[4Q2-GS-10] AI application:

[4Q2-GS-10-04] AI Werewolf gamelog generation using Transformer Decoder and its application to role estimation

Junya INABA¹, 〇Takashi OTSUKI¹ (1. Yamagata University)

Keywords:AIWolf, Transformer Decoder, Gamelog generation, Role estimation

In the case of role estimation of AI Werewolf, many studies have used feature vectors extracted from game information and trained classifiers to estimate the roles. In this paper, we report on a method to estimate the roles of others using a generation model of the game log. The standard game log of the AI Werewolf platform, which aims to reproduce the game, is not suitable for this study because it contains information that the game participants cannot know. Therefore, we convert it into a "subjective game log" limited to information from the participants' perspective, and train the Transformer Decoder by treating the subjective game log as an artificial language. In addition, a role disclosure section is placed at the end of the subjective game log, and the text before the role disclosure section is given as a prompt to generate the role disclosure section, thereby estimating the roles of others. Experiments were conducted using game logs from past AI Werewolf Competition protocol division finals, and it is shown that this method, which does not extract features, can estimate the roles of others with high accuracy.