Advances in Lifted Importance Sampling

We consider lifted importance sampling (LIS), a previously proposed approximate inference algorithm for statistical relational learning (SRL) models. LIS achieves substantial variance reduction over conventional by using various lifting rules that take advantage of the symmetry in representation. However, it suffers from two drawbacks. First, does not some important symmetries representation and may exhibit needlessly high on models having these symmetries. Second, uses an uninformative proposal distribution which adversely affects its accuracy. propose improvements to address limitations. we identify new SRL define rule taking this symmetry. The reduces LIS. new, structured approach constructing dynamically updating via adaptive sampling. demonstrate experimentally our improved is substantially more accurate than algorithm.