Evaluation of Distance Measures for Speciated Evolutionary Neural Networks in Pattern Classification Problems

Recently, evolutionary neural networks are hot topics in a neural network community because of their flexibility and good performance. However, they suffer from a premature convergence problem caused by the genetic drift of evolutionary algorithms. The genetic diversity in a population decreases quickly and it loses an exploration capability. Based on the inspiration of diversity in nature, a number of speciation algorithms are proposed to maintain diverse solutions from the population. One problem arising from this approach is lack of information on the distance measures among neural networks to penalize or discard similar solutions. In this paper, a comparison is conducted for six distance measures (genotypic, phenotypic, and behavioral types) with representative speciation algorithms (fitness sharing and deterministic crowding genetic algorithms) on six UCI benchmark datasets. It shows that the choice of distance measures is important in the neural network evolution.