Supervised Classification by RBF Networks With Self-Constructing Clustering

Radial basis function (RBF)networks have been widely used in a variety of applications, including supervised classification. Two issues are often encountered in the applications. First, the number of hidden nodes has to be decided. Second, the settings of the basis functions have to be set. In this paper, we propose a novel RBF network approach for supervised classification applications. Given a set of training patterns, the number of hidden nodes in the hidden layer is determined by applying a self-constructing clustering algorithm on the patterns. Normalized Gaussian functions are taken to be basis functions, and their centers and deviations are set according to the clusters obtained from the clustering algorithm. The optimal values for the weights associated with the output layer are derived by adapting them to the training patterns. Experimental results are shown to demonstrate the effectiveness of the proposed approach.