Hierarchical radial basis function networks

Ersoy (1991) and Ersoy and Hong (1990) have constructed a neural network architecture called the parallel, self-organizing, hierarchical neural network (PSHNN) that contains a number of stage neural networks. In their papers, the stage networks are one-layer networks with delta rule learning. They report the result by using PSHNN in solving some classification problems, but how effective it is compared with other methods was not reported. In this paper we construct a hierarchical network where stage networks are radial basis function networks (HRBFN) and using the nearest neighbor method as decision rule instead of the approximation method used in Ersoy´s paper. As applications, we use our method to solve medical diagnosis problems and some other difficult classification problems. While PSHNN is very sensitive to the number of iterations used in each stage network to train the network, it seems that our HRBFN does not depend on the number of centers for the starting stage network