Performance Analysis of Optical Reservoir Computer with Two Delayed Inputs

Abstract

In this paper, a novel reservoir computing architecture based on the single nonlinear node with two delayed inputs is presented. Its performance is investigated on time-series forecasting tasks, for two different profiles of activation functions implemented in the optical domain. Simulation results show reduced NRMSE value to around 3% for Mackey-Glass time series prediction 11 steps in advance, when sine squared optical activation function based on Mach-Zehnder modulator is used.