Runtime Assurance for Intelligent Cyber-Physical Systems

The designers of safety-critical CPS that are intelligently automated using machine learning (ML) are encouraged to define invariants and utilize metrics to quantify the uncertainty of ML decisions in addition to focusing on the performance and functionality of the algorithm. Wheatman et al. [11] present Runtime Assurance for Distributed Intelligent Control Systems (RADICS) that extends the Simplex architecture [9] to provide runtime assurance for Cyber-Physical Systems (CPS) being controlled by machine learning al-gorithms. RADICS can thus allow designers to guarantee some minimum level of system performance via a safety controller while simultaneously allowing for greater average performance via an artificial intelligence (AI) controller. Existing implementations of RADICS have focused on simulated applications such as vehicular traffic control using the Simulation of Urban Mobility (SUMO) [5] and Flow [12] environments. The aim of this project is to implement RADICS in a physical environment in order to investigate and understand the limitations and challenges of physical deployments. We have selected a water treatment testbed as the application to conduct this evaluation. As a demonstration at ICCPS, we hope to use this testbed deployment to study the impact of real-world issues such as communication latencies, sensor failures, and incomplete information on the RADICS runtime assurance system. We also plan to extend the physical testbed into a hardware-in-the-loop smart city environment where the fidelity of the physical testbed will complement the scalability and flexibility of simulated components. This will allow further evaluation of assurance capabilities such as RADICS before they are deployed in the real world to ensure the safe and reliable operation of intelligent CPS. This work�s novel contributions include an extension of RADICS towards real-world use in cyber-physical systems, analysis of problems inherent to the shift to physical domains, and the introduction of an ensemble-like method for calculating confidence in the RADICS white box monitor.