Applying Monte Carlo CFR to the 5-player Werewolf game

Abstract

In this paper, we applied the MCCFR, a method for calculating strategies of imperfect infor- mation games, which is successfully used in poker, to Werewolf game, a multiplayer imperfect information game. The Werewolf game, a multiplayer imperfect information game, has gotten much attention in recent years. We reduced game states based on the established tactics of the 5-player Werewolf game and applied the Monte Carlo Counterfactual Regret Minimization (MCCFR) algorithm to the Werewolf game. The MCCFR is a variation of Counterfactual Regret Minimization (CFR), one of the methods to calculate strategies of imperfect information games.Furthermore, we constructed an agent played according to the learned strategy and evaluated the performance of MCCFR in the Werewolf game by obtaining the winning rate against agents who participated in the AIWolf competition. As a result of learning with MCCFR, we confirmed that the learned strategies obtained a higher winning rate than the unlearned strategies. It showed that the CFR-based algorithm was effective in Werewolf games. As a result of learning by MC- CFR, we confirmed that the learned strategies had a higher winning rate than the unlearned strategies. Therefore, it is suggested that the CFR-based algorithm is effective in Werewolf games.