Probabilistic Theorem Proving: A Unifying Approach for Inference in Probabilistic Programming

Inference is the key bottleneck in probabilistic programming. Often, the main advantages of probabilistic programming – simplicity, modularity, ease-of-use, etc. – are dwarfed by the complexity and intractability of inference. In fact, one of the main reasons for the scarcity/absence of large applications and real-world systems that are based in large part on probabilistic programming languages (PPLs) (c.f. [6, 8]) is the lack of fast, scalable and accurate inference engines for them. Therefore, in this paper, we consider probabilistic theorem proving (PTP) [5], a recently proposed scalable, general-purpose algorithm for inference in probabilistic logic, and extend it to yield a general-purpose inference algorithm for PPLs.