Photonic decision making for solving competitive multi-armed bandit problem using semiconductor laser networks

Abstract

Multi-armed bandit problems concern decision making when selecting a slot machine among many slot machines with initially uncertain hit probabilities to maximize the total reward; this is a fundamental problem of reinforcement learning. Furthermore, competitive multi-armed bandit problems involve multiple agents in play, manifesting fundamental concerns regarding social figures, not just individual rewards. A representative issue is selection conflict, in which multiple players select the same slot machine and may miss the total reward as a whole. This study proposes a scheme for solving the competitive multi-armed bandit problem using semiconductor laser networks by introducing an exclusive selection mechanism. We numerically implement our method and compare it with conventional algorithms. We show that our method outperforms conventional algorithms in solving the competitive multi-armed bandit problem.