(FEN: 2b3k1/7p/2Qp2pb/3Pp3/2N1P3/5N1P/5KP1/2q5 b - - 1 46)
In the diagram above, we have the position from Anand-Carlsen, Bilbao 2010 after Anand’s 46.Kf2. Carlsen now played the very accurate 46…Bd7!, which essentially forces a draw after 47.Qxd7 Qxc4, as White’s king is too exposed for White to have time to grab the d6- pawn safely.
But on ICC, there was a little bit of a debate over the merits of 46…Qc2+. At first, it seemed a few people assessed 47.Ncd2 Qxc6 48.dxc6 as better for White. Drahacik (an FM, I don’t know his real name) then kibitzed, saying that the endgame was better for Black.
I jumped pretty quickly at that comment, as I didn’t think it was better for Black. He clarified to say that it was Rybka’s assessment, not necessarily his own, that he had posted. In any case, GM Tomi Nyback (a pretty strong GM who hovers around the mid-2600s FIDE) first said joined the discussion by saying White was better in that endgame, but after Drahacik’s comment, he changed tack and said Black was better as well!
I was shocked – how could the position above be assessed as better for Black? A draw seemed the most likely result, but I couldn’t believe that White was in any danger. Nyback, though, said that …Kf7-e7-d8-c7 comes pretty quickly.
That was a bit mysterious, as that’s 4 moves, and unless I fall asleep at the wheel, my four moves are not going to be Kg1-h1-g1-h1. Instead, I’m going to put my knight on d5, either via c4-b6 or b1-c3.
The endgame is actually drawn after something like 48…Kf7 49.Nc4 Ke7 50.Nb6 Kd8 (not 50…Ba6?? 51.c7 and wins) 51.Nd5 Ba6 (not 51…Bg7 52.Ng5 and wins), and now White can draw with any of 52.Nf6, 52.Nh2, 52.h4, or 52.g4. White has a pretty healthy margin for error in this endgame, but Black is not in much danger either.
Nyback mentioned all these ideas, but seemed adamant at first that the endgame was actually more pleasant for Black. After some more discussion, he seemed to changed his mind. Still, it was a bit surprising to see how quickly he changed his mind after seeing the computer’s evaluation.
[Lest you think Nyback is some sort of automaton, he can play some strong chess – in the 2008 Olympiad, he crushed a 2786-rated Magnus Carlsen with a piece sacrifice that the engines completely overlook and incorrectly evaluate.]
To make my point clearer, take a look at the position below. There’s not much going on – 3 pawns apiece, knight versus bishop. Who do you think is better?
(FEN: 6k1/5p2/5bp1/7p/7P/5NP1/5P2/6K1 w - - 0 3)
How do the engines assess this position? I don’t have Rybka 3 or Rybka 4, so I don’t know if this has been addressed, but Rybka 2.3.2 settles in around an advantage of 0.10 for Black at depth 24. FireBird 1.2, the engine I’ve used for the past few months, likes Black by 0.06 at the same depth. Deep Shredder 10, an engine I used around 2005 or 2006 I think, is about as wrong as Rybka, settling in at -0.11.
I’d like to think I’m right in saying that Black is not even symbolically better. With all the pawns on one-side of the board, the bishop isn’t particularly great. In fact, bishop and 3 pawns versus knight and two pawns is generally a draw.
But all 3 engines like Black, and I’m pretty sure it’s not because of anything actually going on in the position – it’s because Black has a bishop. As an experiment, simply switch the knight and bishop for each (black knight on f6, white bishop on f3). At depth 24, my Rybka gives White a 0.17 advantage.
It’s ok, everybody – including chess engines – make mistakes.
Going back to the white knight against black bishop bishop, add a pawn for each side on e2 and e7. The evaluation drops a bit for Rybka (down to -0.04 at the same depth). After the moves 1.e3 e6 2.e4 e5, though, the evaluation goes up to -0.12!. What’s changed in the position? Nothing that would help Black, as far as my human eyes can tell. My preferred engine, FireBird, laughs at Rybka in this instance, showing 0s (0.00, dead equal) in all those 4 on 4 positions.
It’s somewhat important to understand your favorite engines flaws, because otherwise you end being susceptible to a specific “bug” in the engine. Let’s say you’re preparing an opening line for somebody and you turn on the engine. It gives you 3 or 4 lines, all of which look reasonable. However, the top choice is ranked 0.10 ahead – that’s not a sizable difference in the evaluation at the board, but with an engine, those 0.10 differences are magnified. You decide that this is the line you want to play. Then you to go the board, blindly follow it, and realize at some point, “Hmm, I really don’t have anything here – why was this line the best again?”
[Addendum: I wrote this blog on Sunday, but waited until today to post it. However, on Monday night, in the USCL, GMs Akobian and Friedel squared off. Friedel played an opening line that I don’t believe in, reaching the following position:
(FEN: r1b2rk1/pp3p2/4p2p/3p2q1/1PP1n1p1/P3P1P1/2Q2PPN/3RKB1R w K - 2 18)
Friedel had just played 18…Qg5, guarding the g4-pawn. This seemed to be part of both players’ preparation, as both had gained about 4 minutes worth of time with the 30 second increment up to now. Arun Sharma, watching the game online, told me that Friedel had looked at this position before the game and decided it was equal. He also said that the engine “confirmed” the position was equal.
Unfortunately, I don’t think that’s the case at all. Black will be saddled with an IQP on d5 and an overextended kingside structure. His best move would be to play …g4-g7, but that’s not how it works. Black isn’t particularly well developed either. On the other side of the ledger, the Nh2 is misplaced, but I don’t think it’s as much of a long-term liability.
After something like 18.Bd3 Nf6 19.cxd5 exd5 20.0-0 Bd7 21.Qc7 (21.Rfe1 Rfc8 22.Qb2 would transpose to the actual game, which Akobian went on to win), it’s hard for me to believe Black is alright. Engines tend to evaluate IQP positions quite favorably, and so even though the evaluation hovers around equality here, I don’t think that’s actually the right evaluation. Of course, maybe this time I’m wrong …]