Brian Milch
Bhaskara Marthi
Stuart Russell
Abstract: In many real-world probabilistic reasoning problems, one of the questions we want to answer is: how many objects are out there? Examples of such problems range from multi-target tracking to extracting information from text documents. However, most probabilistic modeling formalisms -- even first-order ones -- assume a fixed, known set of objects. We introduce a language called BLOG for specifying probability distributions over relational structures that include varying sets of objects. In this paper we present BLOG informally, by means of example models for multi-target tracking and citation matching. We discuss some attractive features of BLOG models and some avenues of future work.
Appeared in: ICML 2004 Workshop on Statistical Relational Learning and Its Connections to Other Fields, Banff, Alberta, Canada, July 2004.
Download: PS version, PDF version