Evolving artificial neural networks using permutation problem free modified celluler enconding

Abstract

This thesis works with a new evolutionary system for feedforward artificial neural networks (ANNs). An indirect encoding scheme, to be particular, modified cellular encoding (MCE) is proposed to represent ANNs. The original cellular encoding is modified in such a way that it does not suffer from the well-known permutation problem or competing conventions problem of genetic algorithms for evolving ANNs. The functionality of some program symbols in cellular encoding is changed; new rules are added. As a consequence, it is possible to apply crossover operator in the genetic search. Radical change of architecture i.e. behaviour from parents to their children is stopped by keeping the application of crossover on genotypes within certain levels. It is shown in this work that addition / deletion of nodes / conncctions can evidently be done by crossover alone. Other attempts are also taken to minimize behavioural disruption between parents and their offspring. In the evolution system, the number of user specified parameters is also decreased. The evolutionary system is also implemented and its performance is tested on some real world problems. The upshot of the genetic search is studied and assessed against the contemporary researches, although direct comparison with other evolutionary approaches to designing ANN is very difficult. It is shown in this thesis that the genetic search can find a reasonable ANN from the search space in considerably short period.