Design of RBF Neural Networks using Critical Centroids

摘要

In this paper, a new method to design EBF-NN has been presented. This method is based on the selection of the centroids which are critical for classification. It may be divided in two parts. In the first phase, a modified Vector Quantization (VQ) procedure is aplied such that a set of prototype samples (centroids) for each class is determined. In a second phase, among these centroids, a reduced number of critical centroids for the classification task is found. In order to find the minimum number of critical centroids, the method to find critical centroids requires that the centroids obtained in the first phase are separable, i.e., the centroids of different classes are not overlaped. To achieve this objective, the SVQ process is controlled by the Ellipsoidal Boundary Methods (EBM). Several simulations over a well know classification problems show the advantages of using critical centroids as centers of a Radial Basis Function Neural Network (RBF-NN).