Healthcare — chronic disease management, public policy
My research asks how individuals and organizations can make better decisions in
complex operational systems—and how analytics and AI can improve not only
performance, but also resilience, fairness, and social and environmental outcomes.
I work across two closely connected streams. The first develops methods for
decision-making under uncertainty, including robust, adaptive, and data-driven
optimization. I am particularly interested in methods that remain tractable,
transparent, and fair when information is incomplete and conditions change. The
second uses these methods, together with economic modeling and empirical data,
to design better operational systems, incentives, and contracts.
Much of my applied work focuses on global supply chains and responsible
operations: improving smallholder livelihoods, reducing child and forced labor,
protecting forests, strengthening supply-chain resilience, expanding access to
clean technologies, and reducing waste. I also study problems at the
intersections of operations with finance, healthcare, and public policy. Across
these domains, the common goal is to use analytics and AI to design decisions
and systems that work better for organizations, individuals, and society.
Browse by topic
Filter the lists below by research theme. Papers often belong to more than one;
tick “match all selected” to see only those at the intersection.
A grocer selling premade food must decide how long items stay on the shelf, whether to sell the freshest or the oldest first, whether to timestamp them, and how to price. In our base model, the best policy is counterintuitive: sell the freshest item first and do not timestamp, which can extend shelf life, increase sales, and reduce waste. Model extensions identify conditions that favor selling the oldest item first and show how customer heterogeneity can make timestamps valuable.
Lin Shi, Adam Brandt, Dan Andrei Iancu, Katharine Mach, Christopher Field, Moon-Jung Cho, S. Chey, Nilam Ram, Todd Robinson, Byron Reeves(2024).
Climate impacts of digital use supply chains.
Environmental Research: Climate, vol. 3, no. 1, 015009.
What’s this about?
Accounts of technology’s climate footprint usually count the emissions from building, shipping, and operating devices. We introduce digital use supply chains—the production and resource consumption recorded or enabled by everyday digital activity—and use moment-by-moment Screenomics data to connect individual behavior with emissions. In a single-case study, one day of behavior-related emissions is estimated to be roughly 1,000 times the device’s own life-cycle emissions, suggesting opportunities for personalized feedback and behavior-change programs.
Dan Andrei Iancu, Nikolaos Trichakis, Do Young Yoon(2021).
Monitoring with Limited Information.
Management Science, vol. 67, no. 7, pp. 4233–4251.
What’s this about?
Sometimes a decision maker must choose when to stop a treatment or trade while being able to observe its state only a few times. We develop a robust-optimization approach that jointly chooses when to monitor and when to stop. Under certain conditions, fixed monitoring times achieve the same worst-case reward as fully adaptive ones, making the dynamic policy much easier to compute. Applied to monitoring heart-transplant patients, the approach substantially improves on current recommendations.
Many systems allocate tasks centrally to providers whose welfare depends on what they receive, creating a tension between provider guarantees and total value. We derive tight bounds on the value lost when such guarantees are imposed and show that the loss is limited. With many providers, fairness is the main driver; with few, differences in providers’ effectiveness matter more. When providers are identical, the loss never exceeds 50%, and experiments with real and synthetic data find much smaller losses in several practical settings.
So Yeon Chun, Dan Andrei Iancu, Nikolaos Trichakis(2020).
Loyalty Program Liabilities and Point Values.
Manufacturing & Service Operations Management, vol. 22, no. 2, pp. 257–272.
What’s this about?
Loyalty points are a currency the firm issues, and the future service they promise appears on its balance sheet as a liability. We show that the optimal total value of outstanding points should track the firm’s “profit potential”—realized cash flows plus deferred revenue. Point values rise with stronger operating performance and greater uncertainty, allowing loyalty programs to buffer fluctuations in future performance and providing a rationale for them beyond marketing or competition.
🏆Finalist, M&SOM Journal Best Paper Prize
(2021 and 2022)
🏆Second place, Best Student Paper in Supply Chain Management, POMS Society (Joann de Zegher)
(2016)
What’s this about?
When a new farming technology helps the buyer but costs the farmer, adoption depends on how the two sides trade. We study when changing the contract, switching from commodity sourcing to direct sourcing, or combining both can turn a one-sided innovation into a mutual gain. The answer hinges on how the technology’s costs scale. Using farm data from Patagonia, Argentina, we estimate that the proposed mechanism could increase average supply-chain profit by 6.9% while also producing environmental benefits.
Firms often borrow to finance inventory, then price that inventory both to earn a profit and to service the debt. We show limited liability leads such sellers to charge higher prices and discount more slowly, and that these distortions compound over time into a downward performance spiral. We then quantify how much of the loss practical debt covenants can recover.
Many chronic-disease medications work only for a subgroup of patients, and no biomarker identifies that subgroup in advance. We develop an adaptive treatment framework that learns from both continuous measures of disease progression and the timing and severity of infrequent events such as relapses, helping physicians decide when to persist and when to stop. Applied to interferon treatment for multiple sclerosis, the resulting policies provide a cost-effectiveness frontier and benchmarks for existing treatment guidelines.
A manufacturer worried about disruptions at its suppliers’ suppliers usually cannot choose those tier-2 firms directly, but it can shape their selection through contracts with tier 1. When tier-1 suppliers share tier-2 sources in a diamond-shaped network, the manufacturer should rely less on excess inventory and multisourcing and more on inducing tier-1 mitigation. Yet manufacturers prefer less overlap while tier-1 suppliers may prefer more; penalty contracts can alleviate this conflict.
Operating flexibility is normally an asset, but under debt it can invite risk-shifting — and we find the resulting borrowing costs can erase more than a third of a firm’s value. We then ask whether the covenants lenders actually write can restore it. Simple financial covenants suffice when the firm can liquidate inventory mid-season; richer forms of flexibility demand more.
There are two natural ways to measure risk across many periods: apply a single risk measure to the total future cost, or compose one-step risk mappings. We characterize when one always dominates the other and introduce a metric for how far apart they are. An asymmetry emerges — the tightest upper bound admits an exact characterization, while the lower bound does not.
A manager running many client accounts cannot treat them independently: trades move prices, so executing one account affects the others. We develop a tractable method that jointly optimizes all trades and divides the resulting market-impact costs across accounts, allowing the manager to balance aggregate gains with equitable treatment. Numerical studies indicate that the approach outperforms methods commonly used in industry or proposed in prior research.
Robust optimization protects you against the worst case — but in doing so it often leaves performance on the table when the worst case does not materialize. Two decisions can be identical under the worst case while one is better in every other scenario. We show how to find the robust decisions that are not dominated this way, at essentially no additional cost.
Two classical approaches to dynamic robust optimization rarely meet: dynamic programming, which is exact but intractable, and simple decision rules, which are tractable but usually approximate. We give conditions — uncertainty sets that are integer sublattices of the unit hypercube, plus a technical condition — under which affine decision rules are exactly optimal, bridging the two.
Local search for binary optimization normally trades solution quality against running time with no principled way to set the dial. We develop a general-purpose algorithm whose single parameter controls both the depth of the search and its computational cost. The method has a formal approximation guarantee for a class of set-packing problems and, on large randomly generated set-covering and set-packing instances, performs competitively with leading general-purpose optimization software.
Multistage decisions under uncertainty are usually attacked with simple policy classes, because the exact problem is intractable. We introduce a hierarchy of polynomial disturbance-feedback policies, each computable from a single semidefinite program and indexed by the polynomial’s degree. Raising the degree buys accuracy at a predictable computational price.
Multistage robust optimization is generally intractable, so the field relies on policies that depend affinely on observed disturbances — a restriction adopted for convenience and assumed to be suboptimal. For one-dimensional, constrained problems with convex state costs and linear control costs, we prove affine policies are exactly optimal. The proof turns on the geometry of the feasible set rather than dynamic programming.
More than 1.5 million children work in cocoa production in Côte d’Ivoire and Ghana, despite decades of intervention. We model child labor as one of the levers smallholder households use to balance production, consumption, and finances under harvest uncertainty. The analysis identifies three forces—resource relief, labor productivity, and consumption expansion—that explain why the same program may reduce child labor for some households and increase it for others. Descriptive survey evidence from an NGO partner in Ghana is consistent with the framework, which shows how observable household characteristics can help target interventions.
🏆Winner, SAWIT Challenge, USAID & Indonesia Business Council for Sustainable Development
(2016)
🏆Finalist, INFORMS George Nicholson Student Paper Competition (J. de Zegher)
(2018)
What’s this about?
Millions of smallholder farmers grow the world’s commodities, and some clear forest to expand their farms and escape poverty. We compare individual incentives with two collective designs: area no-deforestation, which rewards farmers if no forest is cleared, and a new “area no-use” condition, which rewards them if no one earns income from recently deforested land. Which approach works best depends on local cooperation, the cost and reliability of blocking forest use, and whether every farmer must be compensated.
During COVID-19, many governments confined people differently by age or activity, but the value of such targeting was contested. We develop a framework that balances mortality against lost economic output and apply it to Île-de-France. Targeting by either age or activity improves on uniform policies, while targeting by both improves on either dimension alone. These gains persist when epidemiological parameters are uncertain and can even increase with greater ambiguity.
🏆Finalist, INFORMS George Nicholson Student Paper Competition (B. Taşkesen)
(2023)
What’s this about?
We generalize LQG control to a broad family of ambiguity sets around a nominal Gaussian distribution. The worst-case distribution is itself Gaussian, with zero mean and an inflated covariance, and the optimal controller remains linear in the observations. These structural results yield a Frank-Wolfe algorithm that outperforms semidefinite-programming reformulations and extend to infinite-horizon control; under Wasserstein ambiguity, they also hold for elliptical nominal distributions.
Digital platforms connecting smallholder farmers and intermediaries promise traceability and better livelihoods, but participants may return to informal networks if the platform provides too little value. Combining an optimization model with data from Indonesia’s palm-oil supply chain, we find that profitability depends critically on controlling transportation costs and learning at least part of the informal network. Payments should go primarily to farmers, and minimum-cost matching is usually best—except when unusually well-connected intermediaries must be prioritized to keep the platform stable.
Smallholder households growing commodities such as cocoa often depend on their children’s labor to secure basic subsistence. We model their borrowing, saving, consumption, and child-labor decisions and uncover sharp trade-offs: better credit access can reduce or increase child labor depending on why a household borrows; better savings access consistently reduces child labor but can lower consumption; and price premiums reduce child labor only when sufficiently large, otherwise they may increase it. Effective interventions must therefore be tailored to household circumstances.
Erick Delage, Dan Andrei Iancu(2015).
Robust Multi-stage Decision Making.
TutORials in Operations Research: The Operations Research Revolution, pp. 20–46, INFORMS.
What’s this about?
Decisions under uncertainty often unfold over time, so a useful plan must preserve the ability to adapt as new information arrives. This tutorial provides a unified view of robust multistage decision-making: the distinction between static, fully adaptive, and partially adjustable decisions; why exact solutions become intractable; connections with robust dynamic programming; when simple policies can work well or even be optimal; how time-consistency problems arise; and how the framework is used in applications.
How much distributional ambiguity can classical LQG control absorb? Across a broad family of divergence-based uncertainty sets, we prove that the worst-case noise distribution remains Gaussian and that a policy linear in the observations remains optimal. The adversary responds to uncertainty by inflating the noise covariance rather than shifting its mean. These structural results preserve the familiar form of LQG control even when the noise distribution is not known precisely.
🏆Spotlight presentation (3.06% of 12,343 submissions)
(2023)
What’s this about?
Classical LQG control assumes the distribution of the noise is known. We instead allow any distribution within a Wasserstein ball around a Gaussian estimate, including non-Gaussian distributions, and optimize against the worst case. Despite this ambiguity, a policy that is linear in the observations remains optimal. An efficient numerical method uses the Frank-Wolfe algorithm to find the least-favorable distributions and Kalman filtering to compute the controller.
Utilities linked by transmission lines must dispatch a changing mix of generators to meet demand as both loads and renewable supply vary over time. We formulate a multi-period dispatch model that accounts for the cost of adjusting generation between periods and derive an efficient solution method. Under conditions often encountered in practice, the solution is globally optimal; numerical experiments quantify the benefits of adjusting thermal generation at finer time scales.
Multistage robust control becomes intractable as the horizon grows, so the field leans on policies that respond affinely to observed disturbances — a practical restriction usually assumed to cost something. For one-dimensional, box-constrained problems with convex state costs and linear control costs, we prove those affine policies are exactly optimal. The argument comes from polyhedral geometry rather than dynamic programming, and yields fast algorithms when the state costs are piecewise affine.
A doctoral thesis on making robust optimization work when decisions unfold over time rather than all at once. It develops adaptive policies that respond to uncertainty as it is observed while remaining computationally tractable, and applies them to inventory and revenue management.
An undergraduate thesis approaching quantum information geometrically. If the state of n qubits lives in a high-dimensional complex space, what do the natural symmetric structures in that space look like? It develops an algorithm for constructing analogues of the Platonic solids there — uniform Hilbertian polytopes — as a step toward measuring entanglement.