서브메뉴
검색
Data Collection for Robust Machine Learning in Multi-Agent Systems
Data Collection for Robust Machine Learning in Multi-Agent Systems
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260311091527.5
- ISBN
- 9798270231125
- DDC
- 620
- 저자명
- Akcin, Oguzhan
- 서명/저자
- Data Collection for Robust Machine Learning in Multi-Agent Systems / Oguzhan Akcin
- 발행사항
- [Sl] : The University of Texas at Austin, 2025
- 형태사항
- 1 electronic resource (203 pages)
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
- 주기사항
- Advisors: Chinchali, Sandeep Committee members: Fridovich-Keil, David; Vikalo, Haris; Vishwanath, Sriram; Stone, Peter.
- 학위논문주기
- - Ph.D. : The University of Texas at Austin, 2025.
- 초록/해제
- 요약Modern robotics systems, such as autonomous vehicles and mobile manipulators, often operate as distributed fleets that collect data to train and improve machine learning models for perception, prediction, and decision-making. These systems encounter several practical challenges in real-world deployments: network bandwidth limits restrict the volume of data that can be uploaded for training, labeling resources are expensive and scarce, human supervision is limited and unreliable, and decentralized operation often leads to redundant or uninformative data collection. These issues demand scalable and intelligent data curation strategies that are adaptive to network conditions, resource constraints, and fleet heterogeneity. This dissertation develops frameworks for distributed data collection and curation in multi-robot systems, grounded in submodular optimization, active learning, and game-theoretic coordination. First, it introduces a game-theoretic framework for decentralized data selection, where robots are modeled as independent players that compute upload strategies that collectively converge to high-quality datasets using only local observations and limited information exchange. Second, it presents a distributed active learning algorithm based on submodular maximization, enabling robots to collaboratively select informative samples while avoiding redundancy. Third, it formulates the allocation of limited human supervision as a stochastic submodular maximization problem, allowing adaptive assignment in the presence of uncertain connectivity. Finally, it proposes a two-stage upload and annotation framework, where robots upload selected data under bandwidth constraints and a central server chooses a subset to label under an annotation budget. Together, these contributions form a scalable and robust data collection paradigm for real-world robotic fleets. The proposed methods are validated across diverse sensing modalities, environments, and downstream tasks-including physical robot experiments, audio and visual classification, and trajectory prediction in autonomous driving-demonstrating consistent improvements in data efficiency, model performance, and human supervision utility. This work offers a principled and practical foundation for building the next generation of adaptive, data-driven multi-robot systems.
- 언어주기
- English
- 일반주제명
- Computer science
- 일반주제명
- Information technology
- 일반주제명
- Robotics
- 키워드
- Machine learning
- 키워드
- Data collection
- 키워드
- Robots
- 기타저자
- The University of Texas at Austin Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260311s2025 us eng d■001000017361176
■00520260311091527.5
■006m o d
■007cr|nu||||||||
■020 ▼a9798270231125
■040 ▼aMiAaPQD▼beng▼cMiAaPQD▼erda
■082 ▼a620
■1001 ▼aAkcin, Oguzhan▼eauthor.
■24510▼aData Collection for Robust Machine Learning in Multi-Agent Systems ▼cOguzhan Akcin
■260 ▼a[Sl]▼bThe University of Texas at Austin▼c2025
■264 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a1 electronic resource (203 pages)
■336 ▼atext▼btxt▼2rdacontent
■337 ▼acomputer▼bc▼2rdamedia
■338 ▼aonline resource▼bcr▼2rdacarrier
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: B.
■500 ▼aAdvisors: Chinchali, Sandeep Committee members: Fridovich-Keil, David; Vikalo, Haris; Vishwanath, Sriram; Stone, Peter.
■5021 ▼bPh.D.▼cThe University of Texas at Austin▼d2025.
■520 ▼aModern robotics systems, such as autonomous vehicles and mobile manipulators, often operate as distributed fleets that collect data to train and improve machine learning models for perception, prediction, and decision-making. These systems encounter several practical challenges in real-world deployments: network bandwidth limits restrict the volume of data that can be uploaded for training, labeling resources are expensive and scarce, human supervision is limited and unreliable, and decentralized operation often leads to redundant or uninformative data collection. These issues demand scalable and intelligent data curation strategies that are adaptive to network conditions, resource constraints, and fleet heterogeneity. This dissertation develops frameworks for distributed data collection and curation in multi-robot systems, grounded in submodular optimization, active learning, and game-theoretic coordination. First, it introduces a game-theoretic framework for decentralized data selection, where robots are modeled as independent players that compute upload strategies that collectively converge to high-quality datasets using only local observations and limited information exchange. Second, it presents a distributed active learning algorithm based on submodular maximization, enabling robots to collaboratively select informative samples while avoiding redundancy. Third, it formulates the allocation of limited human supervision as a stochastic submodular maximization problem, allowing adaptive assignment in the presence of uncertain connectivity. Finally, it proposes a two-stage upload and annotation framework, where robots upload selected data under bandwidth constraints and a central server chooses a subset to label under an annotation budget. Together, these contributions form a scalable and robust data collection paradigm for real-world robotic fleets. The proposed methods are validated across diverse sensing modalities, environments, and downstream tasks-including physical robot experiments, audio and visual classification, and trajectory prediction in autonomous driving-demonstrating consistent improvements in data efficiency, model performance, and human supervision utility. This work offers a principled and practical foundation for building the next generation of adaptive, data-driven multi-robot systems.
■546 ▼aEnglish
■590 ▼aSchool code: 0227
■650 4▼aComputer science
■650 4▼aInformation technology
■650 4▼aRobotics
■653 ▼aModern robotics systems
■653 ▼aMachine learning
■653 ▼aMulti-agent systems
■653 ▼aData collection
■653 ▼aRobots
■7102 ▼aThe University of Texas at Austin▼bElectrical and Computer Engineering.▼edegree granting institution.
■7201 ▼aChinchali, Sandeep▼edegree supervisor.
■7730 ▼tDissertations Abstracts International▼g87-06B.
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17361176▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
