Human-Side Strategies in the Werewolf Game Against the Stealth Werewolf Strategy

Abstract

The werewolf game contains unique features, such as persuasion and deception, which are not included in games that have been previously studied in AI research. Studying the werewolf game could be one of the next challenging targets for AI research. In this paper, we concentrate on a werewolf-side strategy called the “stealth werewolf” strategy. With this strategy, each of the werewolf-side players behaves like a villager, and the player does not pretend to have a special role. Even though the strategy is thought to be suboptimal, so far this has not been proved. In this paper, we limit the human-side strategies such that the seer reveals his/her role on the first day and the bodyguard never reveals his/her role. So, the advantage of the werewolves in determining the player to be eliminated by vote is nullified. We calculated the $\varepsilon$-Nash equilibrium of strategies for both sides under this limitation. The solution shows that the winning rates of the human-side are more than half when the number of werewolves is assigned as in common play. Since it is thought to be fair and interesting for the winning rate to stay near 50%, the result suggests that the “stealth werewolf” strategy is not a good strategy for werewolf-side players. Furthermore, the result also suggests that there exist unusual actions in the strategies that result in an $\varepsilon$-Nash equilibrium.