Theoretical research of the policy gradient reinforcement learning without Markov properties and its application to games

2016 Fiscal Year Final Research Report

• Project/Area Number: 26330419
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Entertainment and game informatics 1
• Research Institution: Shibaura Institute of Technology
• Principal Investigator: Harukazu Igarashi 芝浦工業大学, 工学部, 教授 (80288886)
• Co-Investigator(Renkei-kenkyūsha): ISHIHARA Seiji 東京電機大学, 理工学部, 准教授 (50351656)
• Research Collaborator: MORIOKA Yuichi ; YAMAMOTO Kazumasa
• Project Period (FY): 2014-04-01 – 2017-03-31
• Keywords: 強化学習 / 方策勾配法 / マルチエージェント / コンピュータ将棋 / ロボカップ / ソフトマックス探索

Outline of Final Research Achievements

In this research project, we have made theoretical and practical research for developing expressions of policy functions and learning methods in the policy gradient reinforcement learning algorithms. Our final goal is constructing a general methodology that can be applied to computer games and engineering fields.　The results of this project are as follows.
(1)Theoretical research on the policy gradient reinforcement learning: we proposed new methods in ①hierarchical reinforcement learning to learn higher strategies of agents, ②learning with separated knowledge of environmental dynamics and action-values in agent policies, and ③learning with a fuzzy controller for policies.
(2) Practical application of the policy gradient reinforcement learning: we applied the proposed learning methods to pursuit games, robot soccer games and computer shogi and examines the efficiency of our methods.

Free Research Field

人工知能