Theory Refinement of Bayesian Logic Programs

Bayesian Logic Programs (BLP) [8][9] is a powerful and elegant framework for combining the expressiveness of first order logic with Bayesian networks. They can represent both Bayesian networks and logic programs, and their kernel in Prolog is an adaptation of an usual Prolog metainterpreter. It has been successfully compared to other such proposals in the literature. In this paper, we present a theory refinement system called RBLP, which minimally modifies a given BLP to make it consistent with the available training data. RBLP unifies FORTE [20], which induces a revision of Logic Programs from data, with an adaptation of EM [10] for learning the entries of the CPTs.