Author: Robert Hyatt
Date: 07:50:52 07/13/02
Go up one level in this thread
On July 13, 2002 at 02:07:17, Vincent Diepeveen wrote: >I still do not understand which positions you talk about which R=2 >is finding and R=3 isn't. Note that he used fixed-depth. This is therefore not surprising since some lines will be searched one ply shallower.. > >What the hell do you do in qsearch, no checks or so? > >On July 11, 2002 at 17:40:07, Omid David wrote: > >>On July 11, 2002 at 17:37:37, Robert Hyatt wrote: >> >>>On July 11, 2002 at 17:27:50, Omid David wrote: >>> >>>>On July 11, 2002 at 17:20:36, Andrew Dados wrote: >>>> >>>>>On July 11, 2002 at 16:38:50, Omid David wrote: >>>>> >>>>>>As part of an extensive research (will be published soon), we tested null-move >>>>>>pruning with fixed depth reductions of R=2 and R=3 on about 800 positions of >>>>>>"mate in 4" (searched to depth of 8 plies) and "mate in 5" (searched to depth of >>>>>>10 plies). The results naturally show that R=2 has greater tactical performance >>>>>>(greater number of checkmate detection). However, we also conducted about >>>>>>hundred self-play matches under 60min/game time control between R=2 and R=3. The >>>>>>outcome is a rather balanced result (R=2 only a little better). Considering that >>>>>>the tests where conducted on a rather slow engine (100k nps), on faster engines >>>>>>R=3 is expected to perform better. >>>>>> >>>>>>So, apparently R=2 is not _by_far_ better than R=3 as some assume. I believe >>>>>>Bruce Moreland had also some good results with R=3 that show it's not too >>>>>>inferior to R=2. Has anyone conducted similar experiments? >>>>> >>>>>Since you didn't specify what engine you have used I assume from your experiment >>>>>it was something which didn't do checks in qsearch. >>>>> >>>>>I think your experiment is of little meaning. >>>>> >>>>>What data would be interesting here is tree size reduction at different depths >>>>>in 'normal' positions; then match result with average depth of 11-14 (this is >>>>>low end of what most engines reach at current hw with tournament tc). >>>>> >>>>>For my engine which does most checks in qsearch R=3 compared to R=2 reduces tree >>>>>size by 15-40% at depth=11. It misses some deep, quiet threats comparing to R=2 >>>>>(forks, some passed pawn combi, etc), but its WAC result at 5s/position (amd 450 >>>>>then) was almost the same (worse by 2 positions if my memory serves). >>>>> >>>>>In very limited nunn-style matches of 32 games g/15 R=3 was never worse then R=2 >>>>>for my engine. >>>>> >>>>>When engine has stripped down qsearch ala Crafty it will need more depth to >>>>>offset additional null reduction. >>>>> >>>>>Please redo your experiment with fritz which does do checks in qsearch. >>>>> >>>>>-Andrew- >>>> >>>>The tree size, etc have been calculated. But that's even not the point. The >>>>point is that in practice by changing the R from 2 to 3, the engine won't lose >>>>too much strength since on many occasions the faster search by R=3 will find the >>>>correct move one ply later but won't waste for that too much. (Although the >>>>research wasn't about this point at all, I just found this point interesting). >>>> >>>>>For my engine which does most checks in qsearch R=3 compared to R=2 reduces >>>>>tree size by 15-40% at depth=11. It misses some deep, quiet threats comparing >>>>>to R=2 >>>> >>>>The fixed depth search on test suites shows that R=2 is clearly far stronger >>>>than R=3, no surprises there. Of course R=3 misses many tactical threats as you >>>>mentioned in fixed depth comparison to R=2. >>> >>> >>>I think this last should be expected. But the point should be that R=3 should >>>actually go deeper, maybe a ply. Which _might_ gain back the tactical losses >>>found at equal depths. >> >>Yes, this is the only explanation why while R=2 is far stronger than R=3 in >>fixed depth, in practical matches R=2 doesn't have such a great dominance over >>R=3.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.