MIT Libraries logoDSpace@MIT

MIT
View Item 
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Doctoral Theses
  • View Item
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Doctoral Theses
  • View Item
JavaScript is disabled for your browser. Some features of this site may not work without it.

Advanced mixed-integer programming formulations : methodology, computation, and application

Author(s)
Huchette, Joseph Andrew
Thumbnail
DownloadFull printable version (1.383Mb)
Other Contributors
Massachusetts Institute of Technology. Operations Research Center.
Advisor
Juan Pablo Vielma.
Terms of use
MIT theses are protected by copyright. They may be viewed, downloaded, or printed from this source but further reproduction or distribution in any format is prohibited without written permission. http://dspace.mit.edu/handle/1721.1/7582
Metadata
Show full item record
Abstract
This thesis introduces systematic ways to use mixed-integer programming (MIP) to solve difficult nonconvex optimization problems arising in application areas as varied as operations, robotics, power systems, and machine learning. Our goal is to produce MIP formulations that perform extremely well in practice, requiring us to balance qualities often in opposition: formulation size, strength, and branching behavior. We start by studying a combinatorial framework for building MIP formulations, and present a complete graphical characterization of its expressive power. Our approach allows us to produce strong and small formulations for a variety of structures, including piecewise linear functions, relaxations for multilinear functions, and obstacle avoidance constraints. Second, we present a geometric way to construct MIP formulations, and use it to investigate the potential advantages of general integer (as opposed to binary) MIP formulations. We are able to apply our geometric construction method to piecewise linear functions and annulus constraints, producing small, strong general integer MIP formulations that induce favorable behavior in a branch-and-bound algorithm. Third, we perform an in-depth computational study of MIP formulations for nonconvex piecewise linear functions, showing that the new formulations devised in this thesis outperform existing approaches, often substantially (e.g. solving to optimality in orders of magnitude less time). We also highlight how high-level, easy-to-use computational tools, built on top of the JuMP modeling language, can help make these advanced formulations accessible to practitioners and researchers. Furthermore, we study high-dimensional piecewise linear functions arising in the context of deep learning, and develop a new strong formulation and valid inequalities for this structure. We close the thesis by answering a speculative question: Given a disjunctive constraint, what can we reasonably sacrifice in order to construct MIP formulations with very few integer variables? We show that, if we allow our formulations to introduce spurious "integer holes" in their interior, we can produce strong formulations for any disjunctive constraint with only two integer variables and a linear number of inequalities (and reduce this further to a constant number for specific structures). We provide a framework to encompass these MIP-with-holes formulations, and show how to modify standard MIP algorithmic tools such as branch-and-bound and cutting planes to handle the holes.
Description
Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2018.
 
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
 
Cataloged from student-submitted PDF version of thesis.
 
Includes bibliographical references (pages 193-203).
 
Date issued
2018
URI
http://hdl.handle.net/1721.1/119282
Department
Massachusetts Institute of Technology. Operations Research Center; Sloan School of Management
Publisher
Massachusetts Institute of Technology
Keywords
Operations Research Center.

Collections
  • Doctoral Theses

Browse

All of DSpaceCommunities & CollectionsBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

My Account

Login

Statistics

OA StatisticsStatistics by CountryStatistics by Department
MIT Libraries
PrivacyPermissionsAccessibilityContact us
MIT
Content created by the MIT Libraries, CC BY-NC unless otherwise noted. Notify us about copyright concerns.