We should investigate the effect of tie-breaking on LM-Cut performance.
Background:
As issue69 showed, LM-Cut tie-breaking has a significant effect on expansions.
At some point we made parts of the LM-cut computation incremental, which
affected tie-breaking in such a way that the number of expansions increased in
many domains (not hugely, but noticeably). Overall coverage also got worse due
to a huge coverage decrease in Airport. At the same time, overall search time
decreased in most domains because the speed-up due to incremental computation
more than balanced the increase in expansions.
Scatter plots that show the main findings from that issue are attached. Some
domains with 0-cost actions and very large costs are omitted (also in the plots
TOTAL-*.png that give all data points in one plot) because the configurations
tested in that issue supported these domains only partially or not at all.
Notes for repository divers:
The "good" tie-breaking (w.r.t. number of expansions) was still present in SVN
revision r3613. The "bad" tie-breaking was already present in SVN revision
r5272. The experiments in the scatter plots used translator revision r3613 and
preprocessor revision r5272 for both search code versions.
|