Summary
International Symposium on Nonlinear Theory and Its Applications
2020
Session Number:C4L-B
Session:
Number:C4L-B-1
Reinforcement Learning Based on Electro-Optic Delay-Based Reservoir Computing
Kanno Kazutaka, Uchida Atsushi,
pp.425-428
Publication Date:2020/11/09
Online ISSN:2188-5079
DOI:10.34385/proc.74.C4L-B-1
PDF download (184.6KB)
Summary:
Machine learning based on photonic implementation has been paid increasing attention for fast information processing and energy efficiency. In this study, we present numerical demonstration of reinforcement learning based on photonic reservoir computing using an electro-optic delay system. Reinforcement learning is one of machine learning schemes to do so as to maximize a reward from environment. Photonic reservoir computing receives states in environment and decides actions, where photonic reservoir computing is trained based on Q-learning to maximize the total reward. Cart-pole problem, which is a famous benchmark task in reinforcement learning, is demonstrated to evaluate the performance of our scheme.