Nonlinear plant identifier using the multiple adaptive RBF network convex combinations

Abstract

Due to its simplicity and robustness, the stochastic gradient (SG) algorithm has become a very popular learning algorithm for adaptive radial basis function (RBF) network filter. However, a difficulty is its inherent compromise between convergence speed and precision, that is imposed by the selection of a fixed value for the step sizes. In this paper, to solve the problem, we propose to generalize the views of the convex combination, which is combining multiple RBF filters (MCRBF) using the SG algorithm with different step sizes. Simulation works with nonlinear system identification have been carried out to illustrate the effectiveness of this approach. © 2012 Springer-Verlag GmbH.