Results for the bug-fixed code are in, and there are no more invalid plans:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-5-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-5-eval-abs-p.html

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-6-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-6-eval-abs-p.html

Results of first experiment with optimized code:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-3-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-3-eval-abs-p.html

Compare with the following to see the non-optimized vs. optimized difference:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-1-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-1-eval-abs-p.html

Coverage and expansions remained essentially the same, but speed is improved.

This still has the invalid plans, but experiments with the fixed code have been
started. (This time I did not run the no-symmetry configs, though. Hope the fix
cannot affect those adversely.)

I uploaded a new version which should fix the bug (tested on at least one problem 
that previously failed with each heuristic).

Results of second experiment with optimized code are in:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-4-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-4-eval-abs-p.html

Compare with the following to see the non-optimized vs. optimized difference:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-2-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-2-eval-abs-p.html

Having looked at the summary coverage and score_total_time data, this indeed
looks much improved! :-)

It's the same with the logistics00 example with LM-CUT.
I think that something is not updated right in the node/node_info when a state 
from the same orbit as a state that is already in the open list is reached with a 
better f-value, but I'm not sure, and still looking into it.

I looked at the invalid plan, and it could be repaired by replacing some of the
operators by symmetric operators, so it looks like this should be fixable (at
least from this example).

Update: I have looked into the Peg Solitaire #25 validation failure with blind
search and can reproduce it, both with the old code (revision 9edab9ae17f9) and
the current one (revision 2bf9bb7ca79e). The problem is solved in less than 10
seconds and only has one generator, so hopefully it should not be too hard to
debug it.

(Nir, if you want to run the standard validator on this, you maybe first have to
need to remove the string "- number" from the domain file or VAL will crash. At
least my copy does, but I'm not sure if it's the newest version. This does not
affect the semantics of the domain, as "- number" is the default anyway.)

Thanks, Jendrik!

Nir, I've started experiments with the optimized code (revision 2bf9bb7ca79e).

In the meantime, the other experiment with the less optimized code (revision
9edab9ae17f9) has completed, but unfortunately it shows validation errors with
the symmetry configs:

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-1-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-1-eval-abs-p.html

The section of interest is "validate_error"; the first, much smaller HTML file
shows per-domain summaries, the second, huge one shows individual results.

Examples of tasks where solutions failed to validate include:
- with blind search: Pegsol #25
- with LM-cut: Logistics00 #10-0, Pegsol #25
- with m&s: Miconic #8-3, Logistics98 #01

(That's not a complete list; I selected ones that are maybe potentially easier
to work with than some of the others.)

This should be fixed now.

A much optimized version is now available.

Jendrik: the score_search_time in the per-domain link in the previous message
looks wrong. For example, in assembly it's 100 even though nothing is solved,
and more generally all the scores look much too high. They should be much closer
to score_total_time.

Results of the second part of the experiment are in and not looking bad, I
think. :-)  mostly looked at the overall coverage results so far. See the first
link for per-domain summaries and the second for per-problem results.

http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-2-eval-abs-d.html
http://www.informatik.uni-freiburg.de/~helmert/exp-issue267-2-eval-abs-p.html