A partial least squares regression method for growing radial basis function networks

A novel partial least squares (PLS) learning algorithm is proposed for constructing radial basis function (RBF) networks. The algorithm grows RBF centers one by one with PLS regression method until an adequate network has been constructed, and the resulting parsimonious radial basis function-partial least squares (RBF-PLS) network demonstrates satisfying generalization performance and noise toleration capability. The proposed learning strategy provides an efficient means for fitting RBF networks, and this is illustrated by modelling nonlinear function and chaotic time series.