Learning for verification in embedded systems: A case study

Verification of embedded systems is challenging whenever control programs rely on black-box hardware components. Unless precise specifications of such components are fully available, learning their structured models is a powerful enabler for verification, but it can be inefficient when the system to be learned is data-intensive rather than control-intensive. We contribute a methodology to attack this problem based on a specific class of automata which are well suited to model systems wherein data paths are known to be decoupled from control paths. We test our approach by combining learning and verification to assess the correctness of grey-box programs relying on FIFO register circuitry to control an elevator system.