Learning Radial Basis Function Networks with the Trust Region Method for Boundary Problems

Abstract

We consider the solution of boundary value problems of mathematical physics with neural networks of a special form, namely radial basis function networks. This approach does not require one to construct a difference grid and allows to obtain an approximate analytic solution at an arbitrary point of the solution domain. We analyze learning algorithms for such networks. We propose an algorithm for learning neural networks based on the method of trust region. The algorithm allows to significantly reduce the learning time of the network.