Statistical verification of learning-based cyber-physical systems

The use of Neural Network (NN)-based controllers has attracted significant attention in recent years. Yet, due to the complexity and non-linearity of such NN-based cyber-physical systems (CPS), existing verification techniques that employ exhaustive state-space search, face significant scalability challenges; this effectively limits their use for analysis of real-world CPS. In this work, we focus on the use of Statistical Model Checking (SMC) for verifying complex NN-controlled CPS. Using an SMC approach based on Clopper-Pearson confidence levels, we verify from samples specifications that are captured by Signal Temporal Logic (STL) formulas. Specifically, we consider three CPS benchmarks with varying levels of plant and controller complexity, as well as the type of considered STL properties - reachability property for a mountain car, safety property for a bipedal robot, and control performance of the closed-loop magnet levitation system. On these benchmarks, we show that SMC methods can be successfully used to provide high-assurance for learning-based CPS.