Experimental demonstration of adaptive model selection based on reinforcement learning in photonic reservoir computing

Abstract

Reservoir computing provides superior information processing ability for a time series prediction based on appropriate learning prior to task execution. The performance of reservoir computing, however, may degrade if the characteristics of the input signal drastically change over time because the internal model of reservoir computing deviates from the subjected input signal trains. We propose a method for adaptive model selection using reinforcement learning in electro-optic delay-based reservoir computing. We experimentally show that an adaptive model selection is effective when different dynamical models for the input signals change dynamically over time.