A very interesting post

[Sylwy]

Take a look here, please:

https://github.com/official-stockfish/S ... -678112885

[Laskos]

Take a look here, please:

https://github.com/official-stockfish/S ... -678112885

Different nets for different stages of the game is worth pursuing. It sure will bring an order of 100 Elo points in less than a year. The nets perform best for very narrow tasks. For example, very regular Chess (regular openings, material etc) is played better by SF NNUE than some variants close to Chess, and divergent variants are played very poorly compared to classical SF. Similarly, concentrating on very narrow tasks as "regular openings" or even "French Defense" or "opposite color Bishops endgame" will benefit greatly many-nets SF NNUE Elo-wise. In some time we will have automated distribution and training for different nets over different parts of the game, with number of different nets in dozens and more.

[towforce]

Different nets for different stages of the game is worth pursuing. It sure will bring an order of 100 Elo points in less than a year. The nets perform best for very narrow tasks. For example, very regular Chess (regular openings, material etc) is played better by SF NNUE than some variants close to Chess, and divergent variants are played very poorly compared to classical SF. Similarly, concentrating on very narrow tasks as "regular openings" or even "French Defense" or "opposite color Bishops endgame" will benefit greatly many-nets SF NNUE Elo-wise. In some time we will have automated distribution and training for different nets over different parts of the game, with number of different nets in dozens and more.

The plan:

1. a "master net" to classify a position and choose which evaluation net to use

2. a large number of evaluation nets to do the actual evaluation, each specialising in a different type of position

Although I wasn't thinking about NNs at the time, I had the same basic idea of a large number of evaluation functions for different position types 23 years ago - link (page 29).

[Laskos]

Different nets for different stages of the game is worth pursuing. It sure will bring an order of 100 Elo points in less than a year. The nets perform best for very narrow tasks. For example, very regular Chess (regular openings, material etc) is played better by SF NNUE than some variants close to Chess, and divergent variants are played very poorly compared to classical SF. Similarly, concentrating on very narrow tasks as "regular openings" or even "French Defense" or "opposite color Bishops endgame" will benefit greatly many-nets SF NNUE Elo-wise. In some time we will have automated distribution and training for different nets over different parts of the game, with number of different nets in dozens and more.

The plan:

1. a "master net" to classify a position and choose which evaluation net to use

2. a large number of evaluation nets to do the actual evaluation, each specialising in a different type of position

Although I wasn't thinking about NNs at the time, I had the same basic idea of a large number of evaluation functions for different position types 23 years ago - link (page 29).

Great, especially now, with NN evals. NN engines are much more specialized players than classical engines with handcrafted evals. Handcrafted eval engine has a degree of abstraction and generalization for many kinds of tasks higher than a NN eval engine. It is very visible when playing Chess variants, where NN based engines lose hundreds of Elo points to classical engines. But if NN engine is trained on this variant, it becomes much stronger than classical engines. The search also matters, Lc0 with PUCT is probably damned to be weak in endgames and deep tactics compared to top AB searchers, no matter what specialized endgame nets are used for Lc0.

[Jhoravi]

The plan:

1. a "master net" to classify a position and choose which evaluation net to use

2. a large number of evaluation nets to do the actual evaluation, each specialising in a different type of position

Although I wasn't thinking about NNs at the time, I had the same basic idea of a large number of evaluation functions for different position types 23 years ago - link (page 29).

Isn't it simpler for separate nets depending on the number of pieces? About 30 Nets for the whole game.

[M ANSARI]

I agree and have been thinking that for a while now. A single NNUE evaluation is very limited at 20mb if you use that evaluation for all the phases of the game. But if you manage to find an algo that can correctly steer into specialized evaluation modules then there should be a tremendous amount of low lying fruit with regards to gains in ELO strength. Actually ... if several evaluation nets can be incorporated and switched on the fly, it would make sense to make the nets smaller (thus run much faster). The thing is how do you correctly decide to automatically switch to the correct net? Should it be by the number of pieces left? Should it be with regards to how many possible moves are in the position? Maybe a more sophisticated system where King safety or passed pawns become a factor? Really a lot of potential ideas that can be tried and for sure there will be a major shift in engine strength soon. This actually might mean that EGTB's are no longer needed as the endgame can be correctly played by specialized nets. There might also be a better way to run these smaller nets and make them run much faster on the hardware side. The next few months might see a major change in how the software AND hardware is used. Maybe the 2080 Ti cards that people purchased for Lc0 can still be put to good use.

For all this to happen there needs to be some good way to go through millions of games and choose specific positions with the specified number of pieces or certain criteria in the position to train specialized nets. I am guessing if the number of pieces is the criteria this should be easy ... but if the criteria is king safety or passed pawns or "dynamic" position etc... probably more difficult. Maybe this software is already available?

[Jouni]

Just when I was about to ask I see in github pull request: the embedding of the nnue default file in the engine executable! No more any difficulties to install NNUE for beginners.