[4E1-GS-2] Machine learning: agents

[4E1-GS-2-01] Construction and Analysis of Strategies Using Reinforcement Learning in Incomplete Information Games

〇Shintaro Abe¹, Takashi Takekawa¹ (1. Kogakuin University of Technology and Engineering)

Keywords:Incomplete information game, Reinforcement Learning, Decision tree

Game AI as Shogi and Go in complete information games have defeated top players and have achieved great success. Strong game AI in imperfect information games is attracting attention as a research target. In this research, one of the imperfect information games of " Hol's der Geier " is used as a subject. We experimented with five cards, which are simplified versions of the original environment. Reinforcement learning was used to repeatedly build strategies from games against opponents, with each learning generation discussing the superiority or inferiority of its strategy to that of older generations.We also evaluated the importance of the game information when the strategy determines the action.As a result, the learned strategy has a win rate of more than 60% against all strategies of the old generation. Decision tree analysis showed that the strong strategy decided their actions according to the cards in the remaining