Newbe question about searches
Henryk Trappmann
--alias seq-sat-fdss-2018 --search-time-limit 10h
The best result was 112 steps, but my question is about how the search works: Several searches have exit code 12 (Search stopped without finding a solution). In this case it also states "Completely explored state space -- no solution!"
But what does that mean, it can not mean that there is no solution at all (because later searches found solutions), but I guess it means considering the particular search algorithm/options there is no solution.
Other searches have exit code 23 (Time limit has been reached.) But how can this be if I provided a search time limit of 10h? How can I specify that no search shall be stopped by a time limit? (Or perhaps only discards the last search if it exceeds the overall time limit)
I also ask myself, how can I specify search options, such that it just continues searching until the number of steps of the found plan is below a given threshold? (Accepting that it runs infinitely if there is no such plan)
Greetings, Henryk
Malte Helmert
> So I wrote these pddl files about khunpan and I know there are solutions
> below 80 steps.
> I let fast downward run with the options
>
> --alias seq-sat-fdss-2018 --search-time-limit 10h
>
> The best result was 112 steps, but my question is about how the search
> works: Several searches have exit code 12 (Search stopped without
> finding a solution). In this case it also states "Completely explored
> state space -- no solution!"
>
> But what does that mean, it can not mean that there is no solution at
> all (because later searches found solutions), but I guess it means
> considering the particular search algorithm/options there is no solution.
sequential portfolios like seq-sat-fdss-2018 are somewhat special in
their behaviour. They run a number of different algorithms for a bounded
amount of time, and the portfolios are designed more for finding any
solution than finding high-quality solutions.
After a first solution is found, all further searches are restricted to
only find solutions that are better than the previously found ones, but
because many of them don't guarantee optimality, the message "completely
explored state space -- no solution" only shows that *this approach* is
unable to find a better solution, not that no better solution exists.
This line only means "there is no better solution" for planner
configurations that guarantee optimality.
> Other searches have exit code 23 (Time limit has been reached.) But how
> can this be if I provided a search time limit of 10h? How can I specify
> that no search shall be stopped by a time limit? (Or perhaps only
> discards the last search if it exceeds the overall time limit)
the portfolio assigns a certain percentage of this to each individual
algorithm. The percentage can be quite small for some configurations, so
if one of them receives, say, 1% of the total timeout, it can easily
time out, since 1% of 10h is only 6 minutes. I think there should be
information about the actual time allocated to each component algorithm
in the logs.
> I also ask myself, how can I specify search options, such that it just
> continues searching until the number of steps of the found plan is below
> a given threshold? (Accepting that it runs infinitely if there is no
> such plan)
could, for example, try the seq-sat-lama-2011 configuration, which will
keep searching forever and only give up after it has found an optimal
solution and explored all options for even better solutions. For
non-portfolio approaches, you can also use the "bound" option for the
search algorithm to say that you're only interested in solutions better
than this.
Perhaps someone can give you better advice on which configurations to
try if you can share your PDDL files with us.
Best,
Malte