Lifted First-Order Probabilistic Inference
                               Eyal Amir

Probabilistic inference algorithms are widely employed in artificial
intelligence.  Among their many applications are tracking partially
observed systems, medical diagnosis, automatic tutoring, and
computational biology.  Such applications use large, complex models,
and are difficult to engineer and learn.  To solve these difficulties,
several methods emerged that use relational probabilistic
specifications.  These specification languages can abstract over large
classes of objects.

Unfortunately, most probabilistic inference algorithms are specified
and processed on a propositional level.  There, random variables are
considered different from others, and real-world structures of objects
and relationships are ignored. In the last decade, many proposals for
algorithms accepting relational specifications have been presented,
but in the inference stage they still operate on a mostly
propositional representation level.

In this talk I will present an exact inference algorithm that operates
directly on a relational level, and that can be applied to any
relational model (specified in a language that generalizes undirected
probabilistic graphical models). I will discuss how further research
can improve our results, and how different applications can gain from
this advance. I will conclude with our experiments, which show
superior performance in comparison with propositional exact inference.

(Joint work with Rodrigo Braz and Dan Roth)