Evaluation and Comparison Criteria for Approaches to Probabilistic Relational Knowledge Representation

In the past ten years, the areas of probabilistic inductive logic programming and statistical relational learning put forth a large collection of approaches to combine relational representations of knowledge with probabilistic reasoning. Here, we develop a series of evaluation and comparison criteria for those approaches and focus on the point of view of knowledge representation and reasoning. These criteria address abstract demands such as language aspects, the relationships to propositional probabilistic and first-order logic, and their treatment of information on individuals. We discuss and illustrate the criteria thoroughly by applying them to several approaches to probabilistic relational knowledge representation, in particular, Bayesian logic programs, Markov logic networks, and three approaches based on the principle of maximum entropy.