Second-order multilayer perceptrons and its optimization with genetic algorithms

Abstract

There have been many efforts to combine multilayer perceptrons (MLP) and radial basis function networks (RBFN). Among these works, circular backpropagation networks (CBPN) achieved both MLP and RBFN's properties by simply modifying MLP. In this paper, CBPN is extended to take all first and second-order terms of data as input. We show that the proposed network can represent not only MLP and RBFN but also ellipsoidal basis function networks (EBFN). Using Baldwin effect-based genetic algorithm, we develop an approach for optimizing this network.