Improving the prediction capability of radial basis function networks

Radial basis function networks (RBFNs) have been well established as a class of supervised neural networks. Because they do not require a priori process information and they possess localized interpolation properties, they are attractive for the empirical modeling of complex nonlinear multivariable processes such as those associated with the chemical industry. In this paper, classical regularization theory is used to develop a technique for improving the prediction capabilities of RBFNs by incorporating process knowledge obtained from physicochemical relationships through modification of the objective function. An analysis of this new objective function is provided and compared to the conventional least-squares objective function. Several chemical process examples are provided to demonstrate the improved predictive capability of this modified RBFN.