서브메뉴
검색
Safety-Assured Autonomy for Learning-Enabled Cyber-Physical Systems
Safety-Assured Autonomy for Learning-Enabled Cyber-Physical Systems
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151122
- ISBN
- 9798382762708
- DDC
- 621.3
- 저자명
- Wang, Yixuan.
- 서명/저자
- Safety-Assured Autonomy for Learning-Enabled Cyber-Physical Systems
- 발행사항
- [Sl] : Northwestern University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 170 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
- 주기사항
- Advisor: Zhu, Qi.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2024.
- 초록/해제
- 요약Modern Cyber-Physical Systems (CPSs) face significant challenges due to the operation in uncertain and dynamically changing physical environments, and increasing complex computational processes in the cyber domain. These complexities often surpass what traditional rule-based approaches can handle. To this end, there has been a shift towards leveraging data-driven machine learning techniques in CPSs, a Learning-Enabled Cyber-Physical Systems (LE-CPSs) paradigm. These techniques, particularly neural networks, offer promising solutions across various CPS functionalities, from sensing and perception to low-level planning and control. In LE-CPSs, neural networks have attracted attention for their ability to improve performance, bypass the need for intricate, error-prone system models, and effectively deal with environmental uncertainties for planning and control tasks.Despite their potential effectiveness, neural networks face challenges such as training errors, limited data coverage, inherent uncertainties, and a lack of comprehensive safety analyses and assurances. These factors have hindered their application in real-world safety-critical CPS deployments. To this end, this thesis presents a safety-assured autonomy framework designed specifically for LE-CPSs to tackle this issue. This framework combines formal verification techniques with machine learning approaches to enable safety-assured planning and control in autonomous systems for offline training and online execution.First, we introduce a verification-guided control learning framework for design time and offline learning. This framework seamlessly integrates verification into the control learning process in a closed-loop manner via gradient guidance, ensuring that the optimized controllers come with formal safety guarantees. Specifically, we employ reachability analysis tools to approximate gradients for safety and goal-reaching metrics in environments where the dynamics are fully understood. To enhance the practicality of this framework, we shift focus to safe reinforcement learning (RL) as an alternative approach for learning safe control policies without complete knowledge of system dynamics. Our innovative bi-level optimization formulation enables our joint learning framework to perform formal safety verification and reinforcement learning optimization concurrently through a differential convex optimization approach. We expand this bi-level joint learning framework for fully unknown and stochastic environments by using a novel generative-model-based soft barrier method. This method allows us to encode hard safety chance constraints with all components parameterized by deep neural networks and provides a probability-bounded safety "guarantee".Then, we propose a control adaptation framework for run-time switching among multiple existing controllers, either model-based or neural-network-based ones. Leveraging a novel control invariant computation and reinforcement learning optimization, this framework features safety assurance and energy efficiency. While prior approaches yield improved safety and performance, challenges of generalizability and explainability persist. Exploiting the interest in large language models (LLMs) and their adept reasoning capabilities for runtime execution, we investigate their potential to bolster driving safety, particularly within autonomous driving, a typical LE-CPS application. Employing an LLM (ChatGPT) as a high-level online behavior decision-maker, we explore its capacity to make planning intentions based on textual descriptions of driving environments, with a concurrent safety verifier facilitating online in-context safety learning.Through this comprehensive framework introduced in this thesis, we aim to propel the integration of verification into learning-based planning and control for offline and online safety assurance in safety-critical CPS deployments, to enhance the overall system reliability.
- 일반주제명
- Computer engineering
- 일반주제명
- Computer science
- 일반주제명
- Information technology
- 키워드
- Machine learning
- 키워드
- Safety assurance
- 기타저자
- Northwestern University Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 85-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017160825
■00520250211151122
■006m o d
■007cr#unu||||||||
■020 ▼a9798382762708
■035 ▼a(MiAaPQ)AAI31146713
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aWang, Yixuan.
■24510▼aSafety-Assured Autonomy for Learning-Enabled Cyber-Physical Systems
■260 ▼a[Sl]▼bNorthwestern University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a170 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-11, Section: B.
■500 ▼aAdvisor: Zhu, Qi.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2024.
■520 ▼aModern Cyber-Physical Systems (CPSs) face significant challenges due to the operation in uncertain and dynamically changing physical environments, and increasing complex computational processes in the cyber domain. These complexities often surpass what traditional rule-based approaches can handle. To this end, there has been a shift towards leveraging data-driven machine learning techniques in CPSs, a Learning-Enabled Cyber-Physical Systems (LE-CPSs) paradigm. These techniques, particularly neural networks, offer promising solutions across various CPS functionalities, from sensing and perception to low-level planning and control. In LE-CPSs, neural networks have attracted attention for their ability to improve performance, bypass the need for intricate, error-prone system models, and effectively deal with environmental uncertainties for planning and control tasks.Despite their potential effectiveness, neural networks face challenges such as training errors, limited data coverage, inherent uncertainties, and a lack of comprehensive safety analyses and assurances. These factors have hindered their application in real-world safety-critical CPS deployments. To this end, this thesis presents a safety-assured autonomy framework designed specifically for LE-CPSs to tackle this issue. This framework combines formal verification techniques with machine learning approaches to enable safety-assured planning and control in autonomous systems for offline training and online execution.First, we introduce a verification-guided control learning framework for design time and offline learning. This framework seamlessly integrates verification into the control learning process in a closed-loop manner via gradient guidance, ensuring that the optimized controllers come with formal safety guarantees. Specifically, we employ reachability analysis tools to approximate gradients for safety and goal-reaching metrics in environments where the dynamics are fully understood. To enhance the practicality of this framework, we shift focus to safe reinforcement learning (RL) as an alternative approach for learning safe control policies without complete knowledge of system dynamics. Our innovative bi-level optimization formulation enables our joint learning framework to perform formal safety verification and reinforcement learning optimization concurrently through a differential convex optimization approach. We expand this bi-level joint learning framework for fully unknown and stochastic environments by using a novel generative-model-based soft barrier method. This method allows us to encode hard safety chance constraints with all components parameterized by deep neural networks and provides a probability-bounded safety "guarantee".Then, we propose a control adaptation framework for run-time switching among multiple existing controllers, either model-based or neural-network-based ones. Leveraging a novel control invariant computation and reinforcement learning optimization, this framework features safety assurance and energy efficiency. While prior approaches yield improved safety and performance, challenges of generalizability and explainability persist. Exploiting the interest in large language models (LLMs) and their adept reasoning capabilities for runtime execution, we investigate their potential to bolster driving safety, particularly within autonomous driving, a typical LE-CPS application. Employing an LLM (ChatGPT) as a high-level online behavior decision-maker, we explore its capacity to make planning intentions based on textual descriptions of driving environments, with a concurrent safety verifier facilitating online in-context safety learning.Through this comprehensive framework introduced in this thesis, we aim to propel the integration of verification into learning-based planning and control for offline and online safety assurance in safety-critical CPS deployments, to enhance the overall system reliability.
■590 ▼aSchool code: 0163.
■650 4▼aComputer engineering
■650 4▼aComputer science
■650 4▼aInformation technology
■653 ▼aControl learning process
■653 ▼aCyber-physical systems
■653 ▼aFormal verification
■653 ▼aMachine learning
■653 ▼aReinforcement learning
■653 ▼aSafety assurance
■690 ▼a0464
■690 ▼a0489
■690 ▼a0984
■690 ▼a0800
■71020▼aNorthwestern University▼bElectrical and Computer Engineering.
■7730 ▼tDissertations Abstracts International▼g85-11B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17160825▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


