Learning from the past: reservoir computing using delayed variables

Abstract

Reservoir computing is a machine learning method that is closely linked to dynamical systems theory. This connection is highlighted in a brief introduction to the general concept of reservoir computing. We then address a recently suggested approach to improve the performance of reservoir systems by incorporating past values of the input signal or of the reservoir state variables into the readout used to forecast the input or cross-predict other variables of interest. The efficiency of this extension is illustrated by a minimal example in which a three-dimensional reservoir system based on the Lorenz-63 model is used to predict the variables of a chaotic Rössler system.