Improvement of Role Estimation using Machine Learning in werewolf AI

Abstract

Role estimation by existing werewolf AI agents had been heuristically decided by the developer. In this research, we introduced an agent using SVR (Support Vector Regression), which is one of machine learning, to the estimated part of the existing agent and verified its effectiveness. As the result, it was possible to realize estimation rates exceeding the existing method after the middle game. Also, we developed an combined agent that used the existing method in the opening game and the proposed method in the middle game and later. The combined agent gave a significantly higher win rate than the existing agent.