To lichess programmers: warning on stockfish 10 evaluation values

Hello,

Lichess is using now Stockfish 10, I think with its default values for parameters. The "contempt" one is a critical one: it is at 24 in default mode (20 for stockfish 9).

It is clear, up to me, that this high level of contempt gives wrong evaluations: the engine can estimate 0.70 a Carlsbad position without problem for Black: this is a WRONG evaluation, because the GM would give 0.2-0.3 for this position. Stockfish 9 gives around 0.5 (at similar depth, around 27), which is more sensible. Stockfish 10 evaluation value is just crazy: once again, it is wrong, we are NOT close to a pawn advantage in this position, I am sure (51% for White in Chessbase Database, so this is equal in practice).

How lichess programmers can answer to that? Thank you very much.

> This is a WRONG evaluation, because the GM would give 0.2-0.3 for this position

Do not forget that the best GM is < 2900 elo, whereas stockfish is above 3400. A GM has close to 0 chance of winning stockfish. Knowing this your statement looks ridiculous as this is similar to: Carlsen's evaluation of the position was wrong, because some low-rated player thinks that it should be different.

@Toadofsky: thanks, but "contempt" is not a word used once in any of these pages!

- what was the contempt used in stockfish 9 for lichess?
- what is the contempt used in stockfish 10 for lichess?

This value influences rather strongly (when we are looking for an advantage of 0.2 in an opening line, 0 is equal but 0.2 is interesting) the evaluation value of a position (about 0.2, so 0.5 slight advantage can become 0.7 big advantage without any good reason, a GM and even stockfish would agree to me, as after some moves he goes to 0.2, and, once again, the Dtabase gives 51% for White, so equal play), so this question is critical.

Same here:

A position can only be drawn (0) or won (infinite), everything else is arbitrary.

I investigated a correspondence game position of mine with a minus pawn but (over-) compensation, probably won.

-Komodo & Fritz: -1
-SF9: -2
-SF10: -3

The material is „+1“ though but the engine will win the position with a high probability. What is right, what is wrong?

Default contempt values are 20 and 24 for SF9/10. Set them to 0 if this fits better in your picture of chess.

Yes, I understand.

But I maintain: we, currently, are making the following correspondence values (approximately):
0.1 to 0.5: slight advantage" +/=" in literature
0.6 to 1.5: big advantage "+/-" in literature
>=1.5: winning "+-" in literature

And I maintain: my position is +/=, not +/- that I associate with 0.71.
If you destroy these corerspondence because of SF10 contempt 24, you should say it to the world of chess players like "we have changes the values: big advantage "+/- is now between 0.8 and 1.7"! (or, better, modifiy the correspondence values in the software to avoid that very annoying change). But if you launch the engine with these parameters as default values, a lot of people will have a lot of surprises! they will rather always believe in a bigger advantage than what there is objectively. Also, you will destroy the literature because +/= and +/- will be no more correlated between literature and computer.

So I maintain: this is a deep and big issue for serious chess players.

Afaik the contempt value doesn't have much of an effect on the eval. It has more effect on how SF plays.

#4
- what was the contempt used in stockfish 9 for lichess? 0.0
- what is the contempt used in stockfish 10 for lichess? 0.0
github.com/niklasf/stockfish.pexe/blob/f3bf2e0666606114584ea86c30ca7778c3c12fff/src/ucioption.cpp#L60
github.com/niklasf/fishnet/blob/f117f07da83ba686e8d88096d78d4d205c0cecf6/fishnet.py#L751
github.com/ornicar/lila/blob/ddb93e04d40ff49a8e0ad5817f6ac18c529808ee/ui/ceval/src/stockfishProtocol.ts#L31

Previous discovery at lichess.org/forum/lichess-feedback/sawtooth-evaluation-back-again seems to validate our theory that contempt doesn't make sense for analysis.

Contempt would affect the evaluation slightly, but it is disabled for analysis on lichess.org (or else there is a bug).

Can you please give a concrete position, the search depth and the exact evaluation at that depth? That would be useful to verify that there is indeed no contempt.

Thanks very much, this is decisive information.

PS: to stockfish developers: please launch your version with contempt 0, because chess players will not know that there is a parameter which affects clearly the evaluation value (giving 0.71 that is +/- for a slight advantage, which is erroneous).