A Constant Trace Algorithm for Local Adaptation of Rbf Models

Radial basis function (RBF) models are often adapted on-line using exponential forgetting with a single forgetting factor. However, this technique applies forgetting uniformly to the past data in the entire operating space and is not appropriate for non-linear systems where dynamics are different in different operating regions. This paper describes a new development in local forgetting for online adaptation of RBF models. Local adaptation of the model is achieved by using different forgetting factors that vary with the excitation of each basis function. The forgetting factors are varied to keep the trace of the RLS information matrix constant. A numerically robust algorithm is also developed for implementation. Results from applying the new method to modelling a real chemical reactor process are presented. Copyright © 2006 USTRATH