Identification and prediction of non-linear models with recurrent neural network

Abstract

Using a neural network to identify models and predict signals allows to go beyond the linear domain. In this paper, we show the advantage of using neural network for these signal processing applications. Thus, the function charactering the cell (sigmoïd or others) allows the study of non-linear models. Using feedback links specific to a recurrent network, the time is taken into account. Two different goals are assigned to the two phases in using this type of network: 1) the neural network training method uses a gradient backward propagation method. During the learning phase, the weights of the network are modified to identify the parameters of the given model. 2) during the test phase, the network predicts the output for each time step. Results are presented in the case of a Non-Linear AutoRegressive filters and they confirm the good responses of neural networks both for identification of parameters and for prediction of output for these non-linear models.