본문

서브메뉴

Saddle Avoidance, Asymptotic Normality, and Exponential Acceleration in Nonsmooth Optimization
Saddle Avoidance, Asymptotic Normality, and Exponential Acceleration in Nonsmooth Optimiza...
Saddle Avoidance, Asymptotic Normality, and Exponential Acceleration in Nonsmooth Optimization

Detailed Information

Material Type  
 단행본
 
0017161412
Date and Time of Latest Transaction  
20250211151353
ISBN  
9798382842714
DDC  
519
Author  
Jiang, Liwei.
Title/Author  
Saddle Avoidance, Asymptotic Normality, and Exponential Acceleration in Nonsmooth Optimization
Publish Info  
[Sl] : Cornell University, 2024
Publish Info  
Ann Arbor : ProQuest Dissertations & Theses, 2024
Material Info  
311 p
General Note  
Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
General Note  
Advisor: Davis, Damek.
학위논문주기  
Thesis (Ph.D.)--Cornell University, 2024.
Abstracts/Etc  
요약Optimization-based algorithms are the foundation for empirically successful methods in modern fields, such as artificial intelligence and data science. Although classical optimization theory provides guarantees for functions with smoothness or convexity, a significant portion of modern problems do not possess any of these. Despite the worst-case examples where efficient algorithms are unavailable, typical nonsmoothness arises with a "partly smooth" structure, meaning that they are well-behaved relative to a smooth "active manifold."This thesis develops and analyzes first-order algorithms based on the aforementioned nonsmooth structure. We first develop two regularity conditions describing how sub-gradients interact with active manifolds and then show that they hold for a broad and generic class of functions. With these cornerstones, we demonstrate that when randomly perturbed or equipped with stochastic noise, subgradient methods only converge to minimizers of generic, Clarke regular semialgebraic problems. When convergence to a certain minimizer is known, we demonstrate that stochastic (projected) subgradient methods have asymptotic normality, making them asymptotically optimal algorithms in the locally minimax sense of Hajek and Le Cam.These findings culminate with a new first-order algorithm-NTDescent-which exhibits local nearly linear convergence on typical nonsmooth functions with quadratic growth. The convergence rate of NTDescent depends only on the function's intrinsic quantities but not the problem's underlying dimension.
Subject Added Entry-Topical Term  
Applied mathematics
Subject Added Entry-Topical Term  
Statistics
Index Term-Uncontrolled  
Asymptotic normality
Index Term-Uncontrolled  
First-order method
Index Term-Uncontrolled  
Linear convergence
Index Term-Uncontrolled  
Nonsmooth optimization
Index Term-Uncontrolled  
Parameter-free
Index Term-Uncontrolled  
Saddle point avoidance
Added Entry-Corporate Name  
Cornell University Operations Research and Information Engineering
Host Item Entry  
Dissertations Abstracts International. 85-12B.
Electronic Location and Access  
로그인 후 원문을 볼 수 있습니다.

MARC

 008250123s2024        us                              c    eng  d
■001000017161412
■00520250211151353
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798382842714
■035    ▼a(MiAaPQ)AAI31243432
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a519
■1001  ▼aJiang,  Liwei.▼0(orcid)0009-0005-3287-9966
■24510▼aSaddle  Avoidance,  Asymptotic  Normality,  and  Exponential  Acceleration  in  Nonsmooth  Optimization
■260    ▼a[Sl]▼bCornell  University▼c2024
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2024
■300    ▼a311  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  85-12,  Section:  B.
■500    ▼aAdvisor:  Davis,  Damek.
■5021  ▼aThesis  (Ph.D.)--Cornell  University,  2024.
■520    ▼aOptimization-based  algorithms  are  the  foundation  for  empirically  successful  methods  in  modern  fields,  such  as  artificial  intelligence  and  data  science.  Although  classical  optimization  theory  provides  guarantees  for  functions  with  smoothness  or  convexity,  a  significant  portion  of  modern  problems  do  not  possess  any  of  these.  Despite  the  worst-case  examples  where  efficient  algorithms  are  unavailable,  typical  nonsmoothness  arises  with  a  "partly  smooth"  structure,  meaning  that  they  are  well-behaved  relative  to  a  smooth  "active  manifold."This  thesis  develops  and  analyzes  first-order  algorithms  based  on  the  aforementioned  nonsmooth  structure.  We  first  develop  two  regularity  conditions  describing  how  sub-gradients  interact  with  active  manifolds  and  then  show  that  they  hold  for  a  broad  and  generic  class  of  functions.  With  these  cornerstones,  we  demonstrate  that  when  randomly  perturbed  or  equipped  with  stochastic  noise,  subgradient  methods  only  converge  to  minimizers  of  generic,  Clarke  regular  semialgebraic  problems.  When  convergence  to  a  certain  minimizer  is  known,  we  demonstrate  that  stochastic  (projected)  subgradient  methods  have  asymptotic  normality,  making  them  asymptotically  optimal  algorithms  in  the  locally  minimax  sense  of  Hajek  and  Le  Cam.These  findings  culminate  with  a  new  first-order  algorithm-NTDescent-which  exhibits  local  nearly  linear  convergence  on  typical  nonsmooth  functions  with  quadratic  growth.  The  convergence  rate  of  NTDescent  depends  only  on  the  function's  intrinsic  quantities  but  not  the  problem's  underlying  dimension.
■590    ▼aSchool  code:  0058.
■650  4▼aApplied  mathematics
■650  4▼aStatistics
■653    ▼aAsymptotic  normality
■653    ▼aFirst-order  method
■653    ▼aLinear  convergence
■653    ▼aNonsmooth  optimization
■653    ▼aParameter-free
■653    ▼aSaddle  point  avoidance
■690    ▼a0796
■690    ▼a0364
■690    ▼a0463
■71020▼aCornell  University▼bOperations  Research  and  Information  Engineering.
■7730  ▼tDissertations  Abstracts  International▼g85-12B.
■790    ▼a0058
■791    ▼aPh.D.
■792    ▼a2024
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161412▼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

    Detail Info.

    • Reservation
    • Not Exist
    • My Folder
    • First Request
    • Non-Book Loan Application
    • Nighttime Book Loan Application
    Material
    Reg No. Call No. Location Status Lend Info
    TF10092 전자도서 대출가능 My Folder 부재도서신고 비도서대출신청 야간 도서대출신청

    * Reservations are available in the borrowing book. To make reservations, Please click the reservation button

    Books borrowed together with this book

    Related Popular Books

    Available after logging in.