Author: Stuart Cracraft
Date: 19:32:31 01/29/06
Go up one level in this thread
On January 29, 2006 at 21:40:18, Robert Hyatt wrote: >On January 29, 2006 at 19:10:52, Stuart Cracraft wrote: > >>On January 29, 2006 at 10:45:28, Robert Hyatt wrote: >> >>>On January 29, 2006 at 09:52:57, Kurt Utzinger wrote: >>> >>>>On January 29, 2006 at 09:38:56, Majd Al-Ansari wrote: >>>> >>>>>I have a completely different result. I see quite a big difference and many >>>>>"won" games are now won instead of drawn. >>>> >>>> Please be good enough to present those results >>>> here. This would be of great interest. >>>> Kurt >>>> >>>>I have checked out quite a few games >>>>>and I will say that EGTB's greately improve endgame play for Rybka, and plug a >>>>>lot of holes. >>>> >>>> This is contrary to long experience with other engines >>>> where you can almost see no difference regarding overall >>>> score after playing some hundred games. >>>> Kurt >>>> >>>> EGTB are especially important if the other side has them. Not >>>>>having them will leak a lot of points. Still there is quite a ways to go for >>>>>Rybka when endgames are concerned. It still plays some endings horribly. But >>>>>the gaps are getting smaller and smaller and I am very interested to see how >>>>>Rybka will be with beta 14 (EG knowledge added). >>> >>> >>>One note: >>> >>>Playing EGTB vs NoEGTB to see if EGTB helps is probably the wrong way to measure >>>the experiment. It is more useful to take a known good program _with_ EGTBs, >>>and play your favorite engine against it, with your engine not using 'em, then >>>playing again with 'em. If the opponent doesn't have 'em, then your not having >>>them might not expose the problem as well as making sure your opponent can >>>always win those tricky cases and you now have to rely only on your eval to hang >>>on... >> >>Also - I think the sample size has to be MUCH larger than what we are talking. >> >>A few hundred games is not going to do it. >> >>Start with a few thousand for the EGTB and work your way up from there. >> >>Stuart > > >The general rule-of-thumb is that the closer two programs are in skill, the more >games you need to really see whether one is better or not.... Yes - I'd be interested to know of any papers that have been done to quanity the magnitude of the difference and the magnitude of games both. Stuart
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.