Marking this one as resolved.

Just for the record: I did some experiments just to make sure I did not mess
anything up with the last few changes.
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-22-issue491-v4-cggl-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-22-issue491-v4-dfp-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-22-issue491-v4-rl-comp.html

Thanks! I'll fix the small things you mentioned and will merge (back on the way
home :-) ).

I looked at the code and made all the comments I have. It's only a few small
things; otherwise, looks ready to merge for me. :-)

Right, results look good. :-) I'll have a look at the code next.

I think this was the mixture of bad luck (some run logs I compared showed no
difference in behavior other than time/memory comsumption which I could not see
any reason for) and a bug which only affected tasks detected as unsolvable by
the translator (empty g_operators).

http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-21-issue491-v3-cggl-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-21-issue491-v3-dfp-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-21-issue491-v3-rl-comp.html

You are right, actually nothing should have changed with the last commit.

The diff is here:
https://bitbucket.org/SilvanS/fd-dev/commits/777cd54388d30258f047578c80f4f22bced5e309

I don't think they looks so good. In the previous version, the time scores were
fine (mostly green), now they look quite a bit red. Do you have an idea where
this comes from? Is it possible to prepare a diff between v1 and v2 for
commenting on?

I did more refactorings (see pull request) and compared against the previous
version (looks all good):
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-19-issue491-v2-cggl-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-19-issue491-v2-dfp-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-19-issue491-v2-rl-comp.html

I created a pull request.

The design sounds good to me for now. Can you send a link to the code?

http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-16-issue491-v1-cggl-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-16-issue491-v1-dfp-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-10-16-issue491-v1-rl-comp.html

Results look rather fine I think, gaining a bit in several configurations.

Should we discuss a more advanced design of handling non-existing labels or is
that fine for the moment?

Experiments for the first version are running.

I simply removed the tree structure, deleting all reduced labels, setting the
entries in the vector of all labels to 0, similar to how the set of all
transition systems is handled in the merge-and-shrink heuristic.

This is part of meta issue432 (item 2 in msg3681).

Get rid of unused labels rather than having a tree-like structure for the set
of all labels. This should be possible under the assumption that abstractions
are always in a valid state.