A new unsupervised learning algorithm for the placement of centers in a radial basis function neural network (RBFNN)

RBFNNs represent a promising tool for the approximation of functions of several variables, and find numerous applications in many engineering fields. Their overall approximation accuracy is critically dependent on the relative positions of centers with respect to the training samples. In this paper the problem of an optimal center placement is formulated in terms of an energy function. A new and robust unsupervised learning technique is then developed, which leads to a center arrangement which is independent of their initial positions and of the order of presentation of the training samples. The performances of this new algorithm are then compared with those of the classical winner-take-all types in the given example