Memristor-based input delay reservoir computing system for temporal signal prediction

Abstract

Reservoir computing (RC) system, featured by its recursive structure, has been utilized for temporal signal processing, offering both low power consumption and high computational speed. This work reports on a novel input delay reservoir computing (ID-RC) system based on the oxide memristors, which can be applied to temporal signal prediction. The particle swarm optimization (PSO) algorithm is employed in the ID-RC system to obtain optimal hyperparameters for multi-step prediction in the Mackey-Glass task, with a normalized root-mean-square error (NRMSE) of only 0.09 at the 20th step. Significantly, by employing the ID-RC system in temporal signal prediction of the Hénon map and the nonlinear autoregressive moving average (NARMA10), small NRMSEs of 0.047 and 0.017 were achieved, respectively. The memristor-based ID-RC system turns out to be highly promising in forecasting of chaotic time series.