Robust echo state network for recursive system identification

Abstract

The use of recurrent neural networks in online system identification is very limited in real-world applications, mainly due to the propagation of errors caused by the iterative nature of the prediction task over multiple steps ahead. Bearing this in mind, in this paper, we revisit design issues regarding the robustness of the echo state network (ESN) model in such online learning scenarios using a recursive estimation algorithm and an outlier robust-variant of it. By means of a comprehensive set of experiments, we show that the performance of the ESN is dependent on the adequate choice of the feedback pathways and that the prediction instability is amplified by the norm of the output weight vector, an often neglected issue in related studies.