Incorporating Argumentation into Reinforcement Learning for Werewolf Game

Abstract

We introduce a method that incorporates an argumentation framework into reinforcement learning (RL) for Werewolf Game. Werewolf Game is a type of hidden role games. Hidden role gams are the games where some player do not know who other teammates are. In hidden role games, conversation is important to know each player's position and mind. However, conducting hidden role game's conversation by natural language is difficult for agent. Therefore, we introduce a conversation that empowers agent's playing hidden role games. In specific, we introduce a method that address the conversation process by using argumentation framework and feeding it as state for a RL agent. In this context, we are assuming a 3-player's Werewolf Game where players are rule-based agents that consider the situation of conversation well when they decide their action. For the sake of evaluation, we show the RL agent's win rate and easy analysis in the current research situation.