Merged. I'll send an email to Gabi and Salomé.

Looks good to merge. Once it's merged, please also let Gabi and Salomé know
about this change because they have started looking at the current search code.

Results are in:
http://ai.cs.unibas.ch/_tmp_files/pommeren/issue468-base-issue468-v1-compare.html

There are still some unexplained errors from issue467, but otherwise all failed
assertions are gone. The new version is even slightly faster in most cases
although the effect is tiny.

Yes. Its called astar_lmcount_lm_merged_rhw_hm in lab, but the config is the same.

Is this seq-opt-bjolp?

I pulled and merged the latest changes from default and updated the experiment.

> Florian, can you check which of the configs in our standard experiments use
> mpd=True (including the built-in aliases like bjolp) and set up an experiment
> for this? (Or find someone else to do this?)

Out of the cofigs in lab that we usally use for testing, only one config uses
mpd=true. The same config is used in some portfolios with different time
settings and in a selmax config which is currently disabled. For now, I'll just
run the config individually on all tasks in the optimal suite.

Erez, I do not have write access to your repo so I sent you a pull request
adding the experiment file and tags. I'll start the experiment with this version.

Thanks, I'll set up the experiment.

Ok, fix is at the repository at:

https://bitbucket.org/batman/downward-fixes/branch/issue468

Let me know if you don't have read permissions

Thanks, Erez, that would be great.

Florian, can you check which of the configs in our standard experiments use
mpd=True (including the built-in aliases like bjolp) and set up an experiment
for this? (Or find someone else to do this?)

I figures this out: it's not closed nodes being closed again, it's dead end 
nodes being closed. The diff below seems to fix this.
Should I create a branch with the issue name in downward-fixes?

diff -r a38b046fe1e5 src/search/eager_search.cc
@@ -264,6 +264,8 @@
 
         if (use_multi_path_dependence) {
             assert(last_key_removed.size() == 2);
+            if (node.is_dead_end())
+            		continue;
             int pushed_h = last_key_removed[1];
             assert(node.get_h() >= pushed_h);
             if (node.get_h() > pushed_h) {

Yes, this happens in the default branch, too.

I'd be happy to try and debug this (now that AAAI submissions are done). Is this 
in the main branch of the public repository?

Changed the title because the bug must be in the search code, not in the
heuristics. It's probably harmless, but we need to reconsider our invariants
here. Added a cross-reference from issue207.

The following assertion fails when searching with
"astar(lmcount(lm_rhw()),mpd=true)" on large airport tasks, e.g. p37. this only
happens with mpd=true.

search_space.cc:117: void SearchNode::close(): Assertion `info.status ==
SearchNodeInfo::OPEN' failed.