Action-Centric Probabilistic Programming

Probabilistic models are dominant in data management, but are often described in a combination of natural and mathematical language, and algorithms need to be engineered for an individual application. Probabilistic programming languages have received considerable attention recently as they contribute formal languages that enable re-use and clarity in the problem specification. Our interest in this paper is the development of a programming language, called ALLEGRO, that is in service of the high-level control of an autonomous system, such a mobile robot. The language’s mathematical foundations rests of Bayesian conditioning, and its programs are interpreted over a sophisticated logical theory of actions supplemented with user-defined axioms. While many existing probabilistic programming languages can easily be shown to capture stochastic transitions, we argue that an actioncentric probabilistic programming has many valuable properties. As a modeling language, ALLEGRO can be seen as a basis for relating high-level control specifications, including plans with loops, on the one hand, and high-level probabilistic models, such as relational graphical models, on the other.