The PITA System for Logical-Probabilistic Inference

Probabilistic Inductive Logic Programming (PILP) is gaining interest due to its ability to model domains with complex and uncertain relations among entities. Since PILP systems generally must solve a large number of inference problems in order to perform learning, they rely critically on the support of efficient inference systems. PITA [7] is a system for reasoning under uncertainty on logic programs. While PITA includes frameworks for reasoning with possibilistic logic programming, and for reasoning on probabilistic logic programs with special exclusion and independence assumptions, we focus here on PITA’s framework for reasoning on general probabilistic logic programs following the distribution semantics, one of the most prominent approaches to combining logic programming and probability. Syntactically, PITA targets Logic Programs with Annotated Disjunctions (LPADs) [9] but can be used for other languages that follow the distribution semantics, such as ProbLog [3], PRISM [8] and ICL [5], as there are linear transformation from one language to the others [1]. PITA is distributed as a package of XSB Prolog and uses tabling along with an XSB feature called answer subsumption that allows the combination of different explanations for the same atom in a fast and simple way. PITA works by transforming an LPAD into a normal program and then querying the program. In this paper we provide an overview of PITA and an experimental comparison of it with ProbLog, a state of the art system for probabilistic logic programming. The experiments show that PITA has very good performances, often being faster than ProbLog.