ENHANCING THE NONLINEAR MODELING CAPABILITIES OF MLP NEURAL NETWORKS USING SPREAD ENCODING

Two methods for representing data in a multi-layer perceptron (MLP) ne ural network are described and the resultant ability of networks, trai ned by the standard back-propagation algorithm, to identify the dynami cs of non-linear systems is investigated. One of the data conditioning methods has been widely used in studies of the MLP network and consis ts of normalising each network input and output variable and applying the normalised data to single network nodes. In the second method, nam ed spread encoding, each network variable is represented as a sliding Gaussian pattern of excitations across several network nodes. The spre ad encoding technique exhibits similarities with conventional algorith ms used in fuzzy logic and a network utilising this method can be cons idered as a fuzzy-neural type network, Neural networks are configured to represent a non-linear, auto-regressive, exogenous (NARX) input-out put model structure and the performance of trained networks is investi gated in applications to modelling a real liquid level process unit an d a simulation of a highly non-linear chemical process. Results show t hat using the data normalisation method, a network can provide accurat e single-step predictions but is incapable of adequate long-range pred ictions. In contrast to this, the spread encoding technique significan tly enhances the performance of a MLP network model enabling accurate single-step and long-range predictions to be achieved.