Safety Assurance of Autonomous Learning-Enabled Cyber Physical Systems
Musau, Patrick
0000-0002-0227-1336
:
2022-05-10
Abstract
There are few technologies that hold as much promise as autonomous Cyber-physical systems (CPS) in re-orienting the way we move around, explore new environments, distribute resources, and conduct complex missions. To bring forth these benefits, we must ensure that CPS meet rigorous standards of correctness both at design time and during operation. This mandates the development of modeling tools and algorithms that can deal with the complexity exhibited by CPS and their environments. Since CPS operate in unstructured and dynamic environments, their design often incorporates opaque data-driven or machine learning methods. We refer to components synthesized using machine learning or other data-driven techniques as Learning Enabled Components. While these methods have demonstrated significant success in solving complex problems in numerous domains, reasoning about the correctness of these components is a notoriously difficult problem and many safety assurance approaches are not well suited to handle their analysis. Thus, there is an urgent need for the design of safety assurance methods that can provide guarantees of the correct operation of these systems in a manner that is both practical and rigorous.
Bearing the above in mind, the following dissertation presents approaches for the safety assurance of autonomous learning-enabled CPS. The question that this document sets out to answer is how do we construct evidence that an autonomous system satisfies its requirements? Moreover, how can we efficiently synthesize this evidence, based on our assumptions about the environment that a CPS is tasked with operating within, and the operation of the components defining its behavior? To this end, we begin by presenting a case study of two leading machine learning methods for training neural network controllers, Reinforcement Learning and Imitation Learning, that motivate the need to monitor learning-enabled components. We then introduce a safety monitoring regime leveraging real-time reachability that allows us to provide provable guarantees of safety for systems that make use of learning-enabled components. Building on this work, we propose a runtime assurance framework based on the seminal simplex architecture for the safety assurance of autonomous systems. In this regime, we abstract away the need to analyze an underlying controller leveraged within a CPS, and instead focus on the effects of its control decisions on the system's future states. Finally, we propose the development of a Model Predictive Control Regime Leveraging Real-Time Reachability, as a demonstration of utilizing these techniques for the design of safe control regimes. We conclude this dissertation with a discussion of the challenges and limitations of this work in an effort to stimulate the development of novel research towards the assurance of autonomous learning-enabled CPS. Additionally, we present a detailed description of the source code and experimental artifacts leveraged within our experiments for those who wish to reproduce and build on this work.