Abstract & Overview
A constructive learning algorithm for multioutput radial basis function networks is presented. Unlike most network learning algorithms, which require a fixed network structure, this algorithm automatically determines an adequate radial basis function network structure during learning. By formulating the learning problem as a subset model selection, an orthogonal leastsquares procedure is used to identify appropriate radial basis function centres from the network training data, and to estimate the network weights simultaneously in a very efficient manner. This algorithm has a desired property, that the selection of radial basis function centres or network hidden nodes is directly linked to the reduction in the trace of the error covariance matrix. Nonlinear system modelling and the reconstruction of pulse amplitude modulation signals are used as two examples to demonstrate the effectiveness of this learning algorithm.