GA with orthogonal transformation for RBFN configuration

Abstract

A genetic algorithm (GA) approach for radial basis function network (RBFN) training is proposed. Although several non-evolutionary training methods for RBFN training are available, orthogonal least squares (OLS) algorithm has a number of desirable properties for robust training. In the OLS algorithm the basis function centers are selected among the training samples and therefore they are predetermined in their totality. In contrast with this, the width parameters of the basis functions are not and in general cannot accurately be predetermined. Therefore they are generally taken to be equal for all centers. In the GA approach, next to the optimally selected centers, the covariance matrix of the width parameters of these centers can also be determined, thereby the network performance is enhanced. Training RBFN by GA can be accomplished in different ways with different merits. The main contribution of this research lies in its introduction of a novel method, which selects the subset optimally among the available centers. For this it uses orthogonal transformation method namely singular value decomposition (SVD)-QR method integrated to the algorithm. Along this line, GA approach for RBFN training as an alternative to OLS algorithm with additional desirable parameter optimization properties is discussed. By means of experiments described, competitiveness of the approach for robust RBFN training is presented.