NEURAL NETWORKS SYNTHESIS BASED ON STOCHASTIC-APPROXIMATION ALGORITHM

A new method for neural networks synthesis is presented. This is an al gorithm of the stochastic approximation type, which is designed to opt imize functions with non-unique stationary points. This method differs from the usual backpropagation algorithm which cannot deal with the o ptimization of multimodal functions. The stochastic approximation meth od presented in this study consists of the Kiefer-Wolfowitz procedure. Numerical experiments related to a phosphate calcinator modelling are presented in order to illustrate the performance of the optimization algorithm associated with neural networks synthesis.