CONTENTS
1. Introduction
1.1 The Problem Considered
1.2 Practical Relevance
1.3 Previous Work
1.3.1 Neural Networks
1.3.2 Genetic Algorithms
1.3.3 Structure Generation
1.4 Objectives
1.5 Working Environment
1.6 Outline of Remaining Report
2. Theory
2.1 Expert Network
2.1.1 Architecture
2.1.2 Training
2.1.3 Structural Operations
2.2 Modular Network
2.2.1 Architecture
2.2.2 Training
2.3 Genetic Algorithm
2.3.1 General Method
2.3.2 Fitness Evaluator
2.3.3 Selection Algorithm
2.3.4 Crossover Operation
2.3.5 Mutation Algorithm
2.4 Modular Breeder
2.4.1 Integration Aims
2.4.2 The Algorithm
3. Design
3.1 Requirements
3.1.1 Flexibility
3.1.2 Ease of Testing
3.1.3 Ease of Extending
3.2 Choice of Language
3.3 Considerations
3.4 Design of the Systems
3.4.1 Network class
3.4.2 Neural Network
3.4.3 Modular Network
3.4.4 Training of the Neural Network Systems
3.4.5 Genetic Algorithm
3.4.6 Modular Breeder
3.4.7 Monitors
3.4.8 Testers
3.4.9 Summary of Design
3.5 Interface
3.6 Packages
4. Testing
4.1 Choice of Task
4.2 What-Where Vision Task
4.2.1 The Task
4.2.2 The Suitable Structures
4.2.3 Crosstalk
4.3 Results
4.3.1 Neural Network Tests
4.3.2 Modular Network Test
4.3.3 Genetic Algorithm Test
4.3.4 Modular Breeder
5. Conclusions
5.1 Achievements
5.1.1 Sub-systems
5.1.2 Modular network breeder
5.1.3 Interface
5.2 Criticisms
5.3 Future Work
5.3.1 Improvements
5.3.2 Testing
5.3.3 Extensions
Appendix A: Public class and method
listing
Appendix B: Sample code output
Appendix C: Test Parameters