Research On Hybrid Optimization Methods In Bayesian Network Structure Learning

Bayesian network is based on probability theory and graph theory, using probabilities to represent the form of uncertainty, which has become an effective tool to uncertainty reasoning and data mining. Currently, Bayesian networks have been widely used in many frontier research fields, but for complex dynamic systems, it is clearly not practical to directly construct Bayesian network by experts using the field‘s professional knowledge. So how to learn a dynamic Bayesian network structure from the sample data set has become a hot and difficult research problem in this field. In a variety of Bayesian network structure learning methods, hybrid optimization algorithms catch researchers‘ attention because of flexible strategy-selection, multiple composition methods and combining both strengths of the constraints and score-search based structure learning algorithms.In this article, we did the research from the current shortcomings existing in swarm intelligence learning algorithms, and proposed a number of hybrid optimization methods and improvements to enhance the performance of structure learning algorithm. The main work includes the following aspects:1. We reviewed basic knowledge and research status of Bayesian network and related structure learning, combined with the specific algorithms‘ analysis, especially the definition of dynamic Bayesian network and the current status of its structure learning;2. We proposed MI-BPSO algorithm based on mutual information and binary PSO. According to current swarm intelligence learning algorithms‘ deficiencies including the randomness of initial search points and the poor search performance in discrete cases, the new proposed algorithm improved basic PSO respectively in the following three aspects: the construction method of initial network structure, the generation of initial particle swarm and the optimization strategies of binary PSO. Experiment results showed that the hybrid application of three strategies effectively improved the efficiency and quality of learning algorithm, especially mutual information method greatly reduced the structure learning computational complexity of transfer network in dynamic Bayesian network;3. We proposed modified MIC-BPSO algorithm by optimizing MI-BPSO algorithm in the initial network construction method and the best value updating strategy of particle swarm algorithm. Firstly, we used maximal information coefficient instead of mutual information to construct initial undirected graph, improving the initial network structure quality; Secondly, we refined BPSO algorithm‘s personal best value updating level and swarm best value updating level to parent-child node set by the decomposition characteristic of scoring functions, improving the quality and efficiency of the learning algorithm. Experiment results showed that when the new algorithm was applied in static and dynamic network structure learning, the learning quality and time efficiency were both greatly improved.