Auxiliary Variables in Probabilistic Programs

We extend the first-order meta-language of Staton et al. with a new language construct, slice-let. This new construct allows a user to define a random variable x, with a potentially intractable distribution, by introducing an auxiliary random variable, u, and specifying only the conditional distributions of x given u and of u given x. In effect, the distribution of x need only be defined up to some level of approximation, determined by the value of u. In the extended abstract, we outline the denotational semantics for slice-let and give some example programs. Finally, we briefly discuss an approximate representation of the program that can be used by an inference algorithm.

Authors: Ekansh Sharma and Daniel M. Roy

Extended Abstract

This entry was posted in Uncategorized. Bookmark the permalink.