贝叶斯网络结构学习方法研究

介绍贝叶斯网络结构学习的两种方法：渐进正确结构学习和启发搜索方法,并重点对最小描述长度测度函数MDL和DGM测度函数进行了分析比较,提出了一个两阶段从数据集中学习贝叶斯网络（BN）结构的有效算法.首先,第一阶段采用以信息论为基础的渐进正确的结构学习方法来有效的搜索可能的网络空间,构造一个候选网络;然后,采用启发搜索的方法来精简这个候选网络.使用这种分两阶段学习的方法比单纯使用一种方法来获取BN结构更精确和有效率.

Research of learning Bayesian networks structure

Two efficient algorithms for learning a Bayesian network (BN) are introduced. A comparison between two BN structure learning-algorithms is made. Two BN structure fitness functions are analyzed. A two-phase algorithm is proposed. The first phase uses asymptotically the correct structure learning for efficient search space exploration and a candidate network is found quickly. Then a heuristic search may be done to refine the network. In this way a BN network is derived with more computational efficiency than a pure heuristic search is used. At the same time a more accurate structure is derived than a pure asymptotically correct structure learning algorithm is used.