Research And Application About A Dynamic Prediction Model Based On Bayesian Networks

Game Theory is a mathematical theory that corresponds the concept of competition in the reality,and recently earns a wide application in the fields of AI(artificial intelligence). Incomplete informa-tion game is one of the largest uncertain case, and the most worthy to study. The significance of thesimulation for this respect of AI is obviously important.Bayesian prediction is a reasoning approach, based on probability distribution, in which combinethe graph model of the Bayesian theory, and through conditional independence assumption, simplif-y complex issues to show the interaction between the incidents that affect each other with a map, fi-nally calculate the prior probability and posterior probability to solve the problem. The features andbenefits of Bayesian methods make it not only to solve problems related to probability theory (suchas classification clustering), but also to apply the graph model for strategy selection problems (suchas reasoning and recognition).However, when Bayesian prediction is applied to the game theory it shows some problems and i-ssues to be extended. In order to compensate for these, we introduce MAS (Multi-Agent System),that has a good prospect in the future. And so we can combine the dynamic prediction model and t-he agent which presents the entity with intelligence to sufficiently predict the future.After integrally considering the good prospect in simulated reality and the good performance ofBayesian methods, we design a dynamic prediction model in the state of incomplete information g-ame, to make it more universal and better map with reality.The main jobs of this dissertation:(1) Construct a dynamic prediction model structure. As for the lack of considering the uncertainfactors of the surrounding environment through Bayesian methods, we make its application to theMAS in incomplete information game by the use of agent structure theory of perception, and thenconstitute the whole framework combined with agent.(2)Execute an improved learning algorithm. Through deeply research on Bayesian methods, wefind that a Bayesian learning network is basically aptotic. Even for Bayesian learning network afterexecuting the dynamic Bayesian methods, because of the fixed relationship between the previous f- actors and the latter factors in time, it’s difficult to update the whole network. So combine the K2a-lgorithm and the related algorithms to form an accurate learning algorithm. The algorithm is simple,feasible, and high accurate. Full account of the conditional independence assumption, enhance thelearning accuracy while consuming in time as little as possible.(3)Advance a method to judge the accuracy of prediction. During the prediction, there is no valu-ation of the prediction results to know whether the results are accurate, and can’t adjust the predicti-on method. As to Bayesian method, no matter dynamic way or not, there is no good feedback, so t-hat the performance of the dynamic prediction results is not very good. So introduce the concept ofprediction error to judge the prediction results to make the method changing itself though it.(4) Introduce the concept of perception rank in the process of Bayesian prediction. There is no p-rimary and secondary reasoning in Bayesian method, so it can’t identify the incident that most needto be predicted. In this case, it can’t predict in time and sufficiently. So we can use the rank to solvethe problem.