Estimating Roles with Utterance Vectors in Werewolf Game

Abstract

Werewolf Game is a team game in which several secret roles are given to the players and the players try to estimate other players’ role through discussion. For developing the game agents, some studies have used the Long Short-Term Memory (LSTM) which is appropriate for handling the sequence of utterances. The inputs given to LSTM in these studies are usually the one-hot vectors of the utterances. However, they have a problem that the dimension of the input vector increases when the variety of utterance increases. In this paper, we propose a game agent that represents the utterances with the embedded vectors derived from the word2vec algorithm. We also implemented the game agent that estimates all players’ roles. The experimental results show that we can reduce the dimension of the utterance vectors without decreasing the role estimation accuracy.