CAREER:Program Analyses for Improving Reliability of Probabilistic Software

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

Many emerging applications operate on noisy data and make decisions under uncertainty. Probabilistic programming languages represent such computations as programs that operate on random variables and probability distributions. While the existing languages open the world of powerful probabilistic inference even to programmers with limited knowledge of statistics, new techniques need to be developed to improve programmer productivity and simplify debugging of probabilistic software. This project investigates the hypothesis that static program analysis, with its sound and rich symbolic reasoning, is a solid foundation for these techniques. This project will lead to new automated tools to help scientists, engineers, and software developers build reliable and robust probabilistic software. The project will integrate research and education by developing courses based on newly developed ideas, with the goal of empowering future software engineers with solid quantitative reasoning skills. The project will investigate both the foundations of automated relational analysis for probabilistic computations and the practical application of probabilistic analysis to help application programmers and developers of probabilistic programming systems. The project will investigate two impactful relational analyses for probabilistic programs: sensitivity analysis and semantic differencing. The project will develop an ecosystem of techniques that leverage these analyses to identify errors in probabilistic programming systems, improve robustness of probabilistic computations through program transformations, and optimize the performance of applications that operate on noisy data. The benefits and key components of the approach (including flexible abstractions, transformations, and solving mechanisms) will extend to various application domains with inherent randomness. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.