A new AI system can devise several ways to solve a problem, then retain the most effective ones.
In a large port, several ships sometimes await the assistance of tugboats. A decision must be made about which tugboat to send, when to send it and to which vessel. Each one has its own availability, journeys take time and certain schedules must be respected. As soon as one element changes, the entire organization may need to be reconsidered.
Tugboat in front of a container ship
Pixabay illustration
To solve this kind of puzzle, specialists write programs that search for a good arrangement without trying every possibility. To do this, they use practical rules: prioritizing certain ships, reducing travel or avoiding unnecessary waiting. The problem is that a method that is effective in one port is not necessarily effective in another.
Researchers therefore designed a system called LACE, which relies on AIs capable of writing code. Instead of asking them directly for the best answer, it has them create small programs proposing different ways of organizing things.
The system also has tools for checking each proposal. For example, it can verify that a tugboat is not assigned to two ships at the same time and compare the quality of the arrangements obtained. Programs that produce good results are kept and improved through successive trials.
The major difference is that LACE does not seek a single method that works everywhere. It retains several that complement one another. One strategy may be better when there are very many ships, while another may be better when they are spread out. The system then chooses the one best suited to the problem at hand.
The researchers first tested it on 36 calculation and organization problems already used to evaluate this type of method. LACE achieved an average score of 0.945, compared with 0.870 for the best competing system based on AI. A simple request made directly to an AI achieved 0.571.
They then tested it on four new problems inspired by port logistics, including the organization of tugboats. LACE achieved scores ranging from 0.97 to 0.99. Under the same conditions, five other AI-based methods failed to produce a program that met all the constraints.
The trials were conducted on problems prepared for the study, with precise rules and verification methods. But the results show that AI can help design specialized calculation methods, a task that usually requires engineers to invest a great deal of time.