Skip to content

Large-scale optimisation

Your model solved one region fine, but the full network never finishes. We split it into pieces that solve in time and still fit together.

  • Free 30-minute call
  • No commitment
  • You own the code
Tools
CPLEX
Used in
3 documented use cases
Delivery
Documented source code, yours to keep

In short: Large-scale optimisation means solving planning problems too big for one model to finish in time, such as all sites and transports of one company. Decomposition methods split the problem into smaller models that solve quickly and still fit together into one plan.

The problem

The pilot covered one region, and the model solved it overnight. Then you added every site, every product and a full year. Now the run never finishes. Or the software runs out of memory. Standard tools can’t finish a problem this size in time.

What changes

Today, you shrink the problem until it fits: fewer sites, a shorter period, rougher data. Our model covers the whole network at full size. It is built to finish in your planning window. We split the model into pieces for this. The plan for one site still fits the plan for the next.

How it works

Decomposition splits one huge model into smaller ones that work together. Benders decomposition settles the big decisions first, such as which sites open. Then it checks the detail and reports back. Dantzig-Wolfe column generation builds the plan from promising pieces, adding only those that improve it. Some rules tie the parts of a model together, such as one budget for all regions. Lagrangian relaxation lets the model violate these rules and charges a penalty each time.

We parallelise metaheuristics, the fast search strategies, so many searches run at once. Rolling-horizon methods split the problem over time. They plan the next weeks in detail and the rest roughly, then move on. Hierarchical planning decides the big picture first, then the detail, and keeps the sub-problems in step.

Which split fits your problem?

The shape of the problem picks the method. If a few hard yes-or-no choices steer many flows, like sites and shipments, Benders fits. If the plan consists of many similar building blocks, like routes or shift patterns, column generation fits. If your regions are independent and share only a few limits, such as one budget, Lagrangian relaxation fits.

Before writing anything custom, we try the automatic version. CPLEX has a Benders setting that builds the split itself. It puts the whole-number decisions in the master problem, the part that makes the big calls, and the rest in subproblems. When that’s enough, you skip the custom code entirely. When it isn’t, the run still shows where the time goes. The honest boundary: if your full model solves in time, decomposition only adds complexity. We don’t add it for show.

When it isn’t the answer

The honest boundary: if your full model solves in time, decomposition only adds complexity. We don’t add it for show.

Common questions.

Something else? Ask us directly.

Benders splits the model in two: the big decisions, such as which sites open, and the detail, such as the shipments. The detail part checks each proposal and reports back. Benders works when each report rules out many bad choices at once. An answer on Operations Research Stack Exchange warns about the weak case: feedback that only says “not this exact combination” leads to far too many rounds. We test that on a small case before building the full version.

Try it, and measure. If the full model finishes in your window with room to spare, that’s the simplest answer and we’ll say so. Often memory runs out first, or run times grow much faster than the problem does. Then more hardware buys little, and a split pays off.

Not by itself. Benders and column generation, run to the end, can still prove that a plan is the best. Other splits give that proof up for speed. A rolling horizon, for example, plans the next weeks in detail and later weeks only roughly. We tell you which kind you’re getting. Where it can be computed, we also report how far the plan may still be from the best possible.

Sometimes they can, and then we’ll say so. Standard tools are built for an average problem, though. With Large-scale optimisation, we start from yours: its size, its rules and the exceptions nobody wrote down. That is why the plan fits your operation better.

You don’t need either to start. We build the model from the data you’ve got, messy or not, and cleaning it up is part of the job. You get a documented system your team can maintain, and we train them until they can run it long after we’ve left.

It starts with a free 30-minute call about your operation and the model that would fit it. Then we build a working prototype and run it on your numbers. You’ll see results before you commit to anything bigger.

Tell us what you’re planning.

In a free 30-minute call with a specialist, you find out whether optimisation can help.

  • Free, no commitment
  • You talk to a specialist directly
  • A clear go or no-go