A self-organizing NARX network and its application to prediction of chaotic time series

Abstract

This paper introduces the concept of dynamic embedding manifold (DEM), which allows the Kohonen self-organizing map (SOM) to learn dynamic, nonlin-ear input-ouput mappings. The combination of the DEM concept with the SOM results in a new modelling technique that we called Vector-Quantized Temporal Associative Memory (VQTAM). We use VQTAM to propose an unsupervised neural algorithm called Self-Organizing N A R X (SONARX) network. The SONARX network is evaluated on the problem of modeling and prediction of three chaotic time series and compared with MLP, RBF and autoregressive (AR) models. Its is shown that SONARX exhibits similar performance when compared to MLP and RBF, while producing much better results than the AR model. The influence of the number of neurons, the memory order, the number of training epochs and the size of the training set in the final prediction error is also evaluated.