Training of Neural Network Wiener Models with Recursive Prediction Error Algorithm

Abstract

Identification of the Wiener system consisting of a linear dynamics in series with a static nonlinearity is considered. A recursive prediction error training algorithm for a recurrent neural network Wiener model is proposed. The gradient of the model output w.r.t. parameters of its linear part is computed with the sensitivity method. The proposed algorithm has superior convergence properties in comparison with gradient methods such as the sensitivity method or the truncated backpropagation through time. Its performance is illustrated with a simulated example of a pneumatic valve.