DO Radar — 3 May 2026

Industry Signals

Air France Stochastic Tail Assignment Achieves 0.28% Optimality Gap on 600-Leg Instances 🔗

Industry SignalAirline OperationsarXiv:2602.10866 · Feb 2026

Foundation: Tail assignment — the problem of matching specific aircraft (tails) to scheduled flight legs — must be solved by every major airline every day. Deterministic formulations assume fixed delay durations, leaving solutions brittle when actual disruptions propagate through the rotation network.

arXiv:2602.10866 presents Air France's stochastic tail-assignment model using Mixed-Integer Programming (MIP) with column generation (a decomposition technique that generates only the most profitable routes at each iteration), explicitly accounting for delay propagation and routing uncertainty. The pricing subproblem is formulated as a stochastic shortest path and solved via a diving heuristic (a rounding technique for LP relaxations) with lattice-ordering bounds. On instances with up to 600 flight legs the method achieves a 0.28% optimality gap within a few hours of computing time, outperforming previous approaches based on minimum turn-time rules.

Why it matters: Every percentage point of delay-cost reduction at scale translates to millions in fuel and crew-recovery savings. The column-generation framework generalises: the stochastic pricing idea applies equally to crew pairing, gate assignment, or vehicle routing wherever delay propagation must be explicitly modelled.

India's Anna Chakra Cuts Grain Logistics Cost ₹2.5 Billion Annually, Reducing Emissions 35% 🔗

Industry SignalSupply ChainEdelman Award Finalist · 2026

Foundation: India's Public Distribution System (PDS), the world's largest food-security network, moves grain from central warehouses through state depots to roughly 500,000 Fair Price Shops (FPS) serving 810 million beneficiaries. Route distances, truck capacities, and depot-allocation decisions are set monthly — making this a recurring network-flow optimisation at national scale.

Anna Chakra is an Operations Research (OR)-based logistics optimisation system deployed by India's Department of Food and Public Distribution (DFPD) across 30 states, optimising 6,700 warehouse-to-retail routes using network flow and vehicle routing formulations. The system reduces transportation distances by 15–50% and cuts annual logistics costs by ₹2.5 billion while reducing carbon emissions by 35%, benefiting 810 million beneficiaries.

Why it matters: Anna Chakra demonstrates that classical OR (not machine learning) can deliver transformative outcomes at government scale when problem structure is well-defined. The outcome data make this one of the most consequential public-sector OR deployments in recorded history. The deployment write-up below details problem, formulation, and operational impact.

Source: Edelman finalist write-up · INFORMS JAA 2026📋 Case Study below ↓

► 📋 Case Study Source: published deployment — Edelman finalist 2026 Expand for full case ▼

Takeaway

How India solved the world's largest food-logistics optimisation and why classical OR outperformed machine learning at national scale. ↓ Expand to see formulation, difficulty, and full takeaway

Problem

India's Department of Food and Public Distribution (DFPD) must move grain from roughly 2,000 central Food Corporation of India (FCI) warehouses through state-level depots to approximately 500,000 Fair Price Shops (FPS) every month. With 30 states, heterogeneous truck capacities, road-connectivity constraints, state-specific regulations, and equity requirements ensuring every beneficiary receives their entitlement, the monthly allocation-and-routing problem involves millions of decision variables. Before Anna Chakra, routes were set by district officers using spreadsheets and local knowledge, leaving systematic efficiency gaps across the entire network.

Formulation

Objective Minimise total transportation cost (distance × load) across all routes subject to supply availability, demand satisfaction, vehicle capacity, and minimum per-FPS equity allocation

Variables x_ij: flow (tonnes) on arc from warehouse i to FPS j; y_ijk: binary indicator for truck k serving arc (i,j)

Constraints Flow conservation at each warehouse node; full demand satisfaction at each FPS; vehicle capacity limits; minimum equity allocation per beneficiary household; state-specific regulatory restrictions

Structure

What made it hard

Three compounding difficulties: (1) Scale: 6,700 state-level origins, 437,000 FPS destinations, monthly re-optimisation across 30 independent state systems; (2) Heterogeneous constraints: state-specific regulations, truck load limits, minimum allocation quotas per beneficiary household; (3) Equity requirement: every beneficiary must receive their full entitlement regardless of route cost, preventing pure cost minimisation and making the problem a constrained multi-commodity flow rather than a simple transport problem.

What was done

The team formulated the problem as a multi-commodity network flow with vehicle routing constraints, solved using a hierarchical decomposition: central-to-state allocation first (network flow LP/MIP), then state-to-FPS routing (capacitated vehicle routing per district). The system runs monthly for each of 30 states, generating optimised route plans distributed to district logistics officers via a government portal. The decomposition-by-geography approach made national-scale deployment feasible without a monolithic solver — a key architectural decision enabling the rollout.

Practitioner Takeaway

Three durable lessons: (1) When problem structure is known and data is reliable, formulation beats prediction: classical OR outperformed machine-learning-based forecasting approaches trialled earlier in the project; (2) Decomposition by geography enables national-scale deployment: the 30-state decomposition allowed independent state-level solve cycles that fit within available compute budgets; (3) Equity constraints are first-class citizens in the formulation, not post-processing filters: embedding them in the objective and constraint set prevented the optimiser from routing around hard-to-reach communities.

Research Papers

RKHS Uncertainty Sets Harden Stochastic Programmes Against Out-of-Distribution Shifts 🔗

Research PaperStochastic ProgrammingarXiv:2604.20147 · Apr 2026

Foundation: Stochastic programmes optimise expected cost under uncertain data, but assume the training distribution matches the deployment distribution — an assumption routinely violated in practice. Distributionally Robust Optimisation (DRO) addresses this by optimising against the worst-case distribution within an uncertainty set, but the set's shape determines how conservative the solution becomes.

arXiv:2604.20147 proposes uncertainty sets defined using Reproducing Kernel Hilbert Space (RKHS) embeddings, enabling the DRO problem to capture complex, non-linear distributional differences that moment-based or Wasserstein-distance uncertainty sets miss. The resulting programmes remain tractable and demonstrate improved out-of-distribution (OOD) robustness on standard stochastic optimisation benchmarks, with performance degrading gracefully as the test distribution diverges from training.

Why it matters: For practitioners deploying stochastic programmes in shifting environments (demand forecasting, supply chains, energy markets), RKHS uncertainty sets offer a principled way to buy robustness without over-conservatism. The kernel formulation can be integrated into existing DRO solvers with modest modification.

RL-Guided Benders Decomposition Cuts MINLP Solve Time 57.5% vs. Classical GBD 🔗

Research PaperDecompositionarXiv:2604.22107 · Apr 2026

Foundation: Benders decomposition is a classical algorithm for mixed-integer programmes that iterates between a master problem fixing integer decisions and a subproblem generating cuts to tighten the master. The key bottleneck is cut quality: poor cuts force many iterations — making large Mixed-Integer Nonlinear Programming (MINLP) instances intractable in practice.

arXiv:2604.22107 trains a Reinforcement Learning (RL) agent on a bipartite graph representation of the master problem to select which subproblem to solve next, while a KKT (Karush-Kuhn-Tucker conditions)-informed neural network trained via self-supervised learning predicts primal-dual solutions to directly generate Benders cuts, bypassing expensive subproblem solves. The combined method achieves a 57.5% reduction in solution time relative to classical Generalised Benders Decomposition (GBD) on a real MINLP case study.

Why it matters: MINLP is the workhorse formulation for process scheduling, chemical plant optimisation, and energy systems. A 57.5% wall-clock speedup on real instances is practitioner-grade: it shifts overnight batch optimisations into intraday windows. The RL + KKT neural-network pattern is likely to generalise to other decomposition methods including Lagrangian relaxation and ADMM (Alternating Direction Method of Multipliers).

Time-Space Network

In a time-space network, the passage of time is just another arc to traverse. — — paraphrase of Dantzig's network-flow framework, 1960s

A Time-Space Network is a directed graph in which each node represents a location at a specific discrete time step, and each arc represents either a movement (travel arc) or a waiting action (holding arc) between consecutive time periods at the same location. Flow conservation at every node — the fundamental network-flow constraint — encodes that every unit arriving must either move forward or wait; it cannot disappear. This representation transforms a scheduling or routing problem across time into a classical network-flow or shortest-path problem, enabling LP/MIP solvers to handle time-dependent feasibility and resource constraints at scale.

A concrete example

Consider an airline with three airports (A, B, C) and four time periods. Each node is a (location, time) pair: A@t1, A@t2, B@t2, B@t3, etc. A travel arc from A@t1 to B@t2 represents a flight departing A at time 1, arriving B at time 2. A holding arc from A@t1 to A@t2 represents an aircraft waiting on the ground at A. An aircraft flow routed through this graph automatically satisfies the constraint that the aircraft is in exactly one place at any given time. The Air France tail-assignment paper (arXiv:2602.10866) covered in today's signals uses exactly this structure.

Why practitioners misread this

The most common misread is treating a Time-Space Network as a 'schedule visualisation tool' (a Gantt chart with arrows) rather than as a mathematical object enabling flow-based optimisation. The structure matters because it makes the scheduling problem solvable as a minimum-cost flow or shortest-path problem, importing decades of efficient exact algorithms. Without the flow-conservation semantics, you have only a picture.

Where this shows up in practice

Airline crew pairing, tail assignment, and gate scheduling; rail timetabling and rolling-stock rotation; urban transit vehicle scheduling; electric vehicle (EV) charging station dispatch; humanitarian logistics pre-positioning (supply pre-stocking before disasters); port container yard management.

Daily Synthesis

Air France tail assignment Column generation + stochastic pricing achieves a 0.28% optimality gap on 600-leg instances; a production-ready framework for any airline modelling delay propagation in rotation planning.
Anna Chakra (India PDS) Classical OR at national scale: ₹2.5 B annual savings and 35% emissions reduction across 30 states prove that network-flow formulations outperform ad-hoc routing when structure is known and data is reliable.
Robust OOD stochastic optimisation RKHS uncertainty sets extend DRO to non-linear distributional shifts, offering a drop-in robustness upgrade for practitioners whose stochastic programmes drift as real-world data diverges from training distributions.
RL-guided Benders decomposition A 57.5% solve-time reduction on MINLP via RL agent selection and KKT-informed neural cuts. The pattern generalises to any decomposition algorithm with a separable subproblem structure.