AI Teaches Itself to Solve Logistics Puzzles in Minutes, Not Days

Every big company runs on hidden math puzzles. Which truck takes which route? Which factory makes which product? How do you fit the most boxes into a shipping container? Getting the perfect answer can take a computer hours or even days, so businesses use shortcuts — rough rules of thumb that give a good-enough answer fast. Those shortcuts are normally hand-written by human experts, one at a time.

HeurEvo automates that. It splits the job among AI helpers: a planner decides which steps to take and how much time to give each one, a coder turns that plan into working software, and a component evolver keeps a shared library of reusable building blocks. A fourth AI then reads the results, spots what went wrong, and feeds suggestions back. The system runs many versions at once and keeps the winners — closer to evolution than to traditional programming.

The results are the interesting part. In tests against standard hard optimization problems, HeurEvo found high-quality answers under tight time limits, often matching or beating the best specialized solvers — but those solvers were given hours or days of computing time to work with. It also improved the best previously known results on tricky geometry problems, such as packing hexagons as tightly as possible.

The honest caveat: this is a research paper, not a product you can buy, and results on published test problems don't always carry over to messy real-world operations. But the direction matters. The AI isn't just running a recipe someone gave it — it's inventing the recipe, testing it, and rewriting it. That's a meaningful shift in what automation can do, and it points toward cheaper, faster planning for the things that move goods, people, and schedules.