Towards Automated Sequential Monte Carlo for Probabilistic Graphical Models

We revisit the idea of using sequential Monte Carlo (SMC) for inference in general probabilistic graphical models. By constructing a sequence of auxiliary target distributions (also known as a sequential decomposition) based on the graph structure we can run a standard SMC sampler on the graph. In this paper we study the impact of the sequential decomposition on the accuracy of the SMC method by computing the asymptotic variance of the estimator as a function of the decomposition. In general the variance will be intractable, so we propose to use a proxy Gaussian Markov random field with a structure that is identical to that of the original problem. Furthermore, based on these results we propose and evaluate some heuristics for automated SMC inference on any given graph structure.