서브메뉴
검색
Sampling and Estimation for the Validation of Safety-Critical Autonomous Systems
Sampling and Estimation for the Validation of Safety-Critical Autonomous Systems
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104744
- ISBN
- 9798290650234
- DDC
- 629.13309
- 서명/저자
- Sampling and Estimation for the Validation of Safety-Critical Autonomous Systems
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 115 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: A.
- 주기사항
- Advisor: Kochenderfer, Mykel.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약There is increasing interest in deploying autonomous systems to domains such as transportation, healthcare, finance, and robotics because of their potential to improve efficiency and safety. These domains are safety-critical; the consequences of system failures can cause loss of human life or other significant costs. Therefore, we require rigorous validation of autonomous systems before deployment in these domains. Traditional safety validation approaches like real-world testing may be extremely expensive and time-consuming. Formal verification techniques are not scalable to complex systems and environments. Falsification methods provide failure examples in simulation, but they do not provide a comprehensive understanding of the system's behavior.In this thesis, we address these limitations by proposing techniques for failure sampling and estimation to validate safety-critical autonomous systems. Given a probabilistic model of system trajectories, our methods aim to sample from the distribution over trajectories that lead to failure (failure sampling) or to estimate the probability of failure (failure estimation). During system development, engineers may use failure sampling to identify failure modes and improve system safety. Failure estimation can be used to quantify the risk of system failure before deployment. Together, these approaches provide a comprehensive understanding of system behavior and enable rigorous safety validation. Previous methods for failure sampling and estimation have been based on Markov chain Monte Carlo, importance sampling, and dynamic programming. Although these methods have been successful in some applications, they have limitations in terms of generality, scalability, and efficiency.Many existing methods for failure sampling and estimation are tailored to specific system models and failure criteria. This makes it difficult to apply these methods in new settings. We present a framework for failure sampling as probabilistic inference. To achieve this, we formulate the simulation of the system under test as a probabilistic program. The probabilistic program easily interfaces with general-purpose inference algorithms, making failure sampling techniques more accessible. Additionally, the probabilistic program easily incorporates some problem-specific knowledge like gradient information, which improves the efficiency of failure sampling.Autonomous systems may have high-dimensional trajectories and exhibit complex failure modes, making failure sampling challenging. Furthermore, many industry-level systems only provide input-output access, meaning that algorithms cannot use any internal system knowledge like gradients to improve scalability. We propose a method for failure sampling in these high-dimensional black-box settings using deep generative models. Our method adaptively trains a diffusion model to sample trajectories that lead to failure. This method scales to problems with thousands of dimensions and multiple failure modes.Methods for estimating the probability of failure often require many samples to achieve accurate estimates. Since failure events are rare and simulations may be complex, this can be computationally expensive. We propose a state-dependent importance sampling method that improves the efficiency of failure estimation. Our method learns a proposal distribution that builds trajectories in a state-dependent manner. The training process focuses on trajectories that are closer to failure.The sampling and estimation algorithms in this work are tested on a variety of problems. A simple inverted pendulum system is used to illustrate important concepts and algorithms. A rule-based autonomous driving policy is tested in a scenario with a pedestrian in a crosswalk. Finally, we validate a model of the ground collision-avoidance autopilot of the F-16 aircraft. Through experiments, we show that the presented methods can improve the generality, scalability, and efficiency of failure sampling and estimation for safety validation of autonomous systems.
- 일반주제명
- Aviation
- 일반주제명
- Design
- 일반주제명
- Autonomous vehicles
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-01A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358732
■00520260202104744
■006m o d
■007cr#unu||||||||
■020 ▼a9798290650234
■035 ▼a(MiAaPQ)AAI32149733
■035 ▼a(MiAaPQ)Stanfordtz122mx1710
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a629.13309
■1001 ▼aDelecki, Harrison.
■24510▼aSampling and Estimation for the Validation of Safety-Critical Autonomous Systems
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a115 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: A.
■500 ▼aAdvisor: Kochenderfer, Mykel.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aThere is increasing interest in deploying autonomous systems to domains such as transportation, healthcare, finance, and robotics because of their potential to improve efficiency and safety. These domains are safety-critical; the consequences of system failures can cause loss of human life or other significant costs. Therefore, we require rigorous validation of autonomous systems before deployment in these domains. Traditional safety validation approaches like real-world testing may be extremely expensive and time-consuming. Formal verification techniques are not scalable to complex systems and environments. Falsification methods provide failure examples in simulation, but they do not provide a comprehensive understanding of the system's behavior.In this thesis, we address these limitations by proposing techniques for failure sampling and estimation to validate safety-critical autonomous systems. Given a probabilistic model of system trajectories, our methods aim to sample from the distribution over trajectories that lead to failure (failure sampling) or to estimate the probability of failure (failure estimation). During system development, engineers may use failure sampling to identify failure modes and improve system safety. Failure estimation can be used to quantify the risk of system failure before deployment. Together, these approaches provide a comprehensive understanding of system behavior and enable rigorous safety validation. Previous methods for failure sampling and estimation have been based on Markov chain Monte Carlo, importance sampling, and dynamic programming. Although these methods have been successful in some applications, they have limitations in terms of generality, scalability, and efficiency.Many existing methods for failure sampling and estimation are tailored to specific system models and failure criteria. This makes it difficult to apply these methods in new settings. We present a framework for failure sampling as probabilistic inference. To achieve this, we formulate the simulation of the system under test as a probabilistic program. The probabilistic program easily interfaces with general-purpose inference algorithms, making failure sampling techniques more accessible. Additionally, the probabilistic program easily incorporates some problem-specific knowledge like gradient information, which improves the efficiency of failure sampling.Autonomous systems may have high-dimensional trajectories and exhibit complex failure modes, making failure sampling challenging. Furthermore, many industry-level systems only provide input-output access, meaning that algorithms cannot use any internal system knowledge like gradients to improve scalability. We propose a method for failure sampling in these high-dimensional black-box settings using deep generative models. Our method adaptively trains a diffusion model to sample trajectories that lead to failure. This method scales to problems with thousands of dimensions and multiple failure modes.Methods for estimating the probability of failure often require many samples to achieve accurate estimates. Since failure events are rare and simulations may be complex, this can be computationally expensive. We propose a state-dependent importance sampling method that improves the efficiency of failure estimation. Our method learns a proposal distribution that builds trajectories in a state-dependent manner. The training process focuses on trajectories that are closer to failure.The sampling and estimation algorithms in this work are tested on a variety of problems. A simple inverted pendulum system is used to illustrate important concepts and algorithms. A rule-based autonomous driving policy is tested in a scenario with a pedestrian in a crosswalk. Finally, we validate a model of the ground collision-avoidance autopilot of the F-16 aircraft. Through experiments, we show that the presented methods can improve the generality, scalability, and efficiency of failure sampling and estimation for safety validation of autonomous systems.
■590 ▼aSchool code: 0212.
■650 4▼aAviation
■650 4▼aDesign
■650 4▼aTraffic accidents & safety
■650 4▼aAutonomous vehicles
■690 ▼a0389
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-01A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358732▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


