Modeling non-stationary dynamic system using recurrent radial basis function networks

Abstract

This paper addresses the problem of continuous adaptation of neural networks in a non-stationary environment. We have applied the extended Kalman filter to the parameter, state and structure estimation of a recurrent radial basis function network. The architecture of the recurrent radial basis function network implements a nonlinear autoregressive model with exogenous inputs. Statistical criteria for structure adaptation (growing and pruning of hidden units and connections of the network) were derived using statistics estimated by the Kalman filter. The proposed algorithm is applied to non-stationary dynamic system modeling.