Analysis of Voting Tactics of Agents Participating in The AIWolf Competition Using Decision Tree Surrogate Model

Abstract

While the ability to cooperate through persuasion is the key technology of AIWolf agent, the current agents do not have this ability yet. For realizing this ability, the agent is required to infer thinking models of others and persuade them according to the inferred models. However, in recent AIWolf Competition in which many agents determine their actions using machine learning, it is difficult for agent to persuade others because of lack of explainability of the determined actions of others and itself. In this study, focusing on the voting action performed by all living agents, we attempt to explain the voting algorithm of the influent roles such as seer and werewolf using a decision tree as a surrogate model. As a result, it is shown that, even in case where tactics are not explicitly programmed such as machine learning agent, the voting tactics can be explained using the decision tree.