A novel backward elimination algorithm for construction of RBF neural networks

摘要

A novel backward elimination algorithm (BEA) based on the energy contributions of the non-orthogonal (coupled) regressor vectors is introduced for radial basis function (RBF) neural network construction. This algorithm builds RBF model by eliminating the minimum contribution regressor term among all candidates according to mean predicted-residual-sums-of-squares (PRESS) error. It first generates an initial model using computationally affordable batch learning and then updated it by a sequent learning with new training samples arriving. During the whole learning, the network architecture always remains the most optimal. This also can assure a better RBF network even if the RBF original basis is non-orthogonal. The effectiveness of new algorithm is demonstrated by the simulated results.