[4Xin1] Poster session 2

[4Xin1-65] AIWolf Agent Using DNN to Predict Voting Behavior of Others

〇Takashi OTSUKI¹, Hirotaka EBE² (1.Graduate School of Science and Engineering, Yamagata University, 2.Department of Informatics and Electronics, Graduate School of Science and Engineering, Yamagata University)

Keywords:AI Werewolf, Werewolf Game, Incomplete Information Games, Behavior Prediction, Deep Neural Network

In the research field of AI that plays the werewolf game, so-called "AIWolf", there have been many research reports on role estimation of others by machine learning. There is no doubt that the role estimation ability is a major factor that determines the outcome of the werewolf game, but there are many other important factors such as deception, persuation, and predictiting the actions of others, that have not been well studied. In this report, we describe the agent that decides the final vote destination by using the prediction whether other players will vote as declared, and show the results of the evaluation experiment. First, we constructed a predictor of voting behavior using a 4-layer DNN. Next, we implemented a tactic in the seer agent and the werewolf agent that uses the voting behavior prediction results to decide who to vote when they are likely to be expelled by voting. Finally, the relationship between the presence or absence of this tactic and the winning rate was clarified by simulation using finalist agents of the 3rd International AIWolf Competition.