New AI Training Trick Could Mean Cheaper, Faster Delivery Routes

Every time you order something online, a computer somewhere decides which truck carries it and in what order it drops things off. That sounds simple, but route planning is a famously brutal math puzzle: add just one extra stop and the number of possible orderings explodes. Getting routes even slightly shorter saves fuel, driver wages, and delivery time — which is why delivery giants spend real money on this.

The researchers work on a system called GPGLS, which uses "genetic programming" — basically, letting a computer try thousands of rule-of-thumb strategies, keep the winners, and mutate them, the way evolution picks the fittest animals. The problem is that testing every strategy on every example is painfully slow, so they train on small batches of practice problems. Older versions sorted those batches mostly by size: easy ones first, big ones later. The new method, DCL-GPGLS, instead estimates how hard each practice problem is by watching how badly the current AI is failing at it, and updates that estimate as the AI improves. Each round it serves up problems at just the right difficulty, with a safeguard so it doesn't keep drilling the same examples.

They tested it against five rival training methods on a standard set of published routing puzzles called CVRPLIB X. Their approach came out on top on average, and produced the cheapest route on 36 of 65 puzzles it had never seen before. It clearly beat the older fixed-order method on 6 of those puzzles and tied on the other 59 — while using exactly the same amount of computing power, so the win is free.

Here's the honest catch: this is a paper, not a shipping product. The test puzzles are simplified maps without traffic jams, one-way streets, driver shift limits, or time windows. Real delivery companies face all of that. Still, the underlying idea — training an optimizer like a good coach who keeps adjusting the difficulty — is general enough that it could help other gnarly scheduling problems, from warehouse robots to factory planning to ride-hailing apps.