The bug has already been resolved by the fix of issue289. We removed the
outdated assertion.

I think this is fine; the assertion can go. It probably came from a time when
there was no intention to support axioms and general action costs. In PSR-Large,
the goal is achieved by axioms (which have no cost), and while the true goal
distance for non-goal states is always greater than 0, the lmcount heuristic
probably won't recognize this because it misses some landmarks.

correction: more exactly the assertion is triggered by this:
if (h == 0 && g_min_action_cost > 0)
        assert(test_goal(state));

This error does not seem to occur anymore. However, if I run debug mode, I get
another assertion error: landmarks/landmark_count_heuristic.cc:100: int
LandmarkCountHeuristic::get_heuristic_value(const State&): Assertion
`test_goal(state)' failed. (this happens if a non-goal state has heuristic value
0). In normal mode, the planner finds a valid goal.

I tried this only for the first three psr-large problems and only validated p29;
as soon as the grid is executing my jobs again I will run a test over those two
domains with this configuration.

Can lmcount assign h=0 for non-goal states?

This has very likely already been resolved with issue289. Salomé will verify this.

PS: I can have a look at this after ICAPS.

We should definitely integrate issue123 first, if we are happy with the results.
The same probably applies for the changes Salomé is currently working on (see
issue289).

Anyone want to have a shot at finding out what is going on there? I can try to
explain the code if necessary, although I don't understand the landmark code
that well myself.

Adding Silvan to the nosy list. For all bugs related to the landmark code, I
wonder if we should resolve issue123 first, since I'm expecting merge conflicts
otherwise.

Indeed they are.

> This error only happens for 16 tasks in psr-large and psr-middle.

Are these the 16 tasks by any chance?

psr-large
p09-s90-n6-l4-f50.pddl
p18-s133-n9-l3-f30.pddl
p29-s177-n30-l2-f30.pddl
p34-s187-n40-l3-f30.pddl
p37-s193-n50-l2-f30.pddl
p40-s199-n55-l3-f30.pddl
p41-s201-n60-l2-f30.pddl
p42-s203-n60-l3-f30.pddl
p46-s211-n80-l3-f30.pddl
p49-s217-n100-l2-f30.pddl
p50-s219-n100-l3-f30.pddl

psr-middle
p12-s50-n4-l2-f50.pddl
p14-s55-n4-l3-f70.pddl
p20-s69-n5-l3-f30.pddl
p30-s90-n6-l4-f50.pddl
p45-s132-n9-l3-f10.pddl

If they are the problematic tasks, this would explain some failed tasks in the
evaluation for issue232. In that experiment they segfaulted for the
configuration seq_sat_lama_2011 which also uses the landmark generation code.

$ ./downward-1-debug --landmarks "lmg=lm_rhw()" --heuristic "hLM,hFF=lm_ff_syn(lmg)" --search "lazy(single(hLM))" < ~/lab/preprocessed-tasks/WORK-WORK/psr-large/p09-s90-n6-l4-
f50.pddl/output
Simplifying transitions... done!
Initializing Exploration...
Generating landmarks using the RPG/SAS+ approach
downward-1-debug: landmarks/landmark_factory.cc:684: bool LandmarkFactory::remove_first_weakest_cycle_edge(LandmarkNode*, std::list<std::pair<LandmarkNode*, edge_type> >&, 
std::list<std::pair<LandmarkNode*, edge_type> >::iterator): Assertion `parent_p != 0 && child_p != 0' failed.
Peak memory: 5848 KB
caught signal 6 -- exiting
Abgebrochen (Speicherabzug geschrieben)

The normal binary without assertions just segfaults. This error only happens for 16 tasks in psr-large and psr-middle.