11 April 2026

Blue Yonder Named Leader in Inaugural 2026 Gartner Magic Quadrant for Supply Chain Planning Solutions: Discrete Industries

Gartner published its inaugural separate Magic Quadrant (MQ) for Supply Chain Planning Solutions: Discrete Industries in April 2026, splitting what was previously a unified supply chain planning category into two distinct reports, one for discrete manufacturing and one for process manufacturing. Blue Yonder was named a Leader based on its Cognitive Solutions platform, which applies AI-native planning across demand forecasting, supply optimisation, inventory management, and production scheduling for manufacturers of assembled goods. The SADA Loop (See, Analyze, Decide, Act) model at the platform core runs continuous planning cycles rather than batch-period refreshes. Blue Yonder's position in the discrete category reflects the company's focus on tight routing constraints, firm lead times, and production sequencing decisions that characterise discrete component manufacturing, semiconductor fabrication, and automotive assembly, which involve fundamentally different combinatorial structures compared to continuous flow or batch process production.

Why it matters: When Gartner splits a Magic Quadrant, it signals that the two sub-categories have diverged enough in their underlying problem structures that a single vendor ranking no longer serves both audiences. Practitioners who used a combined supply chain planning Magic Quadrant to select planning software for discrete factories should now re-evaluate: vendor positions differ materially between the inaugural discrete-industries ranking and the prior unified report. Any vendor comparison document that does not reference the 2026 discrete-industries Magic Quadrant is benchmarking against outdated criteria.

Column-Generation Solves Real Electricity Technician Routing with Lexicographic Objectives in Under 5 Minutes

Elise Bangerter, David Schindl, Meritxell Pacheco Paneque, Nour Elhouda Tellache, Rodolphe Griset

Electric utilities must schedule daily technician interventions for grid maintenance and fault repair. When resource constraints prevent completing all planned work in a single day, the utility must prioritise: completing as many interventions as possible takes absolute precedence, and only once that ceiling is reached does minimising operational cost become relevant. Bangerter et al. model this as a multi-objective technician routing and scheduling problem with a lexicographic objective structure, where the first criterion (maximise intervention completions) is never traded off against the second (minimise cost). They develop two formulations: a compact Mixed Integer Linear Program (MILP) and an extended set-packing formulation. The core algorithm is a Column GenerationAn algorithmic technique solving large Linear Programs (LPs) and MIPs by iterating between a Restricted Master Problem and a pricing subproblem that identifies the most improving variable to add. — first explained April 9 2026. approach with dynamic programming-based labelling for the pricing subproblem, targeting practical deployment: it delivers high-quality solutions on real French utility instances in under 5 minutes, outperforming the compact formulation on larger instances with lower optimality gaps.

What problem it solves: Standard multi-objective VRP formulations either use weighted objectives (which require practitioners to calibrate trade-off weights) or enumerate the full Pareto front (computationally prohibitive for production use). Lexicographic objectives eliminate both problems by encoding the utility's actual policy: completions are non-negotiable, cost is secondary. The set-packing formulation enables column generation to generate only routes that can realistically be executed given technician skill sets and time windows.

Why it matters: Emergency repair and maintenance scheduling is structurally identical across electric utilities, telecoms field service, and healthcare home-care routing: completions (or patient visits) dominate cost as the primary objective, and any formulation that blends them with a weighted sum misrepresents the actual operating policy. This paper provides a replicable template for encoding lexicographic priority directly into a production-ready column generation algorithm, with benchmark results from a live utility deployment.

Sparse Learning Outperforms Deep Neural Networks on MIP Branching Using Fewer Than 4% of Parameters

Selin Bayramoğlu, George L. Nemhauser, Nikolaos V. Sahinidis — Georgia Institute of Technology

The variable branching decision inside branch-and-bound is the central performance lever in MIP solving: choosing which variable to branch on at each node determines the shape of the search tree. Strong branching (evaluating all candidates) is optimal but prohibitively expensive; learned policies based on graph neural networks (GNNs) are fast but require Graphics Processing Unit (GPU) infrastructure and large training sets. Bayramoğlu, Nemhauser, and Sahinidis propose sparse interpretable models that approximate strong branching scores using fewer than 4% of GNN parameters. The resulting CPU-only models remain effective with small training sets and outperform both SCIP's default branching rules and GPU-accelerated GNN approaches across diverse problem classes in extensive experiments. The interpretable structure of the sparse model also enables inspection of which variable features drive branching quality for a given problem class.

What problem it solves: Production MIP deployments that cannot justify GPU infrastructure for solver acceleration have been unable to benefit from learned branching. Sparse CPU-only models with small training requirements are directly deployable in cost-constrained cloud environments, inside open-source solver pipelines such as SCIP or HiGHS, and on-premises without specialised hardware.

Why it matters: The GNN branching literature has created an implicit assumption that learned branching requires deep models and GPU compute. This paper directly falsifies that assumption: for practical MIP instances, interpretable sparsity is a better inductive bias than depth. Teams running SCIP on CPU have a drop-in path to improved branching quality with no infrastructure change, requiring only a small set of training instances from their own problem class.

Lexicographic Optimisation

A lexicographic objective ranks multiple criteria in strict sequential order and optimises them one at a time. The first criterion is solved to its best achievable value. The feasible region is then restricted to solutions that achieve exactly that optimal first-criterion value, and the second criterion is solved within this restricted set. The process continues until all criteria are exhausted or no further restriction remains. At no stage is any trade-off permitted: a solution that achieves criterion 1 at its optimum but worsens any lower-ranked criterion by any amount is always preferred over a solution that slightly degrades criterion 1 in exchange for large gains on criterion 2.

In practice, lexicographic objectives appear wherever an organisation has a non-negotiable primary goal before any secondary goal applies: completing all emergency utility interventions before minimising cost (today's arXiv:2604.05153), maximising patient coverage before minimising travel time in home-care routing, dispatching all must-respond incidents before balancing workload in emergency service scheduling, and clearing all safety constraints before minimising carbon emissions in energy dispatch. The structure is particularly natural in regulated or safety-critical domains where one objective represents a legal or contractual floor that cannot be compromised for efficiency gains.

The weighted-sum gap in production models: Many planning models that claim to "prioritise" service level over cost actually implement a weighted sum with a high service-level coefficient. When demand spikes cause the service objective to become infeasible at its stated weight, the solver degrades service level to maintain feasibility — which is exactly the behaviour a lexicographic formulation would prevent. Before accepting any multi-objective planning model, ask whether the stated priority is enforced as a sequential constraint or as a weight, and whether the solver has ever returned a solution that traded off the primary objective for gains on a secondary one.

Related:

For practitioners: If you are evaluating supply chain planning software for a discrete manufacturing context, any vendor comparison built on the old unified Gartner Magic Quadrant (MQ) is benchmarking the wrong ranking — request a shortlist specifically against the inaugural 2026 Discrete Industries MQ, where vendor positions differ materially.