Initializing RBF-Networks with Small Subsets of Training Examples

Abstract

An important research issue in RBF networks is how to determine the gaussian centers of the radial-basis functions. We investigate a technique that identifies these centers with carefully selected training examples, with the objective to minimize the network's size. The essence is to select three very small subsets rather than one larger subset whose size would exceed the size of the three small subsets unified. The subsets complement each other in the sense that when used by a nearestneighbor classifier, each of them incurs errors in a different part of the instance space. The paper describes the example-selection algorithm and shows, experimentally, its merits in the design of RBF networks. Introduction Radial-basis-function (RBF) networks, such as the one depicted in Figure 1, are used to approximate functions f : R m ! R p by appropriate adjustments of the parameters in the formula f j (x) = \\Sigma ij w i ' i (x), where x = (x 1 ; : : : ; xn ) is an input vector and each ' i...