NONLINEAR OPTIMIZATION OF RBF NETWORKS

This paper describes the application of advanced nonlinear optimizatio n strategies to Radial Basis Function (RBF) networks. Standard trainin g procedures for RBFs are briefly reviewed and the need for adaptation of the nonlinear centre and width parameters discussed. The failure o f current nonlinear gradient based optimization strategies in this con text is established and linked to the highly ill-conditioned nature of these networks. A new hybrid training algorithm, which combines linea r optimization of the basis function heights with nonlinear optimizati on of centres and widths, is presented and shown to yield significantl y superior training performance.