Approximation by Growing Radial Basis Function Networks Using the Differential-Evolution-Based Algorithm

Abstract

The differential evolution (DE) algorithm is a floating-point-encoded evolutionary algorithm for global optimization. We applied a DE-based method to training radial basis function (RBF) networks with variables including centers, weights, and widths. This algorithm consists of three steps – initial tuning focusing on finding the center of a one-node RBF network, local tuning, and global tuning both using cycling schemes to find RBF network parameters. The mean square error from desired output to actual network output is applied as the objective function to be minimized. Network training is shown by approximating a set of functions and reconstructing the spectra of oil samples and classification. Net performance is compared to approaches reported in the literature, and the resulting network generally performs better based on the tests performed. Results show that DE-based Gaussian RBF growth method improves approximation results reported.