System identification via optimised wavelet-based neural networks

Abstract

Nonlinear system identification by means of wavelet-based neural networks (WBNNs) is presented. An iterative method is proposed, based on a way of combining genetic algorithms (GAs) and least-square techniques with the aim of avoiding redundancy in the representation of the function. GAs are used for optimal selection of the structure of the WBNN and the parameters of the transfer function of its neurones. Least-square techniques are used to update the weights of the net. The basic criterion of the method is the addition of a new neurone, at a generic step, to the already constructed WBNN so that no modification to the parameters of its neurones is required. Simulation experiments and comparison with neural nets having different activation functions for the neurones are also presented.