Simplification of Inferences in Multiply Sectioned Bayesian Networks

Multiply sectioned Bayesian Networks （MSBN） support objected-oriented modeling and modularly modeling, which have become an efficient tool for modeling complex giant systems. At present, how to simplify the time and space complexity of local and global inferences in MSBN has become a key problem constraining their applications. In this paper, two classical exact inference algorithms for local inferences in MSBN are analyzed, the identity of the two algorithms is proved, and a unified explanation is given. Then it is proved that the factor determining the complexity of the inferences is the induced width of the induced graph and the class of Bayesian networks on which exact inference can be performed is found. Finally, the feasibility of reducing the complexity of global inferences in MSBN is discussed and some basic principles to simplify global inferences in MSBN are also given.