The Research Of Bayesian Networks Learning Based On Domain Knowledge

The Bayesian network is the directly graphic representation, it vividly describe the corresponding field thing's degree of relative. The Bayesian network can reason and predict some thing, and it is a main method dealing with uncertain fact. Usually experts can construct a fitting network according to their knowledge, however, constructing network is time-consuming and labor-consuming for complicated networks. If an expert was careless, he will construct a wrong network. We can not acquire enough data by observation in the world. Although enough dada is obtained, we couldn't make the learning algorithm converge to the correct Bayesian network without domain knowledge. Because of the method considered above, domain knowledge is introduced in the learning algorithm in this dissertation in order to improve the precision and the speed effectively. The main contents of this dissertation are as follows:(1) This dissertation makes a survey about the research on Bayesian networks, including the background, the current research state and development trend of Bayesian networks, the basic principle of Bayesian networks, the introduction and analysis of the classic algorithms for inferring and learning of Bayesian networks.(2) The EM algorithm is the main one for Parameter Learning of Dynamic Bayesian Networks, but the shortage are its low convergence and low efficiency .To overcome the drawbacks, large temporal datasets is divided into small blocks .Thus an accelerating EM algorithm based on partial E-step, DA-EM, is proposed . By iterating through the blocks in a cyclic way, DA-EM can guarantee the precision of EM, make it converge faster and compute more efficiently. Our experimental results justify the method.(3) The SEM algorithm can be utilized for learning Bayesian network structure with incomplete data, but great training dada is required to improve some traits, such as lower precision, slow convergent speed and halting at local optima. Moreover, enough data can hardly be collected in reality. KB-SEM is proposed by introducing domain knowledge in this dissertation. A tabu list derived from collecting domain expert knowledge synthesized by using D-S evidence theory is embedded in the SEM to constrain and guide the searching path, and to shorten the searching space. Experimental result show that KB-SEM can improve the precision and the speed effectively, that it can to some extent avoid subjective bias and disturbance of noises in the data sets. Thus, a Bayesian network satisfied by experts is acquired finally.