Specialized NNUE trained on classical eval?

• Material NNUE
• Imbalance NNUE
• Mobility NNUE
• King safety NNUE
• Passed pawn NNUE
• Space NNUE

Obviously NNUE doesn't perform so well in games with unusual positions. So it seems natural to me the logical conclusion to solving these problems is letting go of the idea that we need a "generalized net" and instead create specialized nets for each classical parameter.

@nodchip @locutus2 @amchess @AlexandreMasta @snicolet @vdbergh @jhonnold @Dboingue

[Sopel97]

material is already included in NNUE output, the rest is hard to add due to lack of truth for learning against

@Sopel97, People who says idea doesn't matter has no attachment to the idea, they are attached to outcome so for them execution matter. I would like to know @Luecx opinion on whether this idea is feasible to do so.

I realize that implementing such an idea can be intimidating, but taking that first step forward is the most important thing we can do to make this idea a reality. I have already trained specialized NNUE on mobility and it works well.

[Luecx]

It isnt just possible, I have done it in Koivisto. I used the features from the classical evaluation and had them as an additional input to the network. It gets a bit annoying since all that data must be brought into the trainer and is probably not feasible for the stockfish team at the moment. A second option would be to reconstruct all those features on the currently existing data. I had moderate to none success with it. I think the classical features are simply redundant. The networks, especially the deeper once figure it out themselves.

I think the classical features are simply redundant. The networks, especially the deeper once figure it out themselves.

8/p4p1p/5kp1/1p6/8/3b2P1/PPp2P1P/2R3K1 w - - 0 34

• NNUE evaluation: 1.70 (white side)
• Classical evaluation: -0.87 (white side)

[Luecx]

You cannot compare NNUE to HCE like that to argue that the deeper networks figure it out.

[Dboingue]

Wouldn't this approach first need to get what is already in the input (feature?) space of all NNue (master, and specialized transformers, or whetever many NN desing SF is now using). So that we can first have some mathematical ideas of already ELO worth formulation those very flexible combinators of input features (likely interpretable math. functions derivable (or vice versa) from those chess theory originating human interpretible ones. If I recall loosely about pre-NNue SF engine type history, the static evaluation functions components were already of that nature at inception, and now SF development ecosystem has been NN-wise statistically embedding ELO worther components into their non-linear and flexible combination function space (while pre-NNue one might have considered all components to have been combined linear (but for things like game phase partitions, jump discontinuities there, but nowhere were such functions part of global optimization schemes, that is now thorough, ASFIU).

However, we might need to decoded from CPU targetting compiler language into symbolic math. language that relate to board qunantities.. I find it hard to read through some such code base, and get a flat (sufficiently symbolized and substituted) set of equations. has this been done. would the new doc on NNue report on such past effort. (not just through instance computations and UCI, but in documentation?).

[lucametehau]

Why the fuck did you ping me

Please be nice in the chat.

[Dboingue]

this is about alignment with humans. Expanding the use from gladiator engine to analytical tool and learning tool.

But assuming that classical was the correct function to use just because the words seem to connect with chess theory, is having statistical problems. It needs research, but one could have some non coder interpretation of the code evolution ( and not just the executable releases but the more "abstract" development version with, a symbolic mathematical version (or worded back in chess board information so that many people not only developper that speak computer can participatein the process, people with logic, curiosity and various level of chess experience including high level, but not unique individiuals, as there is no chess theory language of scientific level yet (some are well defined others need refinement and statistical analyses).

There is also the motivation of the developpers to take into account, and the tangible end points that they have always been used to, which are not best discussed here as limited audience diversity of background. yet it would have to also be discussed here. i don't know how this can progress. but there is a big huge conceptual and imaginatino hole given all the factors I understand so far.

With a rigorous but higher level documentation that considers not only the executable but the full development model, and then a non CPU language model (that ought to have been in chunks distributed in the minds of the dev through SF evolution, perhaps buried in small paragrpahs amist text in programming wiki or their ancestral dev forum, or even dispersed in chats and discussions here or on discord) or in the documentation about the models in usage, not just limited assumption of usage user documentation, one could then shed some lights on the past contributions of classical to the emerging fishtest and engine ecosystem ELO as some empirical arm of this research project.

that would be using this current limited purpose techonology based on such ELOs metrics to the study of chess and chess leanring theories, and developpiing a better chess theory languages so that finally we could have some kind of wheel of science scoping the chess board game world. (not the cultural world, words are lame sometimes).

I have seen arguements here and there that make me think that some review work of docs and chats and whatever would contain nuggest of such useful thing if we are in agreement with expanding the goal of this techonology development.

but as long as the oracle mentality on the user side and the gladiator tangible objective function for the human developmen optmization mindset or freamwork are as strong as they are. This might be stalling in "dialogues de sourds"

technically. i would not use any of the coefficients of the classical evalution function defitions.. I would say go back to each component from chess theory (or develpment ideas) that can be written in chess board information without numerical ridden code syntax that all can read and critical thinking debate, and keep the symbolic algebraic formulation including the parameter intent in relation to the chess board understanding or concepts at the origin of those development steps at that function level.

then let the NNue (and research how) be a better statistical framework to include those algreabrais chess theory infused equations, and find out how they can fit in the big picture of chess win ratios.. (e.g. ELO stat).