New Rules for Domain Independent Lifted MAP Inference

摘要: Lifted inference algorithms for probabilistic first-order logic frameworks such as Markov networks (MLNs) have received significant attention in recent years. These use so called lifting rules to identify symmetries the representation and reduce problem over a large model an much smaller model. In this paper, we present two new rules, which enable fast MAP class of MLNs. Our first rule uses concept single occurrence equivalence logical variables, define paper. The states that assignment MLN can be recovered from MLN, each variable is replaced by constant (i.e., object domain variable). second safely remove subset formulas if all classes variables remaining are tautology evaluate true) at extremes assignments with identical truth value groundings predicate). We prove our sound demonstrate via detailed experimental evaluation approach superior terms scalability solution quality state art approaches.