Centre and Variance Selection for Gaussian Radial Basis Function Artificial Neural Networks

Abstract

The quality of the response of a RBF neural network depends strongly on the calculation method of the centres and the variance matrices. This paper describes an algorithm which combines the calculation of the centres and variances of the Gaussian nodes to improve the response of a RBF neural network. The selection of the centres is made using a modified version of the J{-means algorithm and the variances are based on the sample variance-covariance matrices of the input values associated with the centres. Applications to classification and function approximation problems are considered.