Construction of radial basis function neural networks via a minimization of its localized generalization error

Abstract

Lots of researchers have been studying on how to construct radial basis function neural networks. To determine the number and location of hidden neurons, a recursive procedure is adopted with a new evaluation criterion based on localized generalization error model (L-GEM). We derive a new sensitivity expression for Gaussian radial basis function neural network based on L-GEM, and get a new localized generalization error bound. The RBF that yields the minimal localized generalization error bound is selected. We compare our approach with minimization of cross validation, and minimization of training mean square error (MSE) methods. The experimental results show that our approach performs much better than the other two methods with reasonable number of centers.