DO Radar — 13 Apr 2026

Industry Signals

Industry Signal Multiple Domains 13 Apr 2026

2026 Franz Edelman Award Finalists Present Six Live OR Deployments at INFORMS Analytics+

The award ceremony for the 2026 Franz Edelman Award, presented by the Institute for Operations Research and the Management Sciences (INFORMS), takes place this week at the INFORMS Analytics+ Conference in National Harbor, Maryland (April 12–14, 2026). Six finalists compete with documented production deployments in supply chain, logistics, energy, and manufacturing.

The six finalists are Chewy, the Department of Food and Public Distribution (DFPD) in India, ECCO Shoes, Google, Microsoft, and NVIDIA. Among the deployments with published quantitative results: ECCO Shoes developed the Intelligent Auto Replenishment (IAR) solution using a large-scale stochastic programmingAn optimisation framework where some or all input data are uncertain, modelled as random variables. Decisions are structured across stages: a here-and-now decision made before uncertainty is revealed, and a recourse decision made after. — first explained 7 Apr 2026. mixed-integer program that optimises replenishment orders for 536 stores in 27 countries, automating nearly 300,000 orders monthly with a 1.09% reduction in key operational costs, translating to several million euros in annual savings. DFPD, in partnership with the United Nations World Food Programme (WFP) India and the Indian Institute of Technology (IIT) Delhi, deployed Anna Chakra, an OR-based decision support system that optimises state-specific food logistics for India's Public Distribution System (PDS), achieving estimated annual savings of 2.5 billion Indian rupees (INR) and a 35% reduction in emissions while serving over 810 million people. Google's entry shifts flexible compute workloads in time and across data-centre locations to reduce both carbon footprint and power infrastructure costs, combining load forecasting with linear and discrete optimisation methods.

Why it matters: All three of these deployments embed a coupling between a forecasting or learning component and a classical OR solver as the execution engine. ECCO's IAR uses forecast uncertainty to drive the stochastic mixed-integer program; Google's dispatch uses load forecasts to initialise linear and discrete optimisers; DFPD Anna Chakra uses OR solvers to execute logistics decisions derived from demand data. The Edelman competition is the highest-evidence benchmark for production-grade OR at scale, and this year's finalists reinforce that the coupling between data-driven inputs and constrained optimisation solvers is the deployment pattern that delivers measurable results.

Source: INFORMS press release, 21 Jan 2026; award ceremony 13 Apr 2026

Research Papers

Research Paper Energy 📋 Case Study arXiv:2604.05167 — 6 Apr 2026

End-to-End Learning of Correlated Operating Reserve Requirements in Security-Constrained Economic Dispatch

Owen Shen, Hung-po Chao, et al.

Security-Constrained Economic Dispatch (SCED) requires grid operators to hold operating reserves that cover the correlated forecast errors of renewable generators. Standard practice fixes this correlation structure from historical sample covariance, decoupling the reserve-design decision from the dispatch task it is meant to protect. This paper identifies that decoupling as the source of excess cost: an ellipsoidal uncertainty set optimised for historical fit is not the same as one optimised to minimise dispatch expenditure while maintaining coverage. The methodological contribution is a reformulation that makes the correlation structure itself a trainable decision variable in an end-to-end bilevel program, eliminating the decoupling entirely. The approach is notable for yielding finite-sample coverage guarantees through split conformal calibration, rather than relying on asymptotic arguments.

Source: arXiv:2604.05167 📋 Case Study below ↓

► 📋 Case Study Source: paper benchmark — Shen, Chao et al., arXiv:2604.05167 Expand for full case ▼

Takeaway

Training the ellipsoidal uncertainty-set shape as a dispatch-task objective, rather than fitting it to historical covariance, reduces total dispatch cost by 4.8% on a real grid while maintaining empirical coverage above the target level. ↓ Expand to see formulation, difficulty, and full takeaway

Problem

Operating reserve requirements in Security-Constrained Economic Dispatch (SCED) must cover correlated forecast errors from renewable generators. The standard approach specifies the correlation structure exogenously by fitting a sample covariance matrix to historical renewable output errors. The authors show that this decoupling from the dispatch task produces unnecessarily large or misshapen uncertainty sets that increase dispatch cost above what is required for coverage.

Formulation

Type End-to-end trainable bilevel robust optimisation program

Upper level Designer decides: shape of the ellipsoidal uncertainty set (the correlation matrix parameter) to minimise robust SCED cost subject to a target coverage requirement

Lower level Dispatcher solves: the SCED Linear Programme (LP) given the chosen uncertainty set, producing the dispatch cost and KKT dual information used as task gradients in training

Reformulation Coverage constraint profiled into a shape-dependent radius, converting the bilevel to a single-stage differentiable objective; no need to differentiate through the LP solver itself

Calibration Four-way train/tune/calibrate/test split; split conformal calibration applied to the calibrate set provides finite-sample marginal coverage under exchangeability

Full variable-level formulation is in the paper — not reproduced here.

Structure

What made it hard

The authors identify two coupled difficulties. First, the joint bilevel structure couples the uncertainty-set design decision to the SCED LP solve, making direct differentiation through the solver intractable. Second, the distribution of forecast errors is unknown at deployment, so coverage guarantees derived from training data do not transfer directly to new operating conditions without additional calibration. Standard asymptotic arguments are insufficient because grid operators require finite-sample guarantees.

What was done

Shen, Chao, et al. profiled the coverage constraint into a shape-dependent radius, converting the bilevel to a single-stage differentiable objective where KKT dual information from the SCED LP serves as task gradients without differentiating through the solver. A smoothed quantile-sensitivity estimator handles gradient estimation during training; split conformal calibration on a held-out calibrate set provides the finite-sample marginal coverage guarantee at deployment. Tested on real grid data, the learned static ellipsoid reduced dispatch cost by 4.8% relative to the Sample Covariance baseline while maintaining empirical coverage above the target level. (Shen, Chao et al., arXiv:2604.05167.)

Practitioner Takeaway

The paper reports that an ellipsoidal uncertainty set whose shape is optimised for the dispatch task reduces total dispatch cost by 4.8% relative to a Sample Covariance baseline on a real grid, while maintaining empirical coverage above the specified target. The key conclusion is that fitting the uncertainty set to historical data and optimising it for the dispatch task are different problems with different solutions; the authors show both are achievable within a single trainable framework using KKT task gradients and split conformal calibration.

Research Paper General OR arXiv:2603.28943 — 30 Mar 2026

Differentiable Initialisation-Accelerated CPU-GPU Hybrid Combinatorial Scheduling

Mingju Liu, Jiaqi Yin, Alvaro Velasquez, Cunxi Yu — University of Maryland & University of Colorado Boulder

Combinatorial scheduling problems are routinely formulated as Mixed Integer ProgrammingAn optimisation framework in which some decision variables are constrained to integer values, combined with continuous variables and linear constraints. Solvers use branch-and-bound with LP relaxations at each node. — first explained 7 Apr 2026. models and solved by commercial ILP solvers. The bottleneck at scale is the quality of the initial LP relaxation: a weak relaxation forces solvers to explore far more branch and boundAn algorithmic framework for solving combinatorial optimisation problems exactly by partitioning the feasible region into sub-problems (branch) and computing bounds to prune those that cannot contain the optimum (bound). — first explained 7 Apr 2026. nodes before reaching a provably optimal or near-optimal solution. This paper introduces the first framework that uses differentiable optimisation to generate a high-quality warm start for exact ILP solvers, running the differentiable presolve on GPU and feeding its partial solution directly to CPLEX, Gurobi, or HiGHS on CPU.

The differentiable presolve uses a Gaussian reparameterisation to relax binary scheduling decisions into continuous variables, solves the relaxed problem on GPU at high speed, and rounds the result to a partial integer assignment. This partial assignment is passed as a warm start to the exact solver, which then completes the search from a substantially narrowed region. On industry-scale scheduling benchmarks, the hybrid approach achieves up to 10x performance gains over baselines with optimality gaps below 0.1%, the threshold at which solution quality is indistinguishable from exact in most planning contexts.

Why it matters: The paper reports that differentiable pre-solving can narrow the optimality gap to below 0.1% on industry-scale benchmarks while reducing solve time by up to 10x versus running the ILP solver cold. The framework is solver-agnostic (demonstrated with CPLEX, Gurobi, and HiGHS) and requires no changes to the ILP formulation, making it applicable as a drop-in acceleration layer for existing scheduling models.

Source: arXiv:2603.28943

Term of the Day

Non-Anticipativity

A structural constraint in multi-stage stochastic programmingAn optimisation framework where some or all input data are uncertain, modelled as random variables. Decisions are structured across stages: a here-and-now decision made before uncertainty is revealed, and a recourse decision made after. — first explained 7 Apr 2026. that forces a model to use only information available at the time a decision is made — not information that will only become known later. Sounds obvious. But without explicitly encoding it, a mathematical model can accidentally "peek" at future scenarios and generate solutions that look optimal on paper but are physically impossible to execute.

A concrete example — warehouse stocking

It is Monday. You must decide how much stock to order. You know Tuesday's demand could be 100 units (scenario A) or 200 units (scenario B), each with 50% probability.

A model without the non-anticipativity constraint might say: "In scenario A, order 100. In scenario B, order 200." That looks perfect — zero waste, zero stockout. But you have to place one order on Monday, before you know which scenario will occur. The model was implicitly assuming you'd know Tuesday's demand when placing Monday's order. That's not possible.

The non-anticipativity constraint forces Monday's decision to be the same value across both scenarios — because at that moment, you cannot distinguish between them. The model must now find the single best order quantity under genuine uncertainty, say 150 units.

Where this shows up in practice

In energy scheduling, a power plant commitment made at 6am must be fixed regardless of which demand scenario plays out that afternoon. In supply chain, a production run started today cannot vary based on next week's price. In finance, a portfolio position taken now cannot depend on tomorrow's market move. When researchers claim near-perfect stochastic model performance, the first question to ask is: was non-anticipativity enforced? If not, the result is optimistic fiction — the model looked ahead.

Upcoming Conferences

Conference	Dates	Location	Key Track
SIAM OP26 Reg. 5 May	Jun 2–5, 2026	Edinburgh, UK	Theory, algorithms, software, and applications of optimisation. Early registration deadline 5 May 2026. siam.org/op26
APPROX/RANDOM 2026 Deadline 6 May	Aug 19–21, 2026	Boston, MA, USA	Approximation algorithms for combinatorial optimisation (APPROX) and randomised methods (RANDOM). Submission deadline 6 May 2026; double-blind review. approxconference.com
EngOpt 2026	Sep 16–18, 2026	Lisbon, Portugal	7th International Conference on Engineering Optimisation, hosted by Instituto Superior Técnico. Paper submission deadline 1 Jun 2026. ifors.org/engopt-2026

Daily Synthesis

Three of the six Edelman finalists achieve gains by coupling a data-driven component (forecast, stochastic model, or load predictor) directly with a classical OR solver as the execution engine: ECCO Shoes' stochastic mixed-integer program automates 300,000 monthly replenishment orders, India's DFPD Anna Chakra optimises food distribution for 810 million people, and Google combines load forecasting with linear and discrete optimisation to reduce data-centre carbon and power costs. Both research papers this week report measurable gains from the same design: the operating reserve paper achieves a 4.8% dispatch cost reduction by training the uncertainty-set shape as a task objective; the scheduling paper achieves a 10x speedup by using differentiable pre-solving to warm-start exact Integer Linear Programming (ILP) solvers.

The operating reserve paper and the ECCO Shoes deployment share the same structural pattern: a stochastic or probabilistic component (forecast error correlation; demand uncertainty) is coupled directly to a solver rather than estimated separately, and both papers report that the coupling delivers cost reductions the decoupled design does not.
The scheduling paper reports that the warm-start quality from differentiable pre-solving, not solver speed, is the binding constraint on performance at scale: optimality gap narrows to below 0.1% only because the partial integer assignment supplied by the GPU presolve is tight enough to anchor the exact solver's search near the optimum from the start.

For practitioners: The Edelman deployments confirm and the two papers demonstrate quantitatively that the component coupling the forecast or stochastic model to the solver is where the largest gains are found: ECCO reports 1.09% cost reduction via the stochastic mixed-integer program, the reserve paper reports 4.8% dispatch cost reduction via task-trained uncertainty set shape, and the scheduling paper reports 10x speedup via differentiable warm starts. In each case the gain comes from treating the interface between the data-driven component and the solver as a design decision, not a handoff.