Combined projection and kernel basis functions for classification in evolutionary neural networks

Abstract

This paper describes a methodology for constructing the hidden layer of a feed forward network using a possible combination of different transfer projection functions (sigmoidal, product) and kernel functions (radial basis functions), where the architecture, weights and node typology is learnt using an evolutionary programming algorithm. The methodology proposed is tested using five benchmark classification problems from well-known machine intelligence problems. We conclude that combined functions are better than pure basis functions for the classification task in several datasets and that the combination of basis functions produces the best models in some other datasets. © 2007 Springer-Verlag Berlin Heidelberg.