BLOG: Probabilistic Models with Unknown Objects

by Brian Milch

Many AI problems, ranging from sensor data association to linguistic coreference resolution, involve making inferences about real-world objects that underlie some data. In many cases, we do not know the number of underlying objects or the mapping between observations and objects. This talk will present a new probabilistic modeling language, called Bayesian logic (BLOG), which allows us to represent such scenarios in a natural way. A well-formed BLOG model defines a unique distribution over model structures of a first-order logical language; these "possible worlds" can contain varying numbers of objects with varying relations among them. I will also present results from a likelihood weighting algorithm that does inference in finite time per samping step on a large class of BLOG models, even those involving infinitely many random variables.

Back to the LCI Forum page