Stable Output Feedback in Reservoir Computing Using Ridge Regression

An important property of Reservoir Computing, and signal processing techniques in general, is generalization and noise robustness. In trajectory generation tasks, we don’t want that a small deviation leads to an instability. For forecasting and system identification we want to avoid over-fitting. In prior work on Reservoir Computing, the addition of noise to the dynamic reservoir trajectory is generally used. In this work, we show that high-performing reservoirs can be trained using only the commonly used ridge regression. We experimentally validate these claims on two very different tasks: long-term, robust trajectory generation and system identification of a heating tank with variable dead-time.