Integrating Logic and Probability: Algorithmic Improvements in Markov Logic Networks
Marenglen Biba · 2009
Signature from head of PhD committee: This dissertation proposes novel algorithms for learning and inference in Markov Logic Networks. Statistical Relational Learning challenges one of the most important problems of Machine Learning since its birth: integrating logic and probability in learning. Markov Logic is a powerful representation formalism that combines full first-order logic with probabilistic graphical models by attaching weights to first-order formulas and viewing these as templates for features of Markov Networks (MNs). Markov Logic Networks (MLNs) together with a set of constants define ground MNs. MLNs preserve the expressivity of first-order logic and take advantage of probabilistic graphical models algorithms being therefore a powerful model for dealing with structured, noisy and uncertain data. The rich expressivity of MLNs comes at the cost of learning and inference. Structure learning is the task of learning the logical clauses together with their weights