Abstract
dc:description.abstractFirst-order, blackbox optimization methods aim to minimize objective functions using only function and gradient evaluations. On smooth, strongly-convex objectives; classical results ensure the linear convergence of these methods relative to the number of function and gradient evaluations. An analogous nonsmooth theory is challenging due to a lack of reliable, local models. As a consequence, convergence guarantees for nonsmooth optimization procedures have generally remained sublinear. In this dissertation, I propose a "survey descent" method - inspired by a new max-of-smooth model - where an iteratively-updated "survey" of points move closer-and-closer to the objective minimizer while guided by (but never touching) an "active manifold" that lies at the boundary of different smooth subdomains of the nonsmooth objective. I prove that, when the nonsmooth objective possesses a "finite-max" structure, survey descent exhibits local linear convergence; and I demonstrate empirically that survey descent holds promise even beyond the finite-max case. I also propose a blackbox initialization procedure for survey descent that finds an initial survey containing exactly one point from each smooth subdomain of a finite-max objective. Of separate interest, I show how the minimization of the numerical radii of square matrices frequently generates solutions that land on the aforementioned active manifold - an interesting example for an "identification" property of proximal operators, a common subroutine used within nonsmooth optimization methods.
Degree
thesis:*- Name thesis:degree_name
- Ph. D., Operations Research and Information Engineering
- Level thesis:degree_level
- Doctor of Philosophy
- Discipline thesis:degree_discipline
- Operations Research and Information Engineering
- Grantor
- Cornell University
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Han, X.Y.
- Chair dc:contributor.chair
-
- Lewis, Adrian
- Committee members dc:contributor.committeemember
-
- Renegar, James
- Chen, Yudong
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-ShareAlike 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Bibid
- 16219241
- Dc Identifier Other
-
ProQuest Submission ID: 13814
ProQuest Publication ID: 30572965