서브메뉴
검색
Bridging Demonstrations and Decisions: Theory and Algorithms for Provable Imitation Learning
Bridging Demonstrations and Decisions: Theory and Algorithms for Provable Imitation Learning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104804
- ISBN
- 9798288863653
- DDC
- 004
- 서명/저자
- Bridging Demonstrations and Decisions: Theory and Algorithms for Provable Imitation Learning
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 184 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Ramchandran, Kannan;Jiao, Jiantao.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Classical supervised learning paradigms typically assume that training data samples are independently drawn from a target distribution. However, real-world scenarios frequently violate this assumption, presenting data that are temporally correlated, dynamically evolving, or a result of strategic interaction. Learning in these settings is often significantly more challenging both from a theoretical and a practical point of view for these reasons. Recent advances in reinforcement learning (RL) have shown that it is possible to train agents which can operate and generalize in settings where the number of possible outcomes is huge. However, there are a number of challenges with running RL algorithms: these approaches rely on collecting a large amount of ``exploration'' data, resulting from interaction with a dynamic environment. This form of active data collection is often prohibitively expensive in practice, making mistakes may be costly, such as in settings involving human interaction, and this form of data collection may be hard to reuse. Mitigating these concerns requires developing new frameworks for RL.In this dissertation, we develop algorithms and analyses for an alternate learning paradigm that aims to utilize static datasets generated by a demonstrator for training policies. This paradigm broadens the applicability of RL to a variety of decision-making problems where historical datasets already exist or can be collected via domain-specific strategies, and which are infinitely reusable. It also brings the scalability and reliability benefits that modern supervised and unsupervised ML methods enjoy into RL. That said, instantiating this paradigm is challenging as it requires reconciling the static nature of learning from offline datasets (against a fixed distribution of problem instances) with the traditionally interactive nature of RL. A major part of this thesis is geared toward addressing precisely how much of a price one must pay to forgo the power of environment interaction.Imitation Learning (IL) techniques have found a home in several areas, from policy initialization in game-solving agents like AlphaGo, and more recently as a fine-tuning backbone in the form of supervised fine‐tuning (SFT) for large language models (LLMs). The key challenge in all these domains is obtaining sufficiently large, diverse, and high‐quality demonstration datasets. While more data typically yields better performance, expert data can be expensive to collect. We see this challenge manifest in several forms: in robotics and control, acquiring teleoperated or human‐guided trajectories often requires specialized hardware (e.g. motion‐capture rigs or force‐feedback devices) and careful calibration, limiting the scale of dataset collection. In autonomous driving, critical "edge‐case'' scenarios (e.g. collision avoidance in unusual weather) are inherently rare, yet essential for safety; collecting them either in simulation or on‐road is time‐consuming and costly. On the other hand, for training LLMs: fine‐tuning large language models relies on human‐annotated data, which is hard to parallelize and incurs substantial annotation time.Thus, it is pertinent to understand how best to utilize the dataset and leverage favorable properties of the environment and the demonstrator. In this thesis, we will build an understanding of these questions via studying Imitation Learning from a theoretical point of view. We will formulate a statistical question and analyze the best achievable statistical error that algorithms can achieve in various models of feedback. We will utilize these insights to scale up these algorithmic ideas and leverage the expressivity and representation power offered by highly expressive function approximators to develop performant, practical algorithms. Along the way, we will develop various insights into the landscape of the IL problem, and build a comprehensive understanding and unification of algorithms that have already been successfully deployed in practice, such as Behavior Cloning and GAIL and provide principled improvements to these approaches.
- 일반주제명
- Computer science
- 일반주제명
- Electrical engineering
- 일반주제명
- Statistics
- 일반주제명
- Computer engineering
- 일반주제명
- Applied mathematics
- 키워드
- Behavior cloning
- 키워드
- Minimax
- 기타저자
- University of California, Berkeley Electrical Engineering & Computer Sciences
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358879
■00520260202104804
■006m o d
■007cr#unu||||||||
■020 ▼a9798288863653
■035 ▼a(MiAaPQ)AAI32165026
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aRajaraman, Nived.
■24510▼aBridging Demonstrations and Decisions: Theory and Algorithms for Provable Imitation Learning
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a184 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Ramchandran, Kannan;Jiao, Jiantao.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aClassical supervised learning paradigms typically assume that training data samples are independently drawn from a target distribution. However, real-world scenarios frequently violate this assumption, presenting data that are temporally correlated, dynamically evolving, or a result of strategic interaction. Learning in these settings is often significantly more challenging both from a theoretical and a practical point of view for these reasons. Recent advances in reinforcement learning (RL) have shown that it is possible to train agents which can operate and generalize in settings where the number of possible outcomes is huge. However, there are a number of challenges with running RL algorithms: these approaches rely on collecting a large amount of ``exploration'' data, resulting from interaction with a dynamic environment. This form of active data collection is often prohibitively expensive in practice, making mistakes may be costly, such as in settings involving human interaction, and this form of data collection may be hard to reuse. Mitigating these concerns requires developing new frameworks for RL.In this dissertation, we develop algorithms and analyses for an alternate learning paradigm that aims to utilize static datasets generated by a demonstrator for training policies. This paradigm broadens the applicability of RL to a variety of decision-making problems where historical datasets already exist or can be collected via domain-specific strategies, and which are infinitely reusable. It also brings the scalability and reliability benefits that modern supervised and unsupervised ML methods enjoy into RL. That said, instantiating this paradigm is challenging as it requires reconciling the static nature of learning from offline datasets (against a fixed distribution of problem instances) with the traditionally interactive nature of RL. A major part of this thesis is geared toward addressing precisely how much of a price one must pay to forgo the power of environment interaction.Imitation Learning (IL) techniques have found a home in several areas, from policy initialization in game-solving agents like AlphaGo, and more recently as a fine-tuning backbone in the form of supervised fine‐tuning (SFT) for large language models (LLMs). The key challenge in all these domains is obtaining sufficiently large, diverse, and high‐quality demonstration datasets. While more data typically yields better performance, expert data can be expensive to collect. We see this challenge manifest in several forms: in robotics and control, acquiring teleoperated or human‐guided trajectories often requires specialized hardware (e.g. motion‐capture rigs or force‐feedback devices) and careful calibration, limiting the scale of dataset collection. In autonomous driving, critical "edge‐case'' scenarios (e.g. collision avoidance in unusual weather) are inherently rare, yet essential for safety; collecting them either in simulation or on‐road is time‐consuming and costly. On the other hand, for training LLMs: fine‐tuning large language models relies on human‐annotated data, which is hard to parallelize and incurs substantial annotation time.Thus, it is pertinent to understand how best to utilize the dataset and leverage favorable properties of the environment and the demonstrator. In this thesis, we will build an understanding of these questions via studying Imitation Learning from a theoretical point of view. We will formulate a statistical question and analyze the best achievable statistical error that algorithms can achieve in various models of feedback. We will utilize these insights to scale up these algorithmic ideas and leverage the expressivity and representation power offered by highly expressive function approximators to develop performant, practical algorithms. Along the way, we will develop various insights into the landscape of the IL problem, and build a comprehensive understanding and unification of algorithms that have already been successfully deployed in practice, such as Behavior Cloning and GAIL and provide principled improvements to these approaches.
■590 ▼aSchool code: 0028.
■650 4▼aComputer science
■650 4▼aElectrical engineering
■650 4▼aStatistics
■650 4▼aComputer engineering
■650 4▼aApplied mathematics
■653 ▼aBehavior cloning
■653 ▼aImitation learning
■653 ▼aMinimax
■653 ▼aReinforcement learning
■653 ▼aSupervised fine‐tuning
■690 ▼a0984
■690 ▼a0544
■690 ▼a0463
■690 ▼a0464
■690 ▼a0364
■71020▼aUniversity of California, Berkeley▼bElectrical Engineering & Computer Sciences.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358879▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


