서브메뉴
검색
Optimization Problems With Matrix Variables in Machine Learning and Control Applications
Optimization Problems With Matrix Variables in Machine Learning and Control Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103130
- ISBN
- 9798288861833
- DDC
- 658
- 서명/저자
- Optimization Problems With Matrix Variables in Machine Learning and Control Applications
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 203 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Lavaei, Javad.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약This dissertation aims to advance the understanding and solution methods for optimization problems involving matrix variables. These problems are inherently complex and require sophisticated solution approaches. Despite their difficulty, they have critical applications in machine learning and control systems. We integrate high-dimensional statistics, machine learning, system theory, and mathematical optimization techniques to address key challenges in analyzing their solution landscapes and computational feasibility. The dissertation is divided into two parts.In the first part, we study low-rank matrix sensing and completion problems, which are fundamental in recommendation systems and collaborative filtering. The restricted isometry property (RIP) is sufficient for solving matrix sensing problems using the Burer-Monteiro (B-M) factorization approach. However, when the RIP condition is not satisfied, these problems are considered hard. A widely studied case of matrix recovery without RIP is the matrix completion problem. First, we identify a class of matrix completion problems that can be solved using simple graph traversal algorithms in polynomial time. Next, we demonstrate that for this class of polynomial-time solvable problems, the B-M factorization approach gives rise to exponentially many spurious local minima, causing gradient descent algorithms to fail in recovering ground-truth matrices almost surely. This highlights a gap between the information-theoretic and optimization complexity of the problem. On the other hand, another widely used convex semidefinite programming (SDP) formulation approach achieves the exact recovery for the same class of problems. However, we also identify another class of matrix completion problems where the SDP formulation fails, whereas the B-M factorization produces an optimization landscape where all local minima are global. Consequently, when the RIP conditions are not met, neither approach strictly outperforms the other for matrix completion.In the second part, we focus on the system identification problem for dynamical systems in the presence of adversarial disturbances. System identification involves learning a system matrix that governs the relationship between system states and inputs across time. In safety-critical systems, such as power grids, autonomous vehicles, and unmanned aerial vehicles, adversarial agents may attempt to manipulate system dynamics. The least-squares estimator, commonly used for learning system dynamics, is highly susceptible to these outliers and adversarial perturbations. To address this vulnerability, we analyze non-smooth estimators and establish their finite-time recovery guarantees. We show that for both linear time-invariant and parametrized nonlinear systems, non-smooth estimators can recover the exact system matrices in finite time with high probability. In contrast, the least-squares estimator fails to do so. We propose a first-order subgradient method to iteratively update estimates at each time step, leveraging information from previous estimates. This approach eliminates the need to solve a new optimization problem from scratch at each time step, significantly reducing computational complexity.
- 일반주제명
- Industrial engineering
- 일반주제명
- Computer science
- 키워드
- Matrix sensing
- 키워드
- Robust learning
- 기타저자
- University of California, Berkeley Industrial Engineering & Operations Research
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357095
■00520260202103130
■006m o d
■007cr#unu||||||||
■020 ▼a9798288861833
■035 ▼a(MiAaPQ)AAI31939478
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a658
■1001 ▼aYalcin, Baturalp.
■24510▼aOptimization Problems With Matrix Variables in Machine Learning and Control Applications
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a203 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Lavaei, Javad.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aThis dissertation aims to advance the understanding and solution methods for optimization problems involving matrix variables. These problems are inherently complex and require sophisticated solution approaches. Despite their difficulty, they have critical applications in machine learning and control systems. We integrate high-dimensional statistics, machine learning, system theory, and mathematical optimization techniques to address key challenges in analyzing their solution landscapes and computational feasibility. The dissertation is divided into two parts.In the first part, we study low-rank matrix sensing and completion problems, which are fundamental in recommendation systems and collaborative filtering. The restricted isometry property (RIP) is sufficient for solving matrix sensing problems using the Burer-Monteiro (B-M) factorization approach. However, when the RIP condition is not satisfied, these problems are considered hard. A widely studied case of matrix recovery without RIP is the matrix completion problem. First, we identify a class of matrix completion problems that can be solved using simple graph traversal algorithms in polynomial time. Next, we demonstrate that for this class of polynomial-time solvable problems, the B-M factorization approach gives rise to exponentially many spurious local minima, causing gradient descent algorithms to fail in recovering ground-truth matrices almost surely. This highlights a gap between the information-theoretic and optimization complexity of the problem. On the other hand, another widely used convex semidefinite programming (SDP) formulation approach achieves the exact recovery for the same class of problems. However, we also identify another class of matrix completion problems where the SDP formulation fails, whereas the B-M factorization produces an optimization landscape where all local minima are global. Consequently, when the RIP conditions are not met, neither approach strictly outperforms the other for matrix completion.In the second part, we focus on the system identification problem for dynamical systems in the presence of adversarial disturbances. System identification involves learning a system matrix that governs the relationship between system states and inputs across time. In safety-critical systems, such as power grids, autonomous vehicles, and unmanned aerial vehicles, adversarial agents may attempt to manipulate system dynamics. The least-squares estimator, commonly used for learning system dynamics, is highly susceptible to these outliers and adversarial perturbations. To address this vulnerability, we analyze non-smooth estimators and establish their finite-time recovery guarantees. We show that for both linear time-invariant and parametrized nonlinear systems, non-smooth estimators can recover the exact system matrices in finite time with high probability. In contrast, the least-squares estimator fails to do so. We propose a first-order subgradient method to iteratively update estimates at each time step, leveraging information from previous estimates. This approach eliminates the need to solve a new optimization problem from scratch at each time step, significantly reducing computational complexity.
■590 ▼aSchool code: 0028.
■650 4▼aIndustrial engineering
■650 4▼aComputer science
■653 ▼aMatrix completion
■653 ▼aMatrix optimization
■653 ▼aMatrix sensing
■653 ▼aRobust learning
■653 ▼aSubgradient method
■653 ▼aSystem identification
■690 ▼a0796
■690 ▼a0546
■690 ▼a0984
■690 ▼a0800
■71020▼aUniversity of California, Berkeley▼bIndustrial Engineering & Operations Research.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357095▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


