Research And Realization Of Game Playing Based On Machine Learning

With the development of the computer technology, the research of artificial intelligence shows the exciting activity. More and more experts pay attention to game playing that is an important aspect of artificial intelligence. But many programs of game playing have low speed to search and have not got machine-learning ability or powerful learning ability, which reduce the value of research and utility of these programs. This paper researches and discusses game playing and improves the search algorithm to make it more quickly and efficiently. Combining BP nerve netwok with genetic algorithm to form a mix learning method that can be used in program of game playing. Designing a Chinese Chess program on the base. This program not only has the fast search speed, but also has excellent machine learning ability. Experiment shows that after a small quantity of training and learning, the program has strong ability to play chess, which achieved the goal.In order to make computer play chess, the paper applies the given data structure to transform the knowledge involved into the form that the computer can understand. Applying the state space method to describe the course of game playing to make computer hold the chang of game state quickly. In aspect of rule of chess moving, the paper concludes the production rule of chess moving according to the kownledge of Chinese chess. Computer can produce chess noving more conveniencely and quickly through these rules.To game playing, the design of search algorithm is an important problem, which affect the search efficiency. This paper researches and analyses some basic search algorithm, for example, maxmini search algorithm and alpha-beta search algorithm, and indicates their shortcoming. Improving alpha-beta search method according to peoples’ kownledege and experience of playing chess. Introducing heuristic imformation involved and optimizes the course of pruning to form two new search algorithms. These algorithms are more quickly and efficiently than primary algorithms.Evaluate function is a standard of evaluating situation and making it have excellent machine-learning ability is a difficulty of game playing. This paper analyses the evaluate function model in existence structured by BP nerve network and indicates its shortcoming. On the base, using victory or defeat and integrated value of different chessman to structure a new evaluate function model. Data quantity of this model is less than the primary model obviously. Then the paper introduces genetic algorithm and combines it with BP nerve network to form a mix algorithm. It can optimize the initialize weights of BP nerve network, quicken the speed of constringency and strengthen the ability of learning. Avoiding the disadvantage that training time of the primary model is very long, weak ability of learning and liable to get into local minimum, which shows superiority of the new model and algorithm.Experiment shows that the program can search quickly and have a good ability of machine learning. After a small quantity of playing with others, this program not only can defeat the amateurish chess players, but also can predominate when playing with chess software (Chinese Chess QiBing3.0) that have greater ability. The program achieves the anticipative goal as a whole.