Bayesian Logic Networks and the Search for Samples with Backward Simulation and Abstract Constraint Learning

With Bayesian logic networks (BLNs), we present a practical representation formalism for statistical relational knowledge. Based on the concept of mixed networks with probabilistic and deterministic constraints, BLNs combine the probabilistic semantics of (relational) Bayesian networks with constraints in first-order logic. In practical applications, efficient inference in statistical relational models such as BLNs is a key concern. Motivated by the inherently mixed nature of models instantiated from BLNs, we investigate two novel importance sampling methods: The first combines backward simulation, i.e. sampling backward from the evidence, with systematic search, while the second explores the possibility of recording abstract constraints during the search for samples.