紀要論文

モンテカルロ木探索に基づいたトレーディングカードゲームプレイヤー

Monte-Carlo Tree Search based Player for the Trading Card Game

jpn

18810667

https://doi.org/10.15002/00010533

法政大学大学院情報科学研究科

9

125

130

2014-03

Version of Record

Artificial intelligence (AI) games have been topic for a long time, even now. However, implementation techniques of AI games for two players: trading card game such as Magic the Gathering, have not been established still. Also, it is hard to say accuracy of the computer player named CPU can reach a high accuracy. I think the trading card games have three problems. First, it's incomplete information game. Second, it's difficult to make the evaluation function, because history is shallow and tactics is complex. Third, it has larger state space than the existing game studied because such as the type of card is enormous. Therefore, this research proposes the implementation of CPU to incorporate the Monte Carlo tree search in trading card games. As mentioned earlier, Monte Carlo tree search is an effective method in a game that is hard to make the evaluation function. By performing the compression of the game attributes, to reduce the state space size, the CPU learns from the data of the competition game results, improves the search speed of the Monte-Carlo Tree Search, and reduces the CPU's think time. As a result, off against the rule-based type CPU and Monte Carlo tree search based CPU, a result of the 2000 races, the winning rate of the Monte Carlo tree search based CPU is 61.7%.

Article

AA12222297