Training Basis Function Networks Including RBF and Multiresolution Wavelet Models with an Adaptive Orthogonal Least Squares Algorithm

This paper concerns the construction and training of basis function networks for the identification of nonlinear dynamical systems. A new adaptive orthogonal least squares (AOLS) algorithm, which integrates basis function (model term) selection, model size determination and model parameter estimation, is developed for basis function network training. The construction and training of a new multiresolution wavelet network, together with a comparison with RBF models, is discussed in detail. A practical three-phase modelling framework to ensure unbiased nonlinear models is obtained using the basis function networks. Several examples are presented to demonstrate the application potential of the new identification techniques.