2026 Harvey J Greenberg Research Award
Winners: Philippe Blaettchen, Andre P. Calmon, Georgina Hall, and Mohit Tawarmalani
For their paper, “From Trees to Treewidth: Inventory Management in Complex Supply Chain Networks”.
This paper develops a new exact optimization approach for the Guaranteed Service Model (GSM), a widely used framework for determining safety-stock placement in complex supply chain networks. The authors establish a novel connection between inventory optimization and treewidth to derive an exact linear programming formulation whose complexity depends exponentially on the network’s treewidth rather than its number of nodes. Beyond its theoretical contributions, the paper develops a practical hierarchy of LP relaxations that provides valid lower bounds and can certify optimality. The methodology enables significant computational speedups on benchmark instances as well as operational enhancements. The committee was impressed by the paper, which embodies the spirit of Harvey J. Greenberg’s legacy by combining fundamental advances in mathematical optimization, computer science, and operations management, translating the structural concept of treewidth into both new theoretical foundations and practical computational tools for managing inventory in complex supply chains.
Honorable Mention: Nicolas Blin, Stefano Gualandi, Christopher Maes, Andrea Lodi, and Bartolomeo Stellato
For their paper, “Batched First-Order Methods for Parallel LP Solving in MIP”.
This paper develops a batched first-order method for solving multiple linear programs in parallel on GPUs, with a focus on computationally intensive subroutines arising in mixed-integer programming. The authors extend primal-dual hybrid gradient methods to exploit matrix-matrix operations across batches of related LPs, enabling GPU architectures to perform full strong branching and optimization-based bound tightening rather than relying on the computational shortcuts and heuristics traditionally used by MIP solvers. The methodology delivers substantial computational improvements on benchmark instances, and provides guidance regarding the tasks that should be allocated to GPUs vs. CPUs. The committee was impressed by the combination of first-order optimization, GPU computing, and mixed-integer programming techniques in the paper toward a new generation of MIP algorithms designed to exploit modern GPU hardware.
Honorable Mention: Sebastián Vásquez, Leonardo Lozano, and Willem-Jan van Hoeve
For their paper, “A Single-Level Reformulation of Binary Bilevel Programs Using Decision Diagrams”.
This paper develops a new single-level reformulation for binary bilevel optimization by using decision diagrams to represent the follower’s feasible solution space. The authors combine a network-flow representation with linear programming duality to obtain an exact reformulation, and develop polynomial-sized restricted decision diagrams that generate valid dual bounds and can be iteratively strengthened into an exact solution method. Results show that the approach returns new or improved bounds for six benchmark instances and closes two instances for the first time. The committee was impressed by the novel theoretical representation of bilevel optimization problems combining decision diagrams and integer programming techniques, which yields a scalable computational framework for a notoriously difficult class of optimization problems.
The 2026 Harvey J Greenberg Research Award members are:
Alexandre Jacquillat, Chair (MIT, alexjacq@mit.edu)
Necdet Serhat Aybat (Penn State)
Omar El Housni (Cornell)
Shiqian Ma (Rice)
Weijun Xie (Georgia Tech)
Ying Cui (Berkeley)
Yongjia Song (Clemson)
2025 Harvey J Greenberg Research Award
Winners: Braden L. Crimmins, J. Alex Halderman, and Bradley Sturt for their paper "Improving the Security of United States Elections with Robust Optimization."
This paper introduces the first formal approach to designing test decks for logic and accuracy testing (LAT), a procedure that election officials have used for over a century to verify the correctness of voting machines. By employing robust optimization, the authors develop test decks that guarantee detection of any machine misconfiguration that could swap votes across candidates, while minimizing the number of ballots required. Their cutting-plane–based algorithm efficiently solves these optimization problems at scale, and a retrospective study of Michigan's November 2022 general election shows that their method achieves rigorous security guarantees with only 1.2% more ballots than current practice. The approach has since been piloted by the Michigan Bureau of Elections, offering a low-cost and practical solution to strengthen election security and public trust in democratic institutions.
The committee believes this work embodies the spirit of Harvey J. Greenberg's legacy by combining rigorous mathematical techniques (specifically, robust optimization) with a timely and socially important challenge in election security. The paper also demonstrates real-world impact through its pilot program with the Michigan Bureau of Elections.
Honorable Mention: Soroush Saghafian. "Ambiguous Dynamic Treatment Regimes: A Reinforcement Learning Approach."
This paper introduces a novel framework for designing treatment guidelines when traditional assumptions in causal inference do not hold. Standard Dynamic Treatment Regimes (DTRs) often fail in real-world applications such as medicine and public policy, particularly when unobserved confounders exist and evolve over time. To address this challenge, the authors propose Ambiguous Dynamic Treatment Regimes (ADTRs), which evaluate treatment policies against a "cloud" of plausible causal models. By connecting ADTRs to Ambiguous Partially Observable Markov Decision Processes (APOMDPs), the paper develops reinforcement learning methods that efficiently learn effective treatment regimes from observational data. Theoretical guarantees, including consistency and asymptotic normality, are established, and the approach is validated both through simulations and a hospital case study on patients who developed New Onset Diabetes After Transplantation (NODAT).
Honorable Mention: Xiao-Yue Gong and David Simchi-Levi. "Bandits atop Reinforcement Learning: Tackling Online Inventory Models with Cyclic Demands."
This paper addresses the long-standing gap between inventory theory and practice by studying online inventory models under unknown cyclic demand distributions. The authors design reinforcement learning algorithms that leverage structural properties of inventory problems to achieve near-optimal regret bounds, surpassing existing theoretical results. They analyze both lost-sales and multi-product backlogging models, introducing episodic formulations and extending them to non-discarding models through a novel bandit-based approach, Meta-HQL. Their algorithms match the regret lower bounds they establish, while removing dependence on the size of the state–action space. Empirical studies with real sales data from Rossmann and synthetic benchmarks demonstrate rapid convergence to optimal policies and significant improvements over methods that assume i.i.d. demand.
Chairs
- Selva Nadarajah, University of Illinois Chicago
- Andre A. Cire, University of Toronto Scarborough & Rotman School of Management
Steering Committee
- Dan Adelman, University of Chicago
- David Brown, Duke University
- Ricardo Fukasawa, University of Waterloo
- Simge Küçükyavuz, Northwestern University
- Siqian Shen, University of Michigan at Ann Arbor
- Golbon Zakeri, University of Massachusetts Amherst
Review Panel
- Yi-Chun Akchen, University College London
- Margarida Carvalho, Université de Montréal
- Margarita Paz Castro, Pontificia Universidad Católica de Chile
- Levi DeValve, University of Chicago
- Ludwig Dierks, University of Illinois Chicago
- Daniel Jiang, Meta
- Carla Michini, University of Wisconsin-Madison
- Raghav Singal, Dartmouth
Other winners and committees are:
2020-2024