A Hybrid Model of Partial Least Squares and RBF Neural Networks for System Identification

A novel learning algorithm is presented to construct radial basis function (RBF) networks by incorporating partial least squares (PLS) regression method. The algorithm selects hidden units one by one with PLS regression method until an adequate network is achieved, and the resulting minimal RBF-PLS (MRBF-PLS) network exhibits satisfying generalization performance and noise toleration capability. The algorithm provides an efficient approach for system identification, and this is illustrated by modelling nonlinear function and chaotic time series.