Back to search

University of Minnesota

Decision-Focused Learning for Efficient and Interpretable Mixed-Integer Optimization

Abstract

dc:description.abstract

Mixed-integer optimization lies at the core of many real-world decision-making problems, ranging from online systems that require rapid updates to operational planning problems solved by human experts.These problems present three important opportunities: first, to accelerate the solution of computationally challenging mixed-integer optimization models; second, to strengthen these models a priori by incorporating additional domain-specific information; and third, to uncover the decision-making logic embedded in historical decisions made by domain experts. In this thesis, we develop data-driven optimizationframeworks that (i) approximate computationally hard mixed-integer problems using easier-to-solve surrogate models, (ii) strengthen optimization formulations through learned valid inequalities, and (iii) infer hidden decision-making objectives from observed decisions. In the first part of this thesis, we propose a decision-focused framework for constructing surrogate optimization models in the form of linear programs (LPs) that approximate computationally expensive MILPs while remaining significantly faster to solve. Unlike many existing learning-based surrogates, the proposed framework preserves the original linear constraints of the MILP, improving the feasibility and quality of predicted solutions. The approach is demonstrated on model predictive control and scheduling problems. In the second part, we develop a data-driven framework for learning parametric valid inequalities for mixed-integer optimization problems. While general-purpose MIP solvers generate generic cutting planes during branch-and-bound, problem-specific valid inequalities can often lead to substantially tighter formulations and faster solution times. We propose a robust optimization-based learning framework that generates valid inequalities directly from data while guaranteeing that feasible integer solutions of the original problem are preserved. In the final part, we study inverse optimization for production planning in manufacturing systems. We develop a framework for inferring the implicit objective functions underlying expert production plans by learning objective weights from historical decisions.The proposed approach is applied to a real industrial case study provided by Dow, where the learned objectives reveal key operational priorities. Overall, this thesis demonstrates how optimization-based learning can be used to develop efficient data-driven frameworks for complex decision-making systems.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Dixit, Shivi

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Dixit, Shivi. (2026). Decision-Focused Learning for Efficient and Interpretable Mixed-Integer Optimization. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/282295.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Dixit, Shivi. Decision-Focused Learning for Efficient and Interpretable Mixed-Integer Optimization. 2026. https://hdl.handle.net/11299/282295