Classification using radial basis function networks with uncertain weights

This paper considers the performance of radial basis function neural networks for the purpose of data classification. The methods are illustrated using a simple two class problem. Two techniques for reducing the rate of misclassifications, via the introduction of an "unable to classify" label, are presented. The first of these considers the imposition of a threshold value on the classifier outputs whilst the second considers the replacement of the crisp network weights with interval ranges. Two network training techniques are investigated and it is found that, although thresholding and uncertain weights give similar results, the level of variability of network performance is dependent upon the training approach.