6 ms·
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 unco
by nvrmnd 3y ago
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 3y agoHow would one go about proving that a learned heuristic (something from an AI model) is in fact admissible?
- jvanderbot 3y agoFor 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 3y agoSo a tiebreaker then?