Industry Signals
Oracle Named Leader in Both 2026 Gartner Magic Quadrant Reports for Supply Chain Planning, Including Inaugural Process Industries Report 🔗
Foundation: Discrete manufacturers (cars, phones, furniture) assemble countable components and can pause a production line mid-run. Process manufacturers (chemicals, pharmaceuticals, food and beverage, oil and gas) follow recipes, measure output in weight or volume, and often cannot stop a batch once started — yield varies, co-products emerge simultaneously, and transformations are irreversible. Until 2026, Gartner evaluated planning software for both types in a single report.
Oracle Fusion Cloud Supply Chain Planning was named a Leader in both the Gartner Magic Quadrant for Supply Chain Planning Solutions: Discrete Industries and the inaugural Gartner Magic Quadrant for Supply Chain Planning Solutions: Process Industries (published 18 March 2026; Oracle announcement 8 April 2026). The Process Industries report is newly separated from Discrete, recognising that chemical, pharmaceutical, food and beverage, and oil-and-gas manufacturers require structurally different planning models. Oracle covers both segments on a single cloud platform with embedded AI agents for disruption detection, demand sensing, and prescriptive reordering.
IonQ + Einride Publish End-to-End Hybrid Quantum-Classical Workflow for Electric Freight: 12% More Shipments Delivered on Real Data 🔗
Foundation: The Vehicle Routing Problem (VRP) is the task of finding the best set of routes for a fleet of vehicles — which truck goes where, in what order, to serve which customers. A warm-start is a technique where you give the solver a good initial solution to begin from rather than searching from scratch. Quantum computers cannot yet solve full VRP problems, but they can generate better warm-starts than classical methods produce quickly.
IonQ and Einride published results on 13 April 2026 from their three-year partnership embedding quantum algorithms in Einride's Saga fleet platform. The workflow splits the problem: IonQ's Quantum Approximate Optimisation Algorithm (QAOA) handles the upstream shipment selection subproblem and passes the result as a warm-start to Einride's classical VRP solver, which finalises the routes. On Einride's own commercial freight data, shipments delivered improved by up to 12% and drive distance per shipment fell by up to 6%, with operational costs unchanged.
Research Papers
Fairness-aware Strategic Design of Station-based Electric Car-Sharing Systems 🔗
Jue Zhou, Zoha Sherkat-Masoumi, Merve Bodur
Existing integrated models for Electric Vehicle (EV) car-sharing station design jointly optimise location and vehicle routing but treat all user groups identically. This paper identifies that formulation as the source of systematic equity gaps: high-demand neighbourhoods get over-served and low-demand ones get under-served because revenue-maximising station placement ignores service-rate disparity. Zhou, Sherkat-Masoumi, and Bodur introduce a bi-objective framework that directly optimises against two fairness metrics simultaneously, using realised group service rates derived from the operational routing model rather than static accessibility proxies. The paper is the first to embed branch-and-price exactly into a bi-objective EV sharing design, enabling computation of the full Pareto frontier rather than single weighted-sum approximations.
Source: arXiv:2604.11732 📋 Case Study below ↓ReSched: Rethinking Flexible Job Shop Scheduling from a Transformer-Based Architecture with Simplified States 🔗
Foundation: The Flexible Job Shop Scheduling Problem (FJSP) asks how to assign and sequence operations across machines so all jobs finish as early as possible, where each operation can run on any of several eligible machines. It is harder than standard job shop scheduling because both the machine assignment and the operation ordering must be decided simultaneously. Deep reinforcement learning methods have dominated this space for several years, but prior models required increasingly complex state representations to handle more problem variants.
Published at ICLR 2026 (International Conference on Learning Representations)
Deep Reinforcement Learning (DRL) methods for the Flexible Job Shop Scheduling Problem (FJSP) have accumulated complex state representations and architectural bloat as researchers stacked features to handle more constraint types. ReSched identifies this complexity as the bottleneck and goes the opposite direction: it reduces the Markov Decision Process (MDP) state to four essential features (job identity, machine identity, available time, and operation index) and builds a Transformer that uses Rotary Positional Encoding (RoPE) for intra-job sequencing alongside a cross-attention module that directly embeds graph-edge features. The result outperforms all prior classical dispatching rules and state-of-the-art DRL methods on FJSP, and the same architecture generalises to the Job Shop Scheduling Problem (JSSP) and Flexible Flow Shop Scheduling Problem (FFSP) without retraining, a property prior specialised DRL methods do not achieve.
Term of the Day
Large Neighbourhood Search
Large Neighbourhood Search (LNS) is a metaheuristic that improves a feasible solution by repeatedly destroying a large portion of it and then repairing the destroyed portion using a local optimisation solver. Each cycle makes a large, structured move through the solution space rather than the single-variable swaps used in standard local search. The result is that LNS can escape local optima that single-variable methods cannot, because the repair solver finds arrangements reachable only by changing many variables simultaneously. The name is literal: the neighbourhood explored at each iteration is large enough to include solutions several variable-values away from the current one.
A concrete example: vehicle routing
A logistics planner has a feasible route plan for 200 deliveries across 12 trucks. Standard local search tries swapping one delivery between two routes, but it quickly gets stuck: no single swap improves the plan further without violating a time window or capacity constraint.
Without LNS: Local search terminates at a local optimum. Any individual swap either violates a constraint or increases cost, so no improvement is found.
With LNS (destroy-repair): Destroy step: remove 40 deliveries (20%) from their assigned routes, leaving gaps. Repair step: run a small Mixed-Integer Programming (MIP) solver on just those 40 displaced deliveries to re-insert them optimally into the existing route structure. Because 40 deliveries form a tractable subproblem, the repair solver finds an arrangement that no single swap could reach. After 200 iterations of destroy-repair, the solution cost falls 10-15% below the local search result.
The key design choice is the destroy operator. A random destroy finds some improvements; a problem-aware destroy (remove deliveries on the most congested routes, or remove deliveries with tight time windows) finds far better repairs. This is why published LNS results vary so widely: the destroy operator is the engineering insight, not the repair.
Where this shows up in practice
Transport: LNS is the dominant metaheuristic for large-scale VRP variants, crew scheduling, and port berth allocation. Manufacturing: Job shop and open shop scheduling use LNS to destroy a machine's schedule and repair it with a priority-rule or CP solver. Healthcare: Nurse rostering destroys a block of shifts (one ward, one week) and repairs with a CP model under contractual constraints. Energy: Storage dispatch optimisation destroys a time window and repairs with a linear program. The first question to ask when evaluating a published LNS implementation: is the repair exact (MIP or Constraint ProgrammingSpecifies constraints over decision variables and uses propagation plus search to find solutions. First explained 7 Apr 2026.) or heuristic (priority rule, greedy)? Exact repair gives optimality guarantees within the repaired subproblem; heuristic repair is faster but weaker.
Upcoming Conferences
Both research signals today decompose a hard decision problem across specialist components rather than handing it to a single monolithic solver: IonQ and Einride route freight by letting a quantum module generate warm-starts for a classical VRP solver, and ReSched schedules manufacturing jobs using a Transformer trained on a four-feature MDP that generalises across three scheduling problem families.
- The IonQ+Einride workflow is reproducible and modular by design: the quantum subproblem (shipment selection as MIQP) is separable from the classical VRP routing layer, so the quantum component can be swapped for an improved algorithm without redesigning the full pipeline.
- ReSched's result that four features suffice for a state-of-the-art FJSP policy challenges the assumption that more complex state representations are necessary for generalisation -- prior DRL scheduling methods that specialise to one problem type cannot match it on the other variants.
For practitioners: The IonQ+Einride paper (arXiv:2604.11758) gives the first quantum-classical VRP benchmark on real commercial data with a reusable modular architecture; the 12% shipment delivery improvement and unchanged operational cost quantify the current marginal value of the quantum component as a warm-start, not a replacement solver.