서브메뉴
검색
Reinforcement Learning via Differentiable Simulation: Applications in Operations Management
Reinforcement Learning via Differentiable Simulation: Applications in Operations Management
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104708
- ISBN
- 9798286498451
- DDC
- 658
- 저자명
- Che, Ethan.
- 서명/저자
- Reinforcement Learning via Differentiable Simulation: Applications in Operations Management
- 발행사항
- [Sl] : Columbia University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 348 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Dong, Jing;Namkoong, Hongseok.
- 학위논문주기
- Thesis (Ph.D.)--Columbia University, 2025.
- 초록/해제
- 요약Reinforcement learning (RL) is a promising approach for solving dynamic decision-making problems that arise in operations management applications, including scheduling, routing, supply chain management, and adaptive experimentation. However, so far RL algorithms have not experienced widespread industrial adoption due to high sample complexity and difficulty in scaling to large-scale problems. To address these issues, this dissertation proposes a model-based approach to RL via differentiable simulation. Differentiable simulation is a technique for computing gradients of a system's dynamics with respect to any parameters or inputs, which has been widely used in physics and robotics, with roots in the stochastic simulation literature dating back to the 1980s-1990s with methodologies such as infinitesimal perturbation analysis (IPA) and generalized likelihood ratio estimation. We take inspiration from these ideas, which have typically been applied to systems with a handful of parameters (e.g. service rates or base-stock levels), and develop frameworks for applying them to reinforcement learning problems in different operations applications.The first study develops a differentiable optimization framework for batched adaptive experimentation. Sampling with large batches is a ubiquitious practice in online experimentation platforms, due to operational constraints as well as concerns over delayed feedback. Using large batch statistical approximations, we develop a differentiable formulation of Bayesian regret for adaptive experimentation which can be directly optimized, in contrast with existing heuristics that optimize proxies of regret. By using a Bayesain model of how additional data affects future posterior beliefs, this explicitly captures the trade-off between exploration and exploitation. Using differentiability of Bayesian regret, we propose a model-predictive control algorithm for adaptive experimentation in batched settings with short time horizons. The algorithm selects the sampling allocations-the number of samples assigned to each treatment arm-to directly minimize Bayesian regret and re-optimizes this after each batch. We find through extensive numerical experiments that our algorithm outperforms existing heuristics (e.g. Thompson Sampling) especially under short time horizons. We also make theoretical connections between model-predictive control and posterior sampling. The second study follows up on the first study and provides a generalization of our framework for contextual settings, multiple-objectives, constraints, and non-stationarity, which are common considerations in online experimentation platforms. We find that under these different considerations, the design of effective adaptive algorithms is highly non-trivial, as even simple strategies like uniform allocation can outperform adaptive algorithms like Thompson Sampling. In contrast, our framework allows the experimenter to flexibly include these considerations in the optimization problem, rather than developing bespoke algorithms for different situations and combinations of considerations. Not only does this optimization-based approach come with new theoretical guarantees in wide range of settings, in several experiments involving non-stationarity, constraints, and multiple-objectives, our algorithm is able to outperform existing heuristics such as Thompson Sampling.The third study develops a model-based reinforcement learning approach for queueing network control through differentiable discrete-event simulation. Model-free RL methods have been applied to queueing networks, but they are sample inefficient and suffer from slow convergence. Despite non-differentiability of the dynamics of queueing networks, we formulate a differentiable approximation of the dynamics and use this to obtain pathwise policy gradients that are orders of magnitude less noisy than standard REINFORCE gradients. Using these pathwise policy gradients, we are able to learn scheduling policies that outperform or match the policies trained with model-free RL with significantly less data. Unlike prior tailored approaches to queueing, our methods can flexibly handle realistic scenarios, including systems operating in non-stationary environments and those with non-exponential interarrival/service times.The fourth chapter is a theoretical study of the convergence of stochastic gradient descent with Markovian feedback. While there are many works on the convergence of policy gradient with REINFORCE-based gradient estimation, in this chapter we provide a convergence analysis for broader gradient estimators, including differentiable simulation or finite-difference based estimators. The core challenge is that the policy changes the environment and thereby affects the data used to update the policy. The influence of previous decisions on the data generated introduces bias in the gradient estimate, which presents a potential source of instability for online learning not present in the iid case. In this paper, we introduce simple criteria for the adaptively generated data stream to guarantee the convergence of SGD, which depend on the mixing time of the policy-induced dynamics. Our Lyapunov-function analysis allows one to translate existing stability analysis of stochastic systems studied in operations research into convergence rates for SGD, and we demonstrate this for queueing and inventory management problems. We also showcase how our result can be applied to study the sample complexity of an actor-critic policy gradient algorithm.
- 일반주제명
- Industrial engineering
- 일반주제명
- Statistics
- 키워드
- Machine learning
- 키워드
- Queuing systems
- 기타저자
- Columbia University Operations Research
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358479
■00520260202104708
■006m o d
■007cr#unu||||||||
■020 ▼a9798286498451
■035 ▼a(MiAaPQ)AAI32117746
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a658
■1001 ▼aChe, Ethan.
■24510▼aReinforcement Learning via Differentiable Simulation: Applications in Operations Management
■260 ▼a[Sl]▼bColumbia University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a348 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Dong, Jing;Namkoong, Hongseok.
■5021 ▼aThesis (Ph.D.)--Columbia University, 2025.
■520 ▼aReinforcement learning (RL) is a promising approach for solving dynamic decision-making problems that arise in operations management applications, including scheduling, routing, supply chain management, and adaptive experimentation. However, so far RL algorithms have not experienced widespread industrial adoption due to high sample complexity and difficulty in scaling to large-scale problems. To address these issues, this dissertation proposes a model-based approach to RL via differentiable simulation. Differentiable simulation is a technique for computing gradients of a system's dynamics with respect to any parameters or inputs, which has been widely used in physics and robotics, with roots in the stochastic simulation literature dating back to the 1980s-1990s with methodologies such as infinitesimal perturbation analysis (IPA) and generalized likelihood ratio estimation. We take inspiration from these ideas, which have typically been applied to systems with a handful of parameters (e.g. service rates or base-stock levels), and develop frameworks for applying them to reinforcement learning problems in different operations applications.The first study develops a differentiable optimization framework for batched adaptive experimentation. Sampling with large batches is a ubiquitious practice in online experimentation platforms, due to operational constraints as well as concerns over delayed feedback. Using large batch statistical approximations, we develop a differentiable formulation of Bayesian regret for adaptive experimentation which can be directly optimized, in contrast with existing heuristics that optimize proxies of regret. By using a Bayesain model of how additional data affects future posterior beliefs, this explicitly captures the trade-off between exploration and exploitation. Using differentiability of Bayesian regret, we propose a model-predictive control algorithm for adaptive experimentation in batched settings with short time horizons. The algorithm selects the sampling allocations-the number of samples assigned to each treatment arm-to directly minimize Bayesian regret and re-optimizes this after each batch. We find through extensive numerical experiments that our algorithm outperforms existing heuristics (e.g. Thompson Sampling) especially under short time horizons. We also make theoretical connections between model-predictive control and posterior sampling. The second study follows up on the first study and provides a generalization of our framework for contextual settings, multiple-objectives, constraints, and non-stationarity, which are common considerations in online experimentation platforms. We find that under these different considerations, the design of effective adaptive algorithms is highly non-trivial, as even simple strategies like uniform allocation can outperform adaptive algorithms like Thompson Sampling. In contrast, our framework allows the experimenter to flexibly include these considerations in the optimization problem, rather than developing bespoke algorithms for different situations and combinations of considerations. Not only does this optimization-based approach come with new theoretical guarantees in wide range of settings, in several experiments involving non-stationarity, constraints, and multiple-objectives, our algorithm is able to outperform existing heuristics such as Thompson Sampling.The third study develops a model-based reinforcement learning approach for queueing network control through differentiable discrete-event simulation. Model-free RL methods have been applied to queueing networks, but they are sample inefficient and suffer from slow convergence. Despite non-differentiability of the dynamics of queueing networks, we formulate a differentiable approximation of the dynamics and use this to obtain pathwise policy gradients that are orders of magnitude less noisy than standard REINFORCE gradients. Using these pathwise policy gradients, we are able to learn scheduling policies that outperform or match the policies trained with model-free RL with significantly less data. Unlike prior tailored approaches to queueing, our methods can flexibly handle realistic scenarios, including systems operating in non-stationary environments and those with non-exponential interarrival/service times.The fourth chapter is a theoretical study of the convergence of stochastic gradient descent with Markovian feedback. While there are many works on the convergence of policy gradient with REINFORCE-based gradient estimation, in this chapter we provide a convergence analysis for broader gradient estimators, including differentiable simulation or finite-difference based estimators. The core challenge is that the policy changes the environment and thereby affects the data used to update the policy. The influence of previous decisions on the data generated introduces bias in the gradient estimate, which presents a potential source of instability for online learning not present in the iid case. In this paper, we introduce simple criteria for the adaptively generated data stream to guarantee the convergence of SGD, which depend on the mixing time of the policy-induced dynamics. Our Lyapunov-function analysis allows one to translate existing stability analysis of stochastic systems studied in operations research into convergence rates for SGD, and we demonstrate this for queueing and inventory management problems. We also showcase how our result can be applied to study the sample complexity of an actor-critic policy gradient algorithm.
■590 ▼aSchool code: 0054.
■650 4▼aIndustrial engineering
■650 4▼aStatistics
■653 ▼aAdaptive experimentation
■653 ▼aMachine learning
■653 ▼aQueuing systems
■653 ▼aReinforcement learning
■653 ▼aInfinitesimal perturbation analysis
■690 ▼a0796
■690 ▼a0546
■690 ▼a0800
■690 ▼a0463
■71020▼aColumbia University▼bOperations Research.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0054
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358479▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


