Author: Kurt Utzinger
Date: 09:03:24 12/20/03
Go up one level in this thread
On December 20, 2003 at 11:44:54, Uri Blass wrote: >On December 20, 2003 at 11:21:26, Kurt Utzinger wrote: > >>On December 20, 2003 at 11:03:04, scott farrell wrote: >> >>>On December 20, 2003 at 10:53:56, Jay Hysenbeg wrote: >>> >>>>hi, >>>>i was wondering what happens when u use low hash for a long time control. to the >>>>point where the hash is normaly filled before the eng has made its move. >>>>thank you >>>>jay h. >>> >>>well, things start to deteroriate, your move ordering isnt as good etc., and >>>your branching factor ends up so bad that it cant get from say 10 ply to 11 ply >>>or whatever, it sort of like hits a brick wall somewhere. >>> >>>To what extent this happends largely depends on your replacement scheme, and how >>>well it deals with overwriting entries. >>> >>>I test changes to my hash by running extremely small hashes, like 16k entries, >>>just to see how it will cater. >>> >>>Scott >> >> Size of hash tables seem to have much less influence >> than I have imagined. This showed a test over 50 games >> between Fritz 8 [96 MB hash] against Junior 8 [8 MB hash] >> and a second 50 games match Fritz 8 [8 MB hash] vs Junior 8 >> [96 MB hash]. The final result was in both matches almost >> identical [25m+10s]. >> Kurt > >It is possible that 8 MB is already almost enough for your time control. Maybe, I have no idea. And furthermore: 2 x 50 games are most probably not enough to prove something. Kurt > >It is possible that 96 MB is only 10 elo better than 8 MB at the time control >that you tested but if you use 12 MB against 1 MB you get bigger >difference of 25 elo. > >I do not expect more than 25 elo based on upper bound of 7 elo for doubling the >hash tables. > >Even if the difference is 24 elo then it means that you can expect only >26.5-23.5 for the stronger side in 50 games and the statistical error in a match >of 50 games is more than 1.5 point. > >Uri
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.