Study On Game-agent Based On Reinforcement Learning

At present the games centralized in artificial intelligence are favored by people. Through applying reinforcement learning in artificial intelligence, a multi-player Chinese checkers game system with the capabilities of intelligent decision-making and autonomic learning is constructed for Chinese checkers game. The architecture of the system is an extended system based on the traditional game pattern, which is different from the traditional Chinese checkers game.According to the functional interfaces, the system can be divided into three parts: the testing subsystem, the agent self-learning subsystem and the human-agent game-playing subsystem. On the early stage of the system construction, a testing subsystem is first constructed in order to periodically understand the cooperative working situation between different components in the process of system construction. With continuous experience accumulation in interactive play, the agent self-learning subsystem is mainly used for autonomic learning of game-agent to improve playing capability. The agent self-learning subsystem implements the agent self-play training mode and the human-agent training mode. Through observing the human's opening move strategy in the human-agent training mode, the system creates an opening move strategy knowledge base and comprehensively applying it later in play. On the basis of theories of reinforcement learning and BP, the system perfects the capability of playing Chinese checkers through online learning mode of self-play in order to increasingly improve the accuracy of the nonlinear evaluation function of the game implemented by a neural network in the agent self-learning subsystem. The human-agent game-playing subsystem uses the different level game-playing modals learned in the training mode to play in multi-player games.The design ideas for the game board and game-agent are mostly discussed in the field of system design. In the game board design, a structure model of the board is given and the board divisional concept and the definition of game phase are brought forward. An intelligent decision-making and learning modal of agent is put forward with confirmation of the deliberative framework of the game-agent in the game-agent design, which discusses the functions of main components in detail respectively. The network structure of the evaluation function of the game implemented by a neural network and the choice board features are also provide in the game-agent design. In the field of system implementation a multi-document application is constructed which is implemented by MFC (Microsoft Foundation Class) of Visual C++ and based on the agent-oriented programming idea.In the last section of this paper a scalable prototype is investigated according to the characteristics of Chinese checkers from the point of view of considering human play, which concerns how a limited rational deliberative agent can observe and learn the human opponent's play mode in the interactive play and integrate them to guide the process of play. The whole prototype is elementary framework for further study. Some pertinent problems are proposed for reference and explore. An embattling modal for decision making of game-agent is put forward. The steps for decision making based on the embattling modal are also summarized.