Nonlinear Autoregressive Model Design and Optimization Based on ANN for the Prediction of Chaotic Patterns in EEG Time Series

Abstract

This paper presents an Artificial Neural Network (ANN) based Nonlinear Auto-Regressive (NAR) model design to generate and predict time series outputs of Lorenz chaotic system. The ANN based chaotic time series generator can be used for the simulation, analysis and prediction of Electroencephalogram (EEG) signals, which demonstrate chaotic features. The Lorenz chaotic system outputs are used to simulate the chaotic dynamics demonstrated by EEG time series signals. The training performances are investigated for various ANN architectures with different numbers of hidden neurons and feedback delays in the NAR model. The training results show that better training performance can be achieved by increasing the number of hidden neurons, or the number of feedback delays, but the computational cost is also significantly increased. The former optimization method is more beneficial since fewer multiplications are required compared to the latter. This reduces the computational cost for both ANN training and hardware implementation of the trained ANN.