The Automatic Model Selection and Variable Width RBF Neural Networks for Chaotic Time Series Prediction

Abstract

This paper investigates the construction of radial basis function(RBF) neural networks, and a new self-adaptive algorithm is presented to achieve chaotic times series prediction. This method is based on an adaptive Orthogonal Least Squares(OLS) algorithm, that the automatic model method can assign an appropriate number of hidden units for the network, and the variable width model may guarantee a natural overlap between kernel functions. The proposed algorithm may specify appropriate number and widths of kernels simultaneously. The augmented algorithms are employed to some examples such as Mackey-Glass mapping in known and unknown noise chaotic dynamical systems. Experimental results show that the proposed algorithm can produce a better prediction.