Neural Network Normalization for Genetic Search^*

Jung-Hwan Kim, Sung-Soon Choi, and Byung-Ro Moon

Abstract. An arbitrary neural network has a number of functionally equivalent other networks. This causes redundancy in genetic representation of neural networks, which considerably undermines the merit of crossover in GAs [1]. This problem has received considerable attention in the past and has also been called the “competing conventions” problem [2].

We transform each neural network to an isomorphic neural network to maximize the genotypic consistency of two parents. We aim to develop a better genetic algorithm for neural network optimization by helping crossover preserve common functional characteristics of the two parents. This is achieved by protecting “phenotypic” consistency and, consequently, preserving building blocks with promising schemata.

^*This work was supported by Brain Korea 21 Project. The ICT at Seoul National University provided research facilities for this study.

LNCS 3103, p. 398 f.

Full article in PDF

lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2004