Bayesian Logic Networks ( Extended Version ∗ )

This report introduces Bayesian logic networks (BLNs), a statistical relational knowledge representation formalism that is geared towards practical applicability. A BLN is a meta-model for the construction of a probability distribution from local probability distribution fragments (as in a Bayesian network) and global logical constraints formulated in first-order logic. An instance is thus a mixed network with probabilistic and deterministic constraints. We provide the formal semantics of BLNs and explain their practical realization as implemented in the open-source software toolbox ProbCog, which supports learning and a wide range of inference algorithms.