Playing Geister by Estimating Hidden Information with Deep Reinforcement Learning

Abstract

A number of attempts have been made to solve imperfect information games using reinforcement learning. Since it is not easy to apply reinforcement learning directly to imperfect information games, methods, such as Neural Fictitious Self-Play (NFSP), have been proposed. In this study, we investigate an alternative method to solve imperfect information games using reinforcement learning. The basic idea is to learn from self-playing the perfect information games where the hidden information in the original game is revealed. Based on the policy obtained from this learning, the current state of the game is estimated from the opponent's moves and the own best move. We applied the proposed method to an imperfect information game called Geister and evaluated its effectiveness through experiments. As a result, we obtained that the winning rate was higher than that of the direct learning of imperfect information games using reinforcement learning.