Yesterday's MSc presentation concluded Elia Hänggi's work on the topic of an efficient implementation of h^2 in Fast Downward. He both optimized the existing implementation of h^m (although, limited to m=2) and implemented P^m compilation. The compilation approach indeed still resulted in higher coverage.

With his results, we could revive this issue and discuss whether we want to integrate some of his findings while they are still somewhat in sync with the current Fast Downward main branch. Despite the limitation to m=2, there's interesting stuff in his thesis that we could use to optimize the general case as well. Or we could use his P^m compilation approach instead.

Note that high coverage is still not expected. Elia found that the blind heuristic is fast enough to make up for the bad heuristic quality and beats all his approaches in terms of coverage when using our standard resource limits (30 min, 3.5 GiB).

In his Bachelor thesis, Raphael implemented the P^m compilation. Experimental
results show a higher coverage with the compilation and h^max than with the
existing h^m implementation, so it might be worth to integrate his implementation.