Reasoning And Learning Of Chain Event Graphs

The main research contents of the probabilistic graphical models include the representation theory, the reasoning theory and the learning theory. Based on different representation theory, people put forward different models to meet their needs of various problems. In the traditional probabilistic graphical model, the nodes are used to represent the variables, and the edges are used to represent the dependence of variables. Chain Event Graphs is a new class of probability graphical model. Compared with the classical Bayesian Networks, Chain Event Graphs can be more precise and comprehensive to describe the problem, more simple and convenient to carry out probabilistic reasoning. In 2005, Jim Q. Smith et al first proposed the concept of Chain Event Graphs, which caused domestic and international attention.The pre research focus of Chain Event Graphs is to improve its representation theory, and the present work focus on the model learning algorithm. The existing structure learning algorithms are based on the evaluation and model optimization to select the optimal model from the structural space. The advantage of this kind of algorithm is that results are accurate and reliable, the disadvantage is that the algorithm complexity is relatively high and the scope of application is small. To solve these problems, the main work of this paper is as follows:A further study of the representation theory of Chain Event Graphs is made, and the related terms and definitions of Chain Event Graphs are introduced from the foreign literature, and the structure principle is explained. In addition, this paper also has some research on the reasoning theory of Chain Event Graphs.In this paper it is proposed that an innovative structure learning algorithm based on the independent detection of the contingency table. The feasibility, correctness and efficiency of the new algorithm is verified by theoretical and practical experiments. Different from the existing AHC algorithm, the new algorithm does not use the idea of scoring learning, rather use the independent testing to determine the stage partition of the event tree. The theory that stage division can be deduced to position division, is used in the new algorithm. Compared with existed algorithms, the proposed algorithm can greatly reduce the time complexity, the average time was about 5.25% of AHC algorithm's. At the same time, the accuracy of thenew algorithm is proved to improve quickly with the increase of the amount of data.