Learning and Inferences of the Bayesian Network with Maximum Likelihood Parameters

In real applications established on Bayesian networks (BNs), it is necessary to make inference for arbitrary evidence even it is not contained in existing conditional probability tables (CPTs). Aiming at this problem, in this paper, we discuss the learning and inferences of the BN with maximum likelihood parameters that replace the CPTs. We focus on the learning of the maximum likelihood parameters and give the corresponding methods for 2 kinds of BN inferences: forward inferences and backward inferences. Furthermore, we give the approximate inference method of BNs with maximum likelihood hypotheses. Premilinary experiments show the feasibility of our proposed methods.