Structurally adaptive RBF network in nonstationary time series prediction

Abstract

A sequentially adaptive radial basis function (RBF) network is applied to the nonstationary, time series prediction. Sequential adaptation of parameters and structure is achieved using an extended Kalman filter criterion for network growing is obtained from the Kalman filter's consistency test. The Optimal Brain Surgeon and Optimal Brain Damage pruning methods are derived for networks which parameters are estimated by the EKF. Criteria for neurons/connections pruning are based on the statistical parameter significance test. Prediction of the nonstationary logistic map and Lorenz time series is considered.