Global and partial function approximation with evolutionary algorithms

Abstract

We present an evolutionary algorithm that evolves a population of local approximators in order to fit an unknown function. The evolutionary algorithm performs a simultaneous learning of simple local approximators together with the regions in which the local approximators are applied. By combining these simple local approximators, the domain is in fact partitioned in the Voronoi diagram that has as centers, the center points of the region in which each local approximator is efficient and useful. Both continuous and non-continuous approaches are considered. The algorithm seems promising in order to develop good neural networks for function approximation.