A Ms. Pac-Man Controller Based on Monte-Carlo Tree Search

• Main Authors: Hui-Wen Cheng, 程慧文
• Other Authors: Jyh-Jong Tsay
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/45457299652813694441

碩士 === 國立中正大學 === 資訊工程研究所 === 101 === Ms. Pac-Man is a popular video game, an advanced version of Pac-Man. Fast player reactions and intelligent playing strategies is needed to get high scores, with lots of hidden game information a player can’t reach directly, Ms. Pac-Man has been a challenging game for both human players and computer controllers, and thus offers good subjects for AI computing study. The Ms. Pac-Man Competition, first held in 2007, offers a “fair” platform for different approaches to compare their performances with each other. In 2011, the first Monte Carlo Tree Search (MCTS) based Ms. Pac-Man Controller won the competition, breaking the record of the Hand-Coded, Rule-Based (often performs better than AI ones in competition) Controller, achieving a great success. Since then, several researches using different forms of the MCTS Algorithms on Ms. Pac-Man or ghost control also contribute very good work. In this paper we try to combine advantages of these different MCTS approaches to make an AI controller that can achieve a score closer to the human record. The maximum score of our agent is 20,810, with an average score of 13,232. Although the results are not as expected, we still proposed possible causes and future directions for improvement.