Local Importance Sampling in Multiply Sectioned Bayesian Networks

The multiply sectioned Bayesian network (MSBN) is a well-studied model for probability reasoning in a multiagent setting. Exact inference, however, becomes difficult as the problem domain grows larger and more complex. We address this issue by integrating approximation techniques with the MSBN Linked Junction Tree Forest (LJF) framework. In particular, we investigate the application of importance sampling in an LJF local junction tree. We propose an LJF local adaptive importance sampler (LLAIS) with improved sampling convergence and effective inter-agent message calculation. Our preliminary experiments confirm that the LLAIS sampler delivers a good approximation of MSBN local posterior beliefs as well as the message calculation over LJF linkage trees.