Learning Bayesian Network Structure Based On Incomplete Data

As a special kind of graph models, Bayesian Networks(BN) have become more and more important for representing and reasoning uncertainly knowledge, and it has been successfully applied to a wide range of tasks, such as financial data analysis, machine learning, artificial intelligence, forecast decision and so on. The primary task is successfully establishing network model for actual problem, so after analyzing relevant theories of Bayesian Networks, this paper focuses on the structure learning and proposed a new algorithm. The main works can be summarized as follows:Firstly, after discussing the existing algorithms for the structure learning of BN, the main learning mechanism and their advantages and disadvantages are analyzed and summarized in detail.Secondly, based on the basic ABC and DE algorithm, we proposed the HABC algorithm to learn structure of Bayesian network. Combining the strong exploration capacity of ABC algorithm and the strong development capacity of DE algorithm, we proposed a new nectar update strategy and employed an adapted selection strategy in the observation bee stage. HABC algorithm accelerates the convergence speed, keeps the strongly global optimizing capacity of ABC algorithm, and guarantees the diversity of population through the adapted selection strategy. The numerical experiment results of classic alarm network showed that: HABC algorithm accelerated the convergence speed of algorithm and increased the algorithm accuracy due to its ability of jumping out the local cycle, proving the accuracy and effectiveness of the algorithm.Finally, combining SEM algorithm ideas and the good performance of HABC algorithm, we proposed EHABC algorithm to learn structure of Bayesian network from missing data set. It achieved through the mutual iteration of EM program and HABC search program, adopted a new method to select the initial optimal network, and took the population of former generation as the initial population of HABC search program. The experiment simulation of classic alarm showed that: EHABC algorithm had a greater precision than EBN algorithm, and EHABC algorithm resulted high precision and short running time comparing other algorithms.