A two-stage algorithm for automatic construction of RBF neural models

This paper proposes a novel algorithm for automatic construction of radial basis function (RBF) neural models, combining a two-stage stepwise regression approach and the predicted-residual-sums-of-squares (PRESS) statistic. The main objective is to improve the generalization capability and compactness of the RBF neural models. This is achieved through a model refining procedure combined with leave-one-out cross validation. First, the neural model is constructed automatically by selecting important RBF centres which minimize the PRESS error. The contribution of each selected term is then reviewed, and insignificant hidden neurons are replaced. Finally, the forward procedure is utilized again to re-order the selected terms, leading to a reduction in the overall model size. The computation of the PRESS statistic is simplified by introducing a residual matrix in the two-stage method. Simulation examples confirm the effectiveness of the proposed technique.