서브메뉴
검색
Multi-Armed Bandits With Probes and Knapsacks: Algorithms, Theory and Applications
Multi-Armed Bandits With Probes and Knapsacks: Algorithms, Theory and Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103147
- ISBN
- 9798314865866
- DDC
- 004
- 서명/저자
- Multi-Armed Bandits With Probes and Knapsacks: Algorithms, Theory and Applications
- 발행사항
- [Sl] : Carnegie Mellon University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 173 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
- 주기사항
- Advisor: Yagan, Osman.
- 학위논문주기
- Thesis (Ph.D.)--Carnegie Mellon University, 2025.
- 초록/해제
- 요약The multi-armed bandit (MAB) framework is a fundamental paradigm in decision theory and machine learning, in which a decision-maker sequentially selects one of K available actions at each round, with the objective of maximizing the cumulative reward over time. The main challenge lies in balancing exploration of new actions with exploitation of known ones-an essential trade-off for making optimal, data-driven decisions in complex, uncertain environments with limited information, such as clinical trials, online advertising, and recommendation systems. Motivated by these challenges and applications, we propose and analyze two different multi-armed bandit models with broad practical relevance, which are i) bandits with probes, and ii) bandits with anytime knapsacks. Additionally, we focus on utilizing the multi-armed bandit framework in iii) dataset labeling with LLMs (large language models).First, we study the MAB model with probes, where before pulling an arm, the decision-maker is allowed to probe one arm at an associated cost to observe its reward for that round. Based on the probe's outcome, the decision-maker can choose to pull the probed arm or any other arm. Alternatively, an arm may be pulled directly without employing the probe. This variation is of particular interest due to its extensive range of applications, including online learning augmented with machine learning advice, hyper-parameter optimization for machine learning models, and wireless communications. The introduction of the probing mechanism significantly expands the action space, adding complexity to the problem, as the decision-maker must judiciously balance exploration and exploitation while also considering whether to probe an arm. To address these challenges, we propose a novel algorithm that efficiently determines the optimal action in each round to maximize the cumulative reward. Through rigorous analysis, we derive a theoretical gap-independent regret bound on the order of O( √ T), and a gap-dependent regret bound on the order of O(log T); and demonstrate the empirical performance of our algorithm through simulation results.Secondly, we study the MAB model with anytime knapsacks. This model is similar to the bandits with knapsacks (BwK) problem, however, there is an anytime constraint on the average cost instead of a fixed total cost budget. This formulation finds broad applications in areas including inventory management, online advertising, and portfolio management. We propose an algorithm that employs upper confidence bounds to efficiently explore the decision space while strategically under-utilizing the available budget to reduce the number of rounds that need to be skipped to satisfy the anytime cost constraint. We derive theoretical gap-dependent regret bounds and demonstrate that, despite the increased constraints relative to the standard BwK model, our formulation achieves the same gap-dependent regret bounds of O(log T). We provide simulation results to corroborate the theoretical findings and illustrate the empirical performance of the proposed algorithm.Lastly, to illustrate the practical utility of the MAB framework in the newest problem settings, we examine the task of labeling a dataset using different LLMs. In this scenario, we have access to various LLMs, each with distinct costs and capabilities, and a model capable of extracting contextual information from individual data instances. We propose an algorithm that leverages this context information to select a subset of LLMs, as opposed to traditional weighted majority voting methods that query all available LLMs for each label. We demonstrate the utility of our proposed algorithm in this problem setting with simulation results.
- 일반주제명
- Computer science
- 일반주제명
- Statistics
- 일반주제명
- Computer engineering
- 일반주제명
- Electrical engineering
- 키워드
- Online learning
- 기타저자
- Carnegie Mellon University Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357203
■00520260202103147
■006m o d
■007cr#unu||||||||
■020 ▼a9798314865866
■035 ▼a(MiAaPQ)AAI31995839
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aElumar, Eray Can.▼0(orcid)0000-0002-2929-3690
■24510▼aMulti-Armed Bandits With Probes and Knapsacks: Algorithms, Theory and Applications
■260 ▼a[Sl]▼bCarnegie Mellon University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a173 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-11, Section: B.
■500 ▼aAdvisor: Yagan, Osman.
■5021 ▼aThesis (Ph.D.)--Carnegie Mellon University, 2025.
■520 ▼aThe multi-armed bandit (MAB) framework is a fundamental paradigm in decision theory and machine learning, in which a decision-maker sequentially selects one of K available actions at each round, with the objective of maximizing the cumulative reward over time. The main challenge lies in balancing exploration of new actions with exploitation of known ones-an essential trade-off for making optimal, data-driven decisions in complex, uncertain environments with limited information, such as clinical trials, online advertising, and recommendation systems. Motivated by these challenges and applications, we propose and analyze two different multi-armed bandit models with broad practical relevance, which are i) bandits with probes, and ii) bandits with anytime knapsacks. Additionally, we focus on utilizing the multi-armed bandit framework in iii) dataset labeling with LLMs (large language models).First, we study the MAB model with probes, where before pulling an arm, the decision-maker is allowed to probe one arm at an associated cost to observe its reward for that round. Based on the probe's outcome, the decision-maker can choose to pull the probed arm or any other arm. Alternatively, an arm may be pulled directly without employing the probe. This variation is of particular interest due to its extensive range of applications, including online learning augmented with machine learning advice, hyper-parameter optimization for machine learning models, and wireless communications. The introduction of the probing mechanism significantly expands the action space, adding complexity to the problem, as the decision-maker must judiciously balance exploration and exploitation while also considering whether to probe an arm. To address these challenges, we propose a novel algorithm that efficiently determines the optimal action in each round to maximize the cumulative reward. Through rigorous analysis, we derive a theoretical gap-independent regret bound on the order of O( √ T), and a gap-dependent regret bound on the order of O(log T); and demonstrate the empirical performance of our algorithm through simulation results.Secondly, we study the MAB model with anytime knapsacks. This model is similar to the bandits with knapsacks (BwK) problem, however, there is an anytime constraint on the average cost instead of a fixed total cost budget. This formulation finds broad applications in areas including inventory management, online advertising, and portfolio management. We propose an algorithm that employs upper confidence bounds to efficiently explore the decision space while strategically under-utilizing the available budget to reduce the number of rounds that need to be skipped to satisfy the anytime cost constraint. We derive theoretical gap-dependent regret bounds and demonstrate that, despite the increased constraints relative to the standard BwK model, our formulation achieves the same gap-dependent regret bounds of O(log T). We provide simulation results to corroborate the theoretical findings and illustrate the empirical performance of the proposed algorithm.Lastly, to illustrate the practical utility of the MAB framework in the newest problem settings, we examine the task of labeling a dataset using different LLMs. In this scenario, we have access to various LLMs, each with distinct costs and capabilities, and a model capable of extracting contextual information from individual data instances. We propose an algorithm that leverages this context information to select a subset of LLMs, as opposed to traditional weighted majority voting methods that query all available LLMs for each label. We demonstrate the utility of our proposed algorithm in this problem setting with simulation results.
■590 ▼aSchool code: 0041.
■650 4▼aComputer science
■650 4▼aStatistics
■650 4▼aComputer engineering
■650 4▼aElectrical engineering
■653 ▼aMulti-armed bandits
■653 ▼aOnline learning
■653 ▼aStatistical decision theory
■690 ▼a0984
■690 ▼a0463
■690 ▼a0464
■690 ▼a0544
■71020▼aCarnegie Mellon University▼bElectrical and Computer Engineering.
■7730 ▼tDissertations Abstracts International▼g86-11B.
■790 ▼a0041
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357203▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


