Quantifying the Severity of the Permutation Problem in Neuroevolution

Abstract

In this paper we investigate the likely severity of the Permutation Problem on a standard Genetic Algorithm used for the evolutionary optimisation of Neural Networks. We present a method for calculating the expected number of permutations in an initial population given a particular representation and show that typically this number is very low. This low expectation coupled with the empirical evidence suggests that the severity of the Permutation Problem is low in general, and so not a common cause of poor performance in Neuroevolutionary algorithms.