Integrated Verification for Robust Computing

This paper argues for an integrated approach to verification, combining offline and online verification techniques used at different stages of a system’s lifetime, in order to detect and correct failures arising from design errors and program bugs. We propose a formal framework for online verification and recovery, including a fault model, within which such an integrated approach can be investigated. As a particular instance, we consider the problem of recovering a class of finite-state systems, at run-time, from failures of safety properties, while leveraging results of offline verification. This class comprises systems whose behavior can be divided into rounds such that each round is largely independent of the others. We give a randomized recovery strategy based on online learning for which the expected number of recovery actions performed in any state is at most logarithmic in the total number of actions, assuming perfect error localization. Results from design-time verification can be used to reduce the search space of the recovery algorithm and deal with imperfections in error localization. We illustrate our approach with case studies. Our results are a step towards building reactive systems that are robust to failures and self-evolve towards correct systems.