Apparently, Rintanen used these domains in his benchmark but ignored the
action costs entirely. With this modification, we can also run these
domains with our planner. This adds back the affected domains to run
them against Madagascar in a simplified form.
This removes all domains where action costs are used. The reason is that
the encodings currently don’t support action costs, which leads to
incorrect results. Hence, Madagascar and our planner can’t be compared
on these domains.
This sets a soft memory limit of 8192 MB on all configurations. This is
a hint to the planners that they shouldn’t surpass this limit. Should
planners still consume more memory, there is a 1024 MB tolerance margin
(configured in the benchmark runner’s configuration) after which the
measurement is counted as a memout.
Note that this 8192 MB is the default for Madagascar, which is why this
addition doesn’t have an influence on prior results obtained for
Madagascar.
This exchanges Madagascar configuration MpC with Mp in the benchmark on
Rintanen’s instance set. This is because MpC uses planner C, while our
competing configuration uses planner B, which would make the results
difficult to compare.
This adds Madagascar MpC and the best configuration for the generate-
and-test planning strategy to the comparison on Rintanen’s instance set
from AIJ 2012.
This adds multiple configuration variations for the generate-and-check
planning strategy on the instances where the first found plan is not
serializable, with the comparison to Madagascar.
Instead of specifying the heuristic encoding manually, this lets the
planner define its location. This is to make sure that an incorrect
encoding isn’t applied accidentally.
Patrick Lühne
Patrick Lühne