Reference: Dagum, P. & Horvitz, E. A Bayesian Analysis of Simulation Algorithms for Inference in Belief Networks. 1993.
Abstract: Belief networks are an expressive representation for encoding expert knowledge about uncertain causal relationships. Both exact and approximation methods for performing inference with belief networks can pose difficult computational problems in the worst case. Nevertheless, approximation procedures hold promise to provide estimates efficiently for a variety of complex networks that resist exact solution. We characterize the performance of algorithms in the important class of inference procedures based on stochastic simulation. We develop terms for the error associated with estimates generated by several simulation methods, including forward simulation, likelihood weighting, and randomized approximation strategies.