Semantics and Implementation of Probabilistic Programming Languages

点击翻译按钮获取中文摘要

Probabilistic programming languages (PPLs) are increasingly used to model and solve real-world artificial intelligence (AI) problems of growing complexity. The proposed research aims to advance both the theory and practice of PPLs, pushing towards a future where PPLs become as regular and well-understood a technology for model-based AI as conventional programming languages are for traditional software development. Despite recent advances in probabilistic programming, many potentially impactful applications still fall outside the comfort zone of PPLs. A classic example is programming a group of robots that reason distrustfully about the environment and about each other to negotiate a course of actions in a safety-critical setting. Solving such challenging problems would require linguistic features that are simply not in the vocabulary of existing PPLs and demand language implementations that achieve compile-time efficiency, inference-time efficiency, and security. The proposed research will identify the key language improvements needed to address the challenges, propose new designs of PPLs, and study their semantics and implementations. This research will impact the way programmers use PPLs and engineer artificial intelligence. The core appeals of PPLs, since its advent, have been expressive modeling languages and the automation of inference. By addressing these appeals in new ways that are motivated by practical concerns and backed by theoretical rigour, the proposed research will give PPLs a competitive edge in terms of expressiveness, assurance, and efficiency needed in modeling and solving AI problems of real-world relevance. This research will produce new programming languages with open-sourced implementations as well as mechanized metatheories, of value to programming languages and AI communities in the large. Trainees involved in this research will grow to become experts in programming languages research and in the burgeoning field of probabilistic programming, capable of leading future technology innovations that involve programming languages as a key ingredient.