If we run the trunk search code of r3348 on the "old translator" on the one hand
(r3199 of everything) and the "new translator" on the other hand (r3348 of
trunk), then the new translator leads to clearly worse results.
This may be due to the old translator being buggy (it may have omitted operators
due to "conflicting effects" that it shouldn't), but it may also be due to the
new translator being stupid (it includes operators that do nothing, e.g. fly
from A to A in Logistics). Here's some comparison of coverage with greedy BFS
with a single heuristic and no pref ops on the STRIPS set:
a1000 (m&s): total -8, worst: miconic -4
a10000 (m&s): total -1, worst: several -1
a100000 (m&s): total -4, worst: several -1
a25000 (m&s): total -4, worst: freecell -5
a50000 (m&s): total -4, worst: freecell -2
b (blind): all domains unchanged
c (causal graph): total -1, worst: pipesworld-notankage -1
d (additive): total -1, worst: satellite -2
f (FF): total -13, worst: airport, logistics98, pipesworld-notankage,
pipesworld-tankage, tpp -2
g (goal count): total -1, worst: zenotravel -1
m (h^max): total +1, none got worse
u (landmark cut): total -2, worst: several -1
y (context-enhanced additive): total -9, worst: airport -4
We should look into what exactly goes on there and see if we should make any
changes to the translator to improve performance again.
|