Stockfish no longer has contempt

[Madeleine Birchfield]

Never understood why they had contempt in the first place, as it interfered with the accuracy of stockfish's evaluation by artificially inflating it by 24 centipawns, especially when it comes in a gui that does not allow the user to change uci options, such as the Lichess Stockfish interface.

Stockfish does not include a GUI, so it is not clear what you are talking about.

If the Lichess developers didn't want contempt, they could have trivially set it to 0 themselves.

https://en.wikipedia.org/wiki/Default_effect

[syzygy]

Never understood why they had contempt in the first place, as it interfered with the accuracy of stockfish's evaluation by artificially inflating it by 24 centipawns, especially when it comes in a gui that does not allow the user to change uci options, such as the Lichess Stockfish interface.

Stockfish does not include a GUI, so it is not clear what you are talking about.

If the Lichess developers didn't want contempt, they could have trivially set it to 0 themselves.

https://en.wikipedia.org/wiki/Default_effect

I gave you a full explanation, but you are free not to want to understand.

[Michel]

However, SF's approach to contempt is incompatible with NNUE.

I don't understand why it would be intrinsically incompatible with NNUE. Both HCE and NNUE are evaluation functions. I don't see a reason for them to behave differently with regard to contempt.

Note that I am talking about objective contempt (against weaker engines). Not self play contempt. The fact that that appeared to work with HCE in SF was likely just an accident.

[carldaman]

Doesn't Komodo Dragon, which is now NNUE based, still have contempt working?
If so, they can't be incompatible.

[cc2150dx]

yes, Dragons' contempt is working with NNUE.

[syzygy]

However, SF's approach to contempt is incompatible with NNUE.

I don't understand why it would be intrinsically incompatible with NNUE. Both HCE and NNUE are evaluation functions. I don't see a reason for them to behave differently with regard to contempt.

SF's eval calculates an evaluation score vector with 2 components (middle game, endgame) and then outputs a weighted average of the two components based on game phase. Contempt skews the middle game component twice as much as the endgame component, which reflects that the advantage of being SF is greater the more pieces there are still on the board. So a 3-rep draw is 0.00 but a completely balanced position with lots of heavy pieces is +something for SF. A completely balanced position with only little material left is close to 0.00, reflecting that SF"s strength advantage is worthless in case of a dead draw.

NNUE does not produce a middle game and an endgame component. So you can't add SF's contempt to it. You can of course still do the old type of content which just changes the draw value in the search, but that has never worked well.

One could think of using two NNs, one for middle game, one for end game, but it would add huge overhead and you can't really train such a thing. The middle game NN would not applied just to middle game positions but it would have to evaluate the position as if it is a middle game position. This hardly makes any sense to begin with, so good luck training it. The reason it can be done for HCE is that it is meaningful to consider the value of an HCE evaluation feature in middle game position and the value of the same feature in endgame positions. With an NN you don't have such isolated features that can be tuned individually.

Perhaps there is a way to add contempt to NNUE, but it will need a new approach.

I guess it should be possible to implement contempt by adding an input feature that indicates which side is being played by SF.

[syzygy]

Doesn't Komodo Dragon, which is now NNUE based, still have contempt working?
If so, they can't be incompatible.

I tend to be precise in my wording: "SF's approach", which was an approach that actually worked.

[syzygy]

I guess it should be possible to implement contempt by adding an input feature that indicates which side is being played by SF.

But then training data has to come from games of SF vs a weaker engine, and there would be just a single contempt setting.

To get a form of adjustable contempt one would have to add input features encoding the relative strength of the opponent engine. Trying to train it would be terribly messy.

[syzygy]

NNUE does not produce a middle game and an endgame component. So you can't add SF's contempt to it. You can of course still do the old type of content which just changes the draw value in the search, but that has never worked well.

Hmm, perhaps this is not a problem at all. Adding the 2-component HCE and contempt scores and then calculating a weighted average (which is what SF does) gives the same result (up to rounding) as first averaging the two and then adding the weighted averages.

With NNUE there is no averaging of the evaluation scores, but there is probably nothing stopping you from adding a "weighted" contempt score (i.e. a contempt score that tapers off as material disappears from the board).

So it seems I was indeed wrong about the fundamental incompatibility! I just focused too much on SF's very particular (and elegant) implementation, which only requires a single addition and then piggybacks on SF's tapered eval.

I suppose someone tried this already but could not make it work, but I have not followed it too closely.
(Or I am once more overlooking something.)

[syzygy]

I suppose someone tried this already but could not make it work, but I have not followed it too closely.

For example:
https://tests.stockfishchess.org/tests/ ... bade2b79d8
It seems there was no clear enough effect.