[19a-Z34-1~11] FS.1 Focused Session "AI Electronics"

[19a-Z34-6] Analysis of state exploration in Photonic Q-learning

〇Takashi Urushibara¹, Nicolas Chauvet¹, Satoshi Sunada², Kazutaka Kanno³, Atsushi Uchida³, Ryoichi Horisaki¹, Makoto Naruse¹ (1.Univ. Tokyo, 2.Kanazawa Univ., 3.Saitama Univ.)

Keywords:Laser chaos, Reinforcement learning

In proceeding study, we solved the fundamental problem in reinforcement learning called Bandit problem with laser chaos fast and precisely, and in following study, we applicated this learning method to general reinforcement learning and named this novel method as photonic Q-learning. Photonic Q-learning is faster than the conventional method for reinforcement learning called Q-learning. In this study, through further analysis we confirm that the autocorrelation in laser chaos effects the process of learning.