Backward Centre Selection Method for On-Line Adaptation of RBF Network Models

Abstract

An approach to tracking process time variations using an adaptive radial basis function network model is described. The method is based on a numerically robust recursive algorithm for updating the network output layer weights. It is shown how network centres contributing least to the network output can be found and removed from model calculations. Thus, both the structure and weights of the network are adaptive. An illustrative example is given, to demonstrate the effectiveness of the algorithm and illustrate its performance, in an application to modelling a real chemical process. Results show more accurate model predictions compared to using a network with only weight updating.