Night Nurse is Bad Gyal in disguise

[dkappe]

I had a fair amount of 800 node test games from bad gyal 7 though 8h sitting around, so I trained up Night Nurse (bonus for anyone who gets the Jamaican connection) from that data. Now anyone who is curious what a nnue net would play like if it were trained by a leela type net, there you go.

Now Bad Gyal isn’t even the strongest 128x10 net (probably -130 elo to t70), but Night Nurse already plays much more dynamically than the sf trained nets. I’m really curious what a nnue trained by a big net with 1600 nodes per move would look like.

The best nnue nets may end up being trained by Leela.

[AdminX]

Will there be a Pi version?

[dkappe]

Will there be a Pi version?

If sf-nnue runs on pi, then yes.

[AdminX]

Will there be a Pi version?

If sf-nnue runs on pi, then yes.

I hope so, I have not heard any news about NNUE and Pi yet, but I will keep my eyes open for it.

[Ovyron]

Just want to commend the people that are still coming up with nice names for things such as "Bad Gyal" or "Night Nurse" which really give character to what they make. Unlike others that continue using boring names like N.N.U.E., StockfiNN, NBSC, or name their nets with hard cold numbers.

The latest technologies have been full of creativity under the rugs, original names are a great way to show imagination on the other side.

[peter]

The best nnue nets may end up being trained by Leela.

You mean trained by games Leela played against SF NNUE or against SF or against different Leel- nets?
That would be trained by SF and by Leela yet still mainly and depend much on Hardware- TC, as for the quality of the games, wouldn't it?

Or games Leela played on her own?
That would still depend mainly on the books used.
By games Leela playes on her own without book would be somewhat narrow- minded as for openings, wouldn't it?

Not so easy to get the right material for training.

I'd still take big opening books with much diversity and many sharp lines, into big depthes combined with corr. databases.
How about that?

[dkappe]

The best nnue nets may end up being trained by Leela.

You mean trained by games Leela played against SF NNUE or against SF or against different Leel- nets?
That would be trained by SF and by Leela yet still mainly and depend much on Hardware- TC, as for the quality of the games, wouldn't it?

Or games Leela played on her own?
That would still depend mainly on the books used.
By games Leela playes on her own without book would be somewhat narrow- minded as for openings, wouldn't it?

Not so easy to get the right material for training.

I'd still take big opening books with much diversity and many sharp lines, into big depthes combined with corr. databases.
How about that?

You’re training an evaluation function based on the evaluation of an engine at some depth. Stockfish is easy, as the generation of training data from sf’s eval is built in. But you can generate training data from any uci engine (as I have done with toga, bad gyal and Komodo). At 800 nodes per move, lc0 even with a relatively weak 128x10 net generates very good evals.

The engine plays games against itself to generate positions with evals. The training data games start essentially from some position arrived at by a random walk and with random moves, especially king moves, thrown in from time to time afterwards, otherwise you won’t cover a big enough variety of positions. It doesn’t really resemble ‘good chess.’

If you want to train against the result in some way, you have to reduce the randomness.

[peter]

The engine plays games against itself to generate positions with evals. The training data games start essentially from some position arrived at by a random walk and with random moves, especially king moves, thrown in from time to time afterwards, otherwise you won’t cover a big enough variety of positions. It doesn’t really resemble ‘good chess.’

If you want to train against the result in some way, you have to reduce the randomness.

But you can train with .pgn- data too, can't you?

My idea was, to take into .pgn- transcribed book- moves, played out or adjudicated as wdl, together with corr.- games in .pgn and maybe 7some- tbs- positions with their lines to wdl.

Wouldn't that resemble "good chess"?

[dkappe]

The engine plays games against itself to generate positions with evals. The training data games start essentially from some position arrived at by a random walk and with random moves, especially king moves, thrown in from time to time afterwards, otherwise you won’t cover a big enough variety of positions. It doesn’t really resemble ‘good chess.’

If you want to train against the result in some way, you have to reduce the randomness.

But you can train with .pgn- data too, can't you?

My idea was, to take into .pgn- transcribed book- moves, played out or adjudicated as wdl, together with corr.- games in .pgn and maybe 7some- tbs- positions with their lines to wdl.

Wouldn't that resemble "good chess"?

In theory it’s possible, but I’m not sure how many games you would need.

[peter]

In theory it’s possible, but I’m not sure how many games you would need.

Well, neither do I of course.
But Herbert Bellmann's corr. DVD only is already about 1.5 Million games, a book like Cerebellum with wdl- adjudicated lines probably already more then this, 7men Lomonosovs probably some more then both others together, yet maybe on average only as long as the opening- lines.

I'd guess at least 10 millions of games, how many positions? Let's say 0.5 billions?