Examination of AIWolf focusing on expression of lies

Abstract

Agents playing werewolf games, called AIWolf, has important roles in artificial intelligent field. Although lies are one of the important factors in werewolf games. few studies have investigated how to tell lies in AIWolf in werewolf games. In this study, we investigated whether the way to lie in AIWolf increase/decrease the winning rate of the game, through simulation experiments. Under the set conditions, three types of lying programs were prepared; lying certainly, lying uncertainly, and not lying. The winning rate for each lying method under each condition was calculated. Logistic regression analysys did not show significant differences in the winning rate between several pairs of each condition to lie. This result implies that AIWolf have not been made as they can consider individual remarks deeply.