
Statistical relational learning addresses one of the central questions
of artificial intelligence: the integration of probabilistic reasoning
with first order logic representation and machine learning. Recently,
this questions has received a lot of attention. Several statistical
relational learning approaches have been developed in related, but
different areas including machine learning, statistics, databases, and
reasoning under uncertainty.
This thesis starts from an inductive logic programming perspective and
firstly develops a general framework for statistical relational
learning: probabilistic inductive logic programming. Based on this
foundation, the thesis shows how to incorporate the logical concepts
of objects and relations among these objects into Bayesian
networks. As time and actions are not just other relations, it
afterwards develops approaches to probabilistic inductive logic
programming over time and for making complex decision in relational
domains. More specifically, Bayesian networks are upgraded to Bayesian
logic programs, hidden Markov models to logical hidden Markov models;
and Markov decision processes to Markov decision
programs. Furthermore, it will be shown that statistical relational
learning approaches naturally yield kernels for structured data. The
resulting approaches will be illustrated using examples from genetics,
bioinformatics, and classical planning domains.
