stockfish evaluates knight as more than rook

[Uri Blass]

I talk about evaluation and not about search so I let stockfish to search to depth 1 and here are the results(rooks equal near 4.5 pawns when knights equal near 12 pawns).

I wonder if the reason is not using nnue with a big material advantage.


[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBN1 w Qkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 142 142k -4.46 e2-e3

[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/1NBQKBNR w Kkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 143 143k -4.61 e2-e3


[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/R1BQKBNR w KQkq - 0 1
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 97 97k -12.11 Ng1-f3

[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 120 60k -11.80 d2-d4

[abgursu]

I talk about evaluation and not about search so I let stockfish to search to depth 1 and here are the results(rooks equal near 4.5 pawns when knights equal near 12 pawns).

I wonder if the reason is not using nnue with a big material advantage.


[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBN1 w Qkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 142 142k -4.46 e2-e3

[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/1NBQKBNR w Kkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 143 143k -4.61 e2-e3


[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/R1BQKBNR w KQkq - 0 1
Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 97 97k -12.11 Ng1-f3

[d]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1

Stockfish_21071317_x64_avx2:
NNUE evaluation using nn-9e3c6298299a.nnue enabled
1/1 00:00 120 60k -11.80 d2-d4

It seems pretty normal to me. I would prefer playing without a rook too.
You see, in traditional engines, we used to give some value to each piece and than force engine to evaluate that position with given values. What's different in NN's?
The answer is NN's evalutions are based on a lot of values. So a good net should be able to realize that without a knight, you can't even counterdevelop your pieces and would lose the game in the opening and give the values from here. But without a rook, you can develop well and face the troubles in midgame, where no one knows what's gonna happen. Main power of rook is that they're usually unstopable in endgames.
Plus, speaking from the point of NN, there are still 16 pawns in the game, so probably in any position that any exchange haven't maden and isn't available you will get results like this one. (And should.)
Best Regards...

[ChickenLogic]

Just use https://hxim.github.io/Stockfish-Evaluation-Guide/ and go to 'NNUE' then you can see what the net scores each piece in which position.

[Uri Blass]

Just use https://hxim.github.io/Stockfish-Evaluation-Guide/ and go to 'NNUE' then you can see what the net scores each piece in which position.

It seems wrong.

I see that the value of the knight is 6.2 or 6.1 but I saw that the evaluation after one ply of search was near 12 pawns against white and it is not logical that a single move by white reduce the evaluation by 6 pawns.

edit maybe this 6.2 is 12 pawns by the output of stockfish for some reason.

I see NNUE = -1272 (-6.12)

[Uri Blass]

If I remove knight b1 NNUE gives me -1272(-6.12)
If I remove both knight b1 and rook a1 NNUE gives me -1271(-6.11)

I used the NNUE that latest stockfish developement use.
Stupid NNUE and it is better if the stockfish stop to use it.

[ChickenLogic]

If it's so bad then beat it. Should be easy.

[Sopel]

If I remove knight b1 NNUE gives me -1272(-6.12)
If I remove both knight b1 and rook a1 NNUE gives me -1271(-6.11)

I used the NNUE that latest stockfish developement use.
Stupid NNUE and it is better if the stockfish stop to use it.

Both evals are about 100% loss chance according to stockfish, so what's your problem? Apart from the fact that you apparently don't like NNUE for some reason and you're trying find any possible reason to hate on it.

[Uri Blass]

If I remove knight b1 NNUE gives me -1272(-6.12)
If I remove both knight b1 and rook a1 NNUE gives me -1271(-6.11)

I used the NNUE that latest stockfish developement use.
Stupid NNUE and it is better if the stockfish stop to use it.

Both evals are about 100% loss chance according to stockfish, so what's your problem? Apart from the fact that you apparently don't like NNUE for some reason and you're trying find any possible reason to hate on it.

The problem is that there is a difference between different losses.
Being a knight down is better than being knight and a rook down.


I believe that this type of behaviour make stockfish a relatively weak program in beating weak chess engines with a big material handicap.
I also guess that stockfish is relatively weaker in finding mates and I find that old versions of stockfish without stupid evaluation find mates faster.

[Joerg Oster]

If I remove knight b1 NNUE gives me -1272(-6.12)
If I remove both knight b1 and rook a1 NNUE gives me -1271(-6.11)

I used the NNUE that latest stockfish developement use.
Stupid NNUE and it is better if the stockfish stop to use it.

Both evals are about 100% loss chance according to stockfish, so what's your problem? Apart from the fact that you apparently don't like NNUE for some reason and you're trying find any possible reason to hate on it.

The problem is that there is a difference between different losses.
Being a knight down is better than being knight and a rook down.


I believe that this type of behaviour make stockfish a relatively weak program in beating weak chess engines with a big material handicap.
I also guess that stockfish is relatively weaker in finding mates and I find that old versions of stockfish without stupid evaluation find mates faster.

But that's not because of stupid evaluations ...

[Sopel]

If I remove knight b1 NNUE gives me -1272(-6.12)
If I remove both knight b1 and rook a1 NNUE gives me -1271(-6.11)

I used the NNUE that latest stockfish developement use.
Stupid NNUE and it is better if the stockfish stop to use it.

Both evals are about 100% loss chance according to stockfish, so what's your problem? Apart from the fact that you apparently don't like NNUE for some reason and you're trying find any possible reason to hate on it.

The problem is that there is a difference between different losses.
Being a knight down is better than being knight and a rook down.


I believe that this type of behaviour make stockfish a relatively weak program in beating weak chess engines with a big material handicap.
I also guess that stockfish is relatively weaker in finding mates and I find that old versions of stockfish without stupid evaluation find mates faster.

I see no point in further discussion as you seem to be fixated on "stupid" evaluation.