The application of Direction basis function neural networks to the prediction of chaotic time series

In this paper we have examined the ability of Direction basis function networks (DBFN) to predict the output of a chaotic time series generated from a model of a physical system. DBFNs are known to be universal approximators, and chaotic systems are known to exhibit "random" behavior. Therefore the challenge is to apply the DBFN to the prediction of the output of a chaotic system, which we have chosen here to be the Mackey-Glass delay differential equation. The DBFN has been trained with off-line supervised learning using a Recursive Least Squares optimization for obtaining weights. Key issues which are addressed are the estimation of the order of the system and dependence of prediction error on various factors such as placement of DBF centers, selection of perceptive widths, and number of training samples. Included in this study is an implementation of Moody and Darken's K Means Clustering approach to optimally place DBF centers and a heuristic nearest neighbor method for determining perceptive widths.