Séminaire de Alexandre Dubray - UC Louvain
Modeling and Solving Probabilistic Inference Model with Projected Weighted Model Counting
21 nov. 2024 - 14:00Computing probability from probabilistic models is a challenging problem, which is #P-Hard in general. One popular technique for solving such problems is to transform the model into a propositional formula in CNF form and then calculate the formula's weighted number of models. However, classical encodings only partially translate the input models' probabilistic features. For example, the distributions are transformed into clauses, but classical model counters do not use the fact that its values sum up to one.
In this talk, I will present Schlandals, a novel modeling CNF-based language and solver specialized for probabilistic inference problems. I'll show that various probabilistic models can be translated into an extended CNF formula in which distributions are first-class citizens. I'll also show that by constraining the clauses to be Horn, the structure of the probabilistic models can be encoded into the CNF, and the counting process can be accelerated. I will also present a novel approximate inference technique that provides deterministic anytime bounds on the weighted model count by taking advantage of the fact that we are working with distributions.