A non-orthogonal and multi-width RBF neural network for chaotic time series prediction

Abstract

A non-orthogonal and multi-width learning algorithm of radial basis function (RBF) neural network is presented for chaotic time series prediction. It is based on an adaptive algorithm, which takes advantages of the good selection capability of the non-orthogonal method for assigning an appropriate number of hidden units for the network and the ability of the multi-width model for guaranteeing a natural overlap between kernel functions. The proposed algorithm may specify the locations and widths of kernels simultaneously. For known and unknown noise chaotic dynamical systems, the novel algorithm can predict them well and its effectiveness is illustrated by results from experimenting on some examples such as Honen chaotic time series. © 2012 Springer Science+Business Media B.V.