Scheduling and route planning: When artificial intelligence reaches its limits

Automatic route planning and AI-supported dispatching: these buzzwords have been circulating in the logistics industry for several years. There are various approaches to solving this problem, some of which have already proven themselves in practice. Some attempt to evaluate the actual route plans of human dispatchers from the past six months.

Algorithm learns

An intelligent algorithm is supposed to learn from this and recognise what a suitable route should look like and which criteria need to be taken into account. Chat GPT works in a similar way. The big difference, however, is that this so-called ‘large language model’ has access to infinite amounts of written text from the internet, whereas the route plans from the last six months represent a comparatively small information base.

This may be sufficient for simpler dispatching tasks such as distributing general cargo. However, complex planning in the field of intermodal transport cannot be meaningfully mapped with it. Apart from the fact that human-generated plans from the past may be flawed, an algorithm that learns from mediocre dispatchers will ultimately only produce mediocre results itself.

Shadow dispatching in the background

A better AI approach is therefore what is known as shadow dispatching, which works permanently in the background. With the help of AI, this creates route plans that deliver the best results based on the current situation, a set of rules and evaluation criteria. Dispatchers can compare their own work with the results of the AI as shadow dispatching at any time. Better ideas can be adopted and impractical suggestions discarded. The dispatcher therefore remains an indispensable factor.

To calculate useful results, AI needs a lot of information, which varies from company to company. This mainly concerns ‘soft facts’ such as driver preferences. While Willi, a newly married family man, wants to spend as much free time at home as possible, Igor prefers jobs that earn him as much expense money as possible. However, he can no longer be sent to Krefeld to unload because he recently got into a fight with the warehouse clerk there.

These and many other sensitivities must be made clear to the AI in a continuous learning process. Dispatchers and software developers are in constant communication to achieve this. The same applies to changes in soft facts – when existing drivers resign and new colleagues with different preferences join the team.

Like a first-year apprentice

Without this knowledge, any AI software behaves like a first-year apprentice who has just started learning about driving and rest times, the road network and the capacities of individual vehicles. Such ‘apprentices’ have long been available in the form of complex AI applications. Due to their enormous scope, they were not programmed by medium-sized software companies, but by IT giants.

At Google, they are called OR tools and are made available to developers free of charge. The challenge is to translate the problem to be solved into one of the AI models available at Google. For example, you have to abstract that driver Willi's tours should end in Vienna if possible due to his current private situation.

Nevertheless, vehicle fleets will never be able to do without dispatchers entirely. Despite all the precautions and programming, there are still completely unpredictable events in everyday life, such as drivers suddenly falling ill, vehicle breakdowns or accidents. In such situations, a human being will always have to decide how to proceed in each specific case. Speaking of specific: At NUFAM, several companies will be available for technical discussions on the topic of AI. Experts include exhibitors Arealcontrol, Samsara, webfleet solutions, BPW Bergische Achsen, Krone and Scania.