NNUE endgame module?

[M ANSARI]

Is it possible to have an endgame NNUE module? I guess it could be trained against existing 7 EGTB's and maybe the NNUE would end up being a much smaller file than actually having to store all 7 EGTB"S. Not sure if this has been tried before but seems like a plausible idea and maybe is an easier route to going the next step to 8 piece EGTB's.

[lithander]

I have been wondering the same.

[Jouni]

Yes Stockfish nnue already has tablebase information in some form.

[daniel71]

I have been hoping for Endgame Module Databases like all the Grandmaster Reuben Fine Endgame Book Postitions, Grandmaster Yuri Averbakh Endgame Works, Dvoretsky's Endgame Manual, Silman's Complete Endgame Course, Karsten Müller and Lamprecht Fundamental Chess Endings, in a training databases to teach the methods to play better Endgames and to use this Knowledge in the Chess Engine Analysis during games. At least have all these Endgames Postitions be Solved by best play with computers. The heuristics for Endgame play would get improved for software and hopefully help humans. I have wrote about using FPGA Cards for a Endgame Module Database where NNUE software would run and improve the software for endgames where the data would grow and steer the games for precise play. I have always thought this way since the days of Hydra by Dr. Chrilly Donninger.

[Graham Banks]

Caissa already has a normal NNUE plus an endgame NNUE.

[dkappe]

Caissa already has a normal NNUE plus an endgame NNUE.

Stockfish has 8 embedded networks, each trained on positions where number of (pieces+pawns)/4. So essentially SF has eight specialty modules. Since its trained largely on leela data, the data has been tuned for months and years towards best play.

So, the specialty endgame nnue wish has already been fulfilled and then some.

[Graham Banks]

Caissa already has a normal NNUE plus an endgame NNUE.

Stockfish has 8 embedded networks, each trained on positions where number of (pieces+pawns)/4. So essentially SF has eight specialty modules. Since its trained largely on leela data, the data has been tuned for months and years towards best play.

So, the specialty endgame nnue wish has already been fulfilled and then some.

Nice.