I remember once playing a pawn sacrifice on the White side of a KID that might have turned into a double pawn sacrifice (if my opponent had taken a second pawn I was offering).  During the game I wasn't sure whether the double sacrifice was sound, but felt that it was what I "had" to play, anyway.

Later, I looked at the double pawn sac line with my little version of Fritz, and he gave the evaluation as completely even: -0.06.  In my opinion, if Fritz doesn't think you're worse when behind in material, chances are, you're actually a LOT better.  After helping Fritz along another 5 moves or so, he could find no defence for Black to my attack.  So basically I agree with Taljechin's comment.

HgMan wrote on 05/26/07 at 14:18:38:


Must have something to do with the horizon effect. The disadvantage of the silicon move only becomes clear after more moves your computer can handle. I have noticed exactly the same in several gambit positions.

With the Nimzo, the Semi-Tarrasch, and the Caro-Kann, I defend against the IQP with some frequency.

But I'm clearly not asking my question clearly.  Rules or no rules, programming or no programming, why is it that my computer wants to play an objectively inferior move?  It crunches more ply than I care to count, but only realizes its mistake after entering an inferior endgame.  And this in a fairly open position.

Well the issue with programming positional understanding is that a lot of positional rules contradict each other and it takes an experienced player to understand when each "rule" applies.

GMTonyKosten wrote on 06/06/06 at 09:06:01:


Both HIARCS and Rybka (3100+) have surpassed 3000 ELO on quad core setups. However, this is only based on competition with other engines, such as Fritz, Shredder, Junior, Zap!Chess, Glaurung, Fruit, Crafty, etc. Still, this is very impressive.

In positional chess, they are the only engines to consider. HIARCS plays more actively and sometimes has more accurate positional insights, but overall, Rybka is about 100 ELO stronger. It is the clear leader. Why? Vasik Rajlich, the programmer of Rybka, is a fairly strong IM (and so is his wife) with an excellent programming background. From what I understand, this combination is what makes Rybka so successful. It has made possible to have more or at the very least more accurate chess knowledge than other programs. Moreover, he is constantly working on improving this aspect and always solicits feedback and bug reports on the Rybka forums. Just recently, IM Larry Kaufman has worked with Rajlich to make Rybka better understand when to exchange pieces and how to handle pawn majorities in the endgame. This work resulted in a 50 ELO point increase and will be available to customers this month in the form of Rybka 2.3.2 (the last version before 3.0). Given all of this, I am confident that both Rybka and HIARCS usually have the positional understanding of at least a 2000 ELO human.

BTW: Rajlich did say that in some cases when he added more positional knowledge, Rybka became weaker. He mentioned that while positional knowledge is the way forward, a lot of what we take for granted may not be accurate in a large number of cases.

I noticed that most chess players have absolutely no understanding of IQP positions, which is why, after reading winning pawn structures, I am always happy to go into one.

or maybe the computers know something we dont