Study Of Evaluation Algorithm In Imperfect Information Game

Game theory is a very important branch in the field of artificial intelligence research. It can be divided into two categories, one is perfect information game, and another is imperfect information game. In perfect information game, players are aware of all the information about the game, while in imperfect information game, players only know part of the information set. Great progress has made in the research of perfect information game. However, for the study of imperfect information game, there is a long way to go. In recent years, artificial intelligence, computer clusters and distributed computing technology continues to develop, the imperfect information game research has a great opportunity for development. Evaluation algorithm is an important part of imperfect information game. In game system, one strategy is good or bad is determined by evaluation algorithm. The performance of the evaluation algorithm is the core factor that can directly affect the level of intelligence in the game system.Texas Hold’em is a typical imperfect information game. In this paper, we take computer Texas Hold’em as experimental subjects, and study the evaluation algorithm in imperfect information game. In the field of game theory, the artificial neural network algorithm is generally used to predict the opponent’s action. On the basis of previous research, this paper improved artificial neural network model and used experts’ playing records to train this model. This model learns expert’s strategy and takes that as a reference for their own actions. Game system can easily be modeled by opponent’s analysis if this system only learn from a single expert. The opponent’s specifically attack to this type of game system will result in poor system performance. Therefore, in this paper, decisions are made by sampling from multiple experts whose weights were dynamically updated. This method can effectively prevent the opponent modeling and improve the intelligence level of the agent. With the improvement of computing, deep learning algorithm which based on artificial neural network algorithm obtains a large-scale application. This paper proposes deep learning algorithm to Texas Hold’em to characterize the game and predict the behavior of the opponents.The final implement of the computer poker agent has a high level of intelligence. The agent’s decisions are made by a combination of a variety of evaluation algorithms discussed in this paper. Then the agent participated in the 2014 AAAI Computer Poker Competition and achieved a bronze medal.