Improving the efficiency of Gibbs sampling for probabilistic logical models by means of program specialization

Abstract

There is currently a large interest in probabilistic logical models. A popular algorithm for approximate probabilistic inference with such models is Gibbs sampling. From a computational perspective, Gibbs sampling boils down to repeatedly executing certain queries on a knowledge base composed of a static part and a dynamic part. The larger the static part, the more redundancy there is in these repeated calls. This is problematic since inefficient Gibbs sampling yields poor approximations. We show how to apply program specialization to make Gibbs sampling more efficient. Concretely, we develop an algorithm that specializes the definitions of the query-predicates with respect to the static part of the knowledge base. In experiments on real-world benchmarks we obtain speedups of up to an order of magnitude.