Capturing proof process

Abstract

Proof automation is a common bottleneck for industrial adoption of formal methods. Heuristic search techniques fail to discharge every proof obligation (PO), and significant effort is spent on proving the remaining ones interactively. Luckily, they usually fall into several proof families, where a single idea is required to discharge all similar POs. However, interactive formal proof requires expertise and is expensive: repeating the ideas over multiple proofs adds up to significant costs. The AI4FM research project aims to alleviate the repetitive effort by “learning” from an expert doing interactive proof. The expert’s proof attempts can give rise to reusable strategies, which capture the ideas necessary to discharge similar POs. Automatic replay of these strategies would complete the remaining proof tasks within the same family, enabling the expert to focus on novel proof ideas. This thesis presents an architecture to capture the expert’s proof ideas as a highlevel proof process. Expert insight is not reflected in low-level proof scripts, therefore a generic ProofProcess framework is developed to capture high-level proof information, such as proof intent and important proof features of the proof steps taken. The framework accommodates branching to represent the actual proof structure as well as layers of abstraction to accommodate different granularities. The full history of how the proof was discovered is recorded, including multiple attempts to capture alternative, failed or unfinished versions. A prototype implementation of the ProofProcess framework is available, including integrations with Isabelle and Z/EVES theorem provers. Two case studies illustrate how the ProofProcess systems are used to capture high-level proof processes in examples from industrial-style formal developments. Reuse of the captured information to discharge similar proofs within the examples is also explored. The captured high-level information facilitates extraction of reusable proof strategies. Furthermore, the data could be used for proof maintenance, training, proof metrics, and other use cases.