Improvement and Application of Monte Carlo Tree Search Algorithm on Computer Games

• Main Authors: Cheng-Wei Chou, 周政緯
• Other Authors: Shi-Jim Yen
• Format: Others
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/42017145719654914276

博士 === 國立東華大學 === 資訊工程學系 === 102 === Monte Carlo Tree Search (MCTS) is the most popular algorithm in computer games field in recent years. This algorithm is very efficient for computer Go and improves the strength of computer Go program amazingly. This algorithm is also used to other games, even the problem of real world, for example, power management. MCTS is applied and ameliorated in several directions in this article. First, this article tries to use MCTS to Dark Chess, an imperfect information game, and very popular in Chinese culture. Second, this article tries to use the framework of MCTS to build a self-learning method. The self-learning method could greatly improve the strength of program in the game of NoGo, by counterbalancing the lack of domain knowledge by self-learning. Finally, this article focuses on a specific weakness of MCTS. MCTS does not share information between different branches. This article proposes an online learning method to ameliorate it.