Stockfish - Initial position, MultiPV Mode=20, Depth=50

[MikeGL]

If white is not careful, other first 3 moves of black can also transpose into
that +0.17 position. But +0.17 means nothing too because I have seen
endgames worth +2.00 for SF but was actually a draw.

[cdani]

Thanks. Is always interesting to see such big analysys.

But from time to time I feel the need to say that Stockfish is very far from perfection. For example many times wining an endgame requires to do non standard things, so general tuning will prevent the engine to see such win, specially from a very previous position.

[Henk]

Before move 1: +0.20 for white
Depth 50: +0.17 for white

SF is holding onto this 'tempo' bonus, but that's it...there's no other improvement of position.
If chess (in theory) is a win for white, I would expect white score to slowly increase.

this is so imperfect: so many moves pruned, so many nodes evaluated incorrectly.

for correctly assessing the start position, as there are no forced lines, one needs at least an evaluation close to perfection, and SF eval is admittedly very poor.

would not make very much sense to repeat the same exercise with SF before 2045.

If you disable speculative pruning like LMR and use a good evaluation function, result might be interesting even on a shallow depth.

[Nay Lin Tun]

Stockfish evaluations are not always perfect but can still be assumed as one of the best evaluation and her overall evaluations are not inferior to the other top engines. Some engines may evaluate better than stockfish in certain positions. For example, Stockfish will wrongly evaluate a Rook ending with one extra pawn( KRPPP vs KRPP in same side) as 0.65+ while Gull 3.0 will evaluate correctly as 0.23+.

[Joerg Oster]

Before move 1: +0.20 for white
Depth 50: +0.17 for white

SF is holding onto this 'tempo' bonus, but that's it...there's no other improvement of position.
If chess (in theory) is a win for white, I would expect white score to slowly increase.

this is so imperfect: so many moves pruned, so many nodes evaluated incorrectly.

for correctly assessing the start position, as there are no forced lines, one needs at least an evaluation close to perfection, and SF eval is admittedly very poor.

would not make very much sense to repeat the same exercise with SF before 2045.

Here I have to agree.
But I don't think this is mainly a problem of correct evaluation.

For example, let's have a look at those 0.00 scores.

Code: Select all

50/80  200:44:26  9.890.822.131k  18.586k   0,00  1.b3 d5 2.Sf3 Lf5 3.Lb2 e6 4.d3 h6 5.g3 c6 6.Sbd2 Sf6 7.Lg2 Le7 8.0-0 0-0 9.Te1 Sbd7 10.h3 Lg6 11.e4 dxe4 12.dxe4 Dc7 13.De2 a5 14.a3 Tfd8 15.Sh4 Lh7 16.Shf3

50/80  200:44:26  9.890.822.131k  18.586k   0,00  1.g3 c5 2.e4 d5 3.exd5 Dxd5 4.Sf3 Sc6 5.d3 Sd4 6.Lg2 Lh3 7.Lxh3 Sxf3+ 8.Kf1 Sd2+ 9.Kg1 Sf3+ 10.Kf1

50/80  200:44:26  9.890.822.131k  18.586k   0,00  1.e3 d5 2.d4 Sf6 3.Sf3 e6 4.b3 b6 5.c4 Ld6 6.La3 dxc4 7.Lxd6 cxd6 8.bxc4 Sc6 9.Le2 Dd7 10.0-0 0-0 11.Sfd2 La6 12.Sb3 Tab8 13.S1d2 Tfc8 14.h3 h6 15.Te1 Lb7 16.Db1 Se7 17.a4 a5 18.Tc1 Lc6 19.Da2 Lb7 20.Db1

50/80  200:44:26  9.890.822.131k  18.586k   0,00  1.c3 Sf6 2.d4 d5 3.Lg5 c5 4.e3 e6 5.Sf3 Sbd7 6.Lf4 Sh5 7.Le5 Shf6 8.Lf4

50/80  200:44:26  9.890.822.131k  18.586k   0,00  1.h3 e5 2.e3 d5 3.d4 Sc6 4.dxe5 Sxe5 5.Sf3 Sxf3+ 6.Dxf3 Sf6 7.Ld3 c6 8.0-0 Ld6 9.Sc3 Lc7 10.Te1 Dd6 11.Df4 De7 12.Dh4 Dd6 13.Df4

Nominal search depth is 50, and we only get 10 moves for 1. g3. and only 8 moves for 1. c3 ?
How reliable is that?

Sure I know that this repetition pruning gains a lot of Elo, but it also makes these kind of calculations almost meaningless.

Another problem is that Stockfish doesn't give the correct search numbers like nodes and seldepth for the different root moves. It simply shows the same numbers for ALL moves.
I don't know how other engines handle this, but this looks like a very basic solution.

[corres]

[quote="Lyudmil Tsvetkov"]

it might be the case that 1.c4 is winning altogether, at least in some 90% of lines.

[/quote]

You are a 1.c4 fan, as we know.
Nowadays 1.c4 is not a modish opening and this is the only one advantage for it.

[Henk]

it might be the case that 1.c4 is winning altogether, at least in some 90% of lines.

You are a 1.c4 fan, as we know.
Nowadays 1.c4 is not a modish opening and this is the only one advantage for it.

I think I am going to play 1.c3 instead of 1. d3. 1.c4 leaves a pawn unprotected. So it forces a move to protect this pawn. Therefor you loose flexibility. Same holds for 1.e4

Also weakening very important square d4.

Knights on c3, c6 are always badly placed in my games.

[corres]

[quote="kranium"]

Interesting...if you subtract bonus for side-to-move (Tempo = 20)...
1. Nf3 score drops to -0.03

Does this mean that SF sees chess as a draw?, ie w/ little or no advantage for either side after a very deep search.

[/quote]

This means that search function of Stockfish is without any notable bug.

[corres]

[quote="kranium"]

If chess (in theory) is a win for white, I would expect white score to slowly increase.

[/quote]

But there is no such theory.
In practice white has a minimally greater advantage to win the game but this is only a tendency and it is not a general rule.

[Henk]

it might be the case that 1.c4 is winning altogether, at least in some 90% of lines.

You are a 1.c4 fan, as we know.
Nowadays 1.c4 is not a modish opening and this is the only one advantage for it.

I think I am going to play 1.c3 instead of 1. d3. 1.c4 leaves a pawn unprotected. So it forces a move to protect this pawn. Therefor you loose flexibility. Same holds for 1.e4

Also weakening very important square d4.

Knights on c3, c6 are always badly placed in my games.

Unless d5, d4 is a weak square of opponent.