Design and Synthesis of Controllers for Societal-Scale Cyber-Physical Systems

Abstract

In this dissertation, a unifying framework for controller design, synthesis, and validation for societal-scale Cyber-Physical Systems (CPS) is proposed. We use vehicular CPS as a case study of societal-scale CPS. These systems require large-scale simulation to reduce the need for physical tests or field experiments. Such large-scale simulations demand reproducibility and repeatability of results---or else the use of simulation provides no insights into the overall system's dynamics. Many current simulation tools for CPS lack the properties of reproducibility and repeatability. Our proposed approach for \Rahul{scalability} and repeatability of existing simulation tools includes offloading dynamics of systems using federated models, message handling and synchronization to operate a simulator at slower than real-time, or synchronously with another system. The approaches do not require rewriting those simulation tools, thus permitting assembly of tools at their interfaces. With such simulation tools, it is now possible to use model-based design and code-generation techniques to deploy the same implementation in simulation that would be deployed in an experiment. Such approaches may be inaccessible for simulation tools that do not support real-time behavior. Our approach permits the validation of novel controllers and algorithms not only through software-in-the-loop (SWIL) or hardware-in-the-loop (HWIL) simulation but also transfer seamlessly for real-world testing. In this way, both model and simulator can be improved iteratively by feeding data from the physical environment. Large-scale CPS are producing a massive amount of data in real-time which is being used for decision-making and control that engage with infrastructure and humans. For Vehicular CPS, data comes in the form of multiple modalities such as CAN bus, GPS, dashcam, LiDAR, etc. We further propose a generic timeseries tool to work with such vehicular data that allows researchers to gain novel insights about driving behavior, discover rare events, and facilitate data-driven applications for vehicular CPS. We present three case studies: (i) Followerstopper, (ii) deployment of a reinforcement learning controller, and (iii) dual-ring-barrier traffic signal controller. The first two case studies are related to Lagrangian control of an autonomous vehicle (AV) in mixed urban traffic consisting of some highly automated vehicles among mostly human-driven vehicles. The third case study presents a co-simulation of an infrastructure controller using SUMO and ROS whose development takes an approach of model-based design. These use cases provide an engineering solution to improving a controller candidate for societal-scale CPS through a data-driven approach, deterministic and repeatable simulation, and their deployment in the real world.