Construction of Radial Basis Function Networks with Diversified Topologies

Abstract

In this review we bring together some of our recent work from the angle of the diversified RBF topologies, including three different topologies; (i) the RBF network with tunable nodes; (ii) the Box-Cox output transformation based RBF network (Box-Cox RBF); and (iii) the RBF network with boundary value constraints (BVC-RBF). We show that the modified topologies have some advantages over the conventional RBF topology for specific problems. For each modified topology, the model construction algorithms have been developed. These proposed RBF topologies are respectively aimed at enhancing the modelling capabilities of; (i)flexible basis function shaping for improved model generalization with the minimal model;(ii) effectively handling some dynamical processes in which the model residuals exhibit heteroscedasticity; and (iii) achieving automatic constraints satisfaction so as to incorporate deterministic prior knowledge with ease. It is shown that it is advantageous that the linear learning algorithms, e.g. the orthogonal forward selection (OFS) algorithm based leave-one-out (LOO) criteria, are still applicable as part of the proposed algorithms.