Reviews: Dynamic Importance Sampling for Anytime Bounds of the Partition Function

The authors present a method for estimating the partition function that alternates between performing heuristic search and importance sampling. The estimated value of the partition function is confidence bounded and improves with additional computation time. Experimental evaluation is performed on problems from 2006 and 2008 UAI competitions by comparing confidence bounds of the proposed method against previous work that uses only sampling [15] or search [16]. The proposed method significantly outperforms sampling on certain problems and search on others, while maintaining performance roughly comparable to or better than either sampling or search across all problems. The originality of the work is a bit limited, as it is a fairly straightforward (but novel as far as I know) combination of two recent papers, references [15] and [16].