Author: Uri Blass
Date: 16:46:15 06/11/02
Go up one level in this thread
On June 11, 2002 at 19:22:01, Uri Blass wrote: >On June 11, 2002 at 18:38:02, Slater Wold wrote: > >>Uri, you asked for some games, so here it is. >> >>Sjeng 12.13 (128MB hash using a Shredder opening book - 1.73Ghz) >>vs >>Movei (32MB hash using it's own random openings - 1.73Ghz) >> >>30 minutes per game. >> >>Sjeng won +14 =0 -1 > >Thanks for the games. >I am not surprised by the result against sjeng. > >I know that Movei is clearly weaker than sjeng. >I guess that the level of it is in the middle between Tscp and Sjeng. > >some reasons(I think that reason 4 is the main reason): >1)Movei has no real opening book except first move. >2)Movei does not ponder(you mentioned it). >3)Movei does not use hash tables in an effective way(it is using it only for >better order of moves and not to prune the tree and even that use is by the >always replace scheme). >4)Movei has inferior evaluation relative to sjeng(for example it still knows >almost nothing about pawn structure and the only thing that it knows about it is >that more than one pawn on the same file is bad) I forgot reason 5 that says that sjeng is better on tactics than Movei even without pondering. It is possible that a good example the first game when I suspect that Movei blundered by Rxh2(I am not sure if there is a move to save the game at that point and if there is no move to save the game it is a positional win of Sjeng because the score of sjeng jumped only after Rxh2. Here is the relevant position: [D]2r2k1r/pppb2q1/2np1n2/6N1/2Q2p2/1B6/PPP2PPB/R3R1K1 b - - 0 21 am Rxh2 > >I already have a later version that I believe that it is better but it is still >not close to sjeng's level. > >If people want to test the latest version they can send me an email. >It is possible to send email by clicking on my name but in that case They will >need to wait at least 7-8 hours for a reply relative to the hour of this post. > >I may reply emails faster(but not promise) if people send email to >uuu0123@hotmail.co.il(I can get there emails everywhere and not only at the >office). > >Note that I am surprised that Movei did not repeat the same loss twice in games >1 and 5 because it has no learning function except learning to change the first >move in order to prevent losing the same game twice in matches of 4 games but I >see that for some reason it played a different move at move 3 with slightly >different score at the same depth. > >I do not know the reason and I can only guess that maybe it initialized the hash >tables in different random numbers and it caused different order of moves that >cause different pruning. > >I also did few tests with the latest version of Movei under winboard and it lost >35-5 against amyan(no book) in 2 minutes per game. > >I did not check the hash tables of amyan for that match and it used the default >hash tables when Movei used 16 Mbytes hash. > >That version of Movei always changes the choice of the first move(not only twice >like the public version) so there were no double games. To be more clear the public version only learn not to repeat the last game first move but it always repeat the first move of the game that was before the last game. The latest version has an option only to play the first same move if you give it 21 games with the same color so it has no choice because there are 20 legal moves. 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.