Page 1 of 4

Stockfish misevaluations:

Posted: Mon Apr 22, 2019 1:11 pm
by Henk
I encountered this position in a real human-human game. Me playing with white.

Stockfish 10 evaluates this as +3.30

[d] 4b3/5k2/p5p1/P2p1p1p/1PpPrP1P/2P2QP1/5K2/8 w - - 6 8

Stupid engines

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 2:37 pm
by hgm
Well, what can you expect from an engine that is only good on average?

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:06 pm
by mar
And what's wrong with that? White will eventually sac a pawn, break free and win the game :shock:
It will take some time with 50 move rule, sure, but white will win.

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:15 pm
by hgm
With the Bishop on b5 the only Pawn you can sac is the g-Pawn, and after fxg4 black will simply put his King on f5, and white has nothing. Actually I would not be surprised if it was lost for white after such a sac.

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:20 pm
by Henk
mar wrote:
Mon Apr 22, 2019 3:06 pm
And what's wrong with that? White will eventually sac a pawn, break free and win the game :shock:
It will take some time with 50 move rule, sure, but white will win.
How? I don't see it. Black can block everything. You mean sacrifice the b pawn. But then Bb5, Rc6 and Ke8 and Queen can't pass.

By the way Komodo 9: +1.82

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:24 pm
by Henk
hgm wrote:
Mon Apr 22, 2019 2:37 pm
Well, what can you expect from an engine that is only good on average?
Maybe best is to evaluate a position by playing some games and taking the average score. I don't know. Looks like alpha beta fails.
Zero method would do better.

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:43 pm
by mar
hgm wrote:
Mon Apr 22, 2019 3:15 pm
With the Bishop on b5 the only Pawn you can sac is the g-Pawn, and after fxg4 black will simply put his King on f5, and white has nothing. Actually I would not be surprised if it was lost for white after such a sac.
Yes, maybe I was too optimistic and this is indeed a draw with perfect play.
Is there actually any (non-buggy) engine to score this fortress as a dead draw?
This doesn't seem like a typical pawn fortress due to the open e file.
I just tried Critter which I know had some fortress detection code and it still scores this as +2.6 for white.

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 3:46 pm
by mar
Henk wrote:
Mon Apr 22, 2019 3:24 pm
Maybe best is to evaluate a position by playing some games and taking the average score. I don't know. Looks like alpha beta fails.
Zero method would do better.
Maybe a random mover would score this position properly as 0.0

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 4:31 pm
by Raphexon
Latest dev version of SF thinks it's winning. :oops:
+5.7 at depth 53/60
Still +5.38 at 62/73
Still +5.38 at 70.

I'll let it run for a few hours, maybe it will reach 127...

Re: Stockfish misevaluations:

Posted: Mon Apr 22, 2019 5:20 pm
by PK
Perhaps it is a draw, perhaps a win, but it is white to move. Does any engine try immediate b5?