Large-scale distributed Monte-Carlo game-tree search algorithm that can employ different evaluation strategies

2020 Fiscal Year Final Research Report

• Project/Area Number: 17H01846
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Research Field: Entertainment and game informatics 1
• Research Institution: Meiji University
• Principal Investigator: Yokoyama Daisaku 明治大学, 理工学部, 専任准教授 (80345272)
• Project Period (FY): 2017-04-01 – 2020-03-31
• Keywords: 人工知能 / アルゴリズム / ゲーム情報学

Outline of Final Research Achievements

Large-scale search problems in the real world are not applicable to exhaustive search; randomized search algorithms have great ability to explore such problems. Game tree search is an example of such a problem; the Monte-Carlo Tree Search algorithm (MCTS) has been widely used. However, this great advance does not help to achieve good performance in Shogi that has a long-narrow path of `correct’ play. We try to evaluate an algorithm that can employ several different evaluation strategies to improve our previously proposed method. We evaluate the applicability of our method and found several difficulties, such as implementing issues. We also research the applicability for large-scale realistic problems.

Free Research Field

ゲーム情報学

Academic Significance and Societal Importance of the Research Achievements

現実世界には複雑な制約のもとで適切な解を見つけることを要求される、大規模な探索問題が多く存在する。例えば、不完全な情報のもとで最適な戦略を見つける問題などがあり、ゲームをプレイするアルゴリズムを研究することでその問題のエッセンスを考えることが可能になる。本研究は、そのような問題を、現在の高性能な計算機を効率よく活用し、精度よく解くことを可能にするための基礎的な技術を確立することを目指したものであり、広い応用範囲を持つ。