Author: Stephen Ham
Date: 08:50:10 12/07/05
Dear readers, Although I'm a fairly strong chess player, I'm a computer dummy, and thus am usually too embarrassed to post here. Nonetheless, it's just possible that I can add some value with a post or two. Here's one that may be of some value. After downloading the 32-bit beta version of Rybka, I immediately started a tournament at rather long time controls (i.e. 3hours/40, etc.) with Shredder 9, Toga II, Fruit 2.2.1, and Junior 9. Fruit played with an improved Fritz 7 book, Toga played with a Shredder 7 book that I've modified to play only the opening variants I play in correspondence chess, including my prepared TNs. I gave an improved Hiarcs 9 book to Rybka, since its default setting is "positional." When I write, "improved" that means I've added data and altered the move selections to some extent to correspond with the opening theory I think is "objectively" best, while still leaving the general pre-set openings about 99% intact. The result was that Toga II won (scoring 100%!), Rybka scored 50% and Junior 9 came last. The results of too few games mean very little. Toga's victories seemed somewhat attributable to its superior book. It got excellent positions after the openings. Still, it's apparent that Rybka is a very strong engine and in the class of these other great engines. So, I examined Rybka's games in some detail. I'm impressed by its apparent positional style and its human-like move selection. Unlike the calculating monsters, I sense that Rybka "knows" what it's doing. However, when I examined Rybka's move selection in tactically sharp positions in post-game analyses, I was disappointed to find that Rybka didn't allow me to analyze in multiple variant format. Also, it won't allow me to use the "Deep Position Analysis" for more than one line. Do others have this problem? If so, I'll speculate that this will be fixed with the commercial version. Again, I had too few games to analyze, but I'm very happy with Rybka. Yes, it seems to consider very few ply and nodes per unit of time, relative to other engines. Nonetheless, it doesn't seem to be adversely affected. Instead, in some sharp positions I've tested, its judgment of the positions were sound. Still, I need more data. So, prior to going to bed last night, I set up a very long time-control game where Rybka was Black versus Junior 9. Junior 9, with its enterprising and risky style, took a pawn that I thought was poisoned. Personally, I never would have even considered that capture, since it looked totally unsound (the capturing Bishop was pinned to Junior's Queen by Rybka's Rook). Nonetheless, I was wrong apparently, since Junior evaluated the capture as giving it a slight edge while Rybka claimed it was about equal! When I checked this morning, the game was about 35-moves deep and Junior was still a pawn ahead. Both engines evaluated White to have a clear advantage. Again, this is too little data. But Junior 9, an engine that normally finishes last or near the bottom of my long time control tournaments with Shredder, Toga, and Fruit (I think some of this poor performance is due to an inferior book, which I'm gradually improving with each game), seems to have evaluated the complex positions resulting from its pawn grab better than Rybka in this game. Tonight, I'll stop the game if it's still going, and try to examine Rybka's move selection in more detail. Still, I'm frustrated by the inability to see it perform in multi-variant and deep position analysis mode. My interest to date has been to see how Rybka evaluates opening theory and how it plays the middle game. So I've given minimal study of its endgame ability. I've read here of several instances where Rybka failed to play simple endgames properly and find mates. I found no such examples in my games, although I did have the perception that its endgame technique was inferior to the other engines. So while no solitary endgame blunders stood out, there did seem to be some slow erosion of Rybka's performance in endgames. I think Rybka is already a very strong engine in beta format and given the feedback here and elsewhere, I expect its author will improve its strength with the commercial version. All the best, Steve
This page took 0.01 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.