I'm referring to the French Defense (C11 ECO code) game where Alpha Zero played 30. Bxg6 ! , sacrificing a bishop for a better position. After Alpha Zero's move 30. , I couldn't find any good moves for Black. However, after the earlier move by White 28. Rc2, Houdini suggests after many hours the reply 28... Qe7 instead of 28... Kd7 which was played by Stockfish. And if 29. Ng5 , then 29... Bh6. All in all, analysis with the Houdini chess engine strongly suggests that 28... Qe7 is a much better move than 28... Kd7 for Black. Am I missing anything?

The position after 28. Rc2 is:

[StartFlipped "0"]
[FEN "3kqb1r/1b3p2/1p2p1p1/pP1pP3/P2P1PQP/3BKN2/2R5/8 b - - 3 28"]

Added: Dec 18, 9am
I'm doing an infinite analysis on 4 threads using Houdini 6, with 2 Gb of hash memory. After 4 hours and a depth ~= 33, Houdini still prefers 28. ... Kd7 for Black, with an evaluation of +0.15 .

Eventually, Houdini 6 had 28. ... Ba3 as its first choice for Black. Moving the Bishop from the f8 square to the a3 square allows the Black Queen and rook to connect on the back rank. It also allows moving the Queen to the f8 or g8 squares...

Perhaps Alpha Zero's play is too deep, some of the time, to get reasonnably accurate evaluations in a few hours. Or, to simulate Alpha Zero's play could take an inordinate amount of time per move...

P.S. I upgraded to Houdini 6 today, and at 4 hours of infinite analysis with it, it's a lot less than the time spent analyzing before the upgrade.

  • 2
    Can someone post a diagram? – SmallChess Dec 17 '17 at 13:06
  • My feeling is that both Kd7 and Qe7 are lost by best play. You might have to go earlier in the game. – SmallChess Dec 18 '17 at 14:19
  • I'd suggest actually entering one of the candidates (Ba3,Qe7, etc) and see how the eval changes then. Again this is a very uncomfortable position to play as black even for an engine. :) – Sint Dec 19 '17 at 8:33
  • One of the most "promising" continuations for Black is: 28. ...Ba3 29. Ng5 Qg8 . What's unusual is that the evaluations keep going up and down as the search depth increases. – David Bernier Dec 19 '17 at 11:47
  • I didn't find any drawing moves for Black in the line 28. ... Ba3. I'm now exploring the other candidate: 28. ... Qe7 . – David Bernier Dec 19 '17 at 21:44

It is quite possible that Qe7 holds black's position better than Kd7.

Qe7 at least gets white to pay attention to a4 pawn.

That said, black is still worse, as that poor bishop on b7 is boxed in forever.

So A0 would be able to evaluate when it is worth giving up a4 pawn and be able to press on for many moves.

At best black(Stockfish) grovels on for a draw.

  • I think Qe7 is better for Black than Kd7. 28... Qe7 prolongs the game and either results in a win for White or a draw. I would say that under optimal play, the position after Qe7 is too complicated to call as a White win or a draw. – David Bernier Dec 25 '17 at 21:17

If you analysed for hours on end with Houdini, chances are its recommendation is stronger than what Stockfish played. There are two main reasons:

1) Stockfish didn't have hours to think during the game with AlphaZero.

2) Stockfish isn't especially renowned for how well it evaluates positions. Its main strength is how deep it can look ahead in such a short time. Larry Kaufman himself said that Komodo would be superior to Stockfish if it weren't for Stockfish's speed (I'm paraphrasing). Since the diagram you posted is very positional and involves quiet maneuvers, Stockfish must rely on its evaluation abilities more. In this domain, Houdini could very well be equal (or possibly better).

You're right though that AlphaZero's play is too deep for a normal engine like Stockfish/Houdini to fully understand. Through machine learning, AlphaZero developed an extremely intelligent intuition, which cannot usually be matched by brute-force calculation (especially in such a strategic position that you posted).

Your Answer

By clicking "Post Your Answer", you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.