Author: Sune Fischer
Date: 12:22:48 07/21/02
Go up one level in this thread
On July 21, 2002 at 14:38:38, Robert Hyatt wrote: >On July 21, 2002 at 03:36:23, Sune Fischer wrote: > >>On July 21, 2002 at 03:29:51, Uri Blass wrote: >> >>>On July 21, 2002 at 03:16:40, Sune Fischer wrote: >>> >>>>On July 20, 2002 at 22:22:22, Robert Hyatt wrote: >>>> >>>>>On July 20, 2002 at 08:13:44, Gian-Carlo Pascutto wrote: >>>>> >>>>>>On July 20, 2002 at 08:04:01, Sune Fischer wrote: >>>>>> >>>>>>>I think it matters "a factor of 2". >>>>>>> >>>>>>>1) it helps you to prune >>>>>>>2) you get better evaluation in the upper plies when you can return a score >>>>>>>based on a deeper search. >>>>>>> >>>>>>>number one will show itself directly because you iterate deeper, the second one >>>>>>>you don't "see", but it does improve depth along some branches in the same way. >>>>>> >>>>>>1) I get +- 10% hash hits (and less prunes) in typical middlegame. Not enough to >>>>>>matter a factor of two (but I didnt check this so not 100% sure). >>>>>> >>>>>>2) Uh? >>>>>> >>>>>>-- >>>>>>GCP >>>>> >>>>> >>>>>Run your program with a tiny hash and a deep search. Then a big hash and >>>>>a deep search. In middlegame positions this will be at least a factor of >>>>>2x. Measure time to depth. Small hash might take 4 minutes to get to depth >>>>>12, then big hash will take around 2 minutes... >>>> >>>>You shouldn't measure time to ply, that would not give you the full benefit of >>>>the hash. You should use time to find the right move. >>>> >>>>I have never seen the hash bring only a factor of 2, even in middlegame. >>>>Last I tested I saw a mate in 5 being solved in 1/4 of the nodes with the hash. >>> >>>Do you use killer moves? >>> >>>I know that at least in version 1.17 you do not do it. >>> >>>Not using killer moves can increase the advantage >>>of hash tables. >> >>That is right, for some reason they increased the size of the tree. >>I found an error in my validate move rutine, perhaps it's time to try them >>again:) >> >>But I still say time-to-ply is not the right way of testing, a "ply" is >>meaningless in this regard. >> >>-S. > >OK.. A statement I can actually live with, finally. :) > >You are correct in the tree searched with hashing and the tree searched without >hashing have differences. > >1. transpositions are eliminated, making the hash tree smaller. > >2. move ordering is better with a hash tree. > >3. The hash tree grafts information from one branch to another, making it >potentially more accurate, for the same ply. > >But, since there is no real way to measure anything else (ie if you measure >time to solution for both it could still be criticized in the same way since >the hash tree will be a bit more accurate in addition to being faster... So >even comparing that is flawed... > Okay, there is an easier way then:) Take a strong program, test its strength with and without hashing. Now we know double speed is about 40-60 elo(*), compare that to the hash gain/loss. That should be far easier to measure, it would also include everything from opening, mid to endgame making the whole discussion more realistic and less academic. As usual I doubt the experiment will be done, because we don't care about the result as long as hashing == more rating! foodnote: (*) (depends on program, pool of players etc. but let's not get into that :)
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.