LLMs Outclass Humans in AI Planning with Evolutionary Heuristics

JUST IN: The world of symbolic AI planning just got a shake-up. Researchers have pulled off a wild feat by using evolutionary search coupled with large language models (LLMs) to craft heuristics that smoke the old-school, hand-engineered ones. And get this, these new domain-independent heuristics aren't just playing catch-up. They're stealing the spotlight.

The Evolutionary Twist

By letting an LLM mutate parent heuristics written in C++, researchers have hit a new milestone. They stored candidate heuristics in a MAP-Elites archive, focusing on informedness and speed. The result? A fitness score that blends coverage with solving time, setting a fresh benchmark.

On unseen testing domains, the evolved heuristics didn't just hold their ground, they dominated. They solved more tasks than even the strongest baseline heuristics, spanning the Pareto frontier of informedness-speed tradeoff. That's a mouthful, but it basically means they're efficient and effective.

Out with the Old, In with the New

Sources confirm: starting from the trivial blind heuristic outperformed starting with the strong FF heuristic, even when the end result was an FF variant. It's a surprising twist that shakes up the conventional wisdom in AI planning.

And just like that, the leaderboard shifts. These evolved programs aren't just pretty faces. They're written in plain C++, making them easy to plug into existing planners. This isn't just a boost in performance. It's a easy integration that retains the soundness and completeness guarantees of the underlying search.

Why This Matters

Why should you care? Simple. This changes AI planning. It's not just about outperforming human-engineered solutions. It's about creating heuristics that adapt and evolve, offering a glimpse into the future of AI development.

The labs are scrambling. Will this lead to a new era where LLMs consistently outsmart human ingenuity in AI planning? It looks that way. The ball's in the court of traditional researchers now. Will they adapt or get left behind?