Open Access Te Herenga Waka-Victoria University of Wellington
Browse
thesis_access.pdf (1.95 MB)

Problem Decomposition and Adaptation in Cooperative Neuro-Evolution

Download (1.95 MB)
Version 3 2023-03-14, 23:29
Version 2 2023-03-13, 23:57
Version 1 2021-11-12, 09:09
thesis
posted on 2023-03-14, 23:29 authored by Chandra, Rohitash

One way to train neural networks is to use evolutionary algorithms such as cooperative coevolution - a method that decomposes the network's learnable parameters into subsets, called subcomponents. Cooperative coevolution gains advantage over other methods by evolving particular subcomponents independently from the rest of the network. Its success depends strongly on how the problem decomposition is carried out. This thesis suggests new forms of problem decomposition, based on a novel and intuitive choice of modularity, and examines in detail at what stage and to what extent the different decomposition methods should be used. The new methods are evaluated by training feedforward networks to solve pattern classification tasks, and by training recurrent networks to solve grammatical inference problems. Efficient problem decomposition methods group interacting variables into the same subcomponents. We examine the methods from the literature and provide an analysis of the nature of the neural network optimization problem in terms of interacting variables. We then present a novel problem decomposition method that groups interacting variables and that can be generalized to neural networks with more than a single hidden layer. We then incorporate local search into cooperative neuro-evolution. We present a memetic cooperative coevolution method that takes into account the cost of employing local search across several sub-populations. The optimisation process changes during evolution in terms of diversity and interacting variables. To address this, we examine the adaptation of the problem decomposition method during the evolutionary process. The results in this thesis show that the proposed methods improve performance in terms of optimization time, scalability and robustness. As a further test, we apply the problem decomposition and adaptive cooperative coevolution methods for training recurrent neural networks on chaotic time series problems. The proposed methods show better performance in terms of accuracy and robustness.

History

Copyright Date

2012-01-01

Date of Award

2012-01-01

Publisher

Te Herenga Waka—Victoria University of Wellington

Rights License

Author Retains Copyright

Degree Discipline

Computer Science

Degree Grantor

Te Herenga Waka—Victoria University of Wellington

Degree Level

Doctoral

Degree Name

Doctor of Philosophy

Victoria University of Wellington Item Type

Awarded Doctoral Thesis

Language

en_NZ

Victoria University of Wellington School

School of Engineering and Computer Science

Advisors

Frean, Marcus; Zhang, Mengjie