Research On Uncertain Knowledge Inference And Its Applications In Bayesian Networks

In daily life, people need to handle a large amount of uncertainty problems. Bayesian networks (BN) is the effective tools to represent uncertainty problems, which visually expresses the causal relationship between probability events by utilizing directed graph on one hand, and calculates probability distribution of events by using Bayesian statistical theory on the other hand. This thesis carried out the deep research on open problems in Bayesian network classifiers and inference algorithms for BN. The main jobs are as follows:(1) This thesis does a systematic analysis and discussion on Bayesian network theory, and summarizes the advantages and characteristic of Bayesian networks compared with other methods of data mining. According to the difference of research approaches and ideas, Bayesian network learning is intensively summarized and the advantages and weaknesses of various kinds of methods are pointed out.(2) Na?ve Bayesian classifier (NBC) with its concise structure and excellent performance has been extensively studied, however, because of its requirement of conditional independence between attributes, its application in classification is limited. This thesis presents the GA-NBC-TAN algorithm which combines feature selection with structure augmentation. Compared with the same number of attributes of NBC, GA-NBC-TAN algorithm improves the classification accuracy.(3) Bayesian network inference is the first problem to be solved in face of application of BN, However, exact inference and approximate inference in BN are NP-hard problems. Junction tree algorithm (JTA) is a commonly used and effective exact inference in BN, but because of the conversion between BN and JT is not unique, finding the optimal JT is proved to be NP-hard problem. An improved adaptive genetic algorithm (AGA) to address the nodes eliminating sequence in triangulating BN is proposed to obtain optimal JT. Compared with the standard genetic algorithm, AGA-Triangulation algorithm shows better performance. Based on the work in the above, LAZY-ARVE algorithm which utilizes optimal JT is discussed and compared with LAZY-AR algorithm, promoting greater efficiency. This method can solve general queries and queries with evidence.