An Extended Approach of a Two-Stage Evolutionary Algorithm in Artificial Neural Networks for Multiclassification Tasks

Abstract

This chapter considers a recent algorithm to add broader diversity at the beginning of the evolutionary process and extends it to sigmoidal neural networks. A simultaneous evolution of architectures and weights is performed with a two-stage evolutionary algorithm. The methodology operates with two initial populations, each one containing individuals with different topologies which are evolved for a small number of generations, selecting the half best individuals from each population and combining them to constitute a single population. At this point, the whole evolutionary cycle is applied to the new population. This idea was previously proposed by us for product unit neural networks, and we now extend to sigmoidal neural networks. The experimentation has been carried out on twelve data sets from the UCI repository and two complex real-world problems which differ in their number of instances, features and classes. The results have been contrasted with nonparametric statistical tests and show that our proposal significantly improves the test accuracy of the models with respect to the obtained ones with a standard methodology based on a single population. Moreover, the new proposal is much more efficient than other methods developed previously by us.