It is sometimes possible to optimize probabilistic programs, either statically or dynamically. We introduce two examples demonstrating the need for both approaches. Furthermore, we identify a set of challenges related to the two approaches, and more importantly, how to combine them.
We present this topic in order to explore new research directions in optimization for probabilistic programming. Our hope is to generate discussions, and to establish connections with other people working in the same general area.
Authors: Daniel Lundén, David Broman, Lawrence M. Murray