This Dagstuhl Seminar focuses on the cross-fertilization between different research fields by bringing together (a) those working on model-based reasoning (MBR) for stochastic systems, (b) those working on probabilistic programming languages (PPLs), and (c) those working on low-level inference techniques for both MBR and PPLs that could benefit from cross-fertilization given their common technical underpinnings.

On one hand, researchers in MBR subfields have developed tailored languages such as RDDL (planning), Prism (verification), and subsets of probabilistic graphical models (PGMs) for specialized probabilistic inference tasks (belief state tracking, anomaly detection, prediction and forecasting, system fault diagnosis, etc.). While these languages have found successful applications, they are purpose-built for their use cases and their associated inference engines are similarly specialized. On the other hand, PPLs such as WebPPL, Dice, or Anglican offer a flexible and expressive paradigm for specifying and reasoning about highly general stochastic systems that subsume the aforementioned MBR subfields. By borrowing constructs from general programming languages, these PPLs are easy to use and highly general. However, inference in these very general PPLs is highly challenging and hence substantial knowledge is required to achieve the necessary efficiency for practical applications. The holy grail of PPLs would be efficient, highly scalable, and easy to use engines that support expressive languages.

The aim of this Dagstuhl Seminar is to bring experts from PPLs, PGMs, probabilistic model checking and program analysis, and AI planning together in order to discuss recent advances and identify opportunities for cross-fertilization. We aim to:

• Taxonomize existing representations and methods. What is a taxonomy of language representations (e.g., propositional vs. lifted, discrete and/or continuous, recursion, etc.)? What is a taxonomy of inference problems (finite, infinite, nested, indefinite horizon, average or worst-case?) What is a taxonomy of algorithmic methodologies (e.g., optimization, decision diagrams, SAT/SMT, weighted model counting, Monte Carlo tree search, etc.)?
• Identify opportunities for cross-fertilization. These include methodological questions, such as: Which representations and inferential problems exist across fields? Is there a common low-level "assembly" representation of PPLs that could become a shared compilation target by high-level languages? What compile-time or run-time optimizations could be applied in this case? How do we support automatic abstraction? Other opportunities may arise in improving usability, sharing ideas to ease the learning curve for users.
• Distill challenge problems. What are the challenge problems, their requisite representations and inferential constraints that define high impact problems for probabilistic programming languages (e.g., safety and robustness in AI, transportation systems and urban mobility, the renewable energy grids, predictive maintenance)? How can we initiate collaborative research efforts between different communities to tackle these problems?

Creative Commons BY 4.0

Sebastian Junges, Joost-Pieter Katoen, Scott Sanner, and Guy Van den Broeck