How to ensure safety of learning-enabled cyber-physical systems?

摘要

ContextModern cyber-physical systems (CPS) integrate machine learning and deep learning components for a variety of tasks, including sensing, control, anomaly detection and learning process dynamics from data.Formal verification of CPS is paramount to ensure their correct behaviour in many safety-critical application domains (including robotics, avionics, health and automotive).However, traditional CPS verification methods are designed to work with mechanistic CPS models, and hence cannot deal in general with data-driven components.Therefore, how to guarantee the correct behaviour of learning-enabled CPS is still an open question which must be addressed in order to deploy these systems in real-world safety-critical settings.Within this research question, we welcome contributions about the formal analysis of learning-enabled CPSs.Examples include but are not limited to: