AI in transport: prediction, adaptation and continuous improvement in route planning

Route planning does not end when an initial plan is generated.

In a real logistics operation, planning is exposed to constant change. An order may be delayed, a customer may change their availability, a vehicle may suffer an incident, an unloading operation may take longer than expected or a route may be affected by traffic, access restrictions or operational problems.

For this reason, artificial intelligence applied to transport should not be limited to calculating routes at a specific moment. Its true value emerges when it helps anticipate problems, adapt decisions and progressively improve planning quality.

In this context, combining machine learning and genetic algorithms is particularly interesting.

Machine learning makes it possible to learn from accumulated experience and detect patterns that are not always obvious. Genetic algorithms make it possible to use that information to search for new solutions when conditions change.

Bringing both techniques together supports logistics planning that is more dynamic, realistic and adapted to day-to-day operations.

Planning as a dynamic process

For a long time, route planning was understood as a static process.

Orders are received, routes are calculated, the plans are given to drivers and the operation is carried out.

However, this approach does not fully reflect reality.

In many companies, planning changes even after it has been generated. New orders may appear, goods preparation may be delayed, drivers may be absent, vehicles may become unavailable, customers may change their schedules or incidents may occur on the road.

In addition, many deviations are not entirely random. Some recur with a certain frequency and can be learnt from historical data.

For example, there may be customers who almost always cause waiting, areas where deliveries take longer at certain times of day or routes that regularly run late even though they appear sound on paper.

Artificial intelligence makes it possible to turn this accumulated experience into useful planning information.

Learning from real operations

A transport or distribution company generates data constantly.

Every completed route contains valuable information: planned time, actual time, service duration, mileage travelled, delays, incidents, undelivered orders, waiting times, vehicle use and customer behaviour.

Traditionally, some of this knowledge remained with planners and traffic managers, who learnt through experience which customers were more complex, which routes were more delicate or which areas required a greater margin.

Machine learning can complement that human experience by analysing large volumes of historical data.

A model can identify patterns such as:

• Customers with above-average unloading times.
• Areas with recurring delays.
• Time bands with a higher risk of non-compliance.
• Orders that usually require more service time.
• Routes that tend to deviate from the initial plan.
• Vehicles or resources better suited to particular operations.

This information can then be incorporated into the optimisation engine to generate more reliable routes.

From theoretical times to real times

One of the most important aspects of route planning is time estimation.

It is not enough to calculate how long it takes to travel from one point to another. Loading, unloading, waiting, access, parking, documentation and customer service times must also be considered.

In many systems, these times are modelled using fixed values or general rules. For example, every customer may be assigned ten minutes of service time, or a standard time may be applied according to order type.

This approach may be sufficient for simple operations, but it often falls short in real scenarios.

Two customers with similar orders may require very different amounts of time. One may offer quick unloading, easy access and available staff. Another may involve waiting, entering a restricted area or carrying out a more complex unloading operation.

Machine learning makes it possible to estimate these times more accurately using historical data.

The system can learn that service time depends on factors such as the customer, area, time of day, product type, delivered volume, vehicle used or history of incidents.

When these estimated times are incorporated into the optimisation engine, routes are no longer based only on theoretical assumptions and become closer to the actual behaviour of the operation.

Predicting operational risks

Artificial intelligence can also help identify risks before they become problems.

A plan may satisfy every constraint on paper yet still have a high risk of failing in practice.

For example, a route may include several deliveries with tight time windows, customers with variable service times and little margin between stops. Although the initial calculation indicates that the route is feasible, the probability of delay may be high.

Machine learning can help detect these situations.

Using historical data, the system can estimate the likelihood that a route, customer or time band will generate delays. It can also identify problematic combinations: certain orders together, particular areas at specific times or routes with little operational margin.

This information can be used in the genetic algorithm's evaluation function.

A solution would be assessed not only by distance or cost, but also by its level of risk.

In this way, the system can prefer a slightly longer plan that is more robust and less exposed to incidents.

More robust routes

In logistics, the best route is not always the shortest.

An overly tight route may be highly efficient in theory but fragile in the face of even a small deviation.

If an unloading operation takes ten minutes longer than expected, the whole route may be affected. If a critical delivery is placed at the end of a sequence with little margin, the risk of non-compliance increases. If several complex customers are grouped on the same route, the plan may become unstable.

AI applied to transport makes it possible to introduce the concept of robustness into planning.

A robust route is one that not only optimises costs, but also retains enough margin to absorb minor incidents.

This may involve spreading higher-uncertainty orders more evenly, avoiding concentrations of risk, reserving margins in certain time bands or selecting sequences that are less sensitive to delay.

Genetic algorithms are particularly well suited to this approach because they can evaluate complete solutions and compare different balances between cost, time and risk.

Adapting to change

When conditions change, the system must be able to respond.

If an urgent order arrives, it is not always desirable to recalculate the entire plan from scratch. If a vehicle becomes unavailable, its orders may need to be redistributed while preserving as much of the existing routes as possible. If a delivery is delayed, changing only part of the sequence may be enough.

Adaptation is one of the major challenges in logistics optimisation.

Genetic algorithms can use existing solutions as a starting point and evolve them towards new alternatives. This makes it possible to preserve valid parts of the previous plan and modify only what is necessary.

This capability is highly useful in dynamic environments, where operations change but rebuilding the whole plan is not always desirable.

The system can search for a new solution that balances several objectives:

• Resolve the incident.
• Keep as many routes unchanged as possible.
• Reduce the impact on customers.
• Avoid missed time windows.
• Control the additional cost.
• Distribute the workload reasonably.

AI is useful not only for planning better at the outset, but also for helping redesign the operation when changes occur.

Continuous improvement

One of the most interesting advantages of applying artificial intelligence to transport is the possibility of continuous improvement.

Every plan generates new data. Every completed route makes it possible to compare what was expected with what actually happened.

If a delivery was planned to take fifteen minutes but ultimately took thirty, that difference provides information. If a route was repeatedly delayed in a particular area, that pattern can be analysed. If a customer regularly fails to comply with its receiving schedule, the system can learn this.

Over time, these data make it possible to adjust estimates, improve the evaluation function and refine optimisation criteria.

Planning stops being an isolated process and becomes a learning cycle:

• A plan is generated.
• Routes are executed.
• Real data are collected.
• Forecasts and results are compared.
• Patterns are learnt.
• The next plan is improved.

This cycle allows the system to adapt progressively to the reality of each company.

The role of the planner

AI applied to transport does not remove the need for human judgement.

On the contrary, it can help planners work with more information and less manual workload.

An intelligent system can calculate thousands of alternatives, detect risks, propose routes, assess scenarios and flag potential incidents. However, people's operational knowledge remains fundamental.

Planning managers know details that are not always present in the data: agreements with customers, exceptional circumstances, commercial preferences, temporary restrictions or strategic decisions.

The best approach is not to replace the planner, but to provide better tools.

AI can handle large-scale calculation, comparison of alternatives and pattern detection. The operations team can focus on validation, adjustment and higher-value decisions.

Application scenarios

The combination of machine learning and genetic algorithms can be applied to many transport scenarios.

In last-mile distribution, it can help forecast more realistic delivery times and build routes that comply more reliably with time windows.

In freight transport, it can optimise the assignment of loads, vehicles and resources while taking account of costs, capacities and actual timings.

In field service or mobile operations, it can help assign jobs to technicians according to location, availability, skills and the estimated duration of each intervention.

In urban logistics, it can help manage access restrictions, time bands, congestion and highly variable areas.

In every case, the objective is the same: to turn data and constraints into better operational decisions.

Avoiding AI disconnected from reality

One risk of applying artificial intelligence in logistics is building models that appear advanced but do not represent the real operation accurately.

A model can be technically sophisticated and still offer little practical value if it does not include the right constraints, works with incomplete data or optimises criteria that do not match business priorities.

For this reason, AI applied to transport must be connected to operational reality.

It must take account of how the company works, which constraints are mandatory, which objectives have priority and which decisions users consider acceptable.

The aim is not to apply machine learning or genetic algorithms simply because they are fashionable technologies. They should be used where they provide real value to the planning process.

Technology must serve the operation, not the other way round.

Benefits of intelligent, adaptive planning

AI-based planning can provide significant benefits.

• It enables more realistic time estimates.
• It helps anticipate delays and incidents.
• It generates more robust routes.
• It reduces improvisation during the working day.
• It improves compliance with time windows.
• It facilitates adaptation to change.
• It makes better use of the available fleet.
• It reduces operating costs.
• It enables learning from accumulated experience.
• It gives the planning team better tools.

These benefits do not depend on one isolated technique, but on the correct integration of data, machine learning, optimisation and operational knowledge.

Conclusion

AI applied to transport makes it possible to move from static planning towards planning that is more intelligent, predictive and adaptive.

Machine learning helps organisations learn from real operations, improve estimates and anticipate risks.

Genetic algorithms make it possible to use that information to generate efficient, robust routes that can adapt when conditions change.

The combination of both techniques means that logistics planning is not limited to calculating routes from theoretical data, but progressively learns from what happens in practice.

After more than 25 years working with optimisation techniques, artificial intelligence, machine learning and genetic algorithms, Evolution Algorithms applies this experience in LOGISPLAN to help companies plan routes that are more efficient, realistic and adapted to their day-to-day operations.