본문

서브메뉴

Statistically Efficient Methods for Computation-Aware Uncertainty Quantification and Rare-Event Optimization
Statistically Efficient Methods for Computation-Aware Uncertainty Quantification and Rare-...
Statistically Efficient Methods for Computation-Aware Uncertainty Quantification and Rare-Event Optimization

Detailed Information

자료유형  
 학위논문 서양
최종처리일시  
20250211151330
ISBN  
9798382797632
DDC  
310
저자명  
He, Shengyi.
서명/저자  
Statistically Efficient Methods for Computation-Aware Uncertainty Quantification and Rare-Event Optimization
발행사항  
[Sl] : Columbia University, 2024
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2024
형태사항  
379 p
주기사항  
Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
주기사항  
Advisor: Lam, Henry.
학위논문주기  
Thesis (Ph.D.)--Columbia University, 2024.
초록/해제  
요약The thesis covers two fundamental topics that are important across the disciplines of operations research, statistics and even more broadly, namely stochastic optimization and uncertainty quantification, with the common theme to address both statistical accuracy and computational constraints. Here, statistical accuracy encompasses the precision of estimated solutions in stochastic optimization, as well as the tightness or reliability of confidence intervals. Computational concerns arise from rare events or expensive models, necessitating efficient sampling methods or computation procedures.In the first half of this thesis, we study stochastic optimization that involves rare events, which arises in various contexts including risk-averse decision-making and training of machine learning models. Because of the presence of rare events, crude Monte Carlo methods can be prohibitively inefficient, as it takes a sample size reciprocal to the rare-event probability to obtain valid statistical information about the rare-event. To address this issue, we investigate the use of importance sampling (IS) to reduce the required sample size. IS is commonly used to handle rare events, and the idea is to sample from an alternative distribution that hits the rare event more frequently and adjusts the estimator with a likelihood ratio to retain unbiasedness. While IS has been long studied, most of its literature focuses on estimation problems and methodologies to obtain good IS in these contexts. Contrary to these studies, the first half of this thesis provides a systematic study on the efficient use of IS in stochastic optimization. In Chapter 2, we propose an adaptive procedure that converts an efficient IS for gradient estimation to an efficient IS procedure for stochastic optimization. Then, in Chapter 3, we provide an efficient IS for gradient estimation, which serves as the input for the procedure in Chapter 2.In the second half of this thesis, we study uncertainty quantification in the sense of constructing a confidence interval (CI) for target model quantities or prediction. We are interested in the setting of expensive black-box models, which means that we are confined to using a low number of model runs, and we also lack the ability to obtain auxiliary model information such as gradients. In this case, a classical method is batching, which divides data into a few batches and then constructs a CI based on the batched estimates. Another method is the recently proposed cheap bootstrap that is constructed on a few resamples in a similar manner as batching. These methods could save computation since they do not need an accurate variability estimator which requires sufficient model evaluations to obtain. Instead, they cancel out the variability when constructing pivotal statistics, and thus obtain asymptotically valid \uD835\uDC61-distribution-based CIs with only few batches or resamples. The second half of this thesis studies several theoretical aspects of these computation-aware CI construction methods. In Chapter 4, we study the statistical optimality on CI tightness among various computation-aware CIs. Then, in Chapter 5, we study the higher-order coverage errors of batching methods. Finally, Chapter 6 is a related investigation on the higher-order coverage and correction of distributionally robust optimization (DRO) as another CI construction tool, which assumes an amount of analytical information on the model but bears similarity to Chapter 5 in terms of analysis techniques.
일반주제명  
Statistics
일반주제명  
Applied mathematics
키워드  
Rare-event estimation
키워드  
Stochastic simulation
키워드  
Uncertainty quantification
키워드  
Machine learning
키워드  
Computation-aware
기타저자  
Columbia University Operations Research
기본자료저록  
Dissertations Abstracts International. 85-12B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008250123s2024        us                              c    eng  d
■001000017161248
■00520250211151330
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798382797632
■035    ▼a(MiAaPQ)AAI31240894
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a310
■1001  ▼aHe,  Shengyi.
■24510▼aStatistically  Efficient  Methods  for  Computation-Aware  Uncertainty  Quantification  and  Rare-Event  Optimization
■260    ▼a[Sl]▼bColumbia  University▼c2024
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2024
■300    ▼a379  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  85-12,  Section:  B.
■500    ▼aAdvisor:  Lam,  Henry.
■5021  ▼aThesis  (Ph.D.)--Columbia  University,  2024.
■520    ▼aThe  thesis  covers  two  fundamental  topics  that  are  important  across  the  disciplines  of  operations  research,  statistics  and  even  more  broadly,  namely  stochastic  optimization  and  uncertainty  quantification,  with  the  common  theme  to  address  both  statistical  accuracy  and  computational  constraints.  Here,  statistical  accuracy  encompasses  the  precision  of  estimated  solutions  in  stochastic  optimization,  as  well  as  the  tightness  or  reliability  of  confidence  intervals.  Computational  concerns  arise  from  rare  events  or  expensive  models,  necessitating  efficient  sampling  methods  or  computation  procedures.In  the  first  half  of  this  thesis,  we  study  stochastic  optimization  that  involves  rare  events,  which  arises  in  various  contexts  including  risk-averse  decision-making  and  training  of  machine  learning  models.  Because  of  the  presence  of  rare  events,  crude  Monte  Carlo  methods  can  be  prohibitively  inefficient,  as  it  takes  a  sample  size  reciprocal  to  the  rare-event  probability  to  obtain  valid  statistical  information  about  the  rare-event.  To  address  this  issue,  we  investigate  the  use  of  importance  sampling  (IS)  to  reduce  the  required  sample  size.  IS  is  commonly  used  to  handle  rare  events,  and  the  idea  is  to  sample  from  an  alternative  distribution  that  hits  the  rare  event  more  frequently  and  adjusts  the  estimator  with  a  likelihood  ratio  to  retain  unbiasedness.  While  IS  has  been  long  studied,  most  of  its  literature  focuses  on  estimation  problems  and  methodologies  to  obtain  good  IS  in  these  contexts.  Contrary  to  these  studies,  the  first  half  of  this  thesis  provides  a  systematic  study  on  the  efficient  use  of  IS  in  stochastic  optimization.  In  Chapter  2,  we  propose  an  adaptive  procedure  that  converts  an  efficient  IS  for  gradient  estimation  to  an  efficient  IS  procedure  for  stochastic  optimization.  Then,  in  Chapter  3,  we  provide  an  efficient  IS  for  gradient  estimation,  which  serves  as  the  input  for  the  procedure  in  Chapter  2.In  the  second  half  of  this  thesis,  we  study  uncertainty  quantification  in  the  sense  of  constructing  a  confidence  interval  (CI)  for  target  model  quantities  or  prediction.  We  are  interested  in  the  setting  of  expensive  black-box  models,  which  means  that  we  are  confined  to  using  a  low  number  of  model  runs,  and  we  also  lack  the  ability  to  obtain  auxiliary  model  information  such  as  gradients.  In  this  case,  a  classical  method  is  batching,  which  divides  data  into  a  few  batches  and  then  constructs  a  CI  based  on  the  batched  estimates.  Another  method  is  the  recently  proposed  cheap  bootstrap  that  is  constructed  on  a  few  resamples  in  a  similar  manner  as  batching.  These  methods  could  save  computation  since  they  do  not  need  an  accurate  variability  estimator  which  requires  sufficient  model  evaluations  to  obtain.  Instead,  they  cancel  out  the  variability  when  constructing  pivotal  statistics,  and  thus  obtain  asymptotically  valid  \uD835\uDC61-distribution-based  CIs  with  only  few  batches  or  resamples.  The  second  half  of  this  thesis  studies  several  theoretical  aspects  of  these  computation-aware  CI  construction  methods.  In  Chapter  4,  we  study  the  statistical  optimality  on  CI  tightness  among  various  computation-aware  CIs.  Then,  in  Chapter  5,  we  study  the  higher-order  coverage  errors  of  batching  methods.  Finally,  Chapter  6  is  a  related  investigation  on  the  higher-order  coverage  and  correction  of  distributionally  robust  optimization  (DRO)  as  another  CI  construction  tool,  which  assumes  an  amount  of  analytical  information  on  the  model  but  bears  similarity  to  Chapter  5  in  terms  of  analysis  techniques.
■590    ▼aSchool  code:  0054.
■650  4▼aStatistics
■650  4▼aApplied  mathematics
■653    ▼aRare-event  estimation
■653    ▼aStochastic  simulation
■653    ▼aUncertainty  quantification
■653    ▼aMachine  learning
■653    ▼aComputation-aware
■690    ▼a0796
■690    ▼a0463
■690    ▼a0800
■690    ▼a0364
■71020▼aColumbia  University▼bOperations  Research.
■7730  ▼tDissertations  Abstracts  International▼g85-12B.
■790    ▼a0054
■791    ▼aPh.D.
■792    ▼a2024
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161248▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

Preview

Export

ChatGPT Discussion

AI Recommended Related Books


    New Books MORE
    Statistics for the past 3 years. Go to brief

    Подробнее информация.

    • Бронирование
    • не существует
    • моя папка
    • Первый запрос зрения
    • Non-Book Loan Application
    • Nighttime Book Loan Application
    материал
    Reg No. Количество платежных Местоположение статус Ленд информации
    TF09650 전자도서 대출가능 My Folder 부재도서신고 비도서대출신청 야간 도서대출신청

    * Бронирование доступны в заимствований книги. Чтобы сделать предварительный заказ, пожалуйста, нажмите кнопку бронирование

    Books borrowed together with this book

    Related Popular Books

    Available after logging in.