Complex Behavior in Evolutionary Robotics

✍ Scribed by Lukas König

• Publisher: De Gruyter Oldenbourg
• Year: 2015
• Tongue: English
• Leaves: 262
• Category: Library

✦ Synopsis

Today, autonomous robots are used in a rather limited range of applications such as exploration of inaccessible locations, cleaning floors, mowing lawns etc. However, ongoing hardware improvements (and human fantasy) steadily reveal new robotic applications of significantly higher sophistication. For such applications, the crucial bottleneck in the engineering process tends to shift from physical boundaries to controller generation. As an attempt to automatize this process, Evolutionary Robotics has successfully been used to generate robotic controllers of various types. However, a major challenge of the field remains the evolution of truly complex behavior. Furthermore, automatically created controllers often lack analyzability which makes them useless for safety-critical applications. In this book, a simple controller model based on Finite State Machines is proposed which allows a straightforward analysis of evolved behaviors. To increase the model's evolvability, a procedure is introduced which, by adapting the genotype-phenotype mapping at runtime, efficiently traverses both the behavioral search space as well as (recursively) the search space of genotype-phenotype mappings. Furthermore, a data-driven mathematical framework is proposed which can be used to calculate the expected success of evolution in complex environments.

• The author is involved in a large project of the German Research Foundation (DFG) devoted to the topic
• The presented research findings are already being used internationally (e.g. in Robotic Gait Orthosis at ETH Zurich and in South Korea)
  

✦ Table of Contents

Acknowledgements
Contents
List of Figures
List of Tables
List of Notations
1 Introduction
1.1 Evolutionary Robotics and Evolutionary Swarm Robotics
1.2 Further Classifications
1.3 Challenges of ER
1.4 Structure and Major Contributions of the Thesis
2 Robotics, Evolution and Simulation
2.1 Evolutionary Training of Robot Controllers
2.1.1 Two Views on Selection in ER and ESR
2.1.2 Classification of Fitness Functions in ER
2.1.3 The Bootstrap Problem
2.1.4 The Reality Gap
2.1.5 Decentralized Online Evolution in ESR
2.1.6 Evolvability, Controller Representation and the Genotype-Phenotype Mapping
2.1.7 Controller Representation
2.1.8 Recombination Operators
2.1.9 Success Prediction in ESR
2.2 Agent-based Simulation
3 The Easy Agent Simulation
3.1 History of the Easy Agent Simulation Framework
3.2 Basic Idea and Architectural Concept
3.2.1 Overview
3.2.2 Preliminaries
3.2.3 Classification of the Architecture
3.2.4 The SPI Architecture from an MVC Perspective
3.2.5 Comparison of the SPI Architecture with State-of-the-Art ABS Frameworks
3.3 Implementation of the SPI within the EAS Framework
3.3.1 Overview
3.3.2 Plugins
3.3.3 Master Schedulers
3.3.4 The classes SimulationTime and Wink
3.3.5 The Interface EASRunnable
3.3.6 “Everything is an Agent”: a Philosophical Decision
3.3.7 Running a Simulation
3.3.8 Getting Started
3.4 A Comparative Study and Evaluation of the EAS Framework
3.4.1 Method of Experimentation
3.4.2 Results and Discussion
3.5 Chapter Résumé
4 Evolution Using Finite State Machines
4.1 Theoretical Foundations
4.1.1 Preliminaries
4.1.2 Definition of the MARB Controller Model
4.1.3 Encoding MARBs
4.1.4 Mutation and Hardening
4.1.5 Selection and Recombination
4.1.6 Fitness calculation
4.1.7 The Memory Genome: a Decentralized Elitist Strategy
4.1.8 Fitness Adjustment after Mutation, Recombination and Reactivation of the Memory Genome
4.1.9 The Robot Platforms
4.2 Preliminary Parameter Adjustment using the Example of Collision Avoidance
4.2.1 Specification of Evolutionary Parameters
4.2.2 Method of Experimentation
4.2.3 Evaluation and Discussion
4.2.4 Concluding Remarks
4.3 A Comprehensive Study Using the Examples of Collision Avoidance and Gate Passing
4.3.1 Method of Experimentation
4.3.2 Experimental results
4.3.3 Concluding remarks
4.4 Experiments With Real Robots
4.4.1 Evolutionary Model
4.4.2 Method of Experimentation
4.4.3 Results and Discussion
4.4.4 Concluding Remarks
4.5 Chapter Résumé
5 Evolution and the Genotype-Phenotype Mapping
5.1 Overview of the Presented Approach
5.2 A Completely Evolvable Genotype-Phenotype Mapping
5.2.1 Definition of (complete) evolvability
5.2.2 Properties of ceGPM-based genotypic encodings
5.2.3 The Translator Model MAPT and the Course of Evolution
5.2.4 Genotypic and Phenotypic Spaces
5.2.5 Evolutionary Operators
5.3 Evaluation of the Proposed Evolutionary Model
5.3.1 First Part – Method of Experimentation
5.3.2 First Part – Results and Discussion
5.3.3 Second Part – An Alternate Completely Evolvable Genotype-Phenotype Mapping and its Effects on Evolvability
5.3.4 Second Part – Method of Experimentation
5.3.5 Second Part – Results and Discussion
5.3.6 Third Part – Method of Experimentation
5.3.7 Third Part – Results and Discussion
5.4 Chapter Résumé
6 Data Driven Success Prediction of Evolution in Complex Environments
6.1 Preliminaries
6.2 A Model Capturing Completely Implicit Selection
6.2.1 Two Parents per Reproduction (CIS-2)
6.2.2 Eventually Stable States (k=2)
6.2.3 Tournament size k
6.2.4 Eventually Stable States (arbitrary k)
6.3 Extending the CIS Model to Capture Explicit Selection
6.4 Experiments
6.4.1 Evolutionary setup
6.4.2 Experimental Results Using the EIS Model
6.4.3 Remarks on Evolution in the scope of the CIS Model
6.5 Chapter Résumé
7 Conclusion
References
Index