A Heuristic Algorithm Of Training RBFNN Based On Localized Generalization Error Model

Firstly, this paper introduces radial basis function neural network (RBFNN) and the result of localized generalization error model (LGEM) proposed by Doctor W.W.Y. NG Wing W.Y.NG. Since the result of LGEM contains the integral operator, when using the result in practice, the computation complexity is high. Then W.W.Y. NG used the Central Limit Theorem to simplify the result of the LGEM based on some assumptions, and used the simplified result on the architecture Selection problem of RBFNN. However, his method applies for the condition when the feature dimension of samples is bigger relatively.This paper analyses the result of the LGEM, and then proposes a heuristic algorithm for training RBFNN based on LGEM. This method avoids the integral operator and does not restrict the number of sample dimension. Experiments show that the heuristic method successfully solves the problem of computation complexity of LGEM, and the training accuracy obtained a good effect.