서브메뉴
검색
Formal and Statistical Evaluation of Robotic Systems with Learned Components
Formal and Statistical Evaluation of Robotic Systems with Learned Components
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211153052
- ISBN
- 9798346379539
- DDC
- 519.93
- 서명/저자
- Formal and Statistical Evaluation of Robotic Systems with Learned Components
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 174 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-05, Section: B.
- 주기사항
- Advisor: Schwager, Mac.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약Machine learning is now an essential tool for robotics which enables components of a robot's perception, prediction, or planning modules to be learned from data. This capability has allowed robots to operate in less structured environments than previously possible. However, this feature is accompanied by additional challenges. Namely, evaluating the safety and performance of robotic systems with learned components is challenging due to the complexity and opacity of learned components. This thesis addresses this problem by proposing new methodologies for both formal and statistical evaluation.In Part I of this thesis, we present a method for formal analysis of neural networks, leveraging insights from analysis of piecewise-affine functions. We show how a common class of neural networks can be equivalently represented as a piecewise-affine function, and provide an efficient algorithm for finding this alternative representation. Then, we show how the piecewise-affine representation lends itself to analyzing and unveiling various properties of the neural network. Specifically, for dynamical systems represented as neural networks, we show how the piecewise-affine representation can be used to compute exact forward and backward reachable sets, invariant sets, and regions of attraction. Furthermore, we show how to use the piecewise-affine representation to verify input-output properties of neural networks. Lastly, we show that this representation allows us to check whether the neural network is invertible, and how invertibility can be leveraged to provide significant (15x) speedups to the computation of backward reachable sets and regions of attraction.In Part II of this thesis, we pivot from formal reachability-based evaluation to iv statistical evaluation. When the complexity of the learned components precludes the use of formal evaluation, statistical techniques can be used to infer the safety and performance of robotic systems. We first use statistical techniques which make no distributional assumptions on the observations or dynamics of the robot to place probabilistic bounds on various safety and performance measures for the robot. We then provide an analysis which details how these probabilistic bounds are affected by distribution shift (e.g., sim-to-real), and how to construct bounds which are robust to specified levels of distribution shift. When robot evaluations are limited to few policy rollouts, we show how to construct optimally sample-efficient bounds on the cumulative distribution function of robot performance. Finally, we provide an opensource Python package for computing sample-efficient bounds on the probability of success.
- 일반주제명
- Control theory
- 일반주제명
- Aircraft
- 일반주제명
- Success
- 일반주제명
- Neural networks
- 일반주제명
- Robots
- 일반주제명
- Dynamical systems
- 일반주제명
- Robotics
- 일반주제명
- Mathematics
- 일반주제명
- Statistics
- 일반주제명
- Systems science
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164827
■00520250211153052
■006m o d
■007cr#unu||||||||
■020 ▼a9798346379539
■035 ▼a(MiAaPQ)AAI31643339
■035 ▼a(MiAaPQ)Stanfordkx343pb4749
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a519.93
■1001 ▼aVincent, Joseph Anthony.
■24510▼aFormal and Statistical Evaluation of Robotic Systems with Learned Components
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a174 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-05, Section: B.
■500 ▼aAdvisor: Schwager, Mac.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aMachine learning is now an essential tool for robotics which enables components of a robot's perception, prediction, or planning modules to be learned from data. This capability has allowed robots to operate in less structured environments than previously possible. However, this feature is accompanied by additional challenges. Namely, evaluating the safety and performance of robotic systems with learned components is challenging due to the complexity and opacity of learned components. This thesis addresses this problem by proposing new methodologies for both formal and statistical evaluation.In Part I of this thesis, we present a method for formal analysis of neural networks, leveraging insights from analysis of piecewise-affine functions. We show how a common class of neural networks can be equivalently represented as a piecewise-affine function, and provide an efficient algorithm for finding this alternative representation. Then, we show how the piecewise-affine representation lends itself to analyzing and unveiling various properties of the neural network. Specifically, for dynamical systems represented as neural networks, we show how the piecewise-affine representation can be used to compute exact forward and backward reachable sets, invariant sets, and regions of attraction. Furthermore, we show how to use the piecewise-affine representation to verify input-output properties of neural networks. Lastly, we show that this representation allows us to check whether the neural network is invertible, and how invertibility can be leveraged to provide significant (15x) speedups to the computation of backward reachable sets and regions of attraction.In Part II of this thesis, we pivot from formal reachability-based evaluation to iv statistical evaluation. When the complexity of the learned components precludes the use of formal evaluation, statistical techniques can be used to infer the safety and performance of robotic systems. We first use statistical techniques which make no distributional assumptions on the observations or dynamics of the robot to place probabilistic bounds on various safety and performance measures for the robot. We then provide an analysis which details how these probabilistic bounds are affected by distribution shift (e.g., sim-to-real), and how to construct bounds which are robust to specified levels of distribution shift. When robot evaluations are limited to few policy rollouts, we show how to construct optimally sample-efficient bounds on the cumulative distribution function of robot performance. Finally, we provide an opensource Python package for computing sample-efficient bounds on the probability of success.
■590 ▼aSchool code: 0212.
■650 4▼aControl theory
■650 4▼aAircraft
■650 4▼aSuccess
■650 4▼aNeural networks
■650 4▼aRobots
■650 4▼aDynamical systems
■650 4▼aRobotics
■650 4▼aMathematics
■650 4▼aStatistics
■650 4▼aSystems science
■690 ▼a0771
■690 ▼a0800
■690 ▼a0405
■690 ▼a0463
■690 ▼a0790
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-05B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164827▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
