Adaptive prediction of nonstationary signals using chaotic neural networks

Describes the nonlinear adaptive prediction of nonstationary time series using modified chaotic neural networks (CNNs). Since the chaotic neuron in the networks contains an internal feedback, the chaotic neural networks used in the paper have inherently the characteristics of highly nonlinear dynamics, which are required for robust prediction of nonstationary signals. For the relative comparison of prediction performance, the CNN based predictor is compared with a conventional ARMA linear predictor and the recurrent neural networks (RNNs) based predictor. These predictors are evaluated using Mackey-Glass time series added sinusoid and speech signals in single-step and multi-step predictions. Simulation results show that the CNN predictor outperforms other predictors with mean square error