How the Optimisation Works

What Goes In, What Comes Out

What you give it:

• The orders — what has to be delivered or collected, where, and between what times.
• The fleet and the people — vehicles, trailers, drivers, their shifts, what each can carry and what each is allowed to do.
• The rules — loading rules, which vehicles can get into which sites, drivers’ hours, site opening times, which products can travel together.
• The costs — what an hour, a mile, a vehicle, and a missed delivery window actually cost you.

What comes back is a plan that obeys every one of those rules, at the lowest cost it could find — measured on what the operation really spends, not on total distance.

Why There Is No Known “Best” Solution

You cannot calculate the best plan directly. One vehicle calling at ten sites can visit them in 3,628,800 different arrangements. Add hundreds of orders, tens of vehicles, delivery windows, loading rules and drivers’ hours, and there are more possible plans than there are atoms in the universe.

Methods that prove they have found the very best answer do exist, but only for small, stripped-back examples — the vehicle routing problem is one of the hardest problems of its kind anyone has studied. So the optimiser goes looking instead, examining thousands of possible changes every second and keeping what works. (It “searches the solution space”).

How the Search Works

Build Starter Plans (Construction Phase)

The optimiser starts by putting every order on a vehicle in a sensible order. It doesn’t build one plan and work on it — a typical run builds hundreds, using different ideas of what matters most: keeping drops close together, getting the tightest delivery windows covered first, and so on, and starting routes from different first “seed” orders. Different starting points lead somewhere different, and it is worth having a lot of them.

Each plan is polished - run through a first round of improvements, the top plans are then chosen and polished some more before moving to the second phase.

Improvements (Local Search)

After construction and in the exploration phase (see below) solutions run through an improvement step. This runs several low-level heuristic to make small changes to those plans and keeps the improvements that save money: calling at two sites the other way round, moving a drop to a different vehicle, shifting a run of drops from one route to another. This is quick and works — up to a point. Sooner or later it reaches a plan where no single small change helps, which is rarely the best plan available (a “local minimum”).

Shake It Up (Exploration Phase)

To get past that, the optimiser takes a good plan apart on purpose: it lifts out whole routes, or every drop in one area, and rebuilds around what is left. The rebuilt plan is often worse to start with. Allowing that, in a controlled way, is exactly what lets the search reach good plans it could never have edged its way towards a step at a time.

Those plans are all worked on together rather than one after another, and a good idea found in one can be passed to the others. Parallel heuristics covers how that is managed.

The Rules Are Checked as It Goes

Every change — every drop moved, every route rebuilt — is checked against what Optrak knows about your operation before it is allowed: how long roads take at that time of day, whether the load still fits and can be unloaded in the right order, whether the driver, vehicle and trailer are all available and allowed.

Anything that breaks a rule is thrown out there and then, rather than turning up as a problem at the warehouse. It is why a plan Optrak gives you does not need checking over afterwards: a plan that does not work was never allowed in.

What “Better” Means

To choose between two plans the optimiser needs one answer to “which of these is better”, and it gets it from what the plans actually cost to run: the fixed and running costs of each vehicle, drivers’ hours, and fuel or energy.

That matters more than it sounds. Because plans are compared on real cost rather than on total distance, the optimiser will happily make a route longer if that saves a whole vehicle, or reorder a run to protect a delivery window — trade-offs that software aiming only at the shortest distance cannot see.

Where to Go Next

• The methods themselves — the roughly thirty ways Optrak has of building and improving a plan, and how they are combined to suit an operation.
• Parallel heuristics — how many attempts are run at once, what decides which to try next, and what decides whether to keep a result.
• Autotuning — how Optrak works out which methods suit your operation, using your own past work, and does it again as things change.
• In-route charging — how charging stops are placed inside the route when an electric vehicle can’t do a round on one charge.

The first two are the engineering detail, written for technical evaluators and for AI search. The other two describe what Optrak does rather than how it is built.