New AI Trick Could Make Your Packages Arrive Faster and Cheaper

Every time you order something online, someone has to decide which truck visits which address in what order. Doing that badly wastes fuel, time and money. Doing it well is a famously hard math problem, so researchers use AI to help. This new paper, from a team at Victoria University of Wellington, looks at a technique where AI learns the rules of thumb for building delivery routes rather than solving each route from scratch.

The AI here works a bit like selective breeding. It generates many candidate rulebooks, tests them on delivery problems, and keeps the winners to make the next generation. The researchers noticed a flaw: two rulebooks can get the same score while solving routes in completely different ways. So the team added a second filter that also measures how a rulebook behaves during routing, keeping a diverse mix. They call it 'behaviour-based niching' — essentially keeping a varied team instead of ten copies of the same player.

They tested five versions of this idea against a standard baseline on 200-stop delivery problems, running each 30 times. The best version ranked highest on a separate set of 90 test problems, though the actual cost differences between methods were small. The clearest win was simpler: all the new versions ended up with smaller, less bloated rulebooks, a difference the statistics backed up. Simpler rules are easier to run and understand.

The honest catch: this is laboratory work on generated delivery problems, not real streets with traffic, one-way roads, tight time windows or driver breaks. Cost savings were marginal in these tests. Think of it as a promising method for making route-planning software leaner, not a proven way to cut your shipping bill tomorrow.