A probabilistic extension for the DDA algorithm

Abstract

Many algorithms to train radial basis function (RBF) networks have already been proposed. Most of them, however, concentrate on building function approximators and only few specialized algorithms are known that concentrate on RBFs for classification. They are based on heuristics that focus on finding areas where relatively few (or no) conflicts occur, but do not try to approximate the underlying probability distribution function (PDF) of the data. In this paper an extension for an already existing constructive algorithm for RBF networks is introduced. The new method uses the dynamic decay adjustment (DDA) algorithm to find conflict free areas and builds more appropriate PDFs inside each such zone. On a dataset which was generated using Gaussian distributions it is demonstrated that this method builds almost optimal classifiers that compare very well with the theoretical Bayes classifier. It is shown, however, that the generalization capability of such networks does not compare favourable to the DDA itself.