Structure selection of neurofuzzy networks based on support vector regression

Abstract

Neurofuzzy networks are being used increasingly to model non-linear dynamic systems, since they have the approximating ability of neural networks and the transparency of fuzzy systems. However, good generalization results can only be obtained if the structure of the network is suitably chosen. It is shown here that the structure of neurofuzzy networks with scatter partitioning can be obtained from the support vectors (SV) of the Support Vector Regression (SVR), since the SVR can be transformed to a neurofuzzy network. The main advantage of this approach is that the structure of the neurofuzzy networks can now be objectively chosen, as the SV are obtained by constrained optimization for a given modelling error bound. Since neurofuzzy networks are linear-in-weights networks, the estimate of the weights of the networks can be obtained by the linear least-squares method. The properties of neurofuzzy networks based on the SV are derived, and its performance is illustrated by a simulation example involving a non-linear system, and the modeling of Southern Oscillation Index.