Approximations of Bayesian networks through KL minimisation

Exact inference in large, complex Bayesian networks is computationally intr actable. Approximate schemes are therefore of great importance for real wor ld computation. In this paper we consider an approximation scheme in which the original Bayesian network is approximated by another Bayesian network. The approximating network is optimised by an iterative procedure, which min imises the Kullback-Leibler divergence between the two networks. The proced ure is guaranteed to converge to a local minimum of the Kullback-Leibler di vergence. An important question in this scheme is how to choose the structu re of the approximating network. In this paper we show how redundant struct ures of the approximating model can be pruned in advance. Simulation result s of model optimisation are provided to illustrate the methods.