Comment by 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.