Improving the performance of radial basis function (RBF) classification using information criteria

Successful implementation of radial basis function (RBF) networks for classification tasks must deal with architectural issues, the burden of irrelevant attributes, scaling, and some other problems. In this paper, we introduce a new class of kernel functions and try to build a three-way hybrid between the EM algorithm, regression trees, and the new kernel functions. Instead of using linear output and least squares error function, we introduce nonlinear sigmoid function and cross entropy error function into radial basis function (RBF) network. The new resulting network is easy to use, and has favorable classification accuracy. Numerical experiments show that our new model has better performance than logistic regression in binary classification and has equivalent performance in multiple classification as compared with MLP and other nonparametric classification procedures.