Evolving sum and composite kernel functions for regularization networks

Abstract

In this paper we propose a novel evolutionary algorithm for regularization networks. The main drawback of regularization networks in practical applications is the presence of meta-parameters, including the type and parameters of kernel functions Our learning algorithm provides a solution to this problem by searching through a space of different kernel functions, including sum and composite kernels. Thus, an optimal combination of kernel functions with parameters is evolved for given task specified by training data. Comparisons of composite kernels, single kernels, and traditional Gaussians are provided in several experiments. © 2011 Springer-Verlag.