서브메뉴
검색
From Representation Learning to Latent Dynamics and Planning
From Representation Learning to Latent Dynamics and Planning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103041
- ISBN
- 9798293886746
- DDC
- 310
- 서명/저자
- From Representation Learning to Latent Dynamics and Planning
- 발행사항
- [Sl] : New York University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 166 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: LeCun, Yann;Cho, Kyunghyun.
- 학위논문주기
- Thesis (Ph.D.)--New York University, 2025.
- 초록/해제
- 요약The field of artificial intelligence (AI) aims to build systems that can act intelligently in a variety of settings. To make decisions, an AI agent must first understand the context of the problem it is facing. This 'understanding' is investigated by the area of representation learning. Once we have a good representation, we can either directly train the system to output the correct decision using reinforcement learning (RL), or we can train a dynamics model and resort to planning. While the first approach is by far the more common one in the literature, the second approach is under-explored. In this thesis, I explore that gap in the literature, and study three components required for building a planning agent: representation learning, dynamics learning, and planning. I first present a representation learning method called X-CLR, which incorporates cross sample similarities to learn image representations. Then, I focus on dynamics learning, and present a way of splitting the dynamics to model the agent and the surroundings separately to improve planning performance. Additionally, I propose a method for learning representations and dynamics jointly without using reconstruction or rewards. Lastly, I show that representations and dynamics trained end-to-end can be used for planning to solve maze navigation tasks. I show that using dynamics learning and planning has several key advantages over direct action prediction commonly used in RL.
- 일반주제명
- Statistics
- 일반주제명
- Computer science
- 일반주제명
- Information technology
- 키워드
- Latent dynamics
- 기타저자
- New York University Center for Data Science
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017356816
■00520260202103041
■006m o d
■007cr#unu||||||||
■020 ▼a9798293886746
■035 ▼a(MiAaPQ)AAI31847770
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a310
■1001 ▼aSobal, Uladzislau.
■24510▼aFrom Representation Learning to Latent Dynamics and Planning
■260 ▼a[Sl]▼bNew York University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a166 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: LeCun, Yann;Cho, Kyunghyun.
■5021 ▼aThesis (Ph.D.)--New York University, 2025.
■520 ▼aThe field of artificial intelligence (AI) aims to build systems that can act intelligently in a variety of settings. To make decisions, an AI agent must first understand the context of the problem it is facing. This 'understanding' is investigated by the area of representation learning. Once we have a good representation, we can either directly train the system to output the correct decision using reinforcement learning (RL), or we can train a dynamics model and resort to planning. While the first approach is by far the more common one in the literature, the second approach is under-explored. In this thesis, I explore that gap in the literature, and study three components required for building a planning agent: representation learning, dynamics learning, and planning. I first present a representation learning method called X-CLR, which incorporates cross sample similarities to learn image representations. Then, I focus on dynamics learning, and present a way of splitting the dynamics to model the agent and the surroundings separately to improve planning performance. Additionally, I propose a method for learning representations and dynamics jointly without using reconstruction or rewards. Lastly, I show that representations and dynamics trained end-to-end can be used for planning to solve maze navigation tasks. I show that using dynamics learning and planning has several key advantages over direct action prediction commonly used in RL.
■590 ▼aSchool code: 0146.
■650 4▼aStatistics
■650 4▼aComputer science
■650 4▼aInformation technology
■653 ▼aRepresentation learning
■653 ▼aLatent dynamics
■653 ▼aReinforcement learning
■653 ▼aSimilarity graphs
■653 ▼aInverse dynamics modeling
■690 ▼a0463
■690 ▼a0489
■690 ▼a0984
■690 ▼a0800
■71020▼aNew York University▼bCenter for Data Science.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0146
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17356816▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
