Author: John Merlino
Date: 13:21:34 12/31/02
Go up one level in this thread
On December 31, 2002 at 05:47:59, Uri Blass wrote: >On December 31, 2002 at 05:32:58, Dana Turnmire wrote: > >>I played a few Fischer Random games between CM9000 and Genius 7 and decided to >>test the "learning" feature of Genius 7. All fives games were started from the >>same position. Notice in the first game how Genius is crushed (mate was >>announced in 16 by CM9000 around move 29). >> >>In game two at move 22 Genius as black decides on 22...fxe5 instead of 22...Qe8 >>but is still crushed a few moves later. >> >>In game 3 Genius 7 tries 22...Ng5 although Chessmaster sees a pawn plus >>advantage several moves prior to move 22. >> >>In game 4 Genius 7 decides to wait until move 23 and to change from 23...Ke8 to >>23...Qe8 and once again gets crushed a few moves later. >> >>In game 5 Genius 7 waits until move 24 to change 24...Qf8 to 24...b5. Again >>same result a few moves later with a crushing defeat for Genius 7. >> >>It seems the crude attempt of Genius 7 to "learn" from its mistakes are >>laughable since its position has already been lost according to CM9000. It >>could continue to make minor move variations and would probably lose the next 20 >>games in a row doing this. The fischer random starting position basically >>forces both programs to play without an opening book and shows one reason why a >>good opening book is vital to a program. >> >>I understand Genius has a learning feature that is not tied to the opening book >>but attempts to vary it's moves when it starts having a losing score. Wouldn't >>it be much more effective to tie the learning feature in so that when a loss >>occurs the program refuses to play the exact variation of that particular >>opening in which the loss took place? >> >>I know that playing with opening books, games aren't duplicated like they would >>be with fischer random chess but in a long match that might be found at SSDF it >>seems CM9000 would be at a tremendous disadvantage if there was no mechanism to >>stop playing losing lines against other programs. The flip side of this little >>experiment is if a position was started and Genius won the game then Chessmaster >>would be on the losing side every single game. > >I think that there is a problem with the evaluation of Genius. > >Movei shows a clear negative score of more than one pawn from the first >iteration thanks to mobility evaluation. > >[D]r1nk2qr/3p1bpn/1ppPpp1p/P3P2P/1B1P4/3N1QP1/5P2/RB1KR3 b - - 0 22 > >I wonder if Genius7 evaluates mobility. > >Uri I agree with Uri. The problem is most likely in Genius' evaluation, or else it wouldn't consistently reach this losing position. And, yes, mobility is probably the culprit -- Genius only has 22 legal moves in this early midgame position! For reference, here is what CM9000, on a P3-733, says about this position: Time Depth Score Positions Moves 0:00 1/5 1.55 13569 22...Ng5 23.Qe2 f5 24.Nf4 Ne4 25.Bxe4 fxe4 26.Qxe4 0:00 1/5 1.49 29184 22...f5 23.Nf4 Ng5 24.Qb3 Qe8 0:00 1/6 1.55 41005 22...f5 23.Nf4 Ng5 24.Qe2 Ne4 25.Bxe4 fxe4 26.Qxe4 0:02 1/7 1.66 109301 22...f5 23.Nf4 Ng5 24.Qe2 bxa5 25.Rxa5 Rb8 0:04 1/8 2.02 243112 22...f5 23.axb6 Rxa1 24.b7 Bxh5 25.Qxh5 Rxb1+ 26.Kc2 Rxb4 27.Nxb4 Nf6 28.bxc8=Q+ Kxc8 29.exf6 gxf6 0:06 1/8 1.68 400170 22...Qe8 23.Nf4 Ng5 24.Qe2 c5 25.dxc5 fxe5 26.Qxe5 Bxh5+ 27.Kc1 Nf3 0:20 1/9 3.04 1419702 22...Qe8 23.axb6 Nxb6 24.Rxa8+ Nxa8 25.Nf4 Kc8 26.Qa3 Kb7 27.Bc5 Nb6 28.Bxb6 Kxb6 0:22 1/9 2.76 1576931 22...f5 23.axb6 Nxb6 24.Rxa8+ Nxa8 25.Nf4 Kc8 26.Qa3 Kb7 27.Bc5 Qb8 28.Bd3 0:52 1/9 2.36 3443611 22...b5 23.Ra3 Ng5 24.Qe2 Qe8 25.Nc5 Rb8 26.g4 f5 27.a6 0:56 1/9 1.74 3697768 22...Ng5 23.Qe2 f5 24.Nf4 bxa5 25.Bxa5+ Ke8 26.g4 fxg4 27.Qxg4 Nb6 28.Bd3 1:35 1/10 1.82 6331597 22...Ng5 23.Qh1 Qe8 24.g4 Bg8 25.axb6 Rxa1 26.b7 Rxb1+ 27.Ke2 Ra1 28.b8=Q Ra7 2:35 2/11 2.11 10629298 22...Ng5 23.Qh1 Qe8 24.g4 Nh7 25.axb6 Nxb6 26.Rxa8+ Nxa8 27.Nc5 Kc8 28.Bd2 f5 29.Bd3 fxg4 jm
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.