Parameter learning of discrete dynamic Bayesian network with missing target data

The difficulty of discrete dynamic Bayesian network parameter learning lies in: obtaining the transition probability of hidden nodes between slices, lack of observational data in varying degrees. Focusing on the above problems, the forward recursive parameters learning algorithm based on target data missing estimation is proposed. The algorithm uses the correspondent relation between the observed variables and hidden variables in discrete dynamic Bayesian network, using support vector machine to establish a nonlinear function between observed variables for completing the missing data estimation. A complete data set and the forward recursive algorithm are applied to complete parameters updating in inter-slice and in-slice. On the background of aerial target recognition, the advantages of the proposed method at efficiency and accuracy are illustrated compared with the expectative maximization method.