Learning-Based Assume-Guarantee Verification (Tool Paper)

Despite significant advances in the development of model checking, it remains a difficult task in the hands of experts to make it scale to the size of industrial systems. A key step in achieving scalability is to “divide-and-conquer”, that is, to break up the verification of a system into smaller tasks that involve the verification of its components. Assume-guarantee reasoning [9, 11] is a widespread “divide-and-conquer” approach that uses assumptions when checking individual components of a system. Assumptions essentially encode expectations that each component has from the rest the system in order to operate correctly. Coming up with the right assumptions is typically a non-trivial manual process, which limits the applicability of this type of reasoning in practice. Over the last few years, we have developed a collection of techniques and a supporting toolset, for performing assume-guarantee reasoning of software in an automated fashion. Our techniques are applicable both at the level of design models, and at the level of actual source code. In the heart of these techniques lies a framework that uses an off-the-shelf learning algorithm for regular languages, namely L* [1], to compute assumptions automatically. The rest of the paper is organized as follows. Section 2 is a high-level description of our techniques for learning-based assume-guarantee reasoning of software. Section 3 discusses the tool support for our techniques and experimental results obtained from the application of our approach to some industrial size case studies, and we conclude the paper with Section 4.