Things "AI" Can Mean in Route Optimisation

1. Algorithm Selection — Choosing the Right Method per Problem

Different real-world routing problems respond best to different algorithms: many drops versus few, tight time windows versus loose, homogeneous versus heterogeneous fleets, single-fuel versus mixed. Choosing which optimisation approach to apply to a given operation is where ML adds clear, measurable value. This is what Optrak’s autotuning does — searching, automatically, for the heuristic combination that performs best on the work a given operator actually runs.

2. Parameter Learning — Tuning the Chosen Method

Once the approach is chosen, its parameters — move counts, acceptance temperatures, phase durations — still have to be set. Learning these against historical performance, rather than having a consultant turn dials by hand, is the second place ML earns its place. Optrak does this as part of the same autotuning campaign, using standard automated algorithm-configuration methods.

3. Performance Prediction — Learning From the Operational Record

Every plan an operator runs leaves a record: how long a leg actually took, how long a vehicle actually sat at a customer, how much energy it actually used. Mining that accumulated record with ML to estimate travel time, dwell time, service time, and consumption — per customer, per road segment, per vehicle type — would produce better inputs for the optimiser. Better estimates mean more accurate feasibility, ETAs, and costs, in both daily plans and strategic scenarios. The ML output would feed the optimisation rather than replace it.

This is the data-warehouse side of the story. The data warehouse behind strategic modelling captures every plan an operation runs, in a fully queryable form. The missing piece is a live source of actual times and consumption to learn from — after which the prediction models can be trained against it. Optrak’s real-time capability is in active development, not currently supported in existing products.

4. End-to-End Neural Optimisation — The Headline No-One Has Cracked

Training a neural network to produce route plans directly is the use that gets the marketing attention and the conference headlines. It is a genuine and active research area. But today it does not match the performance of well-engineered traditional optimisation on production-scale problems with many hard constraints — mixed fleets, time windows, compatibility rules, driver hours, charging. It is worth watching. It is not yet worth building a production planning operation around, and a vendor that leads with it is selling the research, not the result.

5. Agent Orchestration — The Genuinely New Piece

There is one place a large language model is genuinely involved, and it is genuinely new: an agent that orchestrates the tuning search — reading the run history, choosing which configuration method to run next, bounding its compute budget, and judging when to stop. The boundary is deliberate and strict: no language model produces a route, evaluates a cost, or sets a parameter value. It decides which classical search to run next. Because it sits only at the orchestration layer, the model behind it can be swapped — hosted or local — without touching the optimisation.

Where Optrak Uses Machine Learning, and Where It Doesn’t

Optrak uses the first two — algorithm selection and parameter learning — at every operation, and the fifth, an LLM agent orchestrating the tuning. The third, performance prediction, is in active development on top of the data warehouse. It does not hand route construction to a neural network. The optimisation stays operations-research-grounded and inspectable; ML improves how it is configured and what it is fed, not what produces the plan.

Related

• AI in route optimisation — the short version, and what to ask when a vendor says “AI”
• Strategic uses of AI — the operational record as an asset
• Autotuning — algorithm selection and parameter learning in practice
• How the optimisation works — the operations research underneath