A novel RBF neural network with fast training and accurate generalization

Abstract

For the reason of all the centers and radii needed to be adjusted iteratively, the learning speed of radial basis function (RBF) neural networks is always far slower than required, which obviously forms a bottleneck in many applications. To overcome such problem, we propose a fast and accurate RBF neural network in this paper. First we prove the universal approximation theorem for RBF neural networks with arbitrary centers and radii. Based on this theory, we propose a new learning algorithm called fast and accurate RBF neural network with random kernels (RBF-RK). With the arbitrary centers and radii, our RBF-RK algorithm only needs to adjust the output weights. The experimental results, on function approximation and classification problems, show that the new algorithm not only runs much faster than traditional learning algorithms, but also produces better or at least comparable generalization performance. © Springer-Verlag 2004.