Bridging Machine Learning and Operations Research: From Decision-Focused Learning to Learning to Optimize

Optimization is central to modern decision-making, with critical applications in energy systems, supply chains, transportation, and healthcare. In these domains, operators must repeatedly solve large-scale optimization problems under tight time constraints and pervasive uncertainty. For example, dispatching power plants to balance fluctuating output, routing thousands of vehicles in delivery networks, or scheduling surgeries with limited medical resources. Despite advances in commercial solvers such as Gurobi and CPLEX, and continued improvements in computing hardware, optimization problems involving integer variables remain NP-hard and computationally intractable at scale. The massive scale and complexity of real-world systems further widen this gap, making exact methods insufficient for practical deployment. Motivated by these limitations, this thesis investigates the integration of machine learning (ML) and operations research (OR) to develop scalable, data-driven decision-making frameworks. The contributions span two complementary lines of research: decision-focused learning (DFL) and learning-to-optimize (L2O). The first part studies decision-focused learning (DFL), a predict-then-optimize paradigm for problems with latent parameters that must be inferred from contextual features. Rather than minimizing prediction error, DFL directly trains models to improve downstream decisions. I develop \pyepo{}, an open-source library that unifies existing DFL methods and benchmarks, and propose \cave{} for binary linear programs that accelerates DFL training by several orders of magnitude. The second part advances learning-to-optimize (L2O) for mixed-integer nonlinear programs (MINLPs), which combine discrete decisions with nonlinear and potentially nonconvex constraints. I design a neural framework with differentiable correction layers and a projection-based postprocessing method that yields high-quality integer solutions with near-instant inference. The projection mechanism additionally provides theoretical guarantees of approximate feasibility under mild assumptions. Together, these contributions offer algorithmic insights, practical software, and empirical evidence that learning-based optimization can significantly enhance decision-making in complex, uncertain, and large-scale systems.

Department dc:contributor.department

Mechanical and Industrial Engineering

Year dc:date.issued

2026

Author dc:creator

• Tang, Bo

Advisor dc:contributor.advisor

• Khalil, Elias B

• Decision-Focused Learning
• Integer Programing
• Learning to Optimize
• Machine Learning
• Optimization

Handle dc:identifier.uri

https://hdl.handle.net/1807/151966

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/151966