Prediction of Execution Probabilities and Attack Probabilities of Werewolf Game Using Machine Learning of Game Logs

Abstract

The topic of this paper is an analysis of Werewolf game. In this game, the players are divided into the villagers’ side and the werewolves’ side. The players erasure the opposite side’s players through discussion and inference. In this paper, execution probabilities and attack probabilities are predicted using machine learning. Five-player games are adopted as the fundamental setting. Game logs were collected from an online chat-style Werewolf game server and analyzed as training data using logistic regression. The authors then analyzed the test data (game logs) with the calculated partial regression coefficients. For both execution and attack probabilities, predictions were appropriate when the self-proclaimed seer appeared. Since the appearance of the seer(s) is an objective piece of information about the situation, it is properly reflected in the prediction. In these predictions, the probabilities indicate which players are certainly executed/attacked otherwise which players are certain to survive. When no seers appeared, because there was little objective information, the execution/attack probabilities of each player changed little. In these situations, since the players can’t decide who to execute/attack, the predictions are not irrational. In short, when the seer(s) (who provide objective information to the field) appear, the predictions tend to be appropriate. In addition, the predictions in this paper are stable with not much (50–80) training data, which is one of the characteristics of this method.