서브메뉴
검색
Dense Learning for Advancing Safety Performance of Autonomous Vehicles
Dense Learning for Advancing Safety Performance of Autonomous Vehicles
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105237
- ISBN
- 9798291567968
- DDC
- 004
- 저자명
- Zhu, Haojie.
- 서명/저자
- Dense Learning for Advancing Safety Performance of Autonomous Vehicles
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 124 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Liu, Henry X.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약Autonomous Vehicles (AVs) have garnered substantial attention and investment over the past two decades, yet their commercial availability remains limited. A significant barrier to widespread deployment is the ongoing struggle to attain safety performance comparable to or surpassing that of human drivers, particularly given the extreme rarity of safety-critical events in complex, real-world driving environments. Existing approaches, which heavily rely on learning from failure data, often encounter a performance plateau caused by the seesaw effect-improvements in some scenarios are counterbalanced by regressions in others, limiting overall progress in AV safety performance. This dissertation introduces Dense Reinforcement Learning, a new learning paradigm designed to overcome the limitations imposed by the Curse of Rarity (CoR). By densifying the information content within training datasets, this approach reduces variance in policy optimization and improves training efficiency without compromising unbiasedness. Dense Reinforcement Learning prioritizes both failure cases and informative successes, selectively sampling data according to their contribution to policy gradients and exposure frequencies, while discarding non-informative examples that dilute learning progress. To support this framework, a learned safety metric is developed to evaluate the risk levels of driving behaviors and assist in identifying informative data samples during training. Trained on diverse autonomous driving trajectories, the learned safety metric demonstrates superior accuracy in predicting safety-critical events compared to widely used baselines. Moreover, it shows strong generalization capabilities across highways, intersections, roundabouts, and other complex driving environments, further enhancing its utility for AV safety training. The Dense Reinforcement Learning approach is validated through the development and training of a safety-critical driving agent in an urban test track environment augmented by mixed reality technology. Extensive experiments demonstrate that dense learning substantially enhances AV safety performance, achieving improvements ranging from one to two orders of magnitude compared to baseline methods. These results highlight the potential of dense learning to break through the safety performance stagnation commonly observed in traditional Reinforcement Learning (RL) frameworks, paving the way for AVs to achieve and even surpass human driving safety performance. Beyond AV applications, the dense learning methodology offers broader implications for machine learning domains involving rare, high-stakes events. In particular, dense learning principles have the potential to improve supervised learning processes and enhance the pre-training and fine-tuning stages of Large Language Models (LLMs), where data sparsity and safety considerations remain critical challenges.
- 일반주제명
- Computer science
- 일반주제명
- Transportation
- 일반주제명
- Computer engineering
- 일반주제명
- Automotive engineering
- 키워드
- Safety
- 기타저자
- University of Michigan Civil Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359928
■00520260202105237
■006m o d
■007cr#unu||||||||
■020 ▼a9798291567968
■035 ▼a(MiAaPQ)AAI32271961
■035 ▼a(MiAaPQ)umichrackham006532
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aZhu, Haojie.
■24510▼aDense Learning for Advancing Safety Performance of Autonomous Vehicles
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a124 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Liu, Henry X.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aAutonomous Vehicles (AVs) have garnered substantial attention and investment over the past two decades, yet their commercial availability remains limited. A significant barrier to widespread deployment is the ongoing struggle to attain safety performance comparable to or surpassing that of human drivers, particularly given the extreme rarity of safety-critical events in complex, real-world driving environments. Existing approaches, which heavily rely on learning from failure data, often encounter a performance plateau caused by the seesaw effect-improvements in some scenarios are counterbalanced by regressions in others, limiting overall progress in AV safety performance. This dissertation introduces Dense Reinforcement Learning, a new learning paradigm designed to overcome the limitations imposed by the Curse of Rarity (CoR). By densifying the information content within training datasets, this approach reduces variance in policy optimization and improves training efficiency without compromising unbiasedness. Dense Reinforcement Learning prioritizes both failure cases and informative successes, selectively sampling data according to their contribution to policy gradients and exposure frequencies, while discarding non-informative examples that dilute learning progress. To support this framework, a learned safety metric is developed to evaluate the risk levels of driving behaviors and assist in identifying informative data samples during training. Trained on diverse autonomous driving trajectories, the learned safety metric demonstrates superior accuracy in predicting safety-critical events compared to widely used baselines. Moreover, it shows strong generalization capabilities across highways, intersections, roundabouts, and other complex driving environments, further enhancing its utility for AV safety training. The Dense Reinforcement Learning approach is validated through the development and training of a safety-critical driving agent in an urban test track environment augmented by mixed reality technology. Extensive experiments demonstrate that dense learning substantially enhances AV safety performance, achieving improvements ranging from one to two orders of magnitude compared to baseline methods. These results highlight the potential of dense learning to break through the safety performance stagnation commonly observed in traditional Reinforcement Learning (RL) frameworks, paving the way for AVs to achieve and even surpass human driving safety performance. Beyond AV applications, the dense learning methodology offers broader implications for machine learning domains involving rare, high-stakes events. In particular, dense learning principles have the potential to improve supervised learning processes and enhance the pre-training and fine-tuning stages of Large Language Models (LLMs), where data sparsity and safety considerations remain critical challenges.
■590 ▼aSchool code: 0127.
■650 4▼aComputer science
■650 4▼aTransportation
■650 4▼aComputer engineering
■650 4▼aAutomotive engineering
■653 ▼aAutonomous vehicles
■653 ▼aSafety
■653 ▼aSafety-critical events
■653 ▼aDense Reinforcement Learning
■653 ▼aLarge Language Models
■690 ▼a0709
■690 ▼a0984
■690 ▼a0464
■690 ▼a0540
■71020▼aUniversity of Michigan▼bCivil Engineering.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359928▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


