Study on the mechanism of structure-variable dynamic Bayesian networks

Traditional dynamic Bayesian networks (DBNs) are essentially models that describe a variety of stable processes. To deal with unstable processes, structure-variable dynamic Bayesian networks are more applicable, flexible, and effective. Currently, however, the various inference algorithms under consideration for structure-variable discrete dynamic Bayesian networks (DDBNs) can only handle hard evidence. In this paper, an in-depth and theoretical analysis is given for the mechanism and key characteristics of structure-variable dynamic Bayesian networks, and on this basis, a fast inference algorithm is proposed. Furthermore, a special class of structure-variable dynamic Bayesian networks, dynamic Bayesian networks with missing data, is defined rigorously along with associated network topology and parameter settings of such networks. Several experimental simulations have shown the effectiveness and efficiency of our fast inference algorithm.