[nvrmnd]

I agree, learning admissible heuristics will retain worst case performance, which has always been the measuring stick for these algorithms. It's not at all uncommon to find faster solutions for the average or even p99 cases that cannot provide guarantees on the worst case.

[SnowflakeOnIce]

How would one go about proving that a learned heuristic (something from an AI model) is in fact admissible?

[jvanderbot]

For something like focal search, it doesn't even need admissibility, you just apply it as a second selection heuristic among the choices your admissable heuristic returns as 'top k' choices.

[SnowflakeOnIce]

So a tiebreaker then?