Learning and approximate inference of DBNs

In this paper, the techniques for learning and approximate inference of dynamic Bayesian networks (DBNs) are studied. With respect to the structure learning of DBNs, the EM-EA algorithm is introduced into the learning process of DBNs and the flow of learning DBNs is given out in the presence of the incomplete data and hidden variables. As for the DBNs inference, the difficulties are analyzed that apply directly the standard inference techniques to DBNs. Furthermore, two improved algorithms of likelihood weighting (LW) algorithm, ER algorithm and SOF algorithm are introduced and assembled. The experimental results show that these algorithms can effectively overcome the disadvantages of the LW algorithm, and especially, the assembled algorithm gives a outstanding performance.