Designing AIWolf agents by rule-based algorithm and by deep Q-learning

• Main Author: Zhang, Shengjing
• Other Authors: Bo An
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/149340
• Institution: Nanyang Technological University

The Werewolf game is a popular party game with imperfect information. Players do not know others’ roles, but they must eliminate all the opponents before their teams are all killed or voted out. Designing an intelligent agent to play such kind of game well is a challenging topic for researchers all around the world. The International AIWolf Competition is held every year for participants to design such AIWolf agents and compete with each other. This report designed two agents by using two different methods and compared their performance with champion agents in the previous competitions. The first method was to use the rule-based algorithm. The proposed agent was required to follow a series of rules set before games. It did role estimation first to deduce other players’ roles and find out its opponents. Next, based on the estimation of others’ roles and rules set in advance, the proposed agent chose the best strategy and made a decision on who to vote, who to attack, what information to exchange with others, and so on. The second method was to use reinforcement learning. The proposed agent would first train a deep Q-network by taking some states as input and outputting Q-values of various actions. The neural network could help the agent find out the optimal actions to take at the current state. By calculating average winning rates of each agent in 200,000 games, the results showed that the proposed agent using deep Q-learning had the best performance among all the other agents, including champions in the previous competitions and the agent using the rule-based algorithm. Reinforcement learning is highly recommended when building intelligent agents for AIWolf games.