Author: Uri Blass
Date: 03:25:13 07/17/05
Go up one level in this thread
On July 17, 2005 at 06:15:23, Uri Blass wrote: >On July 17, 2005 at 05:22:47, Sandro Necchi wrote: > >>I have been laughing a lot (maybe crying on the ignorance would have been more >>appropriate?)reading many wrong statements about testing and Elo lists. >> >>so, for those who are new and do not know, SSDF list is the best for the >>following reasons: >> >>1. They use 2 computers and the program complete with own book and ETG, with own >>gui and best setting as suggested by the programmer. >>2. They use long time controls (40/2h 20/1h; international level) only. >>3. They use the same hardware for all programs. >>4. They use a very wide range of programs and not only the new ones to get more >>reliable results. >>5. Ponder on and learning are activated. >> >>Now, even if some people do not agree, the use of own book is the best because >>that book has been developed specifically for that engine and in some cases the >>engine has been developed specifically on that book too. > >I think that in order to prove it you need to test programs also with different >book. > >I do not know of these type of testing > >I do not know if Shredder with your book perform better than Shredder with >Fritz8.ctg >I do not know if Fritz with Fritz8.ctg performs better than Fritz with your >book. > > > This means that the use >>of a different book and the same for all programs would damage or favor some >>programs over others. > > >I think that a lot of people are interested in a different question. >The question which program is stronger in ssdf match is one question and the >question which program to analyze positions is a different question. > >using books and learning in tests does not help to decide which program is >better to use in overnight analysis. > >I prefer the CEGT rating list and not the ssdf rating list in decision which >programs to use in correspondence games. > >The only advantage of the ssdf is longer time control but the CEGT has other >advantages like having more programs(Fruit was not tested in the ssdf list). > >Uri I can add that book learning also cause problems to get statistical analysis of the results because games are not independent events. I hope that games in the CEGT are independent events(I am not sure because it is possible that some programs still earn there from positional learning). When games are not independent events because of learning then it is not clear what is the standard error in the result and not having learning from previous games can produce smaller error with the same number of games. imagine that there is program A and program B when B is deterministic without learning. If A is lucky to win the first 2 games it may win a SSDF match 40-0 by repeating the first 2 games If A is unlucky then the result may be 21-19 for B because B may win the first 2 games and draw games 3 and 4 when A will learn that B is probably stronger than it so it will repeat the draw lines again and again in the rest of the games. Of course this is extreme condition that does not happen but the point is that I have no idea how to analyze possible error in the ssdf list because of the noise of learning. 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.