Construction of compact RBF network by refining coarse clusters and widths

It is known that centers, widths, and weights are three mainly considered factors in constructing a radial basis function (RBF) network. This paper aims at constructing a compact RBF network with two main steps. In the first step, the coarse clusters computed from triangle inequalities are refined to obtain the locations of centers by the defined maximum degree spanning tree (MDST). Meanwhile the coarse widths are obtained. In the second step, a learning algorithm referred to as anisotropic gradient descent method is presented to further refine the above coarse widths. Experiments of the proposed algorithm show its great performance in times series prediction and classification.