Industry Signals
Quickbase Acquires Solvice: Constraint-Based Route Optimisation and Workforce Scheduling Enter a No-Code Enterprise Platform 🔗
Foundation: Solvice is an optimisation-as-a-service company whose APIs solve two classes of problem: Vehicle Routing Problems (VRPs), which find the lowest-cost set of routes for a delivery fleet subject to vehicle capacity, driver hours, and customer time windows; and workforce scheduling problems, which assign staff to shifts under qualification requirements, legal working-hour limits, and workload-balance targets. Quickbase is a no-code application platform used by approximately 12,000 enterprise customers to build custom operational workflows without dedicated engineering teams. Embedding those two solver APIs into a no-code platform means the optimisation runs inside business applications built by operational managers, without any solver integration work.
The 2 April 2026 acquisition embeds those two solver APIs directly into Quickbase's no-code workflow layer. OnRoute handles vehicle routing under vehicle capacity, crew skills, time windows, multi-day horizons, and real-time changes; OnShift handles workforce scheduling under qualification constraints, working-hour limits, and workload-balance requirements. Published benchmarks show up to 35% reduction in drive time for OnRoute customers. Integration into Quickbase Business/Enterprise and FastField Pro is planned for H2 2026.
NVIDIA cuOpt v26.04.00: Host-Resident Problem Data, Mixed-Precision PDLP, and Remote Execution 🔗
Foundation: NVIDIA cuOpt is a GPU-accelerated optimisation engine covering mixed-integer linear programming (MILP), linear programming (LP), quadratic programming (QP), and vehicle routing. PDLP (Primal-Dual LP) is an LP algorithm that solves by iterative gradient steps rather than the matrix decompositions that classical interior-point solvers (which find the optimum by navigating through the interior of the feasible region) require; gradient steps parallelise well on GPUs, where dense matrix decomposition is a bottleneck. Mixed precision means using 32-bit floating-point (FP32) instead of 64-bit (FP64) arithmetic, which is faster on GPUs without FP64 hardware and sufficient for most LP solutions in practice.
Version 26.04.00 extends GPU-native LP solving and vehicle routing on two fronts: a host-resident memory model that allows problem instances exceeding GPU memory to be held in CPU RAM and streamed to the GPU during solve, removing the hard size ceiling; and mixed-precision PDLP support that opens the gradient-step LP solver to FP32-only GPUs lacking 64-bit arithmetic acceleration. Remote execution via cuopt_grpc_server enables cuOpt to act as a containerised microservice in distributed OR pipelines. The release also includes improved lexicographic search and 30+ bug fixes for variable bounds, race conditions, and numerical stability.
| Field | Content |
|---|---|
| Version | v26.04.00 — 9 April 2026 |
| What changed |
|
| Impact | Practitioners can now deploy cuOpt on a broader range of GPU hardware (FP32-only or mixed) and integrate it as a containerised microservice. The host-memory model removes a hard problem-size ceiling for vehicle routing instances. |
| Migration | Breaking change: the solution object now includes the LP solver used -- code that inspects solution objects directly may need updating. |
Research Papers
Graph-RHO: Critical-Path-Aware Heterogeneous Graph Network for Long-Horizon Flexible Job Shop Scheduling 🔗
Foundation: The Flexible Job Shop Scheduling Problem (FJSP) is the task of deciding which machine processes each manufacturing operation and in what order, across a factory floor where each operation can run on multiple compatible machines; the combination of assignment and sequencing decisions grows combinatorially as the number of operations increases. Deep Reinforcement Learning (DRL) approaches to scheduling train a policy to dispatch operations step by step rather than enumerate the full combinatorial search space, but policies trained on one problem type learn structural patterns that fail to transfer to related variants with different machine topologies.
Graph-RHO bridges that transfer gap by combining a topology-aware heterogeneous graph neural network (which encodes operation-machine relationships with multi-relational edges for precedence, compatibility, and current load) with a rolling horizon optimisation framework and a critical-path-aware training mechanism that assigns higher learning weight to bottleneck operations (the steps that gate overall makespan) than to non-critical ones. The combination achieves over 30% solve-time reduction on large-scale instances and generalises zero-shot across FJSP (Flexible Job Shop Scheduling Problem), JSSP (Job Shop Scheduling Problem), and FFSP (Flexible Flow Shop Problem) benchmarks without retraining, a result prior DRL schedulers specialised to a single problem type do not achieve on the other two families simultaneously. 📋 Case Study below ↓
arXiv:2604.10073 — Li et al., Apr 2026Day-Ahead Offering for Virtual Power Plants: A Stochastic Linear Programming Reformulation with 100x Speedup over Column-and-Constraint Generation 🔗
Foundation: A Virtual Power Plant (VPP) aggregates distributed energy resources (DERs) — rooftop solar, batteries, electric vehicles, small generators — into a single entity that bids into the wholesale electricity market as if it were a conventional power station. The day-ahead offering problem is the daily decision of what quantity of electricity to commit to delivering at each hour of the following day, given that actual DER output and market prices will both be uncertain. This paper handles both sources of uncertainty together using stochastic adaptive robust optimisation — a method that finds a bid performing acceptably across a range of plausible price paths and DER output scenarios, not just the most likely one.
The stochastic adaptive robust formulation turns out to have two exploitable structural properties: the second-stage pricing problem (what to do given a realised price path) reduces to a nested resource allocation that a simple greedy pass solves exactly, and the first-stage bid region is isotonic, meaning the optimal hourly commitments always rank in a consistent order from lowest to highest, hour by hour. Together these properties allow Meng, Li, and Cui to replace the standard column generationAn LP solution method that works with a small subset of variables, adding promising columns only when they improve the objective. First explained 9 Apr 2026. solver with a projected subgradient descent algorithm that exploits the ordering directly. On real-world data, this delivers a roughly 100x speedup over the column-and-constraint generation (C&CG) baseline with no loss in solution quality.
Term of the Day
Rolling Horizon Optimisation
"In preparing for battle I have always found that plans are useless, but planning is indispensable." — Dwight D. Eisenhower
Rolling Horizon Optimisation (RHO) is a method for tackling long-horizon planning problems by solving a sequence of short windows rather than the whole horizon at once. At each step, you solve a problem spanning the next w periods, implement the first k decisions, slide the window forward by k periods, and re-optimise. The key mechanism: each re-solve incorporates updated information and revised forecasts that were not available at the previous step. Sounds like a complete solution to long-horizon intractability. It is not -- the decisions near the end of each window can be systematically suboptimal because the model has no information about what comes after the window closes.
A concrete example -- manufacturing scheduling
A job shop has 500 operations to sequence over a 5-day horizon. An exact solver is intractable, so the scheduler uses a rolling window of 8 hours. Without RHO awareness: the model schedules jobs optimally within the 8-hour window but treats the window boundary as if the world ends there. Jobs that start near the end of hour 8 are scheduled as if they have unlimited machine availability for their remaining operations -- because the model cannot see hours 9 onwards. The result is a schedule that looks good within the window but creates a machine bottleneck in the next window when those jobs reappear as partially complete. With RHO awareness (Graph-RHO's approach): the policy explicitly identifies which operations are on the critical path -- the sequence that determines the minimum makespan -- and biases scheduling decisions to protect those operations from the end-effect, even if they fall near the window boundary.
Where this shows up in practice
Manufacturing: Job shop and flow shop scheduling over multi-day or multi-week horizons, where exact solvers scale poorly beyond a few hours of operations. Energy: Unit commitment and economic dispatch solvers typically run in rolling 24-hour or 48-hour windows updated every hour; decisions made near the end of the window can leave the system in a state that the next window cannot efficiently recover from. Supply chain: Multi-period replenishment and production planning models that re-run nightly with a rolling forecast horizon can over-order inventory in the last few periods of each horizon because no holding cost appears beyond the horizon. Transport: Multi-day crew and vehicle scheduling uses rolling windows to handle late changes; schedules created near the boundary of each window may strand crews or vehicles at hard-to-recover positions. The first question to ask when evaluating a published rolling horizon model is: was the end-effect explicitly modelled or measured?
Upcoming Conferences
- Quickbase + Solvice Route and schedule optimisation, previously requiring dedicated engineering to deploy, is now built into a no-code business platform serving ~12,000 companies.
- cuOpt v26.04 GPU-accelerated optimisation no longer requires expensive data-centre graphics hardware: standard lower-cost GPUs now deliver comparable performance.
- Graph-RHO AI scheduling models fail on unfamiliar factory layouts because they treat every production step as equally important during training. Placing more learning weight on the bottleneck steps that gate the entire schedule eliminates this failure and improves performance by 30%+.
- VPP Day-Ahead The optimal energy bids under uncertainty always rank in a consistent order from lowest to highest value; recognising this ordering structure replaces the standard solving method with one that is 100x faster, with no loss in solution quality.