Active Requirement Mining of Bounded-Time Temporal Properties of Cyber-Physical Systems

This paper uses active learning to solve the problem of mining bounded-time signal temporal requirements of cyber-physical systems or simply the requirement mining problem. By utilizing robustness degree, we formulates the requirement mining problem into two optimization problems, a parameter synthesis problem and a falsification problem. We then propose a new active learning algorithm called Gaussian Process Adaptive Confidence Bound (GP-ACB) to help solving the falsification problem. We show theoretically that the GP-ACB algorithm has a lower regret bound thus a larger convergence rate than some existing active learning algorithms, such as GP-UCB. We finally illustrate and apply our requirement mining algorithm on two case studies, the Ackley's function and a real world automatic transmission model. The case studies show that our mining algorithm with GP-ACB outperforms others, such as those based on Nelder-Mead, by an average of 30% to 40%. Our results demonstrate that there is a principled and efficient way of extracting requirements for complex cyber-physical systems.