서브메뉴
검색
Mathematical Aspects of Deep Reinforcement Learning
Mathematical Aspects of Deep Reinforcement Learning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105626
- ISBN
- 9798265428790
- DDC
- 794.8
- 저자명
- Dong, Kefan.
- 서명/저자
- Mathematical Aspects of Deep Reinforcement Learning
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 285 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: A.
- 주기사항
- Advisor: Ma, Tengyu;Brunskill, Emma;Roy, Benjamin Van.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약The recent decade has witnessed a wave of empirical success in deep reinforcement learning (deep RL)-RL with neural networks as the backbone function approximators-in solving complex decision-making problems and achieving superhuman performance in domains such as video games, board games, competitive programming, and mathematical reasoning. This thesis studies the mathematical aspects of deep reinforcement learning.In Part I of the thesis, we focus on the theoretical analysis of deep RL algorithms. The main technical challenge is that, unlike supervised learning, analyzing the RL algorithms in the traditional sense seems to require much stronger guarantees on the extrapolation power of function approximations, such as, e.g., l_inf error guarantees for learning the Q-function, which may not hold even for simple two-layer neural networks. In fact, we construct hard instances where the reward function is represented by a single ReLU neuron and show that finding a near-optimal policy is statistically hard in these cases.Hence, Chapter 2 proposes an alternative paradigm for analyzing deep RL algorithms inspired by the nonconvex optimization literature: we first design an algorithm, ViOlin, that converges to a local maximum of the reward function with a sample complexity determined by the sequential Rademacher complexity of the model class, and then prove that for particular instances, all local maxima are also global. This new paradigm allows us to design algorithms that achieve better regret on instances with better reward landscapes while gracefully giving up on fundamentally hard instances. In Chapter 3, we develop asymptotic instance-optimal algorithms for general bandit and RL problems with finite action spaces. Our algorithm adapts to the complexity of each problem instance and achieves lower regret on easier instances compared to the worst-case scenario. We determine the exact leading term of the optimal asymptotic regret for every instance, and prove that our algorithm, T2C, matches the optimal regret asymptotically. When applied to concrete problems, our results recover the classical gap-dependent bounds for multi-armed bandits and prior work on linear bandits, and improve upon the previous best instance-dependent upper bounds for reinforcement learning.In Part II of the thesis, we apply deep RL algorithms to train large language model formal theorem provers. The main technical challenge is the reward sparsity issue, which significantly slows down the training process. In particular, the model can only observe a positive reward when every step of the proof is correct. Therefore, the complexity of finding a correct proof increases exponentially with the number of required proof steps.In Chapter 4, we develop an algorithm, ProD, that encourages the model to decompose the proof into lemmas, and then recursively prove them. Inspired by how mathematicians train themselves, we reward correct novel lemmas that are proposed and proved even if the original theorem is too challenging to be proved by the current model, providing dense intermediate training signals. In Chapter 5, we design STP, Self-play Theorem Prover, that simultaneously assigns the model two roles: conjecturer and prover. The conjecturer is trained iteratively on previously generated conjectures that are barely provable for the current prover, incentivizing it to generate increasingly challenging conjectures over time. The prover attempts to prove these conjectures using standard expert iteration. We empirically evaluate both methods and demonstrate their efficiency on carefully curated test datasets and standard formal proof benchmarks.
- 일반주제명
- Computer & video games
- 일반주제명
- Deep learning
- 일반주제명
- Human performance
- 일반주제명
- Large language models
- 일반주제명
- Visualization
- 일반주제명
- Distance learning
- 일반주제명
- Decision making
- 일반주제명
- Neural networks
- 일반주제명
- Educational technology
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-05A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017360833
■00520260202105626
■006m o d
■007cr#unu||||||||
■020 ▼a9798265428790
■035 ▼a(MiAaPQ)AAI32316551
■035 ▼a(MiAaPQ)Stanfordnj670jf3453
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a794.8
■1001 ▼aDong, Kefan.
■24510▼aMathematical Aspects of Deep Reinforcement Learning
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a285 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: A.
■500 ▼aAdvisor: Ma, Tengyu;Brunskill, Emma;Roy, Benjamin Van.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aThe recent decade has witnessed a wave of empirical success in deep reinforcement learning (deep RL)-RL with neural networks as the backbone function approximators-in solving complex decision-making problems and achieving superhuman performance in domains such as video games, board games, competitive programming, and mathematical reasoning. This thesis studies the mathematical aspects of deep reinforcement learning.In Part I of the thesis, we focus on the theoretical analysis of deep RL algorithms. The main technical challenge is that, unlike supervised learning, analyzing the RL algorithms in the traditional sense seems to require much stronger guarantees on the extrapolation power of function approximations, such as, e.g., l_inf error guarantees for learning the Q-function, which may not hold even for simple two-layer neural networks. In fact, we construct hard instances where the reward function is represented by a single ReLU neuron and show that finding a near-optimal policy is statistically hard in these cases.Hence, Chapter 2 proposes an alternative paradigm for analyzing deep RL algorithms inspired by the nonconvex optimization literature: we first design an algorithm, ViOlin, that converges to a local maximum of the reward function with a sample complexity determined by the sequential Rademacher complexity of the model class, and then prove that for particular instances, all local maxima are also global. This new paradigm allows us to design algorithms that achieve better regret on instances with better reward landscapes while gracefully giving up on fundamentally hard instances. In Chapter 3, we develop asymptotic instance-optimal algorithms for general bandit and RL problems with finite action spaces. Our algorithm adapts to the complexity of each problem instance and achieves lower regret on easier instances compared to the worst-case scenario. We determine the exact leading term of the optimal asymptotic regret for every instance, and prove that our algorithm, T2C, matches the optimal regret asymptotically. When applied to concrete problems, our results recover the classical gap-dependent bounds for multi-armed bandits and prior work on linear bandits, and improve upon the previous best instance-dependent upper bounds for reinforcement learning.In Part II of the thesis, we apply deep RL algorithms to train large language model formal theorem provers. The main technical challenge is the reward sparsity issue, which significantly slows down the training process. In particular, the model can only observe a positive reward when every step of the proof is correct. Therefore, the complexity of finding a correct proof increases exponentially with the number of required proof steps.In Chapter 4, we develop an algorithm, ProD, that encourages the model to decompose the proof into lemmas, and then recursively prove them. Inspired by how mathematicians train themselves, we reward correct novel lemmas that are proposed and proved even if the original theorem is too challenging to be proved by the current model, providing dense intermediate training signals. In Chapter 5, we design STP, Self-play Theorem Prover, that simultaneously assigns the model two roles: conjecturer and prover. The conjecturer is trained iteratively on previously generated conjectures that are barely provable for the current prover, incentivizing it to generate increasingly challenging conjectures over time. The prover attempts to prove these conjectures using standard expert iteration. We empirically evaluate both methods and demonstrate their efficiency on carefully curated test datasets and standard formal proof benchmarks.
■590 ▼aSchool code: 0212.
■650 4▼aComputer & video games
■650 4▼aDeep learning
■650 4▼aHuman performance
■650 4▼aLarge language models
■650 4▼aVisualization
■650 4▼aDistance learning
■650 4▼aDecision making
■650 4▼aNeural networks
■650 4▼aEducational technology
■690 ▼a0800
■690 ▼a0710
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-05A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360833▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


