Fast automatic architecture selection in RBF networks

Abstract

Fast automatic methods of architecture selection are of great interest for use in local, dynamical modeling. A general procedure to select an optimal network architecture is proposed in the case of RBF nets. Taking as a starting point the universal approximation properties of RBF networks and the natural interpretation of their weights, a search for the simplest, best generalizing network is done. This requires some measure of goodness of fit, which allow to discard initial ovexweighted nets and to stop once too lean ones do not provide good generalization; in our case we will analyze both the error evolution and the statistical nature of the network residues. Since when going down in the number of network units fast network retraining is needed, we will also discuss this issue.