Logic-based Formalisms for Statistical Relational Learning

This chapter provides a selective overview of logic-based approaches to statistical relational learning. Issues of representation, inference, and learning are addressed with an emphasis on representation. A distinction is drawn between “directed” representations with connections to Bayesian nets and “undirected” ones related to Markov nets. Within directed representations a further distinction is made between using conditional probabilities and using logical rules to define probability distributions. Among the formalisms discussed are: the independent choice logic, probabilistic logic programuming, and stochastic logic programs. The PRISM system is used to provide concrete examples of probabilistic inference and parameter estimation. The use of “possible worlds™ to provide semantics is described and its role in connecting differing formalisms is analyzed.