Timefold Ships Q1 2026 Field-Service Release: Area Preferences, Visit Grouping, Labour-Law Breaks 🔗

Foundation: A gas engineer who lives in Manchester does not want to be routed to Sheffield every morning, and two boiler jobs in the same cul-de-sac should not be split across different vans; encoding those preferences inside a Vehicle Routing Problem (VRP)The task of finding the best set of routes for a fleet of vehicles serving a set of customers under capacity and time constraints. First explained 10 Apr 2026. solver is what separates a research VRP from a field-service dispatch product, and Timefold sells a managed Constraint Programming (CP) platform that bundles those encodings so in-house teams do not have to write them.

The Q1 2026 release adds four capabilities to Timefold's Field Service Routing and Employee Shift Scheduling models: technician area preferences biasing routing toward each engineer's home territory, co-located visit grouping that keeps same-address jobs on one vehicle, break scheduling that enforces jurisdiction-specific labour-law rest windows, and an "Efficiency X-Ray" diagnostic attributing lost productive time to specific constraints or data gaps. The company is pitching the release at enterprise dispatchers currently running bespoke CP or VRP stacks in-house.

Feasibility-Aware Imitation Learning Accelerates Benders Decomposition 1.5-3x on Stochastic MILP Benchmarks 🔗

Foundation: A stochastic facility-location model that sizes plants today before demand is known has to consider hundreds of demand scenarios at once, which makes the full mathematical programme enormous; Benders DecompositionA technique that splits a large optimisation problem into a master problem plus per-scenario subproblems, feeding dual information back to the master as cuts. First explained 12 Apr 2026. splits it into a small master that fixes the plant decisions and a subproblem per scenario that evaluates them. The subproblems return cuts that tighten the master; picking which cuts to add is the bottleneck, and imitation learning offers a way to train a neural policy on an expert solver's choices.

The authors train a cut-generation policy by imitating an expert Benders solver, then add an explicit penalty whenever the learned policy proposes a cut that would violate master-problem feasibility. The resulting policy speeds Benders convergence by 1.5 to 3 times wall-clock on stochastic Mixed Integer Linear Programming (MILP) benchmarks, without sacrificing optimality. The feasibility penalty distinguishes the approach from prior imitation work that could silently cut off feasible master solutions.

Satellite Schedulers Can Learn the Operator's Unwritten Rules Rather Than Pre-Specify Them 🔗

Foundation: Every operations team has a collection of rules nobody wrote down ("do not slew the satellite across the sun vector", "leave twenty minutes between polar imaging passes"), and a scheduler that does not know them keeps producing plans the operator quietly rejects. The authors call this active constraint acquisition: rather than demand a complete constraint model up front, the scheduler proposes candidate plans, the operator accepts or rejects each one, and the system infers which unstated rule was violated. Constraint Programming (CP)A paradigm that models problems as sets of variables and constraints and searches the feasible region via propagation plus backtracking. First explained 9 Apr 2026. with a Boolean Satisfiability (SAT) backend is used as the ground-truth oracle.

The approach formulates Earth-observation satellite scheduling as a combinatorial problem in which operational constraints are initially unknown and must be elicited through plan-then-feedback interaction. A CP-SAT solver defines the ground-truth feasible region; the active-acquisition loop is evaluated by how quickly it converges on that region as feedback accumulates. Empirically the loop reaches near-CP-SAT feasibility using a small fraction of the elicitation effort a human modeller would need to write the constraints explicitly.

Chance Constraint

"Almost feasible, almost always" is a very different promise than "feasible, always." — A. Charnes & W. W. Cooper, Management Science (1959)

A chance constraint says the model must satisfy a requirement with at least probability 1 − α rather than in every possible scenario. It is the device for building a model that is allowed to miss in rare conditions in exchange for being achievable, cheaper, or both in normal conditions. The parameter α is the risk tolerance the modeller picks explicitly (typically 1–5%), which becomes the single most consequential number in the model.

One-line version: A chance constraint lets rare scenarios violate the requirement in exchange for a model that is achievable in the common case, with the risk tolerance α as the single most consequential number the modeller picks.

Related terms: Stochastic ProgrammingOptimisation under uncertainty where uncertain parameters are modelled as random variables with known distributions. · Two-Stage Robust OptimisationMake a first-stage decision before uncertainty is revealed, then recourse decisions afterward, optimising against the worst case in a defined uncertainty set. · Budgeted Uncertainty SetA robust uncertainty set that bounds how many parameters can simultaneously take worst-case values, via a budget Γ.

→ Full concept page