δ-NARMA neural networks: a new approach to signal prediction

This article presents a new connectionist architecture for stochastic univariate signal prediction. After a review of related statistical and connectionist models pointing out their advantages and limitations, we introduce the-NARMA model as the simplest non-linear extension of ARMA models. These models then provide the units of a MLP-like neural network: the-NARMA neural network. The associated learning algorithm is based on an extension of classical back-propagation and on the concept of virtual error. Such networks can be seen as an extension of ARIMA and ARARMA models and faces the problem of non-stationary signal prediction. A theoretical study brings understanding of experimental phenomena observed during the-NARMA learning process. The experiments carried out on three railroad related real-life signals suggest that-NARMA networks outperform other studied univariate models.