Implementations of Deep Reinforcement Learning on Board Games
- 주제(키워드) Reinforcement learning , DQN , Double DQN , Dueling DQN , board games
- 발행기관 서강대학교 일반대학원
- 지도교수 김종락
- 발행년도 2019
- 학위수여년월 2019. 2
- 학위명 석사
- 학과 및 전공 일반대학원 수학과
- 실제URI http://www.dcollection.net/handler/sogang/000000063994
- UCI I804:11029-000000063994
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권보호를 받습니다.
초록/요약
Recently the reinforcement learning has increased in popularity especially in the game field. However, most of the researches are on video games such as Atari game or Action Real-Time Strategy game. Board games are less featured research fields. This is due to the lack of useful board game environments for the reinforcement learning. This paper suggests an efficient way to construct environments of six online mathematical board games (Cube Net Game, Rectangle Game, Right Triangle Game, etc.) on MaTricKing[8]. Furthermore, DQN, Double DQN and Dueling DQN with several variations in the structure of networks are applied on the games of a board size 5 times 5. For each game, we figure out which structures of networks and hyper parameters are suitable for deep reinforcement learning. We also find a way to shorten learning time by training models only with last few moves. Finally, for six games, we implement deep reinforcement learning so that trained models perform as well as user-level of MaTricKing.
more