Variational bayes inference for logic-based probabilistic models on BDDs

Abstract

Abduction is one of the basic logical inferences (deduction, induction and abduction) and derives the best explanations for our observation. Statistical abduction attempts to define a probability distribution over explanations and to evaluate them by their probabilities. Logic-based probabilistic models (LBPMs) have been developed as a way to combine probabilities and logic, and it enables us to perform statistical abduction. However non-deterministic knowledge like preference and frequency seems difficult to represent by logic. Bayesian inference can reflect such knowledge on a prior distribution, and variational Bayes (VB) is known as an approximation method for it. In this paper, we propose VB for logic-based probabilistic models and show that our proposed method is efficient in evaluating abductive explanations about failure in a logic circuit and a metabolic pathway. © 2012 Springer-Verlag Berlin Heidelberg.