Variable structure radial basis function network and its application to on-line chaotic time series prediction

To improve the accuracy and the speed of on-line chaotic time series prediction via radial basis function (RBF) network, a sequential learning algorithm is presented for on-line constructing variable structure RBF network. A sliding window is constructed. By learning real-time updated data in the window, the parameters of the connecting weights, number of hidden units and center locations are dynamically tuned. The algorithm achieves parsimonious RBF network quickly, while only a small number of tuning parameters are employed. The variable structure network is applied to Mackey-Glass chaotic time series on-line prediction. The results demonstrate that network possesses satisfactory on-line dynamic identification and prediction performance.