New AI Could Plan Complex Projects Faster by Skipping Slow Test Runs

Anyone who has run a big project — a construction site, a software launch, a hospital shift roster — knows the headache. You have limited people and equipment, tasks that must happen in a certain order, several ways to do each job, and timelines that slip. Companies use software to plan this, but the software's rules of thumb are often hand-written and hard to adapt when things change.

The research here uses genetic programming, which is AI that breeds better rules over time, like a brutal gardening contest. Each candidate rule is tested by simulating the entire project. That's accurate but slow, and computing time is limited. So the team built a 'surrogate' — a fast stand-in model that guesses how a rule will perform without running the full simulation, the way a weather app gives you the forecast instead of waiting for tomorrow.

The interesting part is how they described each rule to that shortcut. They tried three levels of detail: the raw numbers the rule outputs, the ranking of tasks it produces, and finally just the binary yes-or-no decisions it makes. The simplest version — final decisions — worked best and most reliably. Their analysis showed why: accuracy alone wasn't the whole story. The description also determined how well the system threw out near-duplicate rules and kept genuinely different ideas alive. Removing clones and preselecting promising rules helped most when combined.

The honest limitation: this is a lab study. The projects were simulated, not real construction sites or factories, and the code isn't a product you can buy. It also assumes the AI is given a fixed budget of test runs — a sensible but artificial constraint. Still, the lesson is broadly useful: sometimes a rougher, faster guess beats a slow, perfect answer, especially when you can run many guesses in the time it takes to run one.