are created by edge crossover in three steps. Firstly, a child inherits all edges which exist in both parents.
7.1 NetDir: A Direct Representation for Trees
219
Then, the oﬀspring gets the edges which exist only in one parent. Finally, the tree is completed with randomly chosen edges concerning the degree constraints. A direct comparison of this approach to other existing approaches for solving the degreeconstrained MST problem is diﬃcult because an additional heuristic for generating good initial solutions was used. Li (2001) presented an implementation of a direct encoding. The implementation is based on predecessor vectors and the eﬀort for crossover and mutation goes with O(d), where d is the length of a path in a tree. The work illustrates that also a direct representation of trees needs to be represented on a computer system. Based on Raidl (2000), Raidl and Julstrom (2003) presented the edgeset encoding which is analyzed in Sect. 7.2. 7.1.2 Properties of Direct Representations We have noticed in Sects. 2.1.2 and 2.1.3 that GEAs using direct representations do not use an additional genotypephenotype mapping fg : Φg → Φp . In contrast to the so called indirect representations, where the genotypic space is diﬀerent from the phenotypic space, Φg = Φp , the operators are directly applied to the phenotypes xp ∈ Φp . This situation is illustrated in Fig. 7.1. Therefore, diﬀerences between diﬀerent implementations of direct encodings are not the used representation (there is no genotypephenotype mapping and all genotypes are trees) but how the genetic operators crossover and mutation are applied to the phenotypes. indirect encodings
direct encodings phenotypes
genetic operators are applied to the genotypes
specific genetic operators are applied to the phenotypes. genotypes 011010101
Figure 7.1. Direct versus indirect representations
When using direct representations, it is neither necessary to deﬁne a representation (genotypephenotype) nor genotypes. At a ﬁrst glance, it seems that the use of direct representations makes life of GEA designer easier as direct representations release us from the pain of designing eﬃcient representations. However, when using direct representations, we are confronted with two other, serious problems: • Often no standard mutation and recombination operators can be used. • It is diﬃcult to design proper problemspeciﬁc search operators.
220
7 Analysis and Design of Search Operators for Trees
We brieﬂy discuss these drawbacks of direct representations. For traditional, indirect representations with standard genotypes, a large variety of diﬀerent genetic search operators with known properties are available. These standard operators are well examined and well understood. However, when using direct representations, standard operators like npoint or uniform crossover can no longer be used. For each direct representation, problemspeciﬁc operators must be developed. Therefore, most of the theory that predicts behavior and performance of GEAs using standard genotypes and standard operators is useless. Furthermore, the development of proper problemspeciﬁc mutation and crossover operators is a diﬃcult task. High quality operators must be able to detect the BBs and propagate them properly. When using direct representations, the design of proper crossover and mutation operators is often demanding as, in general, the phenotypes are not only strings but more complicated structures like for example trees. Furthermore, to use more advanced GEA methods like estimation of distribution algorithms (EDA) or probabilistic model building GAs (PMBGA) become almost impossible. These types of GEAs no longer use standard genetic search operators but build new generations according to a probabilistic model of the parent generations (M¨ uhlenbein and Paaß 1996; M¨ uhlenbein and Mahnig 1999; Harik 1999; Pelikan et al. 1999; Pelikan et al. 1999; Larranaga et al. 1999; Bosman 2003). These search methods are developed for a few standard genotypes (binary or continuous) and result in better performance than traditional simple GAs for decomposable problems. However, because direct representations with nonstandard phenotypes and problemspeciﬁc genetic operators can hardly be implemented in EDAs or PMBGAs, direct representations can not beneﬁt from these new GEA types. It is diﬃcult to design highquality representations when using an indirect representation and standard search operators. However, the task of creating eﬃcient GEAs does not become easier when using direct representations because standard GEA operators can not be used any more and the design of problemspeciﬁc operators that can be directly applied to the phenotypes is diﬃcult. 7.1.3 Operators for NetDir When using direct representations for trees, problemspeciﬁc operators must be developed. In the following paragraphs, we present mutation and crossover operators for the NetDir encoding. Mutation Section 3.3.2 illustrated that, in general, mutation operators should create oﬀspring which are similar to the parent. Therefore, most mutation operators create oﬀspring with a minimal distance to the parent.
7.1 NetDir: A Direct Representation for Trees
221
Applying the mutation operator of the NetDir representation to an individual results in a neighboring phenotype. The mutation operator is applied directly to a phenotype xp ∈ Φp and results in an oﬀspring y p ∈ Φp with phenotypic distance dpxp ,yp = 1. Therefore, mutation randomly changes one link in the tree. We illustrate the mutation operator in Fig. 7.2. The link e1,5 is randomly chosen for deletion. After deleting this link we have two unconnected subtrees. Finally, a node is randomly chosen from each of the two unconnected subtrees and the link connecting the two nodes is inserted (e2,4 ). 1
1
1
5 0
5 0
3 2
4 choose random edge e 1,5
5 0
3 2
4 remove link edge e 1,5
3 2
4 insert randomly choosen edge e2,4
Figure 7.2. The mutation operator for the NetDir representation. The phenotypic distance between parent (left) and oﬀspring (right) is dp = 1.
When using linear genotypes and applying mutation to genotypes, an allele of the string is mutated with mutation probability pm . Therefore, the probl ability for an individual to remain unchanged by mutation is P = (1 − pm ) , where l denotes the length of the string. The situation is diﬀerent for the NetDir encoding because no linear genotype exists. Mutation for the NetDir encoding is deﬁned as randomly mutating an individual n times with probability pm , where n is the number of nodes in the graph. Therefore, the probability n that the individual remains unchanged is P = (1 − pm ) . Crossover The situation becomes slightly more complicated for the crossover operator. In Sect. 3.3.5, we wanted crossover operators to create oﬀspring that are similar to the parents. The oﬀspring should inherit the highquality substructures of their parents. In terms of metric, crossover operators should ensure that the distances between an oﬀspring and its parents are smaller than the distance between both parents. In terms of schemata, highquality crossover operators should be able to detect the linkage between the alleles in the string (Harik and Goldberg 1996) and oﬀspring should inherit the highquality schemata from their parents. Consequently, the crossover operator of the NetDir representation only uses links that exist in the parents for the creation of the oﬀspring. Therefore, the oﬀspring have similar properties than the parents and the schemata are propagated properly. We denote a complete undirected graph as G = (V, E), where v ∈ V denotes the n diﬀerent nodes and ei,j ∈ E denotes the edge between node i ∈ V and j ∈ V . Two parents are denoted as G1 = (V, E1 ) and G2 = (V, E2 ).
222
7 Analysis and Design of Search Operators for Trees
The two oﬀspring are denoted as Go1 = (V, Eo1 ) and Go2 = (V, Eo2 ). The crossover goes with the following scheme: 1. The set of all nodes V is randomly separated into two subsets V1 and V2 , where V1 ∩ V2 = {} and V1 ∪ V2 = V . 2. All edges ei,j ∈ E1 , where i, j ∈ V1 are added to Go1 . All edges ei,j ∈ E1 , where i, j ∈ V2 are added to the second oﬀspring Go2 . 3. All edges ei,j ∈ E2 , where i, j ∈ V2 are added to Go1 . All edges ei,j ∈ E2 , where i, j ∈ V1 are added to the second oﬀspring Go2 . 4. Do the following steps for each oﬀspring individual separately. 5. There are at least two unconnected subtrees Gs1 = (Vs1 , Es1 ) and Gs2 = (Vs2 , Es2 ), where Gs1 = Gs2 . Add randomly an edge ei,j ∈ (E1 ∪ E2 ) to the oﬀspring, where either i ∈ Vs1 ∧ j ∈ Vs2 or i ∈ Vs2 ∧ j ∈ Vs1 . 6. If the oﬀspring is not fully connected, go to 5. The crossover operator consists of two parts. At ﬁrst, complete substructures are passed from the parent to the oﬀspring (item 13). Then, the yet unconnected subtrees are connected by adding links that exist in one of the two parents (item 46). There are several choices for dividing the set of all nodes V into two subsets V1 and V2 (item 1). If we assume that the n nodes are numbered, we can use uniform, onepoint, or npoint crossover. For uniform crossover the probability that each node belongs to either V1 or V2 is 0.5. For onepoint crossover we have to choose a crossing point c ∈ {1, 2, . . . , n − 1}. The nodes with numbers smaller than c belong to V1 ; the nodes with numbers equal or larger than c belong to V2 . Figure 7.3 illustrates the crossover operator with a small 6node example. In a ﬁrst step, the 6 nodes are separated according to uniform crossover into two subsets V1 = {0, 2, 3} and V2 = {1, 4, 5}. Then, the edges e0,3 and e2,3 from parent 1 and the link e1,5 from parent 2 are added to oﬀspring 1. Analogously, e1,5 and e4,5 from parent 1 and e2,3 from parent 2 are used for the construction of oﬀspring 2. After copying the subtrees from the parents to the oﬀspring, the remaining separated subtrees must be connected. We do this by randomly copying edges which are able to connect the separated subtrees from the parents to the oﬀspring until the oﬀspring are completely connected. Oﬀspring 1 has three unconnected subtrees (G1 = ({0, 2, 3}, {(0, 3), (2, 3)}), G2 = ({1, 5}, {(1, 5)}), and G3 = ({4}, {})). Therefore, the edges e0,1 , e1,3 , e3,5 , e3,4 , and e4,5 can be used for completion of oﬀspring 1. After randomly choosing e3,5 and e3,4 , oﬀspring 1 is fully connected and we can stop. Oﬀspring 2 also has three unconnected subtrees (G1 = ({1, 4, 5}, {(1, 5), (4, 5)}), G2 = ({2, 3}, {(2, 3)}), and G3 = ({0}, {})). For oﬀspring 2, the edges e0,1 , e0,3 , e1,3 , e3,4 , and e3,5 can be used for completion. With choosing e0,1 and e3,4 , oﬀspring 2 is fully connected and we can terminate the algorithm. When measuring the distances between the individuals, the distance between the parents is dpG1 ,G2 = 3. The distance of oﬀspring 1 to parent 1 is
7.1 NetDir: A Direct Representation for Trees parent 1 1
parent 2
steps
1 5
0
5 0
3
parents 2
3 2
4
1. 4
1
1
5
5 0
0
2
2. and 3.
3
3 2
4
4
1
1
5
5 offspring
223
0
0
3
3 2
2
4
4. 4
1
1
5
5 0
0
3
3 2
4
offspring 1
2
5. 4
offspring 2
Figure 7.3. The crossover operator of the NetDir encoding. The oﬀspring only inherit information from their parents. No randomly created links are used for the construction of the oﬀspring.
dpG1 ,Go1 = 2, and to parent 2 is dpG2 ,Go1 = 1. The distance between oﬀspring 2 and parent 1 is dpG1 ,Go2 = 2, and to parent 2 is dpG2 ,Go2 = 2. We see that the distances between the oﬀspring and their parents are smaller or equal to the distances between the parents. The oﬀspring exist mostly of substructures of their parents. 7.1.4 Summary This section presented the direct NetDir representation. After a short historical review of direct tree representations, in Sect. 7.1.2 we discussed the properties of direct representations. Because direct representations directly encode the structure of the problem, standard mutation and crossover operators can often not be used any more. Therefore, we presented in Sect. 7.1.3 treespeciﬁc mutation and crossover operators for the NetDir representation.
224
7 Analysis and Design of Search Operators for Trees
The purpose of this section was not just to present another, new representation but to illustrate that the design of eﬃcient GEAs does not become easier when using direct representations. When using direct representations, we need an engineers’ intuition and knowledge not for the design of a genotypic search space – it is the same as the phenotypic search space and it is determined a priori by the structure of the problem – but for the design of proper problemspeciﬁc genetic operators. Therefore, direct representations do not provide eﬃcient GEAs for free, but in comparison to indirect representations, the overall diﬃculty of designing eﬃcient GEAs remains the same or even increases. In this section, we presented the NetDir representation as an example of a direct representation for trees. The NetDir representation directly encodes trees and standard crossover and mutation operators can not be used any more. As most of the existing knowledge about GEA behavior is based on standard operators and standard genotypes, the existing knowledge does not hold any more for direct representations and it is diﬃcult to design highquality search operators.
7.2 The EdgeSet Encoding Raidl and Julstrom (2003) proposed another direct representation for trees denoted as the edgeset encoding. There are two diﬀerent variants of the edgeset encoding: heuristic variants where the encodingspeciﬁc search operators consider the distance weights of the edges, and nonheuristic variants. Results from applying the edgeset encoding to two sets of degreeconstrained MST problem instances indicated that the heuristic variants of the encoding show a higher performance in comparison to other tree encodings such as the Blob code, NetKeys, and weighted encodings (Raidl and Julstrom 2003, p. 238). This section analyses the bias of the edgeset encoding. A bias of a direct encoding means that the encodingspeciﬁc initialization, crossover, and mutation operators prefer a speciﬁc type of solution and push a population in this direction. As the heuristic variants of the edgeset encoding prefer edges with low cost, these variants are expected to show a bias towards the MST. In the second part of the section, the performance of edgesets is investigated for random instances of the optimal communication spanning tree (OCST) problem from Sect. 8.2.1. In contrast to the degreeconstraint MST problem used in Raidl and Julstrom (2003), there are no additional constraints regarding the structure of solutions and all possible trees are feasible. As optimal solutions of the OCST problem are biased towards the MST (Rothlauf et al. 2003), heuristic versions of the edgeset encoding are expected to show good performance. The following subsection summarizes the functionality of the edgeset encoding with and without heuristics. Section 7.2.2 investigates the bias of the
7.2 The EdgeSet Encoding
225
encoding, and Sect. 7.2.3 examines its inﬂuence on the performance of evolutionary search for the OCST problem. The section ends with concluding remarks. 7.2.1 Functionality The edgeset encoding directly represents trees as sets of edges. Therefore, encodingspeciﬁc initialization, crossover, and mutation operators are necessary. The following sections summarize the functionality of the diﬀerent variants with and without heuristics (Raidl and Julstrom 2003). The EdgeSet Encoding without Heuristics Initialization The purpose of the initialization algorithms is to create an unbiased initial solution. Raidl and Julstrom (2003) proposed and investigated three different initialization strategies: PrimRST, RandWalkRST, and KruskalRST. PrimRST overrepresents starlike trees and underrepresents trees similar to lists. RandWalkRST has an average running time of O(n log n), however, the worstcase running time is unbounded. Therefore, Raidl and Julstrom (2003) recommended the use of the KruskalRST which is based on the algorithm from Kruskal (Kruskal 1956). In contrast to Kruskals’ algorithm, KruskalRST chooses edges ei,j not according to their corresponding distance weights dij but randomly. KruskalRST has a small bias towards starlike trees (which is lower than the bias of PrimRST). procedure KruskalRST(V, E): T ← ∅, A ← E; //E is the set of available edges ei,j while T  < V  − 1 do choose an edge {(uv)} ∈ A at random; A ← A − {(eu,v )}; if u and v are not yet connected in T then T ← T ∪ {(eu,v )}; return T . A spanning tree T of an undirected graph G(V, E) with the set E of edges and the set V of nodes is a subgraph that connects all vertices of G and contains no cycles. Recombination To obtain an oﬀspring Tof f from two parental trees T1 and T2 with the edge sets E1 and E2 , KruskalRST is applied to the graph Gcr = (V, E1 ∪ E2 ). Instead of KruskalRST, in principle PrimRST and RandWalkRST can be also used. The crossover operator has high heritability as in the absence of
226
7 Analysis and Design of Search Operators for Trees
constraints, only parental edges are used to create the oﬀspring. Crossover becomes more complicated for constraint MST problems as it is possible that the RST algorithm can create no feasible tree from Gcr = (V, E1 ∪ E2 ). Then, additional edges have to be chosen randomly to complete an oﬀspring. Raidl and Julstrom (2003) distinguished two diﬀerent recombination operators: the variant previously described is denoted KruskalRST crossover. The second variant is denoted KruskalRST* crossover. When using this variant, in a ﬁrst step all edges (E1 ∩ E2 ) are included in the oﬀspring Tof f . Then Tof f is completed by applying KruskalRST to the remaining edges (E1 ∪E2 )\(E1 ∩E2 ). Results from Raidl and Julstrom (2003) indicate a better performance of KruskalRST* for the degreeconstraint MST problem. Mutation The mutation operator randomly replaces one edge in the spanning tree. This replacement can be realized in two diﬀerent ways. The ﬁrst variant of the mutation operator randomly chooses one edge that is not present in T and includes it in T . Then, one edge from the cycle is randomly chosen and removed (“insertion before deletion”). The second variant ﬁrst randomly deletes one edge from T and then connects the two disjoint connected components using a random edge not present in T (“deletion before insertion”). The running time is O(n) if there are no additional constraints. The EdgeSet Encoding with Heuristics The following paragraphs describe how heuristics that rely on the distance weights dij can be included in the edgeset encoding. Raidl and Julstrom (2003) introduced these variants of the edgeset encoding due to the assumption that in weighted tree optimization problems optimal solutions often prefer edges with low distance weights dij . Heuristic Initialization To favor lowweighted edges when generating the initial population, the algorithm KruskalRST starts by sorting all edges in the underlying graph according to their distance weights dij in ascending order. The ﬁrst spanning tree is created by choosing the ﬁrst edges in the ordered list. As these are the edges with lowest distance weights, the ﬁrst generated spanning tree is a MST. Then, the kd edges with lowest distance weights are permuted randomly and another spanning tree is created using the ﬁrst edges in the list. The heuristic initialization results in a strong bias towards the MST. With increasing kd , the bias of randomly created trees towards the MST is reduced. The number of edges which are permuted increases according to kd = α(i − 1)n/N,
7.2 The EdgeSet Encoding
227
where N denotes the population size, i is the number of the tree that is actually generated (i = 1, . . . , N ) and α, with 0 ≤ α ≤ (n − 1)/2, is a parameter that controls the strength of the heuristic bias. Heuristic Recombination The heuristic recombination operator is a modiﬁed version of KruskalRST* crossover. Firstly, the operator transfers all edges E1 ∩ E2 that exist in both parents T1 and T2 to the oﬀspring. Then, the remaining edges are chosen randomly from E = (E1 ∪E2 )\(E1 ∩E2 ) using a tournament with replacement of size two. This means, the distance weights dij of two randomly chosen edges are compared and the edge with the lower distance weight is inserted into the oﬀspring (if no cycle is created). If the underlying optimization problem is constrained, it is possible that the oﬀspring has to be completed using edges not in E . Heuristic Mutation The heuristic mutation operator is based on mutation by “insertion before deletion”. In a preprocessing step, all edges in the underlying graph are sorted according to their weights in ascending order. Doing this, a rank is assigned to every edge. The rank one is assigned to the edge with the lowest weight. To favor lowweighted edges, the edge that is inserted by the heuristic mutation operator is not chosen randomly but according to its rank R = N (0, βn) mod m + 1, where N (0, βn) is the normal distribution with mean zero and standard deviation βn and m = n(n − 1)/2. β is a parameter that controls the bias towards lowweighted edges. If a chosen edge already exists in T , the edge is discarded and the selection is repeated. 7.2.2 Bias As we have seen in Sect. 3.1 a redundant representation is unbiased if all possible phenotypes are represented by, on average, the same number of genotypes (compare also Sect. 6.3.3). Consequently, a search operator is unbiased if it does not overrepresent speciﬁc solutions, and the application of the search operator alone does not modify the statistical properties of a population. An unbiased search operator allows a uniform, nondirected search through the search space. A biased representation or operator should only be used if it is known a priori that the optimal solution of the underlying optimization problem is similar to the overrepresented solutions (compare Sect. 3.1.4). In contrast, unbiased representations or operators should be used if no a priori problemspeciﬁc knowledge is available. Then, the probability of ﬁnding the optimal solution is independent of the structure of the optimal solution.
228
7 Analysis and Design of Search Operators for Trees
The following paragraphs investigate the bias of the edgeset encoding for randomly created trees with n = 10 and n = 16 nodes. To every edge ei,j a nonnegative distance weight dij is associated. Two possibilities for choosing the distance weights dij are considered: • Random weights: The realvalued weights dij are generated randomly and are uniformly distributed in ]0, 100]. • Euclidean weights: The nodes are randomly placed on a 1000 × 1000 grid. The distance weights dij between the nodes i and j are the Euclidean distances between the two nodes. As the distance weights dij are randomly created and dij = dkl , ∀i = l, j = l, we can assume that there is an unique minimum spanning tree (MST) for every problem instance. T is the MST if c(T ) ≤ c(T ) for all other spanning trees T , where c(T ) = ei,j ∈T dij . The similarity between two spanning trees Ti and Tj can be measured using the distance dpTi ,Tj ∈ {0, 1, . . . , n − 1} (compare i j i − luv , where luv is 1 if eu,v exists Sect. 6.1.2) as dpTi ,Tj = 12 u,v∈V, u 0.5, this means Tibest is found in more than 50% of the runs in round i. n(Tibest ) denotes the number of runs that ﬁnd the best solution Tibest in round i.
Euclidean weights random weights
number of problem instances
number of problem instances
500 400 300 200 100 0 0
2
4 dopt,MST
(a) 10 node
6
8
Euclidean weights random weights
300 250 200 150 100 50 0 0
2
4
6
8
10 12 14
dopt,MST
(b) 16 nodes
Figure 7.6. We randomly generated 1000 OCST problems and show the distribution of the problem instances over the distance dopt,M ST between the optimal solution Topt and the MST. Results are presented for 10 and 16 node problems using either random or Euclidean weights. The plots show that the optimal solutions for OCST problems are biased towards the MST.
For ﬁnding the optimal solutions we use a standard GA with traditional parameter settings. The problem was encoded using the NetKey representation (compare Sect. 6.5). The GA uses uniform crossover and tournament selection without replacement. The size of the tournament is three. The crossover probability is set to pcross = 0.8 and the mutation probability (assigning a
7.2 The EdgeSet Encoding
233
random value [0, 1] to one allele) is set to pmut = 0.02. For the GA we started with N0 = 100 and set niter = 20. Each GA run is stopped after a maximum of 200 generations. The computational eﬀort for the experiments is high. Figure 7.6 presents the results of our experiments. We show the number of problem instances over the distance dopt,M ST between the optimal solution Topt and the MST for 1000 randomly created OCST problems with 10 (Fig. 7.6(a)) and 16 (Fig. 7.6(b)) nodes. The OCST problems are created randomly using either random weights in ]0,100] or placing the nodes randomly on a 1000 × 1000 twodimensional grid and calculating the weights as the Euclidean distances between the nodes (details are described in Sect. 7.2.2). The demands rij between the nodes are random and uniformly distributed in ]0,100]. Comparing the results to the average distance dprnd,M ST between a randomly created tree and the MST (Table 7.1) reveals that the optimal solutions for OCST problems are biased towards the MST. Furthermore, OCST problems with random weights show a stronger bias than OCST problems with Euclidean weights. Due to the bias of the optimal solutions towards the MST, the problem should be easy to solve for GEAs using the edgeset encoding. The Performance of the EdgeSet Encoding for Randomly Generated OCST Problems After determining optimal solutions as described in the previous paragraphs, we examine the performance of GEAs using the edgeset encoding. We use the same randomly generated problem instances as before and investigate how the GEA performance depends on the distance dopt,M ST between the optimal solution and the MST. We use a generational GA with tournament selection without replacement of size two and no mutation. Each run is stopped after the population is fully converged or the number of generations exceeds 200. We perform 50 GA runs for each of the 1000 problem instances. In our experiments we compare the performance of GAs using • nonheuristic KruskalRST crossover (p. 225) with nonheuristic KruskalRST initialization (p. 225) (indicated as “KruskalRST”), • nonheuristic KruskalRST* crossover (p. 225) combined with nonheuristic KruskalRST initialization (indicated as “KruskalRST*”), • nonheuristic KruskalRST* crossover combined with heuristic initialization (p. 226) with α = 1.5 (indicated as “heur ini”), • heuristic crossover (p. 227) combined with nonheuristic KruskalRST initialization (indicated as “heur xover”), • heuristic crossover combined with heuristic initialization with α = 1.5 (indicated as “heur ini & xover”), and • as benchmark the unbiased network random key encoding with uniform crossover (indicated as “NetKey”).
234
7 Analysis and Design of Search Operators for Trees
The population size N which is constant in all experiments, is chosen with respect to the performance of the nonheuristic KruskalRST* crossover operator. The aim is to ﬁnd the optimal solution with a probability of about 50 %. Therefore, we choose for the 10 node problems a population size of N = 60 (random weights) and N = 100 (Euclidean weights) and for the 16 node problems a population size of N = 200 (random weights) and N = 450 (Euclidean weights).
1
NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
0.8 0.6
percentage of optimal solutions found
percentage of optimal solutions found
1
0.4 0.2 0 0
1
2
3
NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
0.8 0.6 0.4 0.2 0
4
0
1
2
dopt,MST
3
4
5
6
dopt,MST
(a) 10 node / random weights (b) 10 node / Euclidean weights 1
NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
0.8 0.6
percentage of optimal solutions found
percentage of optimal solutions found
1
0.4 0.2 0 0
1
2
3 dopt,MST
4
5
6
NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
0.8 0.6 0.4 0.2 0 3
4
5
6
7
8
9
10
dopt,MST
(c) 16 node / random weights (d) 16 node / Euclidean weights Figure 7.7. The ﬁgures compare the performance of a GA using diﬀerent combinations of crossover and initialization operators for randomly generated 10 (left) and 16 (right) node OCST problems. The plots show the average percentage of optimal solutions that can be found over dopt,M ST . The heuristic crossover operator outperforms the nonheuristic version only if the optimal solution is very similar to the MST (dopt,M ST ≈ 0). If dopt,M ST > 1 the heuristic crossover results in low GA performance. In contrast, when using the nonheuristic KruskalRST* crossover, GA performance remains about constant.
The results of the experiments are presented in Figs. 7.7 and 7.8. Figure 7.7 shows the percentage of GA runs that ﬁnd the correct optimal solutions
7.2 The EdgeSet Encoding 20
NetKey KruskalRST KruskalRST* 12 heur ini heur xover heur ini& xover 10
NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
18 16 14 gap (in %)
gap (in %)
14
8 6
235
12 10 8 6
4
4 2
2
0
0 0
1
2
3
4
0
1
2
dopt,MST
4
5
6
dopt,MST
(a) 10 node / random weights
(b) 10 node / Euclidean weights
20
25
NetKey KruskalRST KruskalRST* heur ini 15 heur xover heur ini& xover
20
gap (in %)
gap (in %)
3
10
15 NetKey KruskalRST KruskalRST* heur ini heur xover heur ini& xover
10
5
5
0
0 0
1
2
3
4
5
6
3
4
5
dopt,MST
(c) 16 node / random weights
6
7
8
9
10
dopt,MST
(d) 16 node / Euclidean weights
Figure 7.8. We show the mean of the gap between the cost of the best found solution and the cost of the optimal solution over dopt,M ST . The results conﬁrm that the heuristic crossover operator outperforms the nonheuristic variants only if the optimal solutions are very similar to the MST (dopt,M ST ≈ 0). c(T
)−c(T
)
opt f ound over dopt,M ST . Fig. 7.8 shows the gap, (in percent), between c(Topt ) the cost of the best found solution and the cost of the optimal solution over dopt,M ST . We show results for 1000 randomly generated problem instances. Results are plotted only for these dopt,M ST , where there are more than 10 problem instances. For example, we show results for 10 node problems with Euclidean weights only for dopt,M ST ∈ {0, . . . , 6} as there are only 8 (out of 1000) instances with dopt,M ST = 7 (compare Fig. 7.6(b)). The results reveal that the heuristic crossover versions of the edgeset encoding (heur xover and heur ini & crossover) always ﬁnd the optimal solution if the optimal solution is the MST. However for dopt,M ST = 0, the performance of GAs using the heuristic version drops sharply and the optimal solution can not be found if dopt,M ST > 2. In contrast, the performance of the nonheuristic KruskalRST* operator decreases only slightly with larger dopt,M ST and allows
236
7 Analysis and Design of Search Operators for Trees
the GA to correctly identify the optimal solution even for larger dopt,M ST . The performance of the nonheuristic crossover combined with an heuristic initialization (“heur ini”) is similar to the heuristic crossover operator. It always ﬁnds the optimal solution if it is the MST, however with increasing dopt,M ST the decrease of performance is slightly less than for the heuristic crossover. In summary, the heuristic crossover operator performs well only for problems where the optimal solution is slightly diﬀerent from the MST. Otherwise, GAs using the edgeset encoding with heuristic crossover fail. The performance of GAs using the nonheuristic variant is similar to the performance of the NetKey encoding with uniform crossover. These results are conﬁrmed when c(Tf ound )−c(Topt ) (Fig. 7.8). Heuristic variants of the enexamining the gap c(Topt ) coding show high performance if the optimal solution is the MST. However, with increasing dopt,M ST the quality of the solutions strongly decreases and the nonheuristic variants outperform the heuristic variants. In the remaining paragraphs, the performance of the edgesetspeciﬁc mutation operator is examined. As before 1000 random problems of diﬀerent types are generated and the optimal solutions are calculated as described on page 232. For comparing the performance of diﬀerent variants of the mutation operator, a simple simulated annealing (SA) strategy (van Laarhoven and Aarts 1988) is used as a representative example of mutationbased search. SA can be modeled as an GEA with population size one and Boltzmann selection (Mahfoud and Goldberg 1995). In each generation a new solution Tof f is created by applying exactly one mutation to the parent solution Tpar . If c(Tof f ) < c(Tpar ), Tof f replaces Tpar . If c(Tof f ) > c(Tpar ), Tpar is replaced with probability P (T ) = exp (−(c(Tof f ) − c(Tpar ))/T ). With lower T , the probability of accepting worse solutions decreases. In our experiments the start temperature Tstart = 50 is reduced in every step by the factor 0.99. Therefore, Tt+1 = 0.99 × Tt . The number of search steps is set to tmax = 300 for 10 node and tmax = 1000 for 16 node problems. We performed 50 independent runs for each problem instance and investigated the performance of an SA using • nonheuristic mutation (p. 226) and nonheuristic initialization (p. 225) (denoted as “no heur mut&ini”), • nonheuristic mutation and heuristic initialization (p. 226) with α = 1.5 (denoted as “no heur mut, α = 1.5”), • heuristic mutation (p. 227) with β = 5 and nonheuristic initialization (denoted as “β = 5, no heur ini”), • heuristic mutation with β = 5 and heuristic initialization with α = 1.5 (denoted as “β = 5, α = 1.5”), • heuristic mutation with β = 0.5 and nonheuristic initialization (denoted as “β = 0.5, no heur ini”), and • heuristic mutation with β = 0.5 and heuristic initialization with α = 1.5 (denoted as “β = 0.5, α = 1.5”).
7.2 The EdgeSet Encoding
237
We performed no experiments for NetKeys as the corresponding mutation operator can not be directly compared. The mutation operator for NetKeys which changes one allele of the genotype, often does not change the corresponding phenotype, whereas the mutation operator of the edgeset encoding always changes one edge. The results of the experiments are presented in Fig. 7.9. It shows the percentage of SA runs that ﬁnd the correct optimal solutions over dopt,M ST . It can be seen that an SA using heuristic initialization always ﬁnds the optimal solution if dopt,M ST = 0. When using heuristic mutation with a low bias (β = 5), SA performance is always higher than when using nonheuristic mutation (for all considered dopt,M ST ). A small bias of the mutation operator does not push the population towards the MST but allows a diversed population and eﬃcient SA search for solutions somehow similar to the MST. However, when increasing the bias of the heuristic mutation to β = 0.5, SA performance becomes lower than for the nonheuristic case even for small dopt,M ST (especially for the Euclidean problem instances). Then, the heuristic bias of the mutation operator is too strong and pushes the population too strongly towards the MST. The results reveal that by increasing the bias of the mutation operator (lowering β) problems where the optimal solutions are similar to the MST can be solved more eﬃciently; however, problems where the optimal solutions are diﬀerent from the MST can be solved less eﬃciently. To summarize our ﬁndings, the heuristic crossover operator of the edgeset encoding does not allow eﬃcient search due to its strong bias towards the MST. Only problems where the optimal solutions are slightly diﬀerent from the MST can be solved. The heuristic mutation operator results in good performance if β is large as the resulting low bias of the mutation operator prefers solutions similar to the MST and does not push a population too strongly towards the MST. However, if the bias towards the MST induced by β becomes stronger only optimal solutions similar to the MST can be found and mutationbased search fails. The results for the heuristic mutation operator show that the proper adjustment of β is important and crucial for the success of local search. 7.2.4 Summary and Conclusions This section investigated the bias of the edgeset encoding which was proposed by Raidl and Julstrom (2003), and examines its performance for random instances of the optimal communication spanning tree (OCST) problem. The edgeset encoding belongs to the class of direct representations for trees. Instead of deﬁning an additional genotypephenotype mapping, encodingspeciﬁc initialization, crossover and mutation operators are directly applied to the trees. Section 7.2.1 described the functionality of the edgeset encoding and Sect. 7.2.2 performs an exhaustive investigation into the bias of the diﬀerent variants of the edgeset encoding. The work is completed by an inves
238
7 Analysis and Design of Search Operators for Trees 1 percentage of optimal solutions found
percentage of optimal solutions found
1 0.8 0.6 0.4
no heur mut&ini no heur mut, α=1.5 β=5,no heur ini β=5,α=1.5 β=0.5,no heur ini β=0.5,α=1.5
0.2 0 0
1
2
3
0.8 0.6 no heur mut&ini no heur mut, α=1.5 β=5,no heur ini β=5,α=1.5 β=0.5,no heur ini β=0.5,α=1.5
0.4 0.2 0
4
0
1
2
dopt,MST
percentage of optimal solutions found
percentage of optimal solutions found
1
0.8 0.6 0.4 0.2 no heur mut&ini no heur mut, α=1.5 β=5,no heur ini β=5,α=1.5 β=0.5,no heur ini β=0.5,α=1.5 1
2
3
5
6
(b) 10 node / Euclidean weights
1
0
4
dopt,MST
(a) 10 node / random weights
0
3
4
5
dopt,MST
(c) 16 node / random weights
6
no heur mut&ini no heur mut, α=1.5 β=5,no heur ini β=5,α=1.5 β=0.5,no heur ini β=0.5,α=1.5
0.8 0.6 0.4 0.2 0 3
4
5
6
7
8
9
10
dopt,MST
(d) 16 node / Euclidean weights
Figure 7.9. The ﬁgures show the performance of an SA using diﬀerent variants of initialization and mutation operators of the edgeset encoding. The plots show the average percentage of optimal solutions that can be found over dopt,M ST for 1000 randomly created OCST problems. The results show that the heuristic variants of the mutation operator outperform the nonheuristic variants for the OCST problem if β is set properly.
tigation into the performance of crossoverbased search and mutationbased search for randomly generated instances of the OCST problem. The investigation into the bias of the edgeset encoding reveals that the heuristic versions of the initialization, crossover, and mutation operators are biased towards the MST deﬁned on the distance weights. The bias is especially strong for the heuristic crossover operator which results in a quick convergence of a population of trees towards the MST. In contrast, the nonheuristic search operators of the edgesets are unbiased and their application results in an undirected and uniform search through the search space.
7.2 The EdgeSet Encoding
239
Due to the strong bias of the heuristic search operators towards the MST, tree optimization problems can easily be solved if optimal solutions are the MST. However, if optimal solutions are only slightly diﬀerent from the MST, the heuristic crossover operator fails due to its strong bias towards the MST. Therefore, the heuristic crossover operator is not appropriate for solving tree optimization problems. Rothlauf and Tzschoppe (2004) proposed an extension of the heuristic crossover operator which allows us to reduce its strong bias. Using the modiﬁed heuristic crossover operators results in higher GEA performance if the optimal solution is diﬀerent from the MST. In contrast to the heuristic crossover operator, the nonheuristic crossover operator of the edgesets results in good performance for OCST problems even if the optimal solutions are quite diﬀerent from the MST. Its performance is similar to the NetKey encoding from Sect. 6.5. For the mutation operator, the strength of the bias towards the MST can be controlled by an encodingspeciﬁc parameter β. With high β, the bias towards the MST is low, with low β it is strong. Therefore, with low β, tree problems can be solved more eﬃciently if the optimal solutions are similar to the MST, but otherwise less eﬃciently. If β is set appropriately, the heuristic mutation operator is a good choice for OCST problems as optimal solutions of this problem are similar to the MST. The problems of the heuristic variants of the edgeset encoding emphasize the diﬃculties of a proper design of direct representations. In contrast to indirect encodings, the behavior of new, problemspeciﬁc search operators is often unknown. The analysis of the edgeset encoding has shown that although optimal solutions for the OCST problems are biased towards the MST (Rothlauf et al. 2003), direct representations such as the heuristic edgeset encoding that use this problemspeciﬁc knowledge and are biased towards the MST, can fail if the bias is too strong. Therefore, unbiased representations should be used if no problemspeciﬁc knowledge is known a priori.
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
In the previous chapters, the presented theory about representations was used for the analysis and design of representations as well as search operators. The investigations into the properties of representations were based on theory and helped us to understand what happens when GEAs use a speciﬁc representation. However, in practice, GEA users are often less interested in theory about representations but want simple instruments for a quick and rough prediction of the expected performance of a representation. They have several representations at hand and want to know which representation they should choose for their problem. We do not want to leave them alone with their problems, but illustrate how they can advantageously use the proposed theory. This chapter illustrates for scalable test tree problems and realworld tree problems how the performance of GEAs using diﬀerent types of representations can be predicted by using the provided framework about representations. Based on the framework, we give qualitative predictions of solution quality and time to convergence for diﬀerent types of tree representations. Doing this, this chapter also provides an exhaustive comparison of the performance of diﬀerent tree representations. For our comparison we choose the indirect representations Pr¨ ufer numbers (Sect. 6.2), characteristic vectors (Sect. 6.3), the link and node biased encoding (Sect. 6.4), and NetKeys (Sect. 6.5) as well as the direct encodings NetDir (Sect. 7.1) and edgesets (Sect. 7.2). These indirect and direct representations are used on scalable test problems like the onemax tree and deceptive trap problem for trees and on various test instances of the optimal communication spanning tree (OCST) problem from the literature. The results show that using the outlined theory makes it easier to select the proper representation for the problem at hand. The test instances of the OCST problem are chosen because the exact speciﬁcations of the problems are either easily available (Raidl 2001; Rothlauf et al. 2002) or published (Palmer 1994; Berry et al. 1995). For summarizing purposes, the exact data regarding the distance weights and the communication demands for the test instances of the OCST problems are listed in Appendix A.
242
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
The following section provides a comparison of GEA performance for scalable test problems. After a brief analysis of representations in Sect. 8.1.1, we present in Sect. 8.1.2 results for the fully easy onemax tree problem and in Sect. 8.1.3 results for the the fully diﬃcult deceptive tree problem. For each of the two test problems, we provide theoretical predictions and empirical evidence. Then, we focus in Sect. 8.2 on the OCST problem. This problem is deﬁned on trees and researchers have proposed some test instances in the literature (Palmer 1994; Berry et al. 1995; Raidl 2001; Rothlauf et al. 2002). For each problem, we deduce predictions of GEA performance and present empirical results. The chapter ends with a brief summary.
8.1 GEA Performance on Scalable Test Tree Problems This section compares the performance of diﬀerent types of indirect and direct representations for the onemax tree problem and the deceptive trap problem for trees. After a brief analysis of tree representations based on the framework from Chap. 4, we present results for the onemax tree problem (Sect. 8.1.2) and for the deceptive tree problem (Sect. 8.1.3). For both problems, we provide brief descriptions, theoretical predictions about GEA performance, and empirical evidence. 8.1.1 Analysis of Representations We brieﬂy summarize the most important properties of the diﬀerent tree representations and operators from the previous chapters. We focus on the results concerning redundancy, bias, scaling, and locality for Pr¨ ufer numbers, characteristic vectors (CV), the LNB encoding, NetKeys, NetDir, and edgesets. For the use of Pr¨ ufer numbers, NetKeys, CV, and NetDir no additional representationspeciﬁc parameters are necessary. In contrast, for LNBs and edgesets, additional representationspeciﬁc parameters must be set by the user. In the following experiments, we use the same variants of the edgesets as in the investigation presented in Sect. 7.2.3: • KruskalRST: nonheuristic KruskalRST crossover (p. 225) with nonheuristic KruskalRST initialization (p. 225), • KruskalRST*: nonheuristic KruskalRST* crossover (p. 225) combined with nonheuristic KruskalRST initialization, • heur ini: nonheuristic KruskalRST* crossover combined with heuristic initialization (p. 226) with α = 1.5, • heur xover: heuristic crossover (p. 227) combined with nonheuristic KruskalRST initialization, and • h ini & xover: heuristic crossover combined with heuristic initialization with α = 1.5.
8.1 GEA Performance on Scalable Test Tree Problems
243
Further details about the edgeset encoding can be found in Sect. 7.2. For the LNB encoding, we use the following variants: • NB (P2 =1): the nodebiased encoding (p. 180) with the nodespeciﬁc bias P2 = 1, • NB (P2 =20): the nodebiased encoding with the nodespeciﬁc bias P2 = 20, • LB (P1 =1): the linkbiased encoding (p. 181) with the linkspeciﬁc bias P1 = 1, • LB (P1 =20): the linkbiased encoding with the linkspeciﬁc bias P1 = 20, and • LNB (P1 =P2 =1): the linkandnodebiased encoding (p. 182) with the linkspeciﬁc bias P1 = 1 and the nodespeciﬁc bias P2 = 1, The following paragraphs review important properties of the representations presented in Chaps. 6 and 7 which are summarized in Table 8.1. The investigation into the redundancy of indirect representations revealed that Pr¨ ufer numbers are nonredundant and that NetKeys are uniformly and synonymously redundant. Furthermore, the LNB encodings are synonymously redundant; however, the over and underrepresentation of solutions depends on the setting of the nodespeciﬁc and linkspeciﬁc bias. Variants of the LNB encoding that use a large linkspeciﬁc bias P1 are approximately uniformly redundant and show the same behavior as NetKeys. Therefore, GEA performance is independent of the structure of the optimal solution. In contrast, the LNB encoding is nonuniformly redundant if the linkspeciﬁc bias is not large enough. If P1 and P2 are small, LNB encodings are biased towards the minimum spanning tree. For a large nodespeciﬁc bias P2 , the encoding is biased towards stars. Therefore, GEA performance depends on the structure of the optimal solution. In contrast, CVs are uniformly redundant but aﬀected by nonsynonymous redundancy. As not all genotypes that represent the same tree are similar to each other, the recombination of two genotypes that encode similar phenotypes can result in an oﬀspring with diﬀerent properties (compare Sect. 6.3.3). The situation is diﬀerent for indirect encodings such as NetDir or edgesets. For direct encodings, it is not the properties of the genotypephenotype mapping which are relevant but the properties of the search operators. The investigation into the NetDir encoding shows (compare Sect. 7.1) that the search operators are nearly unbiased which results in the same GEA performance as with nonredundant or nonuniformly redundant encodings. In contrast, the edgesets show a bias towards the minimum spanning tree if heuristic variants of the initialization and search operators are used. A biased search operator is equivalent to a nonuniformly redundant encoding. Due to the design of the search operator, both direct representations show high locality as the application of a mutation operator always results in a similar solution. The investigation into the locality of tree representations shows that the Pr¨ ufer number representation has in general low locality. However, the locality
244
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
Table 8.1. Summary of important properties of direct and indirect representations for trees redundancy & bias Pr¨ ufer
nonredundant
NetKey
uniformly and synonymously redundant uniformly and nonsynonymously redundant
CV
NB (P2 =1 nonuniformly (bias towards stars) and P2 =20) and synonymously redundant LB (P1 =1) nonuniformly (bias towards MST) and synonymously redundant LB (P1 =20) nonuniformly and synonymously redundant LNB nonuniformly (bias towards stars and (P1 =P2 =1) MST) and synonymously redundant NetDir nonredundant, search operator unbiased KruskalRST nonredundant, search operator unbiased KruskalRST* nonredundant, search operator unbiased heur. ini nonredundant, initialization biased towards MST (equivalent to nonuniform redundancy) heur. xover nonredundant, search operator biased towards MST (equivalent to nonuniform redundancy) h. ini & xover nonredundant, initialization and search operator biased towards MST (equivalent to nonuniform redundancy)
BBscaling high locality uniformly around stars, low scaled locality elsewhere high locality uniformly scaled low locality due to uniformly nonsynonymous scaled redundancy high locality uniformly scaled high locality uniformly scaled high locality uniformly scaled high locality uniformly scaled high locality locality
high locality

high locality

high locality

high locality

high locality

of Pr¨ ufer numbers is not low everywhere. When encoding stars, the locality of Pr¨ ufer numbers is high; when encoding nonstars, the locality of Pr¨ ufer numbers is low. For redundant representations, low locality is equivalent to nonsynonymity. As for lowlocality representations, similar genotypes do not encode similar phenotypes if nonsynonymously redundant representations are used. Therefore, all encodings that are synonymously redundant can be viewed as highlocality encodings since small changes in the genotype always result in small changes in the phenotype. The nonsynonymously redundant CV encoding has low locality as small changes of the genotype do not necessarily
8.1 GEA Performance on Scalable Test Tree Problems
245
result in a similar phenotype. For direct encodings, the locality is always high as the mutation operators are designed such that trees with similar properties are created. The investigation into the scaling of the BBs has revealed that all examined indirect tree representations have uniformly scaled BBs. There are no BBs that have a higher contribution to the ﬁtness of a tree. Therefore, the dynamics of genetic search are not changed and all alleles are solved implicitly in parallel. Table 8.2. We generated random genotypes xgrnd of diﬀerent sizes n and calculated the minimum phenotypic distance min(dprnd,star ) to stars, and the phenotypic distance dprnd,M ST to the MST. The numbers conﬁrm that CVs, NetKey, NetDir, LB (P1 = 20), and edgesets without initialization heuristics are unbiased (uniformly redundant) as the distances are about the same as for the nonredundant Pr¨ ufer numbers. Edgesets with heuristic initialization are biased towards the MST. The LNB encoding is biased towards stars for large P2 and towards the MST for low P1 and P2 . problem size n distance d min(dprnd,star ) dprnd,M ST min(dprnd,star ) NetKey dprnd,M ST min(dprnd,star ) CV dprnd,M ST min(dprnd,star ) NB (P2 =1) dprnd,M ST min(dprnd,star ) NB (P2 =20) dprnd,M ST min(dprnd,star ) LB (P1 =1) dprnd,M ST min(dprnd,star ) LB (P1 =20) dprnd,M ST min(dprnd,star ) LNB (P1 =P2 =1) dprnd,M ST min(dprnd,star ) NetDir dprnd,M ST KruskalRST, KruskalRST*, min(dprnd,star ) dprnd,M ST heur xover min(dprnd,star ) heur ini, heur ini & xover dprnd,M ST Pr¨ ufer number
8 µ (σ) 3.66 (0.64) 5.17 (1.01) 3.75 (0.62) 5.22 (1.01) 3.66 (0.64) 5.18 (1.01) 2.46 (1.1) 3.31 (1.05) 0.24 (0.68) 5.08 (0.80) 3.74 (0.63) 3.54 (1.15) 3.74 (0.61) 5.15 (1.06) 3.06 (0.89) 3.90 (1.07) 3.66 (0.65) 5.17 (1.02) 3.75 (0.61) 5.26 (1.04) 4.06 (0.52) 0.35 (0.58)
20 µ (σ) 14.67 (0.79) 17.05 (1.26) 14.77 (0.78) 17.10 (1.25) 14.68 (0.80) 17.05 (1.27) 10.72 (2.26) 11.65 (1.77) 1.64 (2.79) 16.55 (1.13) 14.79 (0.76) 12.64 (1.86) 14.77 (0.76) 16.79 (1.34) 12.55 (1.47) 13.88 (1.69) 14.67 (0.79) 17.05 (1.26) 14.78 (0.77) 17.10 (1.27) 15.75 (0.45) 1.13 (1.29)
40 µ (σ) 34.07 (0.84) 37.03 (1.34) 34.15 (0.80) 37.06 (1.32) 34.07 (0.83) 37.02 (1.34) 26.90 (3.14) 26.85 (2.50) 5.98 (7.09) 35.88 (1.53) 34.24 (0.76) 28.93 (2.53) 34.14 (0.80) 36.36 (1.53) 30.18 (1.96) 31.74 (2.21) 34.07 (0.83) 37.02 (1.34) 34.17 (0.81) 37.04 (1.33) 35.53 (0.53) 2.29 (2.30)
To investigate and verify the bias of the diﬀerent representations, we can measure the average distance of randomly generated solutions to a star or to the MST. Consequently, Table 8.2 presents results for randomly created genotypes that encode trees of diﬀerent size n. For each representation, we
246
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
randomly generate 10,000 genotypes and measure the mean µ and the standard deviation σ of the minimum phenotypic distance min(dprnd,star ) towards a star, and the phenotypic distance dprnd,M ST towards the MST. For the different variants of the LNB encoding, we randomly placed the n nodes on a twodimensional grid of size 1,000×1,000. As the distance weights dij are relevant for the construction of the phenotype (compare (6.5)), we used Euclidean distances between the nodes i and j as distance weights dij . For the edgesets, we created 100 random populations of size N = 100, and placed for each population the nodes randomly on the 1,000×1,000 square. The results conﬁrm the ﬁndings from above. The direct encodings NetDir as well as variants of the edgesets that use nonheuristic initialization (KruskalRST, KruskalRST*, and heur. xover) are about unbiased as the distances are similar to Pr¨ ufer numbers. Also, the uniformly redundant encodings NetKeys, CVs, and the LB encoding with large linkspeciﬁc bias are unbiased because they have about the same distances as nonredundant representations (Pr¨ ufer numbers). In contrast, the NB and LNB encoding show a bias ufer towards stars as min(dprnd,star ) is lower in comparison to the unbiased Pr¨ numbers. Variants of the edgesets that use heuristic initialization as well as the LB (P1 = 1) and LNB encoding show a bias towards the MST as dprnd,M ST is lower in comparison to Pr¨ ufer numbers. 8.1.2 OneMax Tree Problem We examine the performance of GAs using diﬀerent representations for the onemax tree problem introduced in Sect. 6.1.5. Problem Description The onemax tree problem was deﬁned in Sect. 6.1.5 as a fully easy tree problem. The problem is fully easy for GEAs as the phenotypic size of the BBs is kp = 1. All tree schemata that contain the optimal solution have higher ﬁtness than their competitors. For the onemax tree problem, an optimal solution Topt is chosen a priori either randomly or by hand. The structure of Topt can either be a random tree, a star, a list, the MST, or any other predeﬁned tree structure. In the following section, we assume a minimization problem and deﬁne the ﬁtness of a tree Ti as the distance di,opt between the optimal solution and the tree. For di,opt = 0, Ti = Topt . Further details regarding the onemax tree problem are provided in Sects. 6.1.5 and 6.1.4. Theoretical Predictions We give predictions on GEA performance for the onemax tree problem. The predictions are based on the results from Sect. 8.1.1.
8.1 GEA Performance on Scalable Test Tree Problems
247
We know that some representations are nonuniformly redundant and are biased towards speciﬁc tree structures. Therefore, GEA performance depends on the structure of the optimal solution. Using the results learnt about redundancy from Sect. 8.1.1, we expect the CV, NetKey, NetDir, edgeset without heuristics, and LB (P1 = 20) encoding to perform independently of the structure of the optimal tree Topt . GEAs using LNB encodings with small biases P1 and P2 will perform better if the optimal solutions are similar to the MST. However, GEAs using these encodings need more generations and ﬁnd less BBs if the optimal solutions are not similar to the MST. GEAs using the NB encoding are biased towards stars and show high performance if the best solutions are stars. If the best solutions are arbitrary trees or lists, GEAs will show lower performance when using the LNB encoding. The edgeset encoding with heuristics introduce a bias towards MSTlike solutions. Therefore, GEAs using edgesets with heuristics are expected to perform well for problems where the optimal solutions are the MST, and show low performance elsewhere. After we have examined the eﬀects of redundancy, we focus on locality. We have seen that Pr¨ ufer numbers have high locality around stars, but low locality elsewhere. This means that the genotypic size of the BBs kg is larger than the phenotypic size of the BBs kp = 1 if the optimal solution is not a star. Therefore, the performance of GEAs using Pr¨ ufer numbers is low if the optimal solution is not a star. CVs are also aﬀected by low locality as the encoding is nonsynonymously redundant. Due to stealth mutation, oﬀspring can be created that are not similar to their parents. In general, we expect that GEAs using CVs show lower performance for the easy onemax problem, as the search is randomized and guided search is no longer possible. However, as we use only crossover and no mutation in our experiments, BBs can not come back into the population once they are extinct. The nonsynonymous redundancy of the CV encodings can partially solve this problem as it works like stealth mutation (compare Sect. 6.3.3) and brings back lost BBs. Therefore, for small problem instances, we expect GEAs using the CV encoding to be able to ﬁnd more BBs than other encodings, but to need much more time to do this. The run duration and the number of solved problems at the end of the run increases when using CVs. This eﬀect of stealth mutation is ampliﬁed by the low diﬃculty of the onemax problem. This problem is especially easy for mutationbased GEAs because the landscape leads GEAs to the correct solution. Therefore, mutation, as well as stealth mutation, increases GEA performance for small problem instances. In contrast, the other representations (NetKey, NetDir, edgesets and LNB) have high locality and the problem diﬃculty is not changed. Empirical Results For our experiments, we use a simple, generational genetic algorithm without mutation, tournament selection without replacement of size 2, and uniform crossover. We randomly generate 200 problems for each problem instance and
248
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
perform for each problem instance and each representation 25 GA runs. We present results for the onemax tree problem with 8, 20, and 40 nodes. The optimal solution is either a randomly chosen star or list, an arbitrary tree, or the MST based on the the distance weights dij . The performance of GEAs is determined by the percentage Psucc of correctly solved problems at the end of the run and the number of generations tconv until the population is fully converged. We use a population size of N = 16 for all n = 8 node problems, N = 100 for all n = 20 node problems, and N = 600 for all n = 40 node problems. The population sizes N are chosen such that some of the representations allow the GA to always ﬁnd the optimal solutions. In general, a higher population size N increases the performance (higher Psucc ), whereas lower N reduces GA performance. The performance diﬀerences between the diﬀerent representations remain about the same when using diﬀerent population sizes. All runs are stopped after the population is fully converged or a maximum of 200 generations is reached. The performance of GAs using diﬀerent representations (compare Sect. 8.1.1) for the onemax tree problem is shown in Tables 8.3 and 8.4. We present results for diﬀerent optimal solutions (Topt is either an arbitrary tree, the MST, a random star, or a random list) and show the percentage Psucc of runs that ﬁnd the optimal solution Topt , the mean and standard deviation of the ﬁtness of the best found solution, and the mean and standard deviation of the number of generations tconv until the population is fully converged or the GA run is stopped after 200 generations. The ﬁtness of the best found solution directly measures the distance between the optimal solution Topt and the best found solution. All results are averaged over 200 randomly created problems and 25 runs for each problem. The numbers show that the performance of GEAs searching for optimal lists is about the same as when searching for optimal arbitrary trees. Furthermore, the performance of GAs using uniformly redundant encodings like CVs, NetKeys, NetDir, edgesets without heuristics, or the LB (P1 = 20) encoding is about independent of the structure of the optimal solution Topt . There are some exceptions for Topt is a star, but, in general, Psucc , the ﬁtness, and Tconv are not aﬀected by the optimal solution being an arbitrary tree, the MST, a star, or a list. The performance of GAs using Pr¨ ufer number is low due to the low locality of the encoding. However, we have seen in Sect. 6.2.4 that the locality of Pr¨ ufer number is high around stars. Therefore, GAs using Pr¨ ufer numbers perform better when searching for optimal stars than when searching for optimal trees, lists, or MSTs. If the problem instances are small (n = 8 and n = 20), GAs using the CV encoding are able to ﬁnd the optimal solution when searching for an arbitrary tree, a list, or the MST, but need a high number of generations tconv . The nonsynonymous redundancy of the encoding continuously introduces new genetic material into the search and prolongs search time allowing the GA to ﬁnd its way to the optimal solutions. However, for larger problem instances, the search
8.1 GEA Performance on Scalable Test Tree Problems
249
MST
arbitrary tree
Topt
Table 8.3. Performance of GAs using diﬀerent types of representations for onemax tree problems of diﬀerent sizes and with diﬀerent Topt (arbitrary tree and MST) 8 nodes 20 nodes 40 nodes ﬁtness tconv ﬁtness tconv ﬁtness tconv Psucc P P µ (σ) µ (σ) succ µ (σ) µ (σ) succ µ (σ) µ (σ) Pr¨ ufer 0.05 1.65(0.8) 18 (4) 0 5.91 (1.2) 77 (9) 0 14.28(1.7) 198 (5) NetKey 0.59 0.47(0.6) 16 (3) 0.91 0.09 (0.3) 40 (4) 1 0 (0) 85 (9) CV 0.95 0.05(0.2) 20 (4) 1 0.01 (0.1) 137(22) 0 17.34(1.0) 200 (0) NB (P2 =1) 0 3.20(0.9) 10 (3) 0 12.44(1.4) 28 (6) 0 29.58(1.8) 62 (12) NB (P2 =20) 0 3.52(0.7) 7 (3) 0 13.18(1.1) 20 (5) 0 30.06(1.5) 38 (9) LB (P1 =1) 0.11 1.63(0.9) 14 (4) 0 5.72 (1.8) 51 (12) 0 14.73(2.7) 150(32) LB (P1 =20) 0.58 0.47(0.6) 16 (3) 0.90 0.11 (0.3) 40 (4) 1 0 (0) 85 (9) P1 =P2 =1 0.13 1.47(0.9) 18 (4) 0 4.51 (1.5) 69 (11) 0 12.22(2.4) 190(15) NetDir 0.39 0.79(0.7) 20 (6) 0.72 0.32 (0.5) 125(35) 0 10.08(0.8) 200 (0) KrukalRST 0.40 0.76(0.7) 25 (9) 0.02 2.14 (0.9) 200 (0) 0 13.75(0.9) 200 (0) KruskalRST* 0.21 1.15(0.8) 11 (2) 0.54 0.58 (0.7) 26 (2) 0.99 0.01 (0) 38 (1) heur. ini 0 4.89(1.1) 2 (2) 0 16.55(1.4) 3 (3) 0 36.43(1.5) 4 (3) heur. xover 0 3.28(0.9) 9 (3) 0 13.60(1.5) 17 (4) 0 33.01(1.8) 20 (4) h. ini & xover 0 5.19(1.0) 1 (2) 0 16.98(1.3) 3 (3) 0 36.88(1.4) 3 (3) Pr¨ ufer 0.03 1.75(0.7) 18 (4) 0 6.15 (1.1) 79 (8) 0 14.44(1.6) 199 (2) NetKey 0.57 0.48(0.6) 16 (3) 0.91 0.10 (0.3) 39 (4) 1 0 (0) 81 (8) CV 0.95 0.05(0.2) 20 (4) 1 0 (0) 132(19) 0 16.76(1.0) 200 (0) NB (P2 =1) 0.55 0.50(0.5) 11 (3) 0.39 0.82 (0.7) 37 (5) 0.51 0.65 (0.5) 95 (12) NB (P2 =20) 0 2.96(0.8) 10 (3) 0 7.19 (1.3) 26 (5) 0 7.79 (1.6) 64 (10) LB (P1 =1) 0.96 0.04(0.2) 12 (2) 1 0 (0) 34 (4) 1 0 (0) 84 (9) LB (P1 =20) 0.61 0.43(0.6) 16 (3) 0.93 0.07 (0.2) 39 (4) 1 0 (0) 81 (8) P1 =P2 =1 0.78 0.23(0.4) 17 (4) 0.87 0.14 (0.3) 63 (8) 0.98 0.02 (0.0) 181(16) NetDir 0.37 0.82(0.7) 20 (6) 0.73 0.33 (0.6) 167(26) 0 9.96 (0.8) 200 (0) KrukalRST 0.42 0.73(0.7) 25 (9) 0.02 1.98 (0.8) 200 (0) 0 13.54(0.9) 200 (0) KruskalRST* 0.23 1.12(0.8) 11 (2) 0.59 0.51 (0.7) 25 (2) 1 0 (0) 36 (1) heur. ini 1 0 (0) 2 (1) 1 0 (0) 4 (1) 1 0 (0) 5 (0) heur. xover 0.65 0.39(0.6) 6 (1) 0.97 0.03 (0.1) 11 (1) 1 0 (0) 14 (0) h. ini & xover 1 0 (0) 1 (0) 1 0 (0) 2 (0) 1 0 (0) 3 (0)
space becomes too large and the random search behavior of GAs using the CV encoding does not allow it to ﬁnd the optimal solution any more. When searching for optimal stars, the performance of GAs using the CV encoding is in general low as the repair process has problems with creating starlike structures. LB encodings with P2 = 20 are strongly biased towards stars and GAs perform very well when searching for the optimal solution is a star, but fail completely when searching for lists, arbitrary trees, or the MST. When using the LB encoding with P2 = 1, the bias towards the star is smaller and there is an additional bias towards the MST. Therefore, this variant of the LB encoding performs slightly better for nonstars especially if the optimal solution is the MST. If the LB encodings only use a small linkspeciﬁc bias P1 = 1, the
250
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
random list
random star
Topt
Table 8.4. Performance of GAs using diﬀerent types of representations for onemax tree problems of diﬀerent sizes and with diﬀerent Topt (star and list) 8 nodes 20 nodes 40 nodes ﬁtness tconv ﬁtness tconv ﬁtness tconv Psucc P P µ (σ) µ (σ) succ µ (σ) µ (σ) succ µ (σ) µ (σ) Pr¨ ufer 0.46 0.72(0.8)14 (3) 0.72 0.54 (1.0) 43 (9) 0.98 0.02 (0.1) 73 (18) NetKey 0.67 0.38(0.6)16 (3) 0.94 0.06 (0.2) 41 (5) 1 0 (0) 94 (9) CV 0.94 0.08(0.4)20 (4) 0 8.03 (0.9)200 (0) 0 27.49(0.8)200 (0) NB (P2 =1) 0.57 0.95(1.4)10 (3) 0.85 0.28 (0.8) 26 (7) 0.98 0.03 (0.1) 47 (10) NB (P2 =20) 0.90 0.66(1.9) 6 (2) 1 0.02 (0.3) 13 (3) 1 0 (0) 27 (5) LB (P1 =1) 0.23 1.68(1.4)14 (4) 0.07 5.78 (3.5) 52 (13) 0.02 15.32(7.7)140(38) LB (P1 =20) 0.66 0.38(0.6)16 (3) 0.94 0.06 (0.3) 41 (5) 1 0 (0) 95 (9) P1 =P2 =1 0.77 0.34(0.8)16 (4) 0.99 0.01 (0.1) 48 (7) 1 0 (0) 135(16) NetDir 0.30 1.05(0.9)17 (6) 0.48 0.72 (0.8) 56 (34) 0.99 0.01 (0.1) 64 (4) KrukalRST 0.26 1.07(0.8)24(11) 0.24 1.22 (1.0)127(41) 0 20.84 (1) 200 (0) KruskalRST* 0.14 1.42(0.9)11 (2) 0.14 1.65 (1.1) 30 (2) 0.69 0.36 (0.6) 54 (2) heur. ini 0 4.94(0.9) 2 (2) 0 16.58(1.0) 3 (4) 0 36.47(1.1) 4 (4) heur. xover 0 3.64(1.1) 9 (3) 0 14.41(1.4) 17 (4) 0 33.99(1.5) 20 (4) h. ini & xover 0 5.23(0.7) 1 (2) 0 17.05(0.7) 2 (3) 0 36.96(0.7) 3 (4) Pr¨ ufer 0.02 1.81(0.7)18 (4) 0 6.20 (1.1) 80 (9) 0 14.35(1.7)199 (2) NetKey 0.55 0.51(0.6)16 (3) 0.90 0.11 (0.3) 39 (4) 1 0 (0) 76 (7) CV 0.96 0.04(0.2)20 (4) 1 0 (0) 114(14) 0 16.31(0.9)200 (0) NB (P2 =1) 0 3.61(0.8)11 (3) 0 13.04(1.3) 29 (6) 0 30.20(1.6) 65 (12) NB (P2 =20) 0 4.51(0.6) 7 (4) 0 14.23(1.0) 25 (5) 0 30.67(1.4) 47 (9) LB (P1 =1) 0.07 1.65(0.8)14 (4) 0 5.68 (1.6) 51 (12) 0 14.47(2.5)151(31) LB (P1 =20) 0.54 0.51(0.6)16 (3) 0.88 0.13 (0.3) 39 (4) 1 0 (0) 77 (8) P1 =P2 =1 0.03 1.83(0.8)19 (4) 0 6.12 (1.3) 70 (11) 0 15.82(2.0)191(14) NetDir 0.37 0.80(0.7)21 (6) 0.63 0.48 (0.7)198(10) 0 9.94 (0.8)200 (0) KrukalRST 0.43 0.71(0.7)25 (9) 0.01 1.93 (0.8)200 (0) 0 13.76(0.9)200 (0) KruskalRST* 0.25 1.07(0.8)11 (2) 0.63 0.44 (0.6) 25 (2) 1 0 (0) 35 (1) heur. ini 0 4.90(1.1) 2 (2) 0 16.53(1.4) 3 (3) 0 36.43(1.5) 4 (3) heur. xover 0 3.23(1.0) 9 (3) 0 13.49(1.6) 17 (4) 0 33.11(1.9) 20 (4) h. ini & xover 0 5.19(1.1) 2 (2) 0 16.97(1.3) 3 (3) 0 36.84(1.4) 3 (3)
encoding is biased towards MSTs. Therefore, GEAs using this encoding show low performance when searching for stars, lists, or arbitrary trees but high performance when searching for the MST. When using the LB encoding with a large linkspeciﬁc bias P1 = 20 the encoding performs the same as NetKeys. When using the LNB encoding with P1 = P2 = 1, there is a bias towards stars and the MST, and GAs show a high performance when searching for optimal stars or for the MST. However, for optimal lists or arbitrary trees, GA performance is low. The performance of the direct encodings NetDir and edgesets is more diﬃcult to predict in comparison to the indirect encodings as the design of the search operator is problemspeciﬁc, and it is often not obvious whether there is a bias. The results for the NetDir encoding show that the crossover
8.1 GEA Performance on Scalable Test Tree Problems
251
operator is biased towards stars. If the optimal solution is an arbitrary tree, a list, or the MST, GAs using the NetDir encoding fail for larger problem instances (n = 40). The results for the edgesets show that GAs using the nonheuristic KruskalRST crossover operator fail; in contrast, when using the slightly modiﬁed nonheuristic Kruskal RST* crossover operator, the encoding shows a good performance (similar to NetKeys). When using the heuristic variants of the edgesets (heur. ini, heur. xover, and heur. ini & xover), the results indicate a complete failure if the optimal solution is a star, a list, or an arbitrary tree even for small problem instances. Due to the strong bias of the heuristic edgesets towards the MST, GAs using the heuristic variants of the encoding fail when searching for nonMSTlike optimal solutions. If the optimal solution is the MST, edgesets with heuristics allow high GA performance and ﬁnd the MST very quickly. We see that the empirical results conﬁrm the theoretical predictions. It is more diﬃcult to predict the performance of direct representation as the properties of problemspeciﬁc search operators (like a possible bias) are more diﬃcult to identify. However, the theoretical investigations into the edgesets (compare Sect. 7.2) illustrated that analyzing the bias of the initialization, crossover, and mutation operators also allow us to make accurate predictions regarding GA performance. The results show that only NetKeys and the LNB encoding with a large linkspeciﬁc bias P1 show high performance for the onemax tree problem independently of the structure of the optimal solution. GAs using the nonheuristic KruskalRST* encoding perform slightly worse but the performance is still high. For the other types of encodings, we conﬁrmed the theoretical predictions and found that GAs using CVs fail for large problem instances due to problems with nonsynonymous redundancy, and GAs using Pr¨ ufer number fail due to the low locality of the encoding. GAs using the NB encoding or NetDir encoding only perform well if the optimal solution is a star, and GAs using heuristic variants of the edgeset encoding fail if the optimal solution is not the MST. 8.1.3 Deceptive Trap Problem for Trees We compare the performance of GAs using diﬀerent representations for the deceptive trap problem for trees. Problem Deﬁnition The deceptive trap problem for trees is deﬁned in Sect. 6.1.5. The problem is fully diﬃcult for crossoverbased GEAs as all tree schemata with k < n−1 that contain the optimal solution Topt have lower ﬁtness than their competitors. For the deceptive tree problem, an optimal solution Topt is chosen a priori either randomly or by hand. As for the onemax tree problem, we assume a minimization problem and the ﬁtness of Topt is fTopt = 0. The ﬁtness fTi of
252
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
other individuals Ti = Topt is deﬁned as di = n − dTopt ,Ti . Therefore, we get for the ﬁtness fi of an individual Ti 0 if Ti = Topt fi = (8.1) n − dTopt ,Ti if Ti = Topt The lower fi of an individual Ti , the less links it has in common with the optimal solution. The size of a deceptive trap problem denotes the number of nodes n. Even small instances of the trap problem for trees can not be solved by GEAs with reasonable population sizes. Therefore, to be able to construct larger problem instances, we concatenate m traps of size n. The size n of a deceptive trap is denoted by the number of nodes that are part of this subproblem. As a result, by concatenating m deceptive trap problems for trees of size n we can construct problems with overall m × n nodes. The ﬁtness of an individual is the sum of the ﬁtness of the deceptive subproblems and m−1 is calculated as i=0 fi , where fi is the ﬁtness of a subtree (8.1). When concatenating deceptive traps to larger problems, we have to consider that the subproblems of size n containing n−1 links must be connected to form a fully connected tree with mn−1 links. Therefore, at least (mn−1)−m(n−1) = m−1 links are not considered for the ﬁtness of an individual.
1
sub−problem 1 fitness: 3
2
sub−problem 2 fitness: 1
4 0
11
3
10 6
7 9
5 8
12
14
13
sub−problem 3 fitness: 0
overall fitness: 4 number of correct sub−problems: 1
Figure 8.1. Calculation of the ﬁtness for three concatenated, deceptive traps of size n = 5. Although a tree with 15 nodes has 14 links, only a maximum of 12 links is considered for the calculation of the ﬁtness.
We illustrate in Fig. 8.1 how the ﬁtness fi of an individual Ti is calculated if we concatenate three deceptive subproblems of size n = 5. The optimal solution should be the star with center 5i, where i ∈ {0, 1, 2}. If we have three concatenated subproblems, there exist three groups of nodes each consist of
8.1 GEA Performance on Scalable Test Tree Problems
253
nodes. For example, subproblem 3 consists of the nodes 5, 6, 7, 8, and 9. Subproblem 3 in Fig. 8.1 shows the optimal solution to the subproblem which is a star with center 5. The ﬁtness of an individual is the sum of the ﬁtness of the deceptive subproblems. Therefore, the minimum ﬁtness is zero (all three subproblems are correct and Ti = Topt ) and the maximum ﬁtness is 12 (each solution of a subproblem has only one link in common with the optimal solutions of the subproblem). In our example, subproblem 2 has no links in common with the optimal solution of the subproblem and its contribution to the overall ﬁtness is f2 = 1. For more information regarding the deceptive tree problem, the reader is referred to Sect. 6.1.5. Theoretical Predictions We predict the performance of GEAs for deceptive trap problems for trees based on the redundancy and locality of the diﬀerent representations. Due to uniform redundancy (compare Sect. 8.1.1), we expect the performance of GEAs using uniformly redundant encodings such as CVs, NetKeys, NetDir, LB (P1 = 20), and edgesets without heuristics to be independent of the structure of the optimal solution. We have seen in Sect. 8.1.1 that the nonuniformly redundant LNB encoding using a small linkspeciﬁc or nodespeciﬁc bias is either biased towards stars or the MST. Therefore, GEA performance would be high if we only had one deceptive subproblem, and the optimal solution is a star or the MST. However, when concatenating m subproblems to a larger problem, the overall optimal solution is diﬀerent from either a star or the MST. Thus, GEAs using LNB encodings with small biases fail for concatenated traps as the optimal solution is neither a star nor the MST. The situation is expected to be similar for edgesets with heuristics. As these encodings have a bias towards the MST, GEAs using these encodings also fail in ﬁnding the optimal solution as the overall optimal solution is not similar to the MST (even if the optimal solutions to the subproblems are MSTs). We have learned that lowlocality encodings make fully diﬃcult problems easier and fully easy problems more diﬃcult (compare Sect. 3.3.4). Therefore, GEAs using Pr¨ ufer numbers will perform well when used on the fully deceptive tree problem where the optimal solution is a nonstar. In comparison to the onemax tree problem where GEAs using Pr¨ ufer numbers failed completely, we expect a better performance for the trap problems. The low locality of the encoding destroys the deceptive character of the problem and makes it easier to solve for GEAs. Furthermore, because Pr¨ ufer numbers have higher locality around stars, GEAs have more diﬃculty in ﬁnding optimal stars than arbitrary trees when solving the trap. Consequently, GEA performance will be higher if the optimal solution of the deceptive trap is a nonstar. Finally, for the nonsynonymous redundant CV encoding, we expect the same eﬀects as for the onemax tree problem. Due to the nonsynonymity of the encoding, we expect GEAs to need more generations but to correctly
254
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
solve a higher number of problems. Although the overall number of ﬁtness evaluations that are necessary to ﬁnd the optimum strongly increases due to stealth mutation (populations do not converge and the GEA runs until it is stopped), CVs are an eﬃcient encoding for solving deceptive traps as the nonsynonymity of the encoding makes the problem easier to solve for GEAs. Experimental Results For our experiments we concatenate four instances of a deceptive trap problem for trees of size 3 and 4. The size of a subproblem denotes the number of nodes. Therefore, the problems for the empirical investigation have either n × m = 12 (size 3) or m × n = 16 nodes (size 4). The minimum ﬁtness of an individual is 0 and the maximum ﬁtness is either 4 × 2 = 8 (4 instances of the 3 node trap) or 4 × 3 = 12 (four instances of the 4 node trap). In our experiments we use a simple genetic algorithm without mutation. For the size 3 problems we use a population size of N = 100 and for the size 4 problems we use N = 1, 200. Trap problems of size 4 are already diﬃcult to solve for GAs and larger problems could hardly be solved using standard GAs. Because uniform crossover would result in high BBdisruption we use twopoint crossover in all runs. Furthermore, we use tournament selection without replacement of size 3. The runs are stopped after the population is fully converged or the number of generations exceeds 200 generations. We generate 200 random problems for each problem instance and perform 25 runs for the diﬀerent representations (compare Sect. 8.1.1) and for each of the randomly generated problems. The optimal solutions for the subproblems are either an arbitrary tree, the MST, or a star. Because we have seen for the onemax tree problem that GAs show the same performance when searching for lists as when searching for arbitrary trees, we neglect the case that the optimal solution is a list. As before, the GA performance is determined by the percentage Psucc of correctly solved problems at the end of the run and the number of generations tconv until the population is fully converged or the GA run is stopped. The performance of GAs using diﬀerent representations for size 3 and 4 trap problems is shown in Table 8.5. We present results for diﬀerent optimal solutions (Topt is either an arbitrary tree, the MST, or a random star) and show the percentage Psucc of runs that ﬁnd the optimal solution Topt , the mean and standard deviation of the ﬁtness of the best found solution, and the mean and standard deviation of the number of generations tconv until the population is fully converged or the GA run is stopped after 200 generations. As predicted, the performance of unbiased and uniformly redundant encodings (CV, NetKey, NetDir and LNB (P1 = 20)) is nearly independent of the structure of the optimal solution. Because LNB encodings with small biases P1 and P2 are biased either towards stars or the MST, the performance of GAs using these types of encodings is low. Because the optimal solutions have nothing in common with a star or the MST, GAs fail. The situation
8.1 GEA Performance on Scalable Test Tree Problems
255
random star
MST
arbitrary tree
Topt
Table 8.5. Performance of GAs using diﬀerent types of representations for deceptive tree problems of diﬀerent sizes and with diﬀerent Topt (arbitrary tree, MST, and star)
Psucc Pr¨ ufer NetKey CV NB (P2 =1) NB (P2 =20) LB (P1 =1) LB (P1 =20) P1 =P2 =1 NetDir KrukalRST KruskalRST* heur. ini heur. xover h. ini & xover Pr¨ ufer NetKey CV NB (P2 =1) NB (P2 =20) LB (P1 =1) LB (P1 =20) P1 =P2 =1 NetDir KrukalRST KruskalRST* heur. ini heur. xover h. ini & xover Pr¨ ufer NetKey CV NB (P2 =1) NB (P2 =20) LB (P1 =1) LB (P1 =20) P1 =P2 =1 NetDir KrukalRST KruskalRST* heur. ini heur. xover h. ini & xover
0.54 0.78 1 0 0 0.09 0.82 0.12 0.94 0.93 0.84 0 0 0 0.49 0.78 1 0 0 0.14 0.84 0.15 0.91 0.93 0.83 0 0 0 0.50 0.78 1 0 0 0.08 0.81 0.10 0.90 0.93 0.84 0 0 0
oder 3 ﬁtness tconv µ (σ) µ (σ) 0.49 (0.6) 26.0 (8.6) 0.23 (0.4) 23.0 (6.0) 0 (0) 199.9 (0.4) 2.58 (0.3) 15.8 (5.8) 2.62 (0.2) 19.7 (6.6) 1.69 (0.8) 19.5 (7.6) 0.18 (0.4) 23.3 (6.1) 1.24 (0.6) 27.4 (8.0) 0.06 (0.2) 107.0 (52.4) 0.07 (0.2) 141.2 (63.9) 0.16 (0.4) 12.2 (1.8) 3.83 (0.1) 0.4 (0.3) 2.63 (0.5) 8.7 (2.4) 3.88 (0.1) 0.4 (0.4) 0.54 (0.5) 26.4 (8.6) 0.23 (0.4) 23.3 (6.1) 0 (0) 199.9 (0.4) 2.53 (0.2) 16.3 (5.2) 2.61 (0.2) 19.4 (6.3) 1.41 (0.6) 20.9 (6.8) 0.17 (0.4) 23.5 (6.1) 1.12 (0.5) 27.9 (7.9) 0.09 (0.3) 108.2 (53.2) 0.07 (0.2) 143.5 (63.9) 0.18 (0.4) 12.2 (1.9) 3.81 (0) 0.4 (0.1) 2.48 (0.4) 9.1 (2.5) 3.85 (0) 0.3 (0.1) 0.54 (0.6) 26.4 (8.6) 0.22 (0.4) 23.1 (5.9) 0 (0) 200 (0.2) 2.58 (0.3) 15.6 (5.6) 2.59 (0.3) 19.3 (6.5) 1.71 (0.8) 19.3 (7.6) 0.20 (0.4) 23.5 (6.2) 1.28 (0.6) 27.3 (8.1) 0.10 (0.3) 108.0 (52.9) 0.07 (0.2) 143.2 (63.6) 0.17 (0.4) 12.2 (1.9) 3.85 (0.1) 0.4 (0.3) 2.62 (0.5) 8.7 (2.5) 3.89 (0.1) 0.3 (0.3)
Psucc 0.17 0.15 0.86 0 0 0 0.16 0.01 0.54 0 0 0 0 0 0.15 0.14 0.88 0 0 0 0.16 0.02 0.47 0 0 0 0 0 0.07 0.18 0.91 0 0 0 0.20 0.03 0.57 0 0 0 0 0
order 4 ﬁtness tconv µ (σ) µ (σ) 1.09 (0.6) 100.0 (44.3) 1.37 (0.8) 95.3 (30.0) 0.14 (0.3) 200 (0) 3.12 (0.4) 62.6 (29.2) 3.06 (0.4) 53.2 (35.5) 3.00 (0.7) 96.4 (34.0) 1.35 (0.9) 96.8 (30.4) 2.19 (0.7) 130.5 (32.0) 0.65 (0.8) 162.2 (40.1) 4.00 (0) 111.3 (37.5) 4.00 (0) 37.6 (7.5) 4.04 (0.1) 1.4 (1.6) 3.98 (0) 11.2 (3.1) 4.07 (0.1) 0.6 (0.8) 1.14 (0.6) 104.1 (43.5) 1.41 (0.8) 96.5 (30.6) 0.12 (0.3) 200 (0) 2.96 (0.1) 54.5 (14.9) 2.98 (0.1) 45.1 (21.6) 2.52 (0.5) 104.9 (29.5) 1.33 (0.8) 96.0 (30.7) 1.97 (0.5) 127.7 (29.8) 0.77 (0.8) 158.6 (40.1) 4.00 (0) 110.0 (36.3) 4.00 (0) 37.3 (7.5) 4.04 (0) 3.1 (0.8) 3.96 (0) 13.0 (1.3) 4.12 (0) 1.3 (0.2) 1.43 (0.7) 111.2 (46.8) 1.28 (0.8) 94.7 (30.1) 0.09 (0.3) 200 (0) 2.84 (0.1) 48.4 (14.2) 2.71 (0.2) 76.5 (24.6) 2.90 (0.7) 94.1 (35.1) 1.21 (0.8) 95.3 (30.3) 1.70 (0.6) 131.6 (30.7) 0.58 (0.8) 160.2 (40.8) 4.00 (0) 120.8 (38.6) 4.00 (0) 39.2 (7.6) 4.04 (0.1) 1.3 (1.5) 3.98 (0) 11.2 (2.7) 4.05 (0.1) 0.6 (0.7)
256
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
is similar when using the heuristic variants of the edgeset encoding. Due to their strong bias towards the MST, GAs fail for deceptive traps independently of the optimal solution of the subproblems, as the overall optimal solution is not similar to the MST. The nonheuristic variants of the edgesets show good performance for the smaller problem instances but fail for the larger problems. In comparison to the fully easy onemax problems, GAs using Pr¨ ufer numbers perform well. The low locality of the encoding helps GAs ﬁnd their way to the optimal solution. As expected, the solution quality is lower when searching for optimal stars (in comparison to arbitrary trees or the MST) because the locality around starlike structures is higher. As we have already seen for the onemax tree problem, the LB encoding with a large linkspeciﬁc bias P1 = 20 results in the same GA performance as when using the NetKey encoding. Although both representations are synonymously redundant, GAs using one of these two encodings are able to solve the problem and only need a few generations. Due to the nonsynonymity of CVs, GAs using the CV encoding perform very well for the diﬃcult deceptive tree problems. However, the nonsynonymity of the encoding increases the number of generations. The results for the NetDir representation are surprising. GAs using this encoding are able to ﬁnd a high proportion of BBs but need a larger number of generations. Coincidentally with the results for the onemax tree problem, NetKeys and LNB encodings with a large linkspeciﬁc bias allow GEAs to reliably solve concatenated trap problems for trees after a few generations. In contrast to the onemax tree problem, GEAs using Pr¨ ufer numbers or CVs show a comparable or higher performance for this diﬃcult problem as the low locality of Pr¨ ufer numbers and the nonsynonymity of CVs make diﬃcult problems easier to solve.
8.2 GEA Performance on Optimal Communication Spanning Tree Problems This section illustrates how we can use the framework about representations for predicting and verifying diﬀerences of GEA performance for some instances of the optimal communication spanning tree (OCST) problem. Consequently, this section also provides a comprehensive comparison for the performance of diﬀerent tree representations on OCST problems. We present results for test instances from Palmer (1994), Raidl (2001), Berry et al. (1995), and some new test problems. In Sect. 8.2.1, we give a brief description of the OCST problem. Section 8.2.2 gives an overview over diﬀerent approaches from the literature that have been used for solving the OCST problem. This is followed in Sect. 8.2.3 by a short description of the test problems (further details can be found in Appendix A). Then, in the remaining subsections, the inﬂuence of representations on
8.2 GEA Performance on the OCST Problem
257
GA performance is studied. Based on an analysis of the properties of representations, which is performed in Sect. 8.2.4, Sect. 8.2.5 gives some predictions on the expected GA performance for the test instances. The predictions and the validity of the developed theoretical concepts for the OCST test instances are veriﬁed in Sect. 8.2.6. 8.2.1 The Optimal Communication Spanning Tree Problem The OCST problem (also known as minimum communication spanning tree problem or simple network design problem (Johnson et al. 1978)) was introduced in Hu (1974). The problem is listed as [ND7] in Garey and Johnson (1979) and Crescenzi and Kann (2003). For the OCST problem, the number and positions of network nodes are given a priori and the cost of the tree is determined by the cost of the links. A link’s ﬂow is the sum of the communication demands between all pairs of nodes communicating either directly, or indirectly, over the link. The goal is to ﬁnd a tree that connects all given nodes and satisﬁes their communication requirements for a minimum total cost. The cost for each link is not ﬁxed a priori but depends on its distance weight and its capacity. A link’s capacity must satisfy the ﬂow over this link, which depends on the entire tree structure. The OCST problem can formally be deﬁned as follows. An undirected graph is denoted as G = (V, E). n = V  denotes the number of nodes and E denotes the number of edges. There are communication or transportation demands between the n diﬀerent nodes. The demands are speciﬁed by an n×n demand matrix R = (rij ), where rij is the amount of traﬃc required between location vi and vj . An n × n distance matrix D = dij determines the distance weights associated with each pair of sites. A tree T = (V, F ) where F ⊆ E and F  = V  − 1 is called a spanning tree of G if it connects all the nodes. The weight c(T ) of the spanning tree is the weighted sum over all pairs of vertices of the cost of the path between all pairs in T . In general, f (dij , bij ) , c(T ) = i,j∈F
where the n × n matrix B = bij denotes the traﬃc ﬂowing directly and indirectly over the edge between the nodes i and j. It is calculated according to the demand matrix R and the structure of T . T is the minimum communication spanning tree if c(T ) ≤ c(T ) for all other spanning trees T . For the OCST problem as proposed by Hu (1974) the cost of a link is calculated as the product of the distance weight dij times the overall traﬃc bij running over the edge. Therefore, f = dij bij and c(T ) = rij × d(pTi,j ), (8.2) i,j∈V, i<j
258
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
where d(pTi,j ) denotes the weight of the unique path from node i to node j in the spanning tree T . The OCST problem seeks the spanning tree with minimal costs among all other spanning trees. The OCST problem becomes the minimum spanning tree (MST) problem if f = dij . No communication requirements rij are considered. Then, T is the minimum spanning tree if c(T ) ≤ c(T ) for all other spanning trees T , where c(T ) = dij (8.3) (i,j)∈E
8.2.2 Optimization Methods for the Optimal Communication Spanning Tree Problem Like other constrained spanning tree problems, the OCST problem is N Phard (Garey and Johnson 1979, p. 207). Furthermore, it was shown in Reshef (1999) that the problem is MAX SN Phard (Papadimitriou and Yannakakis 1991) which means it cannot be solved using a polynomialtime approximation scheme, unless P = N P. Therefore, the OCST problem belongs to the class of optimization problems that behave like MAX3SAT (Garey and Johnson 1979). Only for a few easy and restricted problem instances have algorithms been developed that return optimal solutions. Hu (1974), who introduced the OCST problem, gave exact algorithms for two speciﬁc versions of the OCST problem. He showed that for the complete unweighted graph version, where dij = 1 for every i and j, the problem can be solved in polynomial time using the GomoryHu spanning tree algorithm (Gomory and Hu 1961; Hu 1974). Hu called this the optimum requirement spanning tree problem. In addition, he showed for the uniform demand version of the OCST, where the communication demands rij between any two sites are equal, that the optimal solution is a star if the distance weights dij satisfy a stronger version of the triangle inequality: for every 1 ≤ i, j, k ≤ n such that dij ≤ dik ≤ djk , we have (djk − dij )/dik ≤ (n − 2)/(2n − 2). If both the communication demands rij and the distances weights dij between any two sites are equal, then the optimal solution is also a star. Later, Johnson et al. (1978) showed that only the uniform demand version, where the dij satisfy this stronger version of the triangle equation, can be solved in polynomial time, and that all other uniform demand versions of the OCST problem, where dij ∈ {1, ∞} are N Phard. Wu et al. (1998) extended this work and showed that the uniform demand version where the weights dij satisfy the triangle inequality is N Phard. For the uniform demand version, Wong (1980) presented a heuristic that ﬁnds a tree T which has a maximum cost which is twice that of the optimal solution, c(T ) ≤ 2c(Topt ). The development of exact optimization methods for the general, nonuniform demand version of the OCST problem showed less success. Some early work (Dionne and Florian 1979; Lin 1982; Gavish 1983; Gavish and
8.2 GEA Performance on the OCST Problem
259
Altinkemer 1990) addressed the general network design problem and developed heuristics for ﬁnding optimal graphs G (not trees T ) for given distance weights dij and demands rij . However, as the assumptions that are made for solving the general network design problem are incompatible with solving the OCST problem (Palmer 1994, p. 10ﬀ), these heuristics can not be applied to the OCST problem. Later, Peleg (1997) showed that the OCST problem is reducible to a problem called minimum average stretch spanning tree (MAST) problem. Therefore, both problems are equivalent to each other and approximation algorithms for the MAST problem can also be used for the OCST problem. In the MAST problem, which was introduced in Alon et al. (1995), a graph G and a distance matrix D is given, and a spanning tree T has to be minimizes the average stretch of the edges (e.g. found that 1 T T d(p )/d minimize n−1 ij , where d(pi,j ) is the sum of all the weights i,j i,j∈E along the path between i and j in the spanning tree T ). Alon et al. presented a randomized algorithm for the MAST problem that constructs a spanning tree √ such that the average cost of the tree is less than, or equal to, exp(O( log n log log n)). Other approximation algorithms for the OCST problem are based on the G volume of communication c(G) = i,j∈E rij d(pi,j ) in the complete graph G, where d(pG ij ) is the sum of all the weights along the shortest path between i and j in G. c(G) represents a trivial lower bound for c(T ) because it considers the full original graph G and not only the links used for the tree T . Bartal (1996) and Wu et al. (1998) presented a randomized algorithm that constructs a spanning tree T with expected communication cost c(T ) = O(log2 n)c(G). This result has been improved by Bartal (1998) to an O(log n log log n) approximation. Around the same time, nonrandomized, deterministic algorithms were developed that ﬁnd a spanning tree with cost c(T ) = O(log2 n)c(G) (Peleg and Reshef 1998; Reshef 1999). Charikar et al. (1998) improved these results and presented a deterministic approximation algorithm that results in c(T ) = O(log n log log n)c(G). When using Euclidean distances as distance weights dij , Charikar et al. (1998) and Reshef (1999) presented deterministic approximation algorithms that output a spanning tree with cost c(T ) = O(log n)c(G). Despite the progress in developing approximation algorithms for the OCST problem that are based on the volume of communication c(G), Alon et al. (1995) showed that such approximation techniques cannot approximate the OCST problem better than Ω(log n). For more detailed information about approximation algorithms for the OCST problem, we refer to Reshef (1999). When summarizing the development of optimization algorithms for the OCST problem, we conclude that no eﬃcient algorithmic methods for solving the OCST problem are available. Some algorithms exist for simpliﬁed versions of the OCST problems (complete unweighted graph problem and uniform demand problems), but there are no eﬃcient methods for standard OCST problems. Similarly, deterministic and randomized approximation algorithms for the OCST problem are available which are based on the volume of commu
260
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
nication c(G), but none of them are able to output optimal or nearoptimal solutions (c(T ) ≥ Ω(log n)c(G)). To overcome the limitations of exact and approximation algorithms, and to be able to ﬁnd optimal or nearoptimal solutions for OCST problems, researchers have used heuristic optimization methods like GEAs, simulated annealing, tabu search, and other approaches. One of the ﬁrst heuristic approaches for the OCST problem was presented by Palmer (1994). When applying GEAs to the OCST problem, he recognized that the design of a proper tree representation is crucial for the performance of GEAs. Other representations that are used for the OCST problem or similar problems are the CV encoding (Berry et al. 1994; Berry et al. 1995), weighted encodings like the LNB encoding (Palmer 1994), the weighted encoding (Raidl and Julstrom 2000), the NetKey encoding (Rothlauf et al. 2002), diﬀerent variants of the LNBencoding (Krishnamoorthy and Ernst 2001), direct representations like edgesets (Raidl and Julstrom 2003; Tzschoppe et al. 2004) or other direct representations (Li 2001), determinant factorization (Abuali et al. 1995), Pr¨ ufer numbers (Palmer 1994; Palmer and Kershenbaum 1994a; Kim and Gen 1999; Zhou and Gen 1997; Krishnamoorthy et al. 1999; Gen et al. 1998; Gen et al. 1998; Gargano et al. 1998; Rothlauf and Goldberg 1999; Gottlieb et al. 2001; Julstrom 2001), or variants of Pr¨ ufer numbers like the Blob Code, the Happy Code and the Dandelion Code (Picciotto 1999; Julstrom 2001; Julstrom 2005). Other work applying GEAs to OCST and related problems (e.g. degree constraint MST problems) have been presented by Premkumar et al. (2001) and Chu et al. (2000)). It was shown in Rothlauf et al. (2003) that on average, optimal solutions for OCST problems are similar to the MST (compare also Sect. 7.2.3). That means the average distance dopt,M ST between the optimal solution and the MST is signiﬁcantly lower than the average distance drnd,M ST between a randomly created tree and the MST (Fig. 7.6). Therefore, as the optimal solution of an OCST problem is biased towards the MST, representations as well as operators that favor or overrepresent trees that are similar to the MST are expected to solve the OCST problem more eﬃciently. In summary, due to the lack of eﬃcient algorithmic methods for ﬁnding optimal or nearoptimal solutions, a large amount of work has applied GEAs to OCST problems. When using GEAs for tree problems, the choice of a proper tree representation is one of the most important factors for success. 8.2.3 Description of Test Problems Several instances of the OCST problem have been presented in the literature. The oldest test instances are from Palmer who introduced OCST test problems with 6, 12, 24, 47, and 98 nodes (Palmer 1994). The demands rij were inversely proportional to the distance weights dij . The nodes correspond to cities in the United States, and the distance weights dij were obtained from a tariﬀ database. In analogy to (8.2) the cost c(T ) of a tree is deﬁned as
8.2 GEA Performance on the OCST Problem
c(T ) =
rij × d(pTi,j ),
261
(8.4)
i,j∈V, i<j
where d(pTi,j ) denotes the weight of the unique path from node i to node j in the spanning tree T . For the exact distance and requirement matrix for the 6, 12 and 24 node problem the reader is referred to Palmer (1994) or Appendix A.1. Unfortunately, the data for the 47 and 98 node problems is no longer available1 . Raidl (2001) proposed several test instances of the OCST problem ranging from 10 to 100 nodes. The distances weights dij and the traﬃc demands rij have been generated randomly. The cost of a tree is calculated in analogy to Palmer’s test instances (compare (8.2) or (8.4)) The distance matrix and traﬃc demands are summarized in Appendix A.2. Berry et al. (1995) presented three instances of the OCST problem. They proposed one 6 node and two 35 node problems. The distance weights dij and the traﬃc demands rij are listed in Appendix A.3. As before, the cost of a tree is calculated according to (8.2) or (8.4)). Both 35 node problems use the same traﬃc demands rij , but diﬀer in the distance weights. One problem has uniform weights with dij = 1 (berry35u), and the other has nonuniform weights (berry 35). The best solution found by Berry et al. (1995) for the problem berry35 has cost c(Topt ) = 30, 467. Li and Bouchebaba (1999) improved these result to 16,915. The problem berry35u is an optimum requirement spanning tree problem and can be solved in polynomial time using the GomoryHu spanning tree algorithm (compare Sect. 8.2.2). The optimal solution has cost 16,273. The fourth group of test problems consists of problems that are derived from a realworld 26node problem from a company with locations all over Germany. For fulﬁlling the demands between the nodes, diﬀerent line types with only discrete capacities are available. The cost for installing a link consists of a ﬁxed share and a share which depends on its distance weight. Both cost components depend on the capacity of the link. The cost are based on the tariﬀs of the German Telecom from 1996. For an exact description on how the cost of a link depends on its weight dij and its capacity bij , the reader is referred to Appendix A.4. There are four diﬀerent problems. Rothlauf1 is a problem with n = 16 nodes and traﬃc demands ending only at node v1 (rij = 0, for i, j = 1). In rothlauf2, there are only n = 15 nodes and some additional traﬃc is added. Rothlauf3 is similar to rothlauf1 but uses a modiﬁed cost function. Finally, in rothlauf4 with n = 16, there is traﬃc between all nodes. The distance weights dij between the nodes are calculated as the Euclidean distances. The known best solutions for the four test problems are shown in Fig. A.1. We summarize the most important properties of the four test problems: 1 The data sets were not listed in the thesis and are not directly available from Palmer.
262
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
• rothlauf1: One headquarter and 15 branch oﬃces: This problem is the original design problem. All 15 branch oﬃces (node 2 to 16) communicate only with the headquarter (node 1). Possible line capacities are 64 kBit/s, 512 kBit/s, and 2048 kBit/s. • rothlauf2: One headquarter and only 14 branches: In this problem, one node is removed from the graph and some additional traﬃc is added. • rothlauf3: One headquarter, 15 branches and cheap lines for everybody: In this scenario, the ﬁxed cost for installing a line is only 10% of the cost in rothlauf1. Therefore, the cost of a link is mainly determined by its distance weight. Hence, the optimal solution is more like a minimum spanning tree. If the link costs would only be determined by the distance weights, and if there was only one possible capacity, the optimal solution would be the MST (compare (8.3)). • rothlauf4: 4 headquarters, 12 branches and all working together: For this problem, the demand matrix is completely ﬁlled. Between every node i and j some traﬃc exists. Between the four headquarters (nodes 1, 2, 3 and 4), the traﬃc is randomly chosen between 256 kBit/s and 512 kBit/s. Every other node communicates with the four headquarters and has a randomly chosen demand between 0 and 512 kBit/s. This demand is split into the headquarters at a ratio of 0.4, 0.3, 0.2, and 0.1 for the nodes 1, 2, 3, and 42 . Between all 12 branch oﬃces the demand is randomly chosen between 0 and 64 kBit/s. To make the problem more realistic, two additional line types are available. It is possible to use a line with 128 kBit/s and 4,096 kBit/s capacity with twice the cost of a 64kBit/s and 2,048 kBit/s line.
8.2.4 Analysis of Representations As we have seen in Sect. 8.1.1, an analysis of the bias of representations is important for evaluating the inﬂuence of representations on GEA performance and to be able to make accurate predictions about the expected GEA performance. Therefore, we perform the same type of analysis as described in Sect. 8.1.1. For each representation, we randomly generate 10,000 genotypes and measure the mean µ and the standard deviation σ of the minimum phenotypic distance towards a star, min(dprnd,star ), and the phenotypic distance towards the MST, dprnd,M ST . For the diﬀerent problems, we use the distance weights from the Appendix A. When using edgesets, the number N of individuals that are generated inﬂuence the structure of the encoded tree (compare Sect. 7.2.1). Therefore, for edgesets, we created 100 random populations of size N = 100 resulting in 10,000 random individuals.
2 Node 1 is the most important node and 40% of the traﬃc of the branches ends there; in node 2, 30% of the traﬃc ends, and so on.
8.2 GEA Performance on the OCST Problem
263
Table 8.6. We generate random genotypes xgrnd for the test instances from Palmer and Raidl and calculate the minimum phenotypic distance towards star networks, min(dprnd,star ), and the phenotypic distance to the MST, dprnd,M ST . The results conﬁrm the results presented in Sect. 8.1.1. problem distance d
Pr¨ ufer number NetKey CV NB (P2 =1) NB (P2 =20) LB (P1 =1) LB (P1 =20) LNB (P1 =P2 =1) NetDir KruskalRST, KruskalRST*, heur xover heur ini, heur ini & xover
palmer6 palmer12 palmer24 raidl10 raidl20 µ (σ) µ (σ) µ (σ) µ (σ) µ (σ) min(dp ) 2.04 (0.6) 7.22 (0.7) 18.50 (0.8) 5.42 (0.7) 14.69 (0.7) rnd,star dprnd,M ST 3.36 (0.9) 9.17 (1.2) 21.05 (1.3) 7.20 (1.1) 17.07 (1.3) min(dp rnd,star ) 2.12 (0.6) 7.30 (0.7) 18.61 (0.8) 5.51 (0.7) 14.78 (0.8) dprnd,M ST 3.34 (0.9) 9.15 (1.2) 21.07 (1.3) 7.21 (0.8) 17.11 (0.8) min(dp rnd,star ) 2.04 (0.6) 7.22 (0.7) 18.50 (0.8) 5.41 (0.7) 14.67 (0.8) dprnd,M ST 3.38 (0.9) 9.16 (1.2) 21.04 (1.3) 7.23 (1.1) 17.05 (1.3) min(dp ) rnd,star 1.00 (0.8) 5.49 (1.3) 7.59 (4.2) 4.12 (1.1) 10.75 (2.2) dprnd,M ST 2.37 (0.8) 6.12 (1.1) 18.68 (1.6) 4.24 (1.1) 11.63 (1.3) min(dp ) rnd,star 0.08 (0.3) 0.64 (1.4) 0.71 (2.1) 0.59 (1.1) 2.62 (3.2) dprnd,M ST 3.26 (0.5) 8.86 (0.7) 20.90 (0.9) 6.91 (0.9) 16.34 (1.4) min(dp rnd,star ) 2.08 (0.6) 7.50 (0.6) 18.61 (0.8) 5.41 (0.7) 14.78 (0.8) dprnd,M ST 2.50 (1.0) 6.55 (1.4) 19.45 (1.6) 4.61 (1.3) 12.74 (1.8) min(dp ) rnd,star 2.13 (0.6) 7.30 (0.7) 18.59 (0.8) 5.51 (0.7) 14.78 (0.8) dprnd,M ST 3.28 (0.9) 9.0 (1.2) 21.01 (1.3) 7.02 (1.1) 16.77 (1.3) min(dp ) rnd,star 1.59 (0.7) 6.22 (1.1) 15.39 (1.7) 4.66 (1.0) 12.48 (1.5) dprnd,M ST 2.68 (0.9) 7.12 (1.3) 20.21 (1.4) 5.11 (1.2) 14.02 (1.7) min(dp rnd,star ) 2.05 (0.6) 7.22 (0.7) 18.50 (0.8) 5.42 (0.7) 14.68 (0.8) dprnd,M ST 3.39 (0.9) 9.19 (1.2) 21.05 (1.3) 7.22 (1.1) 17.04 (1.3) min(dp rnd,star ) 2.12 (0.6) 7.30 (0.7) 18.60 (0.8) 5.50 (0.7) 14.78 (0.7) dprnd,M ST 3.30 (0.9) 9.17 (1.1) 21.07 (1.3) 7.20 (1.1) 17.11 (1.2) min(dp ) rnd,star 2.86 (0.3) 8.64 (0.5) 19.79 (0.41) 5.90 (0.3) 15.0 (0.1) dprnd,M ST 0.31 (0.6) 0.66 (0.8) 1.72 (1.8) 0.15 (0.3) 2.22 (0.6)
In Tables 8.6 (test instances from Palmer (1994) and Raidl (2001)) and 8.7 (test instances from Berry et al. (1995) and realworld test instances) we show the results for the diﬀerent representations. The results conﬁrm the ﬁndings from Sect. 8.1.1. The use of uniformly redundant encodings (NetKeys, CVs, or LB (P1 = 20)) results in about the same distances as when unbiased and nonredundant Pr¨ ufer numbers are used. The situation is the same for unbiased direct encodings (NetDir and edgesets with nonheuristic initialization). The situation is diﬀerent for nonuniform redundant or biased encodings. The two variants of the NB encoding, which are nonuniformly redundant, show a bias towards stars. As expected, the bias towards stars increases with increasing P2 (compare Sect. 6.4.3). For the LB encoding, the distances between randomly generated solutions and the MST decreases with lower P1 , as variants of the LB and LNB encoding with a low linkspeciﬁc bias P1 overrepresent MSTlike trees. Finally, the results for variants of the edgesets that use heuristic initialization (heur. ini, heur. ini & xover) conﬁrm the strong bias
264
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
Table 8.7. We generate random genotypes xgrnd for the realworld test instances and the test instances from Berry and calculate the minimum phenotypic distance towards star networks, min(dprnd,star ), and the phenotypic distance to the MST, dprnd,M ST . problem distance d
berry6 berry35u berry35 rothlauf2 rothlauf1, 3, 4 µ (σ) µ (σ) µ (σ) µ (σ) µ (σ) Pr¨ ufer min(dp 10.9 (0.8) rnd,star ) 2.03 (0.6) 29.2 (0.8) 29.2 (0.8) 9.99 (0.8) number dprnd,M ST 3.51 (0.8) n.a. 32.05 (1.3) 12.1 (1.1) 13.1 (1.2) min(dp ) 2.11 (0.6) 29.3 (0.8) 29.3 (0.8) 10.1 (0.7) 11.0 (0.8) rnd,star NetKey dprnd,M ST 3.34 (0.9) n.a. 32.0 (1.3) 12.1 (1.0) 13.1 (1.2) min(dp ) 2.04 (0.6) 29.2 (0.8) 29.2 (0.8) 9.98 (0.8) 10.9 (0.8) rnd,star CV dprnd,M ST 3.54 (0.8) n.a. 32.0 (1.3) 12.1 (1.0) 13.1 (1.2) min(dp ) 1.48 (0.6) 0 (0) 21.6 (3.5) 7.6 (1.8) 7.91 (1.8) NB (P2 =1) p rnd,star drnd,M ST 1.78 (0.9) n.a. 22.1 (2.7) 8.06 (1.4) 8.66 (1.5) min(dp ) 0.16 (0.4) 0 (0) 3.84 (4.7) 0.00 (0) 0.00 (0.1) NB (P2 =20) p rnd,star drnd,M ST 3.20 (0.8) n.a. 30.91 (1.7) 12.13 (0.6) 13.12 (0.6) min(dp ) 2.15 (0.5) 29.3 (0.8) 28.9 (0.9) 10.0 (0.8) 11.0 (0.8) LB (P1 =1) p rnd,star drnd,M ST 1.86 (0.9) n.a. 23.9 (2.4) 8.85 (1.6) 9.6 (1.7) min(dp 11.0 (0.8) rnd,star ) 2.11 (0.6) 29.3 (0.8) 29.3 (0.8) 10.1 (0.8) LB (P1 =20) p drnd,M ST 3.25 (0.9) n.a. 31.4 (1.5) 11.9 (1.3) 12.9 (1.3) LNB min(dp 8.4 (1.3) 9.20 (1.3) rnd,star ) 1.77 (0.6) 24.6 (1.9) 25.4 (2.0) (P1 =P2 =1) dprnd,M ST 2.14 (0.9) n.a. 26.54 (2.2) 9.75 (1.4) 10.5 1.5) min(dp 10.91 (0.8) rnd,star ) 2.04 (0.6) 29.2 (0.8) 29.2 (0.8) 9.98 (0.8) NetDir dprnd,M ST 3.50 (0.8) n.a. 32.0 (1.3) 12.1 (1.1) 13.1 (1.2) KruskalRST, min(dp 11.0 (0.7) rnd,star ) 2.12 (0.6) 29.3 (0.8) 29.2 (0.8) 10.1 (0.7) KruskalRST*, p d 3.34 (0.9) n.a. 32.0 (1.3) 12.1 (1.2) 13.1 (1.2) rnd,M ST heur xover heur ini, min(dp 2.0 (0) 23.6 (0.7) 30 (0) 11.0 (0) 12.0 (0) rnd,star ) heur ini & dprnd,M ST 0.12 (0.3) n.a. 0 (0) 0.66 (0.7) 0.45 (0.6) xover
of these encodings towards the MST. Randomly created solutions are only slightly diﬀerent from the MST. The ﬁndings for the diﬀerent representations are consistent for the diﬀerent test problems. The reader should bear in mind that it is not meaningful to calculate dprnd,M ST for the problem berry35u as all distance weights dij = 1, for i, j ∈ V . Therefore, no unique MST exists for this problem. 8.2.5 Theoretical Predictions on the Performance of Representations It is diﬃcult to predict GEA performance on a problem of unknown complexity. If we have no information regarding the structure of the optimal solution and the diﬃculty of the problem, we are not able to make any predictions about GEA performance. However, we know from experience that many realworld problems are easy. Therefore, the problems are often decomposable and the BBs are not fully deceptive, but of lower order.
8.2 GEA Performance on the OCST Problem
265
Therefore, due to the expected low diﬃculty of OCST problems (compare Sect. 3.3.4) we expect GEAs using Pr¨ ufer numbers to fail when used on these problems. The low locality of the encoding will increase the diﬃculty of easy problems and the use of GEAs results more in a random than a guided search. Furthermore, we know that the performance of GEAs using either uniformly redundant encodings (NetKeys, NetDir, or LNB encodings with large linkspeciﬁc bias P1 ), or unbiased direct representations (NetDir or the nonheuristic variants of the edgesets), is independent of the structure of the optimal solution. For the CV encoding, its nonsynonymous redundancy results in a higher evolvability and allows GEAs to reach a higher number of diﬀerent phenotypes. Therefore, we expect for smaller problems a better solution quality but a much higher number of search steps. For larger problems instances, we expect GEA failure due to the nonsynonymous redundancy of CVs. An important problemspeciﬁc property of the OCST problem was found by Rothlauf, Gerstacker, and Heinzl (2003) (compare also Sect. 7.2.3). This work performed a statistical analysis on the properties of optimal solutions Topt for randomly generated OCST problems using random demands rij and either Euclidean (on a twodimensional grid), or random distance weights dij . They compared the average distances µ(dM ST,rnd ) of randomly created trees towards the MST to the average distances µ(dmst,opt ) of the optimal solutions towards the MST. The results show that the average distance between the optimal solution Topt and the MST is signiﬁcantly smaller than the average distance between a randomly created tree and the MST (compare Fig. 7.6). Therefore, optimal solutions for the OCST problem are biased towards the MST. We have seen in Sects. 8.2.4 and 8.1.1 that the LB and LNB encoding using a small linkspeciﬁc bias P1 as well as the edgeset encoding with heuristics are biased towards the MST. When using these types of encodings, the average distance between a randomly created solution and the MST is smaller in comparison to a nonredundant and unbiased encoding. As the results presented in Fig. 7.6 indicate that in comparison to randomly created solutions optimal solutions are more similar to the MST, we expect good GEA performance for variants of the LNB encoding using a low linkspeciﬁc bias. Such variants of the LB and LNB encoding are biased towards the MST and can make use of this problemspeciﬁc property of the OCST problem. Furthermore, Sect. 7.2.3 has shown that the performance of edgesets with heuristics is low if the optimal solution is not the MST. As for edgesets with heuristics the bias towards the MST is too strong, problems can only be solved where the optimal solution is the MST. Therefore, we expect diﬃculties when using edgesets with heuristics for OCST problems as only problems can be solved where the optimal solution is slightly diﬀerent from the MST. In analogy, we expect low performance for variants of the NB encoding, as these encodings overrepresent trees that are similar to stars.
266
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
8.2.6 Experimental Results As in Sects. 8.1.2 and 8.1.3, for our experiments we use a simple, generational genetic algorithm without mutation. For all test problems, tournament selection without replacement of size 2 and uniform crossover is used. Due to the diﬀerent diﬃculty of the test problems, we use diﬀerent population sizes N for the diﬀerent test instances. The used population sizes N are listed in Table 8.8 and are chosen independently for each test problem. They are chosen such that for some representations the GA ﬁnds the optimal solution in more than half of the runs (Psucc > 0.5). In general, higher N increases the performance (higher Psucc ), whereas lower N reduces GA performance. However, the performance diﬀerences between the diﬀerent representations remain about the same when using diﬀerent population sizes N . All runs are stopped after the population is fully converged, or a maximum of 200 generations is reached. We perform 250 runs for each representation and problem instance.
palmer6 palmer12 palmer24 raidl10 raidl20 berry6 berry35u berry35 rothlauf1 rothlauf2 rothlauf3 rothlauf4
n
N
6 12 24 10 20 6 35 35 16 15 16 16
16 300 800 70 800 16 2,000 300 800 2,000 1,200 800
Properties of optimal dmst,opt min(dstar,opt ) 1 2 5 7 12 17 3 4 4 14 0 2 28 0 30 7 9 4 8 6 9 9 7
solutions c(Topt ) 693,180 3,428,509 1,086,656 53,674 157,570 534 16,273 16,915 60,883 58,619 28,451 112,938
Table 8.8. Population size N used for the diﬀerent test problems and properties of the optimal (or known best) solutions Topt of the test instances
Furthermore, Table 8.8 presents the properties of the optimal (or best found) solutions Topt for the test problems. The results show that the optimal solutions have about the same minimal distances towards a star than randomly created unbiased individuals (compare Tables 8.6 and 8.7). However, the optimal solutions have a smaller distance towards the MSTs than a randomly generated individual. This means that the good solutions are biased towards the MST. As a result, we expect GAs that use encodings which introduce a modest bias towards the MST to have high performance. For these types of encodings (like the LB (P1 = 1) encoding), MSTlike individuals are overrepresented. The performance of GAs using diﬀerent representations (compare Sect. 8.1.1) for the diﬀerent test instances is shown in Table 8.9. We show the mean and standard deviation of the distance dbestf,opt between the best found solution and the optimal solution, the percentage Psucc of runs that ﬁnd the
8.2 GEA Performance on the OCST Problem
267
optimal solution Topt , the mean and standard deviation of the cost c(Tbestf ) of the best solution Tbestf that have been found after tconv generations, and the mean and standard deviation of the number of generations tconv until the population is fully converged or the GA run is stopped after 200 generations. dbestf,opt and Psucc are related as a high percentage Psucc results in a low distance dbestf,opt . All results are averaged over 250 runs.
palmer12 (c(Topt =3,428,509)
palmer6 (c(Topt =693,180)
Table 8.9. Performance of GA using diﬀerent types of representations for the test problems described in Sect. 8.2.3.
Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover
dbestf,opt µ (σ) 1.92 (0.93) 1.32 (1.03) 1.09 (1.02) 0.38 (0.71) 2.66 (0.68) 0.56 (0.85) 1.14 (0.97) 0.70 (0.88) 1.57 (0.83) 1.62 (1.00) 1.58 (0.95) 1 (0) 1 (0.70) 1 (0) 6.49 (0.95 2.03 (1.78) 4.10 (1.16) 0 (0) 1.09 (1.52) 0.73 (1.18) 1.99 (1.76) 1.78 (1.64) 1.72 (1.55) 2.53 (1.32) 1.90 (1.52) 6.00 (0) 4.81 (0.39) 5.00 (0)
Psucc 0.056 0.268 0.392 0.720 0.004 0.660 0.328 0.532 0.088 0.152 0.136 0 0.224 0 0 0.360 0 1 0.604 0.628 0.352 0.432 0.268 0.068 0.292 0 0 0
c(Tbestf ) µ (σ) 733,820 (39,464) 716,171 (26,364) 706,908 (17,693) 697,842 (12,041) 780,191 (36,189) 698,782 (8,561) 708,900 (18,926) 700,173 (12,790) 727,717 (42,784) 731,675 (48,865) 728,206 (36,666) 709,770 (0) 711,885 (21,340) 709,770 (0) 4,111,091 (100,102) 3,452,988 (27,768) 3,700,839 (115,314) 3,428,509 (0) 3,463,583 (56,789) 3,434,632 (14,913) 3,450,763 (27,265) 3,448,563 (22,482) 3,456,236 (38,339) 3,462,151 (44,551) 3,449,270 (22,488) 3,727,552 (0) 3,744,861 (36,297) 3,876,488 (0)
tconv µ (σ) 14.1 (3.5) 15.4 (3.3) 18.4 (5.6) 10.7 (2.8) 7.7 (3.1) 12.1 (2.6) 15.2 (3.3) 16.2 (3.5) 14.3 (3.8) 17.1 (5.4) 10.6 (2.6) 1.5 (0.6) 6.3 (1.8) 1.1 (0.3) 46.1 (6.9) 60.4 (7.2) 200 (0) 35.5 (5.2) 33.6 (6.7) 52.6 (7.0) 60.2 (7.9) 77.3 (11.1) 61.5 (8.8) 88.5 (10.7) 36.5 (8.3) 10.7 (1.0) 16.7 (2.9) 4.4 (0.6)
berry6 (c(Topt =534)
raidl10 (c(Topt =53,674)
palmer24 (c(Topt =1,086,656)
268
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover
dbestf,opt µ (σ) 7.46 (1.06) 0.26 (0.44) 12.48 (1.59) 2.52 (0.73) 10.07 (2.74) 0.56 (0.56) 0.26 (0.44) 0.16 (0.39) 4.10 (1.84) 10.65 (1.34) 1.80 (0.89) 11.00 (0) 10.99 (0.32) 12.00 (0) 1.81 (0.89) 0.34 (0.71) 0.16 (0.38) 0 (0.06) 1.03 (0.49) 0 (0) 0.27 (0.59) 0.04 (0.21) 0.26 (0.52) 0.34 (0.57) 0.79 (0.92) 2.04 (0.21) 2.76 (0.45) 3.00 (0) 1.87 (0.69) 0.52 (0.67) 0.35 (0.62) 0.04 (0.22) 1.78 (0.67) 0.02 (0.13) 0.46 (0.61) 0.09 (0.29) 0.90 (0.86) 0.63 (0.75) 1.08 (0.94) 0 (0) 0.22 (0.45) 0 (0)
Psucc 0 0.736 0 0 0 0.476 0.740 0.852 0 0 0.068 0 0 0 0.080 0.760 0.844 0.996 0.052 1 0.788 0.956 0.772 0.708 0.480 0 0 0 0.012 0.568 0.720 0.960 0.016 0.984 0.604 0.908 0.380 0.524 0.324 1 0.800 1
c(Tbestf ) tconv (σ) µ (σ) 1,584,450 (171,348) 90.7 (22.9) 1,086,866 (353) 87.3 (7.8) 2,769,489 (286,982) 200 (0) 1,120,965 (40,601) 55.3 (11.2) 1,627,326 (196,475) 40.2 (10.9) 1,087,090 (436) 85.4 (8.8) 1,086,863 (350) 86.6 (9.2) 1,092,012 (16,559) 135.3 (16.2) 1,119,077 (43,933) 199.8 (1.3) 1,547,827 (91,086) 200 (0) 1,088,152 (778) 54.3 (2.3) 1,884,444 (0) 6.9 (0.4) 1,855,160 (76,119) 21.7 (3.6) 1,959,790 (0) 6.7 (0.7) 68,867 (9,914) 29.0 (5.2) 54,514 (1,967) 33.1 (4.6) 53,876 (520) 81.2 (16.7) 53,693 (308) 16.6 (3.3) 59,262 (5,361) 16.8 (4.1) 53,674 (0) 24.8 (3.7) 54,363 (1,776) 32.0 (4.4) 53,888 (999) 30.9 (4.9) 54,415 (2,275) 32.4 (4.0) 54,208 (1,398) 51.8 (9.2) 56,013 (4,286) 23.1 (3.0) 55,942 (355) 7.4 (1.4) 57,261 (864) 13.6 (3.7) 58,352 (0) 2.1 (0.8) 677 (75) 13.4 (3.3) 557 (36) 14.5 (3.1) 549 (33) 16.9 (4.0) 535 (7) 9.0 (2.7) 628 (59) 9.8 (2.9) 535 (5) 9.2 (2.2) 553 (29) 14.0 (3.0) 536 (8) 13.2 (3.3) 588 (69) 13.9 (3.1) 563 (41) 15.2 (3.7) 592 (68) 10.3 (2.2) 534 (0) 0.7 (0.6) 545 (26) 5.4 (1.1) 534 (0) 0.6 (0.5)
rothlauf1 (c(Topt =60,883)
berry35u (c(Topt =16,273)
berry35 (c(Topt =16,915)
8.2 GEA Performance on the OCST Problem
Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover
dbestf,opt µ (σ) 25.37 (1.25) 4.11 (1.66) 24.51 (1.69) 0.16 (0.36) 8.58 (2.83) 0.44 (0.50) 2.78 (1.34) 1.88 (1.09) 14.06 (2.28) 18.47 (1.93) 6.87 (1.96) 0 (0) 0.03 (0.18) 0 (0) 30.68 (0.62) 8.88 (2.01) 25.94 (1.61) 30 (0) 30 (0) 8.48 (2) 8.54 (1.99) 25.56 (1.62) 20.56 (1.98) 22.82 (2.34) 7.08 (0.97) 32.56 (1.03) 7.1 (0.86) 32.74 (1.07) 6.94 (0.42) 1.34 (1.16) 8.75 (1.20) 8.47 (0.97) 8.98 (0.20) 1.39 (1.22) 1.50 (1.21) 2.77 (0.95) 3.88 (1.24) 7.91 (1.47) 1.19 (1.08) 9.00 (0) 7.25 (0.62) 7.00 (0)
Psucc 0 0.028 0 0.988 0 1 0.112 0.256 0 0 0 1 0.972 1 0 0 0 0 0 0 0 0 0 0 0.42 0 0.32 0 0 0.296 0 0 0 0.312 0.272 0.004 0 0 0.332 0 0 0
269
c(Tbestf ) tconv (σ) µ (σ) 59,553 (2,795) 104.7 (12.3) 18,083 (763) 99.1 (8.6) 120,397 (12,749) 200 (0) 16,916 (12) 56.1 (6.4) 20,769 (1,976) 45.6 (5.9) 16,915 (0) 75.2 (6.8) 17,519 (526) 94.0 (8.0) 17,181 (250) 128.2 (12.7) 27,975 (4,019) 200 (0) 39,806 (3,512) 200 (0) 19,913 (1,268) 74.8 (4.5) 16,915 (0) 0 (0) 16,923 (56) 15.0 (0.7) 16,915 (0) 0 (0) 21,553 (230) 190.2 (29.1) 17,111 (326) 166.6 (30.2) 34,989 (1,558) 200 (0) 21,513 (0) 27.3 (3.7) 21,513 (0) 28.5 (4.1) 17,061 (329) 169.3 (28.9) 17059 (362) 169 (30.5) 20,093 (353) 189.4 (18.1) 22,264 (1,441) 200 (0) 26,929 (1053) 200 (0) 16,299 (73.2) 131.22 (47.9) 42,129 (5,711) 79.48 (87.48) 16,323 (74.7) 125.58 (44.84) 42,803 (6,585) 91.4 (89.87) 63,680 (236) 62.1 (14.1) 61,144 (345) 86.6 (9.1) 71,657 (2,592) 200 (0) 64,654 (684) 34.2 (9.6) 64,997 (106) 28.6 (5.8) 61,202 (424) 85.1 (8.5) 61,161 (372) 85.5 (8.4) 61,233 (123) 119.3 (17.5) 61,709 (523) 110.9 (11.1) 64,539 (1,411) 200 (0) 61,315 (474) 54.2 (5.0) 61,817,796 (0) 6.0 (0.3) 628,632 (182,178) 20.9 (3.1) 71,864,456 (0) 5.1 (0.5)
rothlauf4 (c(Topt =112,938)
rothlauf3 (c(Topt =28,451)
rothlauf2 (c(Topt =58,619)
270
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KruskalRST KruskalRST* heur. ini heur. Xover heur. ini & xover Pr¨ ufer NetKey CV NB (P2 = 1) NB (P2 = 20) LB (P1 = 1) LB (P1 = 20) LNB (P1 =P2 =1) NetDir KrukalRST KruskalRST* heur. ini heur. Xover heur. ini & xover
dbestf,opt µ (σ) 8.16 (0.91) 0.46 (0.83) 7.55 (1.23) 8.74 (1.52) 9.93 (0.49) 0.54 (0.88) 0.43 0.79) 2.06 (0.33) 1.43 (0.98) 2.50 (1.01) 0.17 (0.48) 6.00 (0) 4.82 (0.40) 5.00 (0) 6.94 (0.84) 1.31 (1.49) 9.32 (1.37) 8.17 (1.16) 8.92 (0.38) 1.42 (1.52) 1.35 (1.59) 4.64 (1.09) 1.80 (1.46) 6.24 (1.34) 2.50 (2.01) 8.00 (0) 4.92 (0.28) 6.00 (0) 9.20 (0.7) 3.20 (1.6) 10.03 (1.4) 10.30 (0.8) 11.76 (0.7) 3.10 (1.5) 3.22 (1.6) 5.80 (1.32) 1.84 (0.9) 5.57 (1.0) 4.20 (1.8) 9 (0) 8.07 (0.2) 9 (0)
Psucc 0 0.756 0 0 0 0.720 0.760 0 0.168 0.004 0.868 0 0 0 0 0.440 0 0 0 0.416 0.480 0 0.384 0 0.240 0 0 0 0 0.128 0 0 0 0.136 0.14 0 0.156 0.004 0.056 0 0 0
c(Tbestf ) (σ) 66,221 (868) 58,651 (57) 69,598 (1,960) 69,901 (2,421) 73,276 (917) 58,664 (114) 58,652 (61) 59,078 (318) 58,872 (231) 59,224 (404) 58,680 (207) 1,281,721 (0) 74,522 (931) 1,284,189 (0) 31,247 (500) 28,722 (352) 34,870 (936) 32,704 (897) 34,286 (573) 28,709 (337) 28,694 (351) 30,386 (298) 28,891 (395) 30,190 (520) 29,182 (542) 61,781,980 (0) 589,340 (168,997) 71,829,016 (0) 128,572 (2,350) 115,016 (2968) 146,387 (4,592) 125,095 (1,856) 128,211 (1,260) 114,741 (2,570) 115,106 (2,790) 117,755 (1,990) 113,086 (438) 119,860 (941) 117,515 (4,107) 186,230 (0) 156,209 (5,134) 197,511 (0)
tconv µ (σ) 63.9 (28.2) 91.1 (9.0) 200 (0) 45.7 (14.2) 34.4 (8.1) 90.9 (9.6) 90.2 (8.5) 132.3 (18.1) 107.9 (7.2) 200 (0) 50.3 (10.1) 6.2 (0.4) 20.0 (2.9) 8.7 (0.8) 79.0 (23.0) 98.2 (9.1) 200 (0) 45.7 (16.2) 34.3 (8.1) 97.6 (9.8) 99.4 (10.7) 157.2 (21.3) 115.2 (13.3) 200 (0) 60.3 (10.4) 6.1 (0.3) 19.0 (2.6) 5.0 (0.4) 68.2 (29.0) 94 (8.2) 200 (0) 52.4 (45.7) 28.6 (17.2) 94 (11.6) 93.6 (11.0) 132.1 (21.9) 92.2 (12.3) 200 (0) 63.98 (28.5) 8.14 (0.4) 19.61 (2.7) 7.32 (0.8)
8.2 GEA Performance on the OCST Problem
271
The results conﬁrm the predictions from Sect. 8.2.5. Due to problems with low locality, GAs using Pr¨ ufer numbers fail and do not ﬁnd the optimal solution even for small problem instances. GAs using the uniformly and synonymously redundant NetKey encoding show a good performance for the diﬀerent test problems independently of the properties of the optimal solution. The CV encoding results in low GA performance as the nonsynonymous redundancy of the encoding does not allow guided search and GAs behave like random search. Only for small six or ten node problem instances can good solutions be found, but large numbers of generations are necessary. The results for the LNB encoding and variants of it show that the use of low values for P1 an P2 results in high GA performance as solutions similar to the MST are overrepresented. The LB encoding with P1 = 1 seems to be an especially good choice, whereas using the NB encoding with large P2 results in a strong bias towards stars and does not allow the GA to reliably solve the test problems. The direct encoding NetDir shows good results and allows GAs to reliably solve the problem instances. The situation is diﬀerent when using edgesets as the performance of GAs depends on the properties of the optimal solution Topt . GAs using heuristic variants of the edgeset encoding (heur. ini, heur. xover, or heur. ini & xover) can only solve test problems where the optimal solution is the MST (dmst,opt = 0). For other test problems, the optimal solution Topt can not be found. In contrast, GAs using the nonheuristic and unbiased variants of the edgeset encoding (KruskalRST or KruskalRST*) often show good performance independently of the structure of Topt . When focusing on the diﬀerent test problems, the results reveal that for berry6 and berry35, where the optimal solution is the MST, GAs using either the LNB encoding with low bias values, or edgesets with heuristics, show very high performance and outperform all other unbiased encodings. However, with increasing dopt,M ST , GAs using heuristic variants of the edgeset fail as the bias towards the MST is too strong. GAs using the LB or LNB encoding with low linkspeciﬁc bias still allow eﬃcient search as the overrepresentation of MSTlike solutions is lower than the bias of the heuristic variants of the edgesets. Furthermore, the results show that the performance of GAs using uniformly redundant or unbiased encodings such as NetKeys, NetDir, edgesets without heuristics, or the LB (P1 = 20) encoding is nearly independent of the structure of the optimal solution Topt . Psucc and tconv are not strongly aﬀected by the properties of the optimal solution. We see that the empirical results conﬁrm the theoretical predictions made in Sect. 8.2.5. As expected, it is more diﬃcult to predict the performance of direct representations as properties of problemspeciﬁc search operators (like a possible bias) are more diﬃcult to identify. However, using the developed theory of representations for a thorough analysis of search operators (compare Sect. 7.2 regarding the edgesets) allows us to validate the predictions about the expected GA performance not only for indirect, but also for direct representations.
272
8 Performance of Genetic and Evolutionary Algorithms on Tree Problems
The experimental results conﬁrmed the results of the theoretical analysis from the previous chapters and showed that only NetKeys, the LNB encoding with a large linkspeciﬁc bias P1 , edgesets without heuristics, and NetDir show good performance for the test problems independently of the properties of the optimal solution. Furthermore, as expected, GAs using CVs fail for large problem instances due to problems with nonsynonymous redundancy, GAs using Pr¨ ufer number fail due to the low locality of the encoding, GAs using the NB encoding show low performance due to their bias towards stars, and GAs using heuristic variants of the edgeset encoding fail if the optimal solution is not the MST.
8.3 Summary This chapter investigated for diﬀerent tree problems how representations inﬂuence the performance of GEAs. Section 8.1 focused on scalable test problems like the onemax tree problem and the deceptive trap problem for trees, and examined how the performance of GEAs depends on the used representation. These two types of test problems for trees allow us to determine a priori the structure of the optimal solution. The experimental results conﬁrmed the theoretical predictions regarding the inﬂuence of representations on the performance of GAs. Section 8.2 focused on OCST problems and investigated for diﬀerent test problems from the literature how the used representation inﬂuences GEA performance. We deﬁned the problem, described existing approaches to solve the problem, and developed theoretical predictions on how the diﬀerent representations inﬂuence GEA performance based on the properties of the encodings from Chaps. 6 and 7. The experimental results conﬁrmed the theoretical predictions. This chapter illustrated how GEA users can use the framework of representations for predicting and explaining the performance of GEAs using diﬀerent types of representations. When applying GEAs to the onemax tree and deceptive trap problem, the optimal solutions are known a priori and it is possible to validate the predictions concerning the expected GEA performance. The experimental results for the scalable test problems conﬁrmed the theoretical predictions. GEAs using Pr¨ ufer numbers fail for onemax tree problems but perform well for deceptive tree problems. Due to problems with nonsynonymous redundancy, GEAs using CVs can only solve small instances of the onemax tree problem. When using CVs for larger onemax tree problems, GEAs fail. The situation is diﬀerent for deceptive trap problems as the nonsynonymity of CVs makes such problems easier and allows GAs to ﬁnd optimal solutions more easily in comparison to other encodings. LNB encodings with a small linkspeciﬁc or nodespeciﬁc bias are biased either towards stars or towards the MST. Therefore, GEAs using such encodings perform well if the optimal solution is either starlike or MSTlike. If the optimal solution is not starlike or MSTlike, GEAs using LNB encodings with a small bias
8.3 Summary
273
show low performance. The situation is similar for edgesets. Edgesets with heuristics show a strong bias towards the MST and only perform well if the optimal solution is the MST. If the optimal solution is slightly diﬀerent from the MST, GEAs using edgesets with heuristics fail. Edgesets without heuristics are nearly unbiased and GA performance is independent of the optimal solution. Network random keys show the same performance as LNB encodings with a large linkspeciﬁc bias P1 . GEAs using these encodings perform reliably well on all test instances and are in general a good choice for the onemax tree and deceptive tree problem. The situation is similar when using GEAs for OCST problems. For many problem instances, the optimal solution is similar to the MST. Therefore, GEAs using encodings that are biased towards the MST show good performance. However, the results show that the bias of the heuristic variants of the edgesets is too strong, as only problems can be solved where the optimal solution is the MST. A better choice are the LNB or LB encoding with a small linkspeciﬁc bias (P1 = 1) as such encodings overrepresent solutions similar to the MST but still allow us to ﬁnd solutions that are diﬀerent from the MST. Furthermore, the results show that NetKeys and LNB encoding with a large linkspeciﬁc bias perform well, whereas Pr¨ ufer numbers, CVs and LNB encodings with a large nodespeciﬁc bias fail. This chapter nicely illustrated the beneﬁts we gain from analyzing the properties of encodings and using the framework about representations for predicting the expected GEA performance. We want to encourage researchers to use the proposed tools for analyzing the properties of representations or search operators for other problem domains. Examining redundancy, scaling, and locality of an encoding a priori allows us to predict GEA performance based on the presented framework about representations.
9 Summary and Conclusions
The purpose of this book is to understand the inﬂuence of representations on the performance of genetic and evolutionary algorithms. This chapter summarizes the work contained in this study and lists its major contributions.
9.1 Summary We started in Chap. 2 by providing the necessary background for examining representations for GEAs. Researchers recognized early that representations have a large inﬂuence on the performance of GEAs. Consequently, after a brief introduction into representations and GEAs, we discussed how the inﬂuence of representations on problem diﬃculty can be measured. The chapter ended with prior guidelines for choosing highquality representations. Most of them are mainly based on empirical observations and intuition and not on theoretical analysis. Therefore, we presented in Chap. 3 three aspects of a theory of representations for GEAs. We investigated how the locality, scaling, and locality of an encoding inﬂuences GEA performance. The performance of GEAs is determined by the solution quality at the end of a run and the number of generations until the population is converged. Consequently, for redundant and exponentially scaled encodings, we presented population sizing models and described how the time to convergence is changed. Furthermore, we were able to demonstrate that highlocality encodings do not change the diﬃculty of a problem; in contrast, when using lowlocality encodings, on average, the diﬃculty of problems changes. Therefore, easy problems become more diﬃcult and diﬃcult problems become easier by the use of lowlocality encodings. For all three properties of encodings, the theoretical models were veriﬁed with empirical results. In Chap. 4, we combined the three elements of representation theory from Chap. 3 to form a framework for a theoryguided design and analysis of representations. The framework describes how the time to convergence and the
276
9 Summary and Conclusions
solution quality of GEAs depends on the redundancy, scaling, and locality of a representation. Using this framework, we could estimate the performance of GEAs using diﬀerent types of representations in a theoryguided manner. We recognized that diﬀerent types of representations can dramatically change the behavior of GEAs. Consequently, we presented major implications when using diﬀerent types of representations. In Chaps. 5, 6, and 7, we used the framework for the analysis of existing representations and the design of new representations in a theoryguided matter. We analyzed the redundancy, scaling, and locality of a variety of diﬀerent representations and made predictions regarding GEA performance based on the framework. These predictions were veriﬁed by experimental results. In Chap. 5, we focused on binary representations for integers. We compared the performance of binary, Gray, and unary encodings for integer problems. Based on the properties of the encodings, the framework was able to explain existing performance diﬀerences. Chapter 6 focused on the analysis and design of tree representations. In analogy to Chap. 5, we analyzed the properties of tree representations, namely Pr¨ ufer numbers, NetKeys, characteristic vector encodings, and the linkandnodebiased encoding. Based on these investigations, we were able to predict GEA performance using the representation framework. Based on the insights into the insuﬃciencies of existing tree representations, we combined in Sect. 6.5 advantageous elements of the CV and the LNB encoding to form the NetKey encoding. The NetKey encoding is a redundant encoding that allows GEAs to perform well independently of the structure of the optimal solution. Furthermore, for the onemax tree problem, we presented a population sizing model for the NetKey encoding and veriﬁed that the time to convergence depends linearly on the problem size. Direct representations do not use an explicit genotypephenotype mapping, but deﬁne search operators directly on the phenotypes. Therefore, in general, standard genetic operators can no longer be used and problemspeciﬁc operators for the phenotypes must be developed. In Chap. 7, we focused on the analysis and design of direct encodings. We developed search operators for a new direct encoding (NetDir) and performed an exhaustive investigation into the properties of the edgeset encoding. The investigation performed into the properties of the edgeset encoding illustrated that the framework for the design of representations can not only be used for indirect representations, but also for search operators and direct encodings. Finally, Chap. 8 illustrated how GEA users can use the provided representation framework for estimating the inﬂuence of diﬀerent representations on GEA performance. Based on the analysis of redundancy, scaling, and locality of tree representations, we compared GEAs performance for the onemax tree problem, the deceptive trap problem for trees, and various instances of the optimal communication spanning tree problem.
9.2 Conclusions
277
9.2 Conclusions We summarize the most important contributions of this work. Framework for design and analysis of representations (and operators) for GEAs. The main purpose of this study was to present a framework which describes how genetic representations inﬂuence the performance of GEAs. The performance of GEAs is measured by the solution quality at the end of the run and the number of generations until the population is converged. The proposed framework allows us to analyze the inﬂuence of existing representations on GEA performance and to develop eﬃcient new representations in a theoryguided way. Furthermore, we illustrated that the framework can also be used for the design and analysis of search operators, which are relevant for direct encodings. Based on the framework, the development of highquality representations remains not only a matter of intuition and random search but becomes an engineering design task. Even though more work is needed, we believe that the results presented are suﬃciently compelling to recommend increased use of the framework. Redundancy, Scaling, and Locality. These are the three elements of the proposed framework of representations. We demonstrated that these three properties of representations inﬂuence GEA performance and presented theoretical models to predict how solution quality and time to convergence changes. By examining the redundancy, scaling, and locality of an encoding, we are able to predict the inﬂuence of representations on GEA performance. The theoretical analysis shows that the redundancy of an encoding inﬂuences the supply of building blocks (BB) in the initial population. r denotes the number of genotypic BBs that represent the best phenotypic BB, and kr denotes the order of redundancy. For synonymously redundant encodings, where all genotypes that represent the same phenotype are similar to kr each other, the probability of GEA failure goes either with O(exp(−r/2 )) k (uniformly scaled representations) or with O(exp(− r/2 r )) (exponentially scaled representations). Therefore, GEA performance increases if the representation overrepresents highquality BBs. If a representation is uniformly redundant, that means each phenotype is represented by the same number of genotypes, GEA performance remains unchanged in comparison to nonredundant encodings. The analysis of the scaling of an encoding reveals that nonuniformly scaled representations modify the dynamics of genetic search. If exponentially scaled representations are used, the alleles are solved serially which increases the overall time until convergence and results in problems with genetic drift but allows rough approximations of the expected optimal solution after a few generations. We know from previous work that the high locality of an encoding is a necessary condition for eﬃcient mutationbased search. An encoding has high locality if neighboring phenotypes correspond to neighboring genotypes. Investigating the inﬂuence of locality shows that highlocality encodings do
278
9 Summary and Conclusions
not change the diﬃculty of a problem. In contrast, lowlocality encodings, where phenotypic neighbors do not correspond to genotypic neighbors, change problem diﬃculty and make, on average, easy problems more diﬃcult and deceptive problems easier. Therefore, to assure that an easy problem remains easy, highlocality representations are necessary. Population sizing and time to convergence models for redundant encodings, exponentially scaled encodings, and NetKeys. Based on a better understanding of redundancy and scaling, we were able to formulate population sizing models and time to convergence models for redundant and exponentially scaled encodings. The models show that for redundant encodings the population size grows with O(2kr /r) and the time to convergence goes with O(const − r/2kr k−1 ), where k denotes the order of building blocks. When using exponentially scaled encodings, we have to distinguish whether we want to consider genetic drift or not. When neglecting genetic drift, the population size N is independent of the length ls of an exponentially scaled BB, but depends scaled BBs √ √ only on the number m of competing exponentially (N = O ( m)). The time to convergence goes with O(ls m). To consider genetic drift, we developed two diﬀerent population sizing models (staircase drift model and approximated drift model) based on the model for the nondrift case. Due to genetic drift, the ability of GEAs to decide well between competing schemata decreases with increasing number of generations. Therefore, the probability of GEA failure increases with larger ls . Instead of using binary genotypes, the NetKey encoding uses continuous variables for representing trees. For NetKeys, we presented a population sizing model for the easy onemax tree problem and showed that the population size N goes with O(n1.5 ), where n denotes the number of nodes. The time to convergence is linearly increasing with O(n). Analysis of binary representations for integers. The framework can be used for explaining the performance diﬀerences of GEAs using diﬀerent types of binary representations for integer problems. The analysis of binary, Gray, and unary encoding has shown that the unary encoding is nonuniformly redundant, that the binary encoding is exponentially scaled, and the Gray and binary encoding have low locality. Therefore, the performance of GEAs using unary encoding depends on the structure of the optimal solution. If the optimal solution is underrepresented, GEAs fail; in contrast, if the optimal solution is overrepresented, GEAs using the unary encoding perform well. When using the binary encoding, all alleles are solved serially and the time to convergence increases. Therefore, some low salient genes are randomly ﬁxed due to genetic drift before they can be reached by the search process. Finally, binary and Gray encoding are lowlocality encodings and change the diﬃculty of the optimization problem. Thus, the resulting problem diﬃculty depends not only on the used representation but also on the considered optimization problem. Some problems like the easy integer onemax problem are easier when using the binary encoding, but there are other problems that are easier when using the Gray encoding.
9.2 Conclusions
279
Analysis of tree representations. The framework about representations can also be used for analyzing the inﬂuence of tree representations and operators on GEA performance. Based on the properties of Pr¨ ufer numbers, characteristic vectors, the link and node biased encoding, and edgesets the proposed framework allowed us to predict GEA performance. The analysis of the Pr¨ ufer number encoding revealed that the locality of the encoding is high around stars and low elsewhere. Therefore, GEA performance is low for easy problems if the optimal solution is not a star, and high if the problem at hand is deceptive and the optimal solution is a nonstar. The link and node biased (LNB) encoding uses a linkspeciﬁc and nodespeciﬁc bias to control the representations inﬂuence on the structure of the represented phenotype. The investigation into the properties of the encoding revealed that the encoding is synonymously redundant. Furthermore, it is uniformly redundant if a large linkspeciﬁc bias is used. If the linkspeciﬁc bias is small the encoding is nonuniformly redundant and biased towards the minimum spanning tree (MST). The use of a nodespeciﬁc bias results in an additional bias of the encoding towards starlike tree structures. Therefore, only GEAs using LNB encodings with a large linkspeciﬁc bias perform independently of the structure of the optimal solution. If the linkspeciﬁc bias is small, GEAs only perform well when searching for optimal MSTlike phenotypes. When using an additional nodespeciﬁc bias, GEAs also perform well when searching for optimal solutions similar to stars. Analyzing the characteristic vector encoding revealed that the encoding is redundant. Because invalid solutions are possible when using characteristic vectors, an additional repair mechanism is necessary which makes the representation nonsynonymously redundant. Due to the uniform redundancy of characteristic vectors, GEA performance is independent of the structure of the optimal solution. However, the repair mechanism results in nonsynonymous redundancy which has the same eﬀect as low locality. Therefore, GEA performance is reduced for easy problems and increased for deceptive problems. With increasing problem size, the repair process generates more and more links randomly and oﬀspring trees do not have much in common with their parents. Therefore, for larger problems, guided search is no longer possible and GEAs behave like random search. The edgeset encoding is an encoding which encodes trees directly by listing their edges. Search operators for edgesets may be heuristic when considering the weights of edges they include in the oﬀspring, or naive, including edges without regard to their weights. Analyzing the properties of the heuristic variants of the search operators reveals a strong bias towards the MST. Therefore, problems where the optimal solutions are diﬀerent from the MST could scarcely be solved. In contrast, the naive variants are unbiased which means genetic search is independent of the structure of the optimal solution. Although no explicit genotypephenotype mapping exists for edgesets and the framework for the design of representations can not be directly applied, it is useful for structuring the analysis of edgesets. Similarly to nonuniformly
280
9 Summary and Conclusions
redundant representations, edgesets overrepresent some speciﬁc types of trees and GEA performance increases if optimal solutions are similar to the MST. Design of new representations for trees. The framework about representations can not only be used for the analysis of existing representations but also for the development of new representations. Based on the insights into representation theory, the NetKey encoding has been proposed. The NetKey encoding is based on the characteristic vector encoding but uses continuous variables for encoding a tree. The encoding is synonymously and uniformly redundant as well as uniformly scaled. In contrast to other encodings, GEAs using NetKeys are able to distinguish between important and unimportant links. The high performance of GEAs using NetKeys was veriﬁed by empirical results and proposes an increasing use of this encoding. Furthermore, we presented the NetDir representation as an example of a theoryguided design of a direct encoding. The search operators for the NetDir representation are developed based on the notion of schemata. Scalable test problems for trees. Last but not least, we presented the fully easy onemax tree problem and the deceptive trap problem for trees as examples for scalable test problems. For these types of problems the optimal solution is determined a priori and the distance of an individual towards the optimal solution determines the ﬁtness of the individual. Both problems, the onemax tree problem and the deceptive trap problem for trees, are easy to implement and can be advantageously used for comparing the performance of GEAs on tree problems. By providing scalable test instances, the performance of diﬀerent types of representations or new, more eﬀective search methods can be analyzed and compared more easily.
A Optimal Communication Spanning Tree Test Instances
Searching the literature for standard test problems for the optimal communication spanning tree (OCST) problem reveals that many researchers use their private test problems which are mostly not published. As a result, the comparison of diﬀerent search algorithms or representations is a diﬃcult and time consuming task. It is not possible to quickly check if a new search method is better than the existing ones. Furthermore, applicants hesitate to use new and eﬃcient search methods or representations if they can not be tested on a variety of diﬀerent test problems and solve these problems well and reliably. Therefore, the building up of a collection of test instances for the OCST problem is necessary. The purpose of this appendix is to go one step in this direction and to present a collection of diﬀerent test instances for the OCST problem. It gives exact details concerning the properties of the problems we used in Sect. 8.2 for the comparison of diﬀerent types of representations. Based on the presented test instances, a fair and standardized comparison of new search techniques or representations becomes possible. For each test problem we present the best known solution, the demands, and the distance weights. The upper right corner of the presented matrices speciﬁes the demands and distance weights. Sect. A.1 summarizes the properties of the test instances from Palmer (1994). We are not able to present data for the 47 and 98 node problems because these are no longer available. Sect. A.2 presents the details for the 10, 20, and 50 node OCST problem from Raidl (2001), Sect. A.3 speciﬁes the berry6, berry35, and berry35u problems presented by Berry et al. (1995), and Sect. A.4 summarizes the speciﬁcations of four realworld test problems from Rothlauf et al. (2002).
A.1 Palmer’s Test Instances We present the details for palmer6, palmer12, and palmer24 test instances presented by Palmer (1994).
282
A Optimal Communication Spanning Tree Test Instances
Table A.1 gives an overview of the best known solutions for the three test problems. The cost of a link is calculated as the traﬃc over a link multiplied by the distance weight of this link (compare (8.4)). Table A.2 lists the demands and distance weights for the palmer6 problem. The properties of the palmer12 problem are speciﬁed in Table A.3 (demands) and Table A.4 (distance weights). Finally, the properties of the palmer24 problem are listed in Table A.5 (demands) and Table A.6 (distance weights). used links DETCHI, PHILDET, PHILNY, HOULA, HOUDET palmer12 3,428,509 SDPHNX, SDLA, SFSD, DETCHI, PHILBALT, PHILDET, PHILNY, DALSD, DALDET, HOUDAL, SANANDAL palmer24 1,086,656 LAPHNX, SDLA, SFLA, SJLA, CHILA, CHIDEN, DETCHI, DETIND, NYBOS, CLEVEDET, COLDET, PHILWDC, PHILJACK, PHILBALT, PHILDET, PHILNY, HOUNO, HOUDET, HOUMEMP, HOUDAL, HOUELPAS, SANANHOU, MILWCHI palmer6
c(Topt ) 693,180
(b) Distance matrix
(a) Demand matrix LA CHI DET NY PHIL HOU
Table A.1. Cost and structure of the best solutions to the test instances from Palmer (1994)
LA CHI DET NY PHIL HOU 0 1 1 1 1 2 0 10 3 4 3 0 5 6 2 0 31 2 0 2 0
LA CHI DET NY PHIL HOU
LA CHI DET NY PHIL HOU 0 16,661 18,083 21,561 21,099 13,461 0 5,658 9,194 8,797 10,440 0 7,230 6,899 11,340 0 4,300 13,730 0 13,130 0
Table A.2. Demand and distance matrix for palmer6
Table A.3. Demand matrix for palmer12 PHNX LA SD PHNX 0 7 8 LA 0 25 SD 0 SF CHI BALT DET NY PHIL DAL HOU SANAN 
SF CHI BALT DET NY PHIL DAL HOU SANAN 4 2 1 2 1 1 3 3 3 7 1 1 1 1 1 2 2 2 6 1 1 1 1 1 2 2 2 0 1 1 1 1 1 2 2 2 0 4 11 4 4 3 3 2 0 6 15 29 2 2 2 0 5 6 3 2 2 0 33 2 2 2 0 2 2 2 0 11 10 0 14 0
A.1 Palmer’s Test Instances
283
Table A.4. Distance matrix for palmer12 PHNX LA SD SF CHI BALT DET NY PHIL DAL HOU SANAN
PHNX LA SD SF CHI BALT DET 0 6,490 5,903 8,484 14,561 18,359 15,976 0 4,523 6,256 16,661 20,618 18,083 0 6,908 16,414 20,292 17,829 0 17,328 21,452 18,714 0 8,425 5,658 0 6,621 0 
NY 19,360 21,561 21,263 22,286 9,194 5,067 7,230 0 
PHIL 18,867 21,099 20,787 21,874 8,797 4,439 6,899 4,300 0 
DAL 10,090 12,639 12,073 14,234 9,603 12,385 10720 13,531 12,967 0 
HOU 10,883 13,461 12,802 15,259 10,440 12,526 11,340 13,730 13,130 4,888 0 
SANAN 9,665 12,236 11,540 14,136 11,237 13,722 12,297 14,912 14,319 5,076 4,478 0
PHNX LA SD SF SJ DEN WDC JACK CHI IND NO BOS BALT DET NY CLEVE COL PHIL MEMP DAL ELPAS HOU SANAN MILW
MILW
SANAN
HOU
MEMP DAL ELPAS
PHIL
COL
CLEVE
NY
DET
IND NO BOS BALT
CHI
DEN WDC JACK
SJ
SF
SD
LA
PHNX
Table A.5. Demand matrix for palmer24
0 100 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0  0 100 100 100 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0   0 0 0 100 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0    0 0 0 0 0 0 0 0 0 0 0 100 0 0 0 0 0 0     0 0 0 0 0 0 0 0 0 0 100 0 0 0 0 0 0      0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 100       0 0 0 0 100 0 0 0 0 0 0 0 0 0 0        0 0 0 0 0 0 0 0 0 0 0 100 0 0         0 0 0 100 0 0 0 0 0 0 0 0 0          0 0 0 0 0 100 0 0 0 0 0 0           0 0 100 100 2 0 0 0 2 0 0           0 0 0 2 0 0 0 0 0 0           0 0 0 0 0 0 0 0 0           0 0 0 0 0 0 0 0           0 0 0 0 0 0 0            0 0 0 100 0 0             0 0 100 0 0              0 100 0 0               0 100 0               0 0               0
284 A Optimal Communication Spanning Tree Test Instances
Table A.6. Distance matrix for palmer24 PHNX LA SD SF SJ DEN WDCJACK CHI IND NO BOS BALT DET NY CLEVE COL PHILMEMP DAL ELPASHOUSANANMILW PHNX 0 649590848824 808 1,810 1,700 1,4561,4891,3362,059 1,835 1,5971,936 1,674 1,588 1,886 1,289 1,009 602 1,088 966 1,461 LA 0 453626599 980 2,034 1,955 1,6611,7171,5952,275 2,068 1,8032,151 1,890 1,813 2,109 1,537 1,269 852 1,341 1,226 1,664 SD  0 698664 967 2,003 1,897 1,6441,6881,5392,241 2,022 1,7892,123 1,864 1,784 2,077 1,483 1,203 791 1,282 1,150 1,644 SF   0 2821,056 2,126 2,106 1,7381,7951,7582,339 2,142 1,8742,226 1,963 1,892 2,184 1,663 1,424 1,051 1,529 1,416 1,715 SJ    0 1,046 2,114 2,088 1,7201,7821,7422,323 2,136 1,8602,228 1,955 1,887 2,176 1,643 1,404 1,031 1,508 1,390 1,719 DEN    0 1,451 1,456 1,0681,1291,1541,674 1,471 1,2031,561 1,295 1,223 1,515 1,005 833 741 975 920 1,065 WDC    0 855 829 754 1,064 675 387 657 524 615 605 463 914 1,206 1,582 1,225 1,348 853 JACK    0 1,036 929 752 1,144 888 983 1,005 979 870 942 817 1,037 1,424 954 1,092 1,097 CHI    0 549 995 1,030 845 568 914 651 600 877 736 963 1,264 1,040 1,127 475 IND    0 919 993 771 567 875 618 528 821 667 932 1,280 995 1,089 592 NO    0 1,373 1,100 1,0471,226 1,074 958 1,161 621 675 1,058 577 714 1,056 BOS    0 655 823 530 813 8,571 585 1,203 1,494 1,851 1,511 1,631 1,038 BALT    0 661 507 633 627 449 940 1,235 1,617 1,256 1,372 862 DET    0 720 463 491 689 812 1,070 1,404 1,130 1,227 576 NY    0 702 722 430 1,068 1,351 1,727 1,370 1,492 934 CLEVE    0 493 660 858 1,123 1,472 1,170 1,279 663 COL    0 671 739 1,019 1,377 1,060 1,162 635 PHIL    0 1,000 1,297 1,663 1,310 1,439 896 MEMP    0 652 1,044 680 795 792 DAL    0 744 488 506 990 ELPAS    0 809 674 1,280 HOU    0 448 1,084 SANAN    0 1,166 MILW    0
A.2 Raidl’s Test Instances
285
A.2 Raidl’s Test Instances This section presents the details for the raidl10, raidl20, and raidl50 test problem proposed by Raidl (2001). We do not list the 75 and 100 nodes test problems herein because they are too extensive to be published. However, the details of the test instances are available and can be directly obtained from the author1 . All demands and distance weights between the nodes were generated randomly and are uniformly distributed. The cost of a link is calculated as the amount of traﬃc over a link multiplied by its distance weight (compare (8.2)). The nodes are labeled with numbers between zero and n − 1. Table A.7 presents the properties of the best known solutions for the raidl10 and raidl20 problems. The demands and distance weights of the raidl10 test problem are speciﬁed in Table A.8. The demands for the raidl20 and raidl50 test problem are shown in Tables A.9 and A.11. The corresponding distance weights can be found in Tables A.10 and A.12. Table A.7. Cost and structure of the best solutions for raidl10 and raidl20
c(Topt ) used links raidl10 53,674 10, 20, 30, 41, 50, 60, 73, 81, 91 raidl20 157,570 20, 75, 96, 97, 100, 110, 124, 130, 131, 133, 134, 1410, 162, 170, 1715, 188, 189, 1810, 1910
Table A.8. Demand and distance matrix for raidl10 (b) Distance matrix (a) Demand matrix 0 1 2 3 4 5 6 7 8 9
1
0 0 
1 34 0 
2 97 79 0 
3 50 65 11 0 
4 93 78 36 23 0 
5 100 81 87 88 80 0 
6 89 82 23 40 16 46 0 
7 24 66 78 91 47 84 53 0 
8 89 98 97 83 96 100 78 98 0 
9 3 72 81 84 9 0 66 58 13 0
0 1 2 3 4 5 6 7 8 9
0 0 
1 8 0 
2 17 47 0 
3 1 31 53 0 
4 41 17 36 53 0 
5 12 87 29 83 64 0 
6 7 59 47 72 39 63 0 
7 16 14 14 6 52 75 21 0 
8 90 5 18 79 16 47 45 89 0 
9 47 9 84 36 31 5 87 31 45 0
Address: G¨ unther Raidl, Institute of Computer Graphics, Vienna University of Technology, Favoritenstraße 911/1861, 1040 Vienna, Austria. EMail:
[email protected] 286
A Optimal Communication Spanning Tree Test Instances Table A.9. Demand matrix for raidl20 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 0 19 7 97 99 22 65 82 53 76 2 44 7 40 67 100 7 40 94 90  0 37 18 98 75 99 90 42 51 4 91 76 91 10 49 53 75 72 17  0 56 91 59 24 34 33 30 0 32 38 6 25 94 43 9 57 18  0 8 13 26 25 17 16 67 74 93 16 26 33 54 10 90 44  0 69 80 44 1 10 10 100 14 16 92 7 26 0 30 44  0 75 43 36 66 26 18 33 100 11 15 26 44 69 2  0 100 79 37 80 22 39 56 32 4 70 48 96 77  0 74 63 73 84 3 16 86 70 8 4 2 8  0 82 84 0 92 52 2 58 30 39 3 18  0 44 59 50 15 28 64 77 71 4 5  0 43 88 9 25 40 79 34 44 47  0 8 92 30 8 83 82 77 40  0 78 82 43 96 93 68 11  0 7 96 75 84 66 79  0 90 76 33 99 0  0 73 43 0 83  0 90 8 74  0 86 83  0 22  0
Table A.10. Distance matrix for raidl20 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 0 30 6 15 67 100 99 34 85 56 3 13 23 2 72 24 53 12 94 57  0 98 30 53 35 38 59 85 82 85 78 16 3 59 73 17 77 73 15  0 62 9 70 65 21 44 18 44 68 71 56 13 79 5 43 83 39  0 80 93 23 61 78 52 28 80 62 1 48 38 25 62 100 33  0 84 16 6 27 85 49 46 7 4 59 37 8 53 19 98  0 32 12 20 92 41 71 20 72 32 72 19 22 96 80  0 73 80 15 88 85 94 72 34 39 79 89 49 15  0 49 1 86 46 32 97 66 76 37 88 47 8  0 72 18 78 93 65 57 65 44 24 4 29  0 17 75 14 55 5 54 56 72 2 56  0 57 100 40 5 17 67 93 4 13  0 52 75 82 29 19 46 37 83  0 41 60 38 21 76 13 86  0 83 69 40 90 40 93  0 97 48 92 36 52  0 8 2 44 12  0 66 95 38  0 99 23  0 57  0
Table A.11. Demand matrix for raidl50 2 56 79 0 
3 24 55 86 0 
4 24 6 30 58 0 
5 29 16 73 14 25 0 
6 17 64 37 12 69 34 0 
7 9 73 61 10 79 17 1 0 
8 97 57 40 48 0 100 36 84 0 
9 47 23 72 75 73 82 9 97 37 0 
10 42 41 46 79 45 49 18 95 38 19 0 
11 5 80 78 70 36 7 60 31 94 55 22 0 
12 25 37 1 36 41 77 31 58 0 30 99 29 0 
13 22 72 10 68 60 22 95 61 83 40 30 89 52 0 
14 11 88 2 69 32 17 93 78 85 40 73 35 25 62 0 
15 59 45 55 40 49 91 8 41 14 41 6 84 36 74 90 0 
16 16 58 41 69 92 38 7 61 57 91 27 85 6 46 88 74 0 
17 97 19 73 63 77 98 70 2 55 53 89 100 58 41 51 80 74 0 
18 87 78 29 32 71 24 15 64 82 77 30 9 100 5 84 76 63 51 0 
19 26 40 78 20 46 37 51 0 38 96 8 45 21 85 89 32 27 46 84 0 
20 65 69 25 53 65 15 90 84 58 9 40 61 25 44 58 28 40 73 44 35 0 
21 7 18 68 97 51 73 34 45 95 65 60 48 10 43 29 23 14 70 13 91 19 0 
22 38 86 53 9 81 51 48 64 21 85 89 54 80 27 93 67 19 62 11 25 24 36 0 
23 9 31 68 4 66 24 62 78 19 7 30 33 32 34 82 52 15 14 24 12 60 57 56 0 
24 66 1 97 67 8 56 0 92 45 31 99 64 73 7 48 26 42 76 89 15 10 96 91 46 0 
25 74 35 2 36 29 77 80 48 50 16 26 49 37 46 68 9 39 51 28 13 3 26 76 95 94 0 
26 94 18 56 1 6 67 17 75 13 58 60 1 72 4 58 61 23 25 50 40 41 17 92 52 77 54 0 
27 2 38 62 49 44 78 33 4 69 82 95 45 43 24 3 20 45 86 74 31 91 73 47 12 92 87 41 0 
28 23 63 90 30 14 47 59 68 39 85 39 59 91 53 77 7 10 67 95 54 43 29 38 50 80 34 84 61 0 
29 14 65 32 42 50 78 22 81 91 29 43 7 98 76 9 3 50 50 66 34 38 24 17 8 26 27 49 86 50 0 
30 65 25 38 66 13 41 40 2 22 80 73 63 52 77 27 71 24 65 5 97 10 63 30 54 4 50 89 99 100 69 0 
31 42 54 34 79 67 43 82 13 89 91 50 57 71 90 99 61 80 20 14 25 44 73 10 7 40 43 37 86 40 70 11 0 
32 42 64 95 22 33 54 15 17 99 39 96 39 17 9 76 13 88 39 4 14 0 29 47 22 79 88 0 70 64 27 78 51 0 
33 87 70 93 55 18 18 98 62 66 72 2 3 7 85 51 70 90 42 13 67 35 73 21 92 16 93 60 46 40 50 16 66 34 0 
34 32 8 86 22 49 31 95 51 66 30 81 88 4 24 87 38 11 82 62 38 58 83 39 25 91 1 91 37 49 89 35 71 0 45 0 
35 66 37 24 81 14 73 46 63 51 97 43 60 86 55 10 65 62 47 51 76 67 29 86 52 87 33 87 50 41 21 42 67 22 3 66 0 
36 16 87 65 35 43 67 18 28 85 13 39 43 71 0 100 65 42 15 28 84 39 74 50 68 70 19 60 87 2 82 40 56 79 39 83 2 0 
37 49 39 89 0 17 32 31 58 2 17 52 14 34 58 11 8 73 93 54 33 0 40 11 38 98 5 85 88 54 43 75 59 86 6 17 96 19 0 
38 75 9 52 91 59 90 13 90 5 11 99 20 8 89 86 53 45 4 21 29 50 97 68 39 8 73 36 99 68 51 94 23 7 23 39 84 90 63 0 
39 79 43 4 83 43 67 20 11 32 10 32 55 69 35 96 96 58 88 83 4 72 12 45 6 62 98 2 76 88 85 84 96 94 86 42 8 5 18 37 0 
40 97 62 27 41 56 13 6 54 6 52 41 34 46 48 71 84 73 17 71 15 29 40 85 90 89 21 38 24 52 59 79 54 65 27 48 86 68 88 57 78 0 
41 83 50 98 36 71 5 32 50 9 30 30 69 16 58 88 50 25 30 52 67 21 12 74 55 27 29 29 66 54 50 86 81 41 40 81 35 73 74 21 14 78 0 
42 50 22 82 18 45 40 80 8 92 65 21 98 33 75 31 34 60 49 31 22 53 51 7 84 96 50 70 35 23 91 18 21 97 44 39 1 45 21 53 29 98 52 0 
43 21 66 28 77 63 56 37 78 73 48 40 63 68 97 76 14 96 71 7 46 62 36 63 23 93 57 71 15 99 89 49 16 66 92 13 25 9 49 81 69 34 15 100 0 
44 71 21 74 3 97 27 93 10 92 37 27 57 68 94 98 1 87 31 49 74 9 100 81 0 32 94 14 52 57 90 13 59 23 39 3 80 20 57 61 63 26 50 67 79 0 
45 62 10 51 53 77 58 72 86 76 71 94 99 24 36 72 22 0 79 10 37 100 3 20 34 2 59 41 95 74 39 34 99 8 3 47 60 92 96 72 96 54 74 77 19 88 0 
46 47 10 49 84 11 13 45 18 88 90 46 46 70 70 31 62 30 61 68 22 2 98 13 63 82 85 9 66 51 29 37 100 83 99 63 30 48 34 36 96 30 34 7 98 53 4 0 
47 87 45 14 38 89 65 67 37 15 27 54 41 40 5 86 41 24 37 86 8 47 9 83 74 88 82 11 55 44 58 0 60 89 44 92 49 0 92 67 24 27 97 44 74 85 84 77 0 
48 24 96 90 15 53 55 7 55 31 90 49 65 48 44 22 69 40 71 47 90 21 69 19 98 52 52 62 63 38 60 15 7 79 3 26 24 61 98 5 24 11 34 57 9 88 6 24 93 0 
49 100 88 44 15 83 37 61 48 69 66 76 21 35 41 81 37 93 78 11 70 10 100 84 9 71 81 39 3 26 97 46 39 16 65 45 74 96 25 8 13 94 72 2 88 80 83 64 98 98 0
287
1 62 0 
A.2 Raidl’s Test Instances
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
0 0 
17 73 64 23 9 34 95 13 57 21 26 84 4 64 32 27 49 70 0 
18 91 30 94 90 67 80 36 68 6 58 46 54 60 54 1 7 59 47 0 
19 32 41 34 66 28 78 58 29 84 9 98 61 91 22 31 70 6 63 76 0 
20 27 38 28 98 28 99 83 17 61 75 9 52 79 51 92 1 50 24 46 44 0 
21 93 98 21 64 74 95 59 63 16 77 7 33 26 44 55 77 65 24 65 57 9 0 
22 37 87 89 61 97 7 36 43 65 26 51 84 86 15 15 81 26 65 40 57 17 43 0 
23 83 56 70 76 72 17 1 67 25 26 42 28 99 85 52 7 3 30 27 35 79 22 92 0 
24 85 82 56 6 29 78 92 65 18 29 73 70 68 65 70 21 6 76 41 36 17 25 60 65 0 
25 8 92 97 62 70 46 2 24 96 33 8 40 20 3 60 1 1 23 14 49 3 11 64 1 33 0 
26 3 6 34 48 73 67 94 44 53 34 35 64 39 85 55 90 7 20 98 48 14 1 34 28 15 47 0 
27 82 85 94 96 29 8 7 57 25 99 97 15 42 9 93 76 39 49 41 33 65 41 33 12 71 83 78 0 
28 90 61 26 29 30 100 52 74 59 26 79 44 34 64 21 5 21 96 61 89 50 13 40 44 64 3 59 27 0 
29 45 7 63 73 76 9 84 88 98 9 72 20 22 94 83 33 74 96 93 8 99 66 45 39 35 57 91 76 45 0 
30 14 91 2 29 51 80 89 41 19 54 24 91 36 52 56 69 77 32 72 78 5 57 85 27 35 48 32 95 4 51 0 
31 92 53 58 13 93 18 8 11 66 82 38 34 49 23 52 68 84 83 43 12 9 14 32 78 17 5 74 14 59 42 86 0 
32 15 98 7 48 82 71 95 82 51 67 57 89 45 61 92 39 35 33 26 3 76 64 9 72 62 42 9 99 98 80 14 52 0 
33 85 80 87 84 53 72 62 65 53 5 37 75 10 77 93 12 86 45 68 3 16 61 51 18 46 14 42 77 43 88 38 44 56 0 
34 47 24 34 43 64 79 51 51 8 94 45 63 10 12 12 10 61 79 73 80 11 75 40 74 60 75 20 84 5 43 99 60 47 1 0 
35 60 33 89 8 77 67 3 79 9 25 5 31 43 32 54 10 45 12 89 75 30 91 75 56 1 5 91 62 12 57 70 93 100 81 78 0 
36 86 6 22 70 19 81 82 77 19 72 5 95 8 10 92 78 79 30 98 43 95 76 66 49 24 48 96 6 88 53 28 92 62 27 57 35 0 
37 23 28 96 45 84 15 25 56 42 85 77 63 68 38 42 10 34 98 70 29 38 37 100 34 37 10 28 42 82 86 25 91 77 49 18 70 54 0 
38 69 4 27 4 88 41 41 64 24 28 39 24 74 61 20 94 68 62 50 45 72 20 49 58 24 21 90 4 90 86 30 75 13 66 91 45 31 33 0 
39 95 91 91 73 98 16 99 46 4 8 70 22 90 63 67 22 99 16 65 92 75 23 57 88 41 61 100 37 64 39 12 97 60 23 16 84 68 20 36 0 
40 19 35 50 2 79 51 45 43 56 34 33 62 20 99 49 72 86 87 9 45 34 74 27 8 62 7 69 67 28 82 6 85 87 31 56 35 62 84 51 33 0 
41 33 8 73 1 55 39 84 8 6 14 30 47 82 79 84 21 90 6 7 5 15 43 85 24 31 81 3 12 55 89 79 97 14 33 56 67 100 51 66 92 67 0 
42 6 65 91 61 91 31 85 2 52 56 43 99 74 67 45 1 44 11 27 99 19 49 28 39 89 13 27 31 14 49 62 83 1 5 69 66 10 7 98 81 81 63 0 
43 98 39 15 52 73 57 42 34 61 49 68 6 77 24 62 71 44 51 59 71 38 7 1 56 96 31 74 86 4 80 47 68 5 79 60 67 77 53 93 49 64 26 80 0 
44 51 78 74 71 18 60 73 42 7 16 49 31 2 8 28 94 13 49 56 63 13 58 11 32 89 69 2 2 39 83 10 82 67 42 22 23 55 95 56 89 20 14 58 90 0 
45 69 30 34 11 11 74 41 7 16 80 41 30 10 13 28 58 92 12 90 6 41 68 23 65 28 88 88 26 27 53 1 96 79 36 47 44 65 90 3 47 27 80 40 20 86 0 
46 50 57 85 55 92 44 98 67 92 82 28 53 19 79 48 23 2 92 3 60 52 96 50 93 3 23 46 66 61 43 42 58 56 95 67 61 97 87 23 87 5 75 54 85 21 25 0 
47 7 33 75 2 10 9 13 83 3 14 48 57 5 62 74 49 42 62 69 65 18 22 17 11 64 31 14 91 32 71 67 80 10 79 83 11 66 13 23 84 58 6 42 76 99 54 65 0 
48 41 93 56 20 96 20 53 15 36 57 24 85 25 44 6 7 13 71 60 74 100 60 80 66 66 19 94 78 30 87 54 65 39 44 37 91 38 7 64 31 39 35 32 55 49 50 64 30 0 
49 92 86 78 96 58 75 34 34 97 92 26 6 30 84 1 30 60 90 92 71 17 48 69 3 11 12 26 34 50 84 27 37 100 41 67 39 96 6 75 58 12 77 87 67 92 15 93 44 67 0
A Optimal Communication Spanning Tree Test Instances
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
16 4 55 50 80 92 97 89 41 45 98 48 26 52 55 95 4 0 
288
Table A.12. Distance matrix for raidl50 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0 71 20 56 94 50 76 57 57 45 65 60 80 19 64 100  0 90 19 36 26 53 20 34 8 53 23 52 19 66 19   0 25 31 36 11 38 17 54 95 1 66 1 91 19    0 9 35 51 97 4 6 45 37 51 71 51 52     0 67 69 11 80 96 16 41 95 12 21 68      0 37 75 83 85 98 64 6 70 44 88       0 86 18 69 44 87 98 89 5 5        0 70 88 92 5 40 83 6 33         0 63 31 48 5 50 13 80          0 87 49 64 3 73 34           0 74 84 68 99 61            0 45 85 32 95             0 69 37 90              0 95 4               0 35               0                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             
A.3 Berry’s Test Instances
289
A.3 Berry’s Test Instances This section presents the details for the berry6, berry35, and berry35u problem instances proposed by Berry et al. (1995). Table A.13 presents the properties of the best known solutions to the diﬀerent problem instances. The demands for the test problems are presented in Tables A.14(a) (berry6) and A.15(a) (berry35 and berry35u). The distance weights are shown in Tables A.14(b) (berry6) and A.15(b) (berry35). For berry35u, the distance weights dij = 1, for i, j ∈ {0, . . . , 34}. The demands are the same for berry35 and berry35u. c(Topt ) used links berry6 534 10, 31, 52, 53, 54 berry35u 16,273 10, 82, 114, 129, 1210,
138, 159, 162,
173, 186, 1915, 209, 218, 251, 253, 2519, 2524, 2622, 2715, 289, 298, 2911, 2925, 305, 3014, 3021, 3022, 317, 3112, 3123, 3218, 3225, 3325, 3429
berry35 16,915
Table A.13. Cost and structure of the best solutions for berry6, berry35u, and berry35
10, 82, 114, 129, 1210, 138, 162, 173, 186, 1915, 209, 2015, 218, 241, 2417, 2418, 253, 258, 2519, 2622, 2715, 289, 2911, 2925, 305, 3014, 3021, 3022, 317, 3112, 3123, 3218, 3324, 3429
Table A.14. Demand and distance matrix for berry6 (a) Demand matrix
(b) Distance matrix
0 1 2 3 4 5 0 0 5 13 12 8 9 1  0 7 4 2 6 2   0 3 10 15 3   0 11 7 4   0 12 5   0
0 1 2 3 4 5 0 0 3 6 5 9 7 1  0 3 2 4 8 2   0 3 7 2 3    0 9 2 4     0 1 5      0
290
A Optimal Communication Spanning Tree Test Instances Table A.15. Demand and distance matrix for berry35 and berry35u (a) Demand matrix for berry35 and berry35u
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 0 639 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 93 0 0 0 0 0 0 0 129 0  0 0000 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 147 0 0 0 0 0 0 0 83 0 0   0 0 0 0 0 0 189 0 0 0 0 0 0 0 99 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 43 0 0    0 0 0 0 0 0 0 0 0 0 0 0 0 0 531 0 0 0 0 0 0 0 623 0 0 0 0 0 0 0 0 0     0 0 0 0 0 0 0 53 0 0 0 0 0 0 0 39 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0      0 43 0 0 0 0 0 0 0 119 0 0 0 0 0 0 0 329 0 0 0 0 0 0 0 651 0 0 0 0       0 0 0 0 0 0 0 0 0 0 0 0 371 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0        0 0 0 0 0 0 0 0 23 0 0 0 0 0 0 0 9 0 0 0 0 0 0 0 171 0 0 0         0 0 0 0 0 41 0 0 0 0 0 0 0 189 0 0 0 0 0 0 0 123 0 0 0 0 0          0 0 0 351 0 0 0 0 0 0 0 61 0 0 0 0 0 0 0 217 0 0 0 0 0 0           0 0 81 0 0 0 0 0 0 0 11 0 0 0 0 0 0 0 0 0 0 0 0 0 0            0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 133 0 0 0 0 0             0 0 0 27 0 0 0 0 0 0 0 161 0 0 0 0 0 0 0 261 0 0 0              0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0               0 0 0 0 0 0 0 0 261 0 0 0 0 0 0 0 639 0 0 0 0                0 0 0 0 147 0 0 0 0 0 0 0 423 0 0 0 0 0 0 0                 0 0 0 0 0 0 0 0 0 69 0 0 0 0 0 0 0 0 0                  0 0 0 0 0 0 0 351 0 0 0 0 0 0 0 117 0 0                  0 0 0 0 0 0 243 0 0 0 0 0 0 0 873 0 0                  0 0 0 0 0 0 639 0 0 0 0 0 0 0 119 0                   0 0 0 0 0 0 0 57 0 0 0 0 0 0 0                    0 91 0 0 0 0 0 0 0 387 0 0 0 0                    0 0 0 0 89 0 0 0 0 0 0 0 0                    0 0 0 0 0 0 0 0 651 0 0 0                    0 0 0 0 0 133 0 0 0 0 0                    0 0 0 21 0 0 0 0 0 0                     0 0 0 0 0 0 0 0 0                      0 0 111 0 0 0 0 0                      0 0 0 0 0 0 0                      0 0 0 0 0 63                      0 0 0 0 0                      0 0 0 0                      0 71 0                      0 0                       0
(b) Distance matrix for berry35 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
0 0 
1 1 0 
2 7 6 0 
3 4 3 3 0 
4 8 7 5 4 0 
5 9 8 4 5 7 0 
6 4 3 7 4 8 9 0 
7 12 11 9 8 10 11 12 0 
8 6 5 1 2 4 3 6 8 0 
9 9 8 6 5 7 8 9 3 5 0 
10 11 10 8 7 9 10 11 3 7 2 0 
11 7 6 4 3 1 6 7 9 3 6 8 0 
12 10 9 7 6 8 9 10 2 6 1 1 7 0 
13 7 6 2 3 5 4 7 9 1 6 8 4 7 0 
14 9 8 4 5 7 2 9 11 3 8 10 6 9 4 0 
15 7 6 4 3 5 6 7 5 3 2 4 4 3 4 6 0 
16 8 7 1 4 6 5 8 10 2 7 9 5 8 3 5 5 0 
17 3 2 4 1 5 6 3 9 3 6 8 4 7 4 6 4 5 0 
18 3 2 6 3 7 8 1 11 5 8 10 6 9 6 8 6 7 2 0 
19 6 5 3 2 4 5 6 6 2 3 5 3 4 3 5 1 4 3 5 0 
20 8 7 5 4 6 7 8 4 4 1 3 5 2 5 7 1 6 5 7 2 0 
21 7 6 2 3 5 2 7 9 1 6 8 4 7 2 2 4 3 4 6 3 5 0 
22 10 9 5 6 8 3 10 12 4 9 11 7 10 5 3 7 6 7 9 6 8 3 0 
23 12 11 9 8 10 11 12 2 8 3 3 9 2 9 11 5 10 9 11 6 4 9 12 0 
24 2 1 5 2 6 7 2 10 4 7 9 5 8 5 7 5 6 1 1 4 6 5 8 10 0 
25 5 4 2 1 3 4 5 7 1 4 6 2 5 2 4 2 3 2 4 1 3 2 5 7 3 0 
26 9 8 4 5 7 2 9 11 3 8 10 6 9 4 2 6 5 6 8 5 7 2 1 11 7 4 0 
27 8 7 5 4 6 7 8 6 4 3 5 5 4 5 7 1 6 5 7 2 2 5 8 6 6 3 7 0 
28 10 9 7 6 8 9 10 4 6 1 3 7 2 7 9 3 8 7 9 4 2 7 10 4 8 5 9 4 0 
29 6 5 3 2 2 5 6 8 2 5 7 1 6 3 5 3 4 3 5 2 4 3 6 8 4 1 5 4 6 0 
30 8 7 3 4 6 1 8 10 2 7 9 5 8 3 1 5 4 5 7 4 6 1 2 10 6 3 1 6 8 4 0 
31 11 10 8 7 9 10 11 1 7 2 2 8 1 8 10 4 9 8 10 5 3 8 11 1 9 6 10 5 3 7 9 0 
32 4 3 7 4 8 9 2 12 6 9 11 7 10 7 9 7 8 3 1 6 8 7 10 12 2 5 9 8 10 6 8 11 0 
33 3 2 6 3 7 8 3 11 5 8 10 6 9 6 8 6 7 2 2 5 7 6 9 11 1 4 8 7 9 5 7 10 3 0 
34 7 6 4 3 3 6 7 9 3 6 8 2 7 4 6 4 5 4 6 3 5 4 7 9 5 2 6 5 7 1 5 8 7 6 0
A.4 Real World Problems
291
A.4 Real World Problems This section presents the properties of four realworld problems. The presented problems are no classical OCST problems as the cost of a link depends nonlinearly on its distance weights dij and the traﬃc bij running over the link. For fulﬁlling the demands between the nodes, diﬀerent line types with only discrete capacities are available. The cost of installing a link consists of a ﬁxed and length dependent share. Both depend on the capacity of the link. The costs are based on the tariﬀs of the German Telekom from 1996 and represent the amount of money (in German Marks) a company has to pay for a telecommunication line of a speciﬁc length and capacity per month. For a detailed description of the four diﬀerent problems the reader is referred to Sect. 8.2.3. In particular, the overall cost of a communication network is calculated as f (dij , capij ), c(T ) = i,j∈F
where F denotes the set of used links, dij are the distance weights of the links between node i to node j, and capij is the capacity of the links. The distance weight dij corresponds to the Euclidean distance between the nodes i and j. The capacity capij of a link must be higher than the overall traﬃc bij running over a link. Therefore, capij ≥ bij , where bij denotes the overall traﬃc over the link between the nodes i and j. This means that to every link between i and j a line is assigned with the next higher available capacity capij . We illustrate this with a brief example. If there are three line types available with capacity 64 kBit/s, 512 kBit/s, and 2048 kBit/s, a line with capacity cap = 64 kBit/s is assigned to all links with less than b = 64 kBit/s of traﬃc. If the traﬃc over a link is between 64 kBit/s and 512 kBit/s the 512 kBit/s line is chosen. If the traﬃc over a link exceeds 512 kBit/s the 2048 kBit/s line must be chosen. Table A.16 and Fig. A.1 present the properties of the best known solutions to the four test problems. Table A.17(a) (rothlauf1 and rothlauf2), Table A.17(b) (rothlauf3), and Table A.18 (rothlauf4) illustrate how the cost of a link depends on the overall traﬃc bij and the distance weight dij of the used line. The largest available line type has capacity cap = 2, 048 kBit/s (rothlauf13) or cap = 4, 096 kBit/s (rothlauf4). If the traﬃc b over a link exceeds this value a large penalty is used. In Table A.19 (rothlauf1 and rothlauf3), Table A.20 (rothlauf2), and Table A.21 (rothlauf4) we present the demands for the diﬀerent test problems. Table A.22 lists thecoordinates of the nodes. The distance weights di,j are calculated as dij = (xi − xj )2 + (yi − yj )2 , where x and y denote the coordinates of the nodes. To get the distances and coordinates in kilometer, the distance weights must be multiplied by 14.87. The factor 14.87 results from the used “Geb¨ uhrenfeldverfahren” of the German Telekom.
292
A Optimal Communication Spanning Tree Test Instances
c(Topt ) used links rothlauf1 60,883 31, 41, 42, 51, 61, 71, 81, 93, 114, 1110, 124, 132, 1411, 154, 166 rothlauf2 58,619 51, 61, 71, 83, 101, 102, 103, 109, 114, 122, 1310, 144, 1410, 155 rothlauf3 28,451 51, 54, 61, 71, 81, 91, 93, 101, 1110, 124, 132, 1410, 152, 154, 166 rothlauf4 112,938 21, 31, 71, 75, 76, 81, 101, 111, 121, 124, 1310, 149, 1410, 151, 167 cost: 60883.71
13
3 10
14
9
4 11
5 16
6
3
10
14
9
13
8
4 12
7
10
(a) rothlauf1
14
9
16
6
4 12
3 10
14
9
5 16
1
6
1 7 8
8
7
(b) rothlauf2
11
15
7
6 8
3
5
1
15
11
15
5 1
2
2
2
11
13
13
12
2
4 12
cost: 112938.45
cost: 28451.76
cost: 58619.43
15
Table A.16. Cost and structure of the best solutions for selected realworld test instances
(c) rothlauf3
(d) rothlauf4
Figure A.1. Best known solutions Topt for the four realworld problem instances Table A.17. Cost of a link for rothlauf1, rothlauf2, and rothlauf3 (a) rothlauf1 and rothlauf2 bij
dij cost [0; 1] 334.58d + 385 ]1; 3] 148.70d + 572