กก

English translation of the abstract and contents of Lernen durch Genetisch-Neuronale Evolution: Aktive Anpassung an Unbekannte Umgebungen mit Selbstentwickelnden Parallelen Netzwerken, ISBN 3-929037-16-6, 268pp, Infix-Verlag, St. Augustin/Bonn, July 1992.

Learning by Genetic Neural Evolution:
Active Adaptation to Unknown Environments
with Self-developing Parallel Networks

Byoung-Tak Zhang

(C) Infix-Verlag, July 1992

กก

Abstract

Artificial neural networks possess, in contrast to symbolic systems, several advantageous properties, such as massive parallelism and fault tolerance. Among various network architectures, the multilayer neural nets are of special importance since they can theoretically realize any arbitrary bounded continuous functions. Previous studies on learning in such networks were usually restricted to the adaptation of connection weights. From the perspectives of artificial intelligence and the theory of system identification, such learning methods have at least two deficiencies. First, their application area is very limited. If the given network architecture is not suitable for the problem, the learning converges very slowly or does not converge at all. Second, this learning paradigm is too passive in the sense that it can model only the systems that are predefined by the given training sets. Networks should, however, be able to actively explore unknown or continuously changing environments and adapt themselves to them.

This work presents an autonomous learning method for neural networks, called GENIAL (Genetic Neural Incremental Autonomous Learning). GENIAL starts with a small training set and a network structure with a single hidden neuron. The learning process consists, on the one hand, in the active expansion of the training set through genetic exploration of the environment on the basis of the network knowledge. On the other hand, GENIAL constructs a neural network model of its environment through adaptation of the structure and weights on the basis of the expanded training set.
The efficiency of these self-developing parallel networks with respect to the learning speed and the generalization performance is analyzed and tested on various function approximation tasks. Practical applicability of the learning method was demonstrated on writer-independent digit recognition and robot control. The experimental results confirm that the GENIAL learning method, confronted with an unknown environment, acquires incrementally and actively new knowledge and thereby builds an effective neural network model of the environment. For the tasks for which a large number of training examples are available, the method finds a problem-specific network structure using a selected subset of the given network. In spite of the computational overhead in network structure optimization the self-developing networks can converge faster than the usual backpropagation networks with pre-optimized structures.

Contents

 
Abstract
Acknowledgements
    i
    ii
  1. Introduction
    1.1 Neural Networks and Artificial Intelligence
    1.1.1 Neural Networks for Artiicial Intelligence
    1.1.2 Artificial Intelligence for Neural Networks
    1.2 Requirements of Learning Methods for Neural Networks
    1.2.1 Construction of Neural Systems
    1.2.2 Quality Criteria for Learning Methods
    1.3 Goals of the Work
    1.3.1 Creativity
    1.3.2 Selectivity
    1.3.3 Adaptivity
    1.3.4 Identification of Black-Box Environments
    1.4 Dissertation Overview

  2. Learning Methods for Neural Networks
    2.1 Neurons
    2.2 A Taxonomy of Neural Nets
    2.3 Some Network Architectures and Learning Procedures
    2.3.1 Perceptrons
    2.3.2 Self-organizing Maps
    2.3.3 Learning Automata
    2.3.4 Relaxation Nets
    2.3.5 Multilayer Nets
    2.3.6 Simulated Annealing and Boltzmann Machines
    2.4 Current Research Issues

  3. The GENIAL Learning Model
    3.1 The Environments
    3.2 A Taxonomy of Associative Learning Methods
    3.2.1 Learning Principles
    3.2.2 Learning Mechanisms
    3.2.3 The 6 Learning Types
    3.2.4 Passive Learning Paradigm
    3.2.5 Active Learning Paradigm
    3.3 Genetic Neural Evolutionary Learning
    3.3.1 The Network
    3.3.2 Genetic Learning
    3.3.3 Neural Learning
    3.3.4 GENIAL Control Algorithm
    3.4 GENIE: An Environment for Developing Artificial Neural Systems
    3.4.1 System Architecture
    3.4.2 Operation Modes

  4. Efficient Gradient Search by Focused Propagation
    4.1 Backpropagation
    4.2 Convergence Properties of Backpropgation
    4.3 Efficient Gradient-Descent Methods
    4.4 Focused Propagation
    4.4.1 Derivation of the Modification Rule
    4.4.2 Training Algorithm FP
    4.4.3 Properties of the FP Algorithm
    4.5 Comparison of Convergence Speeds
    4.5.1 Nonlinear Functions
    4.5.2 Experimental Results
    4.6 Remarks on Choosing Learning Parameters

  5. Minimization of Training Set Complexity
    5.1 Generalization
    5.2 Why Small Training Sets?
    5.3 Selective Incremental Learning
    5.3.1 Algorithm SEL
    5.3.2 Convergence Criterion
    5.3.3 Approximation-Theoretical Considerations
    5.4 Relationship among Generalization, Learing Time and Training Set Size
    5.4.1 Quality Criteria for Learning
    5.4.2 SEL Learning Curves
    5.4.3 Generalization, Learning Speed and Example Selection
    5.5 Minimal Training Sets
    5.5.1 Linear Mappings with Binary Inputs
    5.5.2 Nonlinear Mappings with Continuous Inputs
    5.6 Summary

  6. Active Exploration of Unknown Environments
    6.1 Necessity of Novel Learning Examples
    6.2 Example Generation by Genetic Search
    6.2.1 Genetic Search for Critical Examples
    6.2.2 Reproduction Plan
    6.2.3 Genetic Operators for Example Generation
    6.3 Creative Incremental Learning
    6.4 Genetic Creation vs. Random Generation
    6.4.1 Preliminary Experiments
    6.4.2 Preliminary Results
    6.5 Summary

  7. Genetic Neural Self-developing Nets
    7.1 What Network Size is Optimal?
    7.1.1 Learning Capability and Network Size
    7.1.2 Problems of Pure Connectionist Approaches
    7.2 Approaches to Optimizing Network Structures
    7.2.1 Destructive Approaches
    7.2.2 Constructive Approaches
    7.3 Learning by Self-development
    7.3.1 Development Process
    7.3.2 Two Learning Algorithms for Self-developing Nets
    7.4 Optimality of Self-development Algorithms
    7.4.1 Convergence and Optimality of Algorithms
    7.4.2 v-Optimality of Self-developing Nets
    7.4.3 Reduction of Training Set Complexities
    7.5 Time Complexity of Self-developing Nets
    7.5.1 Selective Developing Networks
    7.5.2 Creative Developing Networks
    7.5.3 Influence of Neural Growth
    7.6 Relation to Previous Approaches
    7.7 GENIAL, Genetic Algorithms and Simulated Annealing

  8. Digit Recognition and Function Approximation
    8.1 Writer-Independent Digit Recognition
    8.1.1 Solutions of Self-developing Nets
    8.1.2 Comparison Results
    8.1.3 Interpretation of Recognition Rates
    8.1.4 Summary
    8.2 Approximation of Complex Functions
    8.2.1 Comparison of Learning Strategies
    8.2.2 Changing Strategies via Neural Growth and Seed Examples
    8.2.3 Summary

  9. Self-developing Nets for Robot Control
    9.1 Basics of Robot Control
    9.2 The Task and the Simulated Robot Arm
    9.3 Learning to Predict Trajectories
    9.3.1 Trajectory Prediction
    9.3.2 Learning Results for Discrete Coding
    9.3.3 Learning Results for Continuous Coding
    9.3.4 Trajectory Prediction for Arbitrary Directions
    9.4 Learning Inverse Kinematics
    9.4.1 Inverse Kinematics Problem
    9.4.2 Creative Development Learning of Inverse Kinematics
    9.4.3 Properties of Genetic Learning
    9.4.4 Learning Results for Continuous Coding
    9.5 Concluding Remarks

  10. Summary and Future Work

    1
    1
    3
    3
    5
    5
    6
    7
    7
    8
    8
    9
    10

    12
    12
    15
    17
    17
    21
    24
    27
    30
    33
    38

    40
    40
    43
    43
    44
    45
    46
    47
    50
    51
    53
    56
    58
    60
    60
    62

    64
    64
    71
    74
    77
    77
    80
    83
    84
    84
    85
    88

    90
    90
    93
    95
    96
    98
    100
    104
    105
    108
    111
    115
    116
    118
    121

    122
    123
    125
    127
    129
    133
    139
    143
    143
    145
    147

    148
    149
    149
    150
    151
    151
    153
    156
    156
    160
    164
    164
    165
    172
    176
    176
    178
    179
    181
    182

    187
    188
    189
    190
    191
    194
    195
    196
    202
    202

    205
    206
    208
    210
    210
    210
    214
    216
    217
    217
    217
    221
    225
    227

    229

A
B




Appendix
Operating Instructions for GENIE
References
Symbol Index
Index
    236
    241
    246
    262
    265
Byoung-Tak Zhang's Home SCAI Lab Dept. of Computer Engineering