Presentation Information
[4Yin-B-06]Vote Prediction of AI Werewolf Using a Transformer Trained on Subjective Game Logs
〇Masaki Takahashi1, Takashi Otsuki1 (1. Yamagata University)
Keywords:
AI Werewolf,Transformer,Behavior Prediction
In AI Werewolf, predicting other players' votes is essential to decision-making. While previous studies on inference of AI Werewolf have mainly relied on feature engineering based on utterance and action, recent advances in large language models have enabled inference directly from history of utterance and action. This study constructs subjective game logs by restricting game logs of AI Werewolf to information available from an individual player's perspective. By treating these logs as an artificial language, a Transformer-based model is trained to predict other players' vote by generating the following text including next action from the preceding text. Experimental results using the logs of past AI Werewolf competitions show that the proposed method can predict other players' votes with high accuracy.