7 more tasks for essentially a one-character change is a nice improvement. :-)

Maybe at some point we should do some more detailed profiling of some of our
heuristics and search algorithms. Who knows which other optimization
opportunities are lurking there...

The experiments are already done: 
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-01-05-issue402-abs-d.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-01-05-issue402-abs-p.html

Sorry for the messy presentation: the revision in the middle column
(252e64b08d68) is the most recent one, with another increase in coverage of 7.
The two other revisions are from the previous experiments: baseline and first
version.

Feel free to mark this one as resolved when you have seen everything you wanted.

Yes, I looked at the changes.

I can do this but only after IPC as the grid's resources are really needed right
now. 

Did you have another look at the changes? I would then already merge the patch
so the others can profit from the faster ipdb procedure as well.

I would be interested in seeing what the performance impact of the latest change
is, but I can run a small experiment myself to look into this if you can tell me
where to get the code. So feel free to merge without an additional experiment.

Sorry, I didn't check my gmail emails. I fixed everything you suggested.
Rietveld was updated.

I don't think we need another experiment, do we?

You can find the patch at https://codereview.appspot.com/39420047/ (ask for
permission to view it).

Excellent! Interestingly, 42 of the 49 added tasks are in the two Peg Solitaire
domains, which appears relevant for issue391. Maybe we changed something related
to the clique computation at some point after we did the SoCS experiments. I'll
add a comment to issue391.

As a rule, I think it's always good to give people a way to look at the patch.
Maybe we should find some way in the future to make it less work to provide a
link. (Shouldn't be too difficult to set up some kind of automation for this.)

There is the per problem report as well, sorry:
http://ai.cs.unibas.ch/_tmp_files/sieverss/2013-12-29-issue402-abs-p.html

(PS: I'll have a look at different reports after the vacations, Malte.)

Here are the results:
http://ai.cs.unibas.ch/_tmp_files/sieverss/2013-12-29-issue402-abs-d.html

As expected, the new version is much better: increase in coverage of 49!
If anyone wants to see the patch, let me know. Otherwise, I assume this is fine
to merge.

I am running some experiments.

In the method PatternGenerationHaslum::is_heuristic_improved, we currently
compute the set of maximal additive subsets for the given PDB and the
CanonicalPDBsHeuristic for every sample. We cannot think of any reason not to
compute this only once for every candidate PDB and reuse it when evaluating the
candidate PDB for every sample.