Reinforcement Learning based Single AI trained from Multi-game
2
조회
0
다운로드
1
인용
0
참고문헌
키워드
초록
Recently, AI Technologies have grown rapidly in various fields including game. There are many approaches to game AI. First, the supervised learning-based approaches are learned from the game play data, and mimic a series of the playing behavior. However, there are limitations in the complicated games due to the linear combination of input features. In order to solve these limitations by the linear combination, the deep neural network-based approaches utilizes more than two neural networks to represent the local characteristics and the global characteristics respectively. Still, these approaches require the enough training data. Without any already prepared training data, the reinforcement learning based approaches can be learned from the relations between the agent action and its reward in the environment, and aimed at that the agent obtains the maximum reward. In this paper, we propose a single AI trained from multi-games by the reinforcement learning. In the proposed AI model, every local agent plays each single game, while the global agent is learned from all local agents. Experimental results show that the agent learned from multi-games is adjusted well on both the training game and a new game, while the agent learned from a single game achieves good performance at only the training game.
Abstract
Recently, AI Technologies have grown rapidly in various fields including game. There are many approaches to game AI. First, the supervised learning-based approaches are learned from the game play data, and mimic a series of the playing behavior. However, there are limitations in the complicated games due to the linear combination of input features. In order to solve these limitations by the linear combination, the deep neural network-based approaches utilizes more than two neural networks to represent the local characteristics and the global characteristics respectively. Still, these approaches require the enough training data. Without any already prepared training data, the reinforcement learning based approaches can be learned from the relations between the agent action and its reward in the environment, and aimed at that the agent obtains the maximum reward. In this paper, we propose a single AI trained from multi-games by the reinforcement learning. In the proposed AI model, every local agent plays each single game, while the global agent is learned from all local agents. Experimental results show that the agent learned from multi-games is adjusted well on both the training game and a new game, while the agent learned from a single game achieves good performance at only the training game.
참고문헌 (12)
- Max Jaderberg, "Human-level performance in 3D multiplayer games with population-based reinforcement learning", Science, 364, pp.859~865 (2019)
- Volodymyr Mnih, "Human-level control through deep reinforcement learning", Nature, 518, pp.529~533 (2015)
- 문성원, "퍼지 확장 기법을 이용한 온라인 게임에 적합한 지능적 AI 기법", 한국게임학회 논문지, 8(3), pp.77~85 (2008), 한국게임학회
- 이상호, "스타크래프트 인공지능의 성능 향상을 위한 적응성있고 비용 효율적인 빌드 오더 선택 기법", 한국컴퓨터게임학회논문지, (14), pp.137~148 (2008), (사)한국컴퓨터게임학회
- Murray Campbell, "Deep Blue", Artificial intelligence, 518, pp.57~83 (2002)
- Jonathan Schaeffer, "A world championship caliber checkers program", Artificial Intelligence, 53, pp.2~3 (1992)
- Gerald Tesauro, "On-line policy improvement using Monte-Carlo search", Advances in NIPS, pp.1068~1074 (1997)
- Brian Sheppard, "World-championshipcaliber Scrabble", Artificial Intelligence, 134, pp.241~275 (2002)
- Markus Enzenberge, "FUEGO – An Open-source Framework for Board Games and Go Engine Based on Monte-Carlo Tree Search", IEEE Transactions on Computational-Intelligence and AI in Games, 2, pp.259~270 (2010)
- David Silver, "Mastering the game of Go with deep neural networks and tree search", Nature, 529, pp.484~489 (2016)
- Yann LeCun, "Deep learning", Nature, 521, pp.436~444 (2015)
- David Sliver, "Mastering the game of Go without human knowledge", Nature, 550, pp.354~359 (2017)