Next-generation Constraint Solvers for Software Engineering and Security

约束求解器，即自动求解数学约束的程序，在工程和科学领域有着广泛的应用。求解器可以被比作瑞士军刀，用于规划机器人的运动、配置汽车或自动发现软件中的错误等应用。工程师使用数学约束对他们的问题建模，然后使用求解器自动解决问题。就在十年前，像文字处理器和操作系统这样的商业软件的可伸缩的自动bug发现被认为实际上是不可行的。由于求解器性能的显著提高（包括我在内的许多研究人员），自动bug查找不仅变得可行，而且在微软等公司也是强制性的。虽然迄今为止的成果很重要，但随着工程师们解决软件合成等更难的应用，对功能更强大、表达能力更强的求解器的需求仍在持续增长。因此，我提出了一个长期的研究计划，以开发新的求解器技术，这些技术比现在的求解器更快，更有表现力，旨在开发软件可靠性和安全性的软件工程工具。更准确地说，我的研究计划有以下三个重点：i)我将探索基于机器学习（ML）的新技术。机器学习理论和技术已经发生了真正的革命。我们可以使用ML和随机推理技术来学习大型约束中的微妙元级模式，从而实现更快的求解（类似于人类如何从数据中识别深层概念），ii)利用无处不在且廉价的多核处理器构建可扩展并行求解器的技术，以及iii)利用特定领域知识作为解锁约束解决方案的关键的求解器技术。拟议的研究将产生深刻的基础科学、技术和商业影响。基础结果将通过参数复杂性和机器学习的思想为求解器启发式提供理论基础。技术和商业影响将是一组新的可扩展和可扩展的求解器，它们有可能改变软件的可靠性和安全性。

Constraint solvers, programs that automatically solve mathematical constraints, are used in myriad applications in engineering and science. Solvers can be likened to swiss-army knives, used in applications such as planning a robot's movement, configuring a car or automatically finding bugs in software. Engineers model their problem using mathematical constraints, and then use solvers to automatically solve them. As little as a decade ago, scalable automatic bug-finding of commercial software like word processors and operating systems was considered practically infeasible. Thanks to impressive gains in solver performance (due to many researchers including myself), not only has automatic bug-finding become feasible but is mandatory in companies like Microsoft. While the gains to-date are important, the demand for ever-more powerful and expressive solvers continues to grow unabated as engineers tackle even harder applications such as software synthesis. Hence, I propose a long-term research program to develop new solver techniques that are orders of magnitude faster and more expressive than today's, aimed at software engineering tools for software reliability and security. More precisely, my research program has the following three thrusts: i) I will explore new techniques based on machine learning (ML). There has been a veritable revolution in ML theory and techniques. We can use ML and stochastic inference techniques to learn subtle meta-level patterns in large constraints that enable faster solving (similar to how humans identify deep concepts from data), ii) techniques that leverage ubiquitous and cheap multi-core processors to build scalable parallel solvers, and iii) solver techniques that leverage domain-specific knowledge as keys to unlock solutions to constraints. The proposed research will have deep fundamental scientific, technical, and commercial impact. The foundational results will provide theoretical underpinning for solver heuristics through ideas from parametric complexity and ML. The technical and commercial impact will be a set of new scalable and extensible solvers which have the potential to transform software reliability and security.