Do we need learning file for engines?

criko · Post by **criko** » Sat Oct 28, 2023 3:27 pm

Jouni · Post by **Jouni** » Sun Oct 29, 2023 5:53 pm

NN is learning

.

criko · Post by **criko** » Sun Oct 29, 2023 7:04 pm

Jouni wrote: ↑Sun Oct 29, 2023 5:53 pm NN is learning .

exact what I think

Frank Quisinsky · Post by **Frank Quisinsky** » Sun Oct 29, 2023 7:09 pm

I vote with yes!

For the first moves on a full board with the idea ...
How I can hold the position complicated or better how I can hold my pieces on board!
Maybe with the second best or third best move.
And if successful engine can save it in a learning file.

So, if a learning file ... for endgames it makes no sense today.

Rebel · Post by **Rebel** » Sun Oct 29, 2023 7:17 pm

criko wrote: ↑Sat Oct 28, 2023 3:27 pm Do we need learning file for engines?

Better not.

1. You can download the games from rating lists and create an embedded book from the pgn in the executable.

2. In the reserve game you can make use of the data of the initial game.

I would say no.

Rebel · Post by **Rebel** » Sun Oct 29, 2023 7:25 pm

Jouni wrote: ↑Sun Oct 29, 2023 5:53 pm NN is learning .

Nevertheless the NN engine will make the same mistakes forever and ever

One of the purposes of an external learning file is to avoid just that.

j.t. · Post by **j.t.** » Sun Oct 29, 2023 9:14 pm

Having the information about which player/other engine the engine is playing available is an interesting thought. Together with some form of persistent memory, it would open the field for some new methods.

Eduard · Post by **Eduard** » Sun Oct 29, 2023 10:57 pm

Jouni wrote: ↑Sun Oct 29, 2023 5:53 pm NN is learning .

and for everything else there are millions of ENG games on Lichess and PlayChess. I prefer to make BIN and CTG books out of that. On Lichess I don't win with learning files but with a good engine and a good BIN book: https://lichess.org/@/SolistaBot After a year break I started new games yesterday.

Frank Quisinsky · Post by **Frank Quisinsky** » Sun Oct 29, 2023 11:44 pm

Eduard:
All the eng-eng games and the stats engines produced should be used better for opening books. I think today the grandmaster are playing openings more risky as for 20 years. With the results that many of the risky moves are for engines directly lost. Often I am thinking such things with many of the games strongest Grandmaster produced today. Each new GM line must be checked with different of strongest engines in detail.

All the Nakamura opening experiments are to find in engine books.
Not good ...

Example:
1. e4, d5, exd5, Qxd5, Nc3, Qa4 (or Qd6 or Qd8) ...

But now ...

Qa4 = popular 1-10, = 3 ... win probably with black = 26%
Qd6 = popular 1-10, = 6 ... win probably with black = 29%
Qd8 = popular 1-10, = 1 ...

This helps a lot to make books much more interesting.

In combination with ...
Opening should be able to give the information ...

Seems to be a new line
An old line

If now ...
An engine should check the three possibilites shortly itself ... before it play the book move.

Qa4 = -0,56
Qd6 = -0,58
Qd8 = -0,52

Most popular is Qd6 ... engine should try to play the popular line at first and ignore the own eval.
But if the line goes deeper and deeper ... engine should switched to the own eval and should use more time for check the possible book moves.

Something like this and much more can make the openings with all the produced eng-eng games (for book statistics genial) much more interesting.
Chessbase do here a lot ... have different very nice features inside.

Have fun with your opening book.
Opening books are more and more important these days!

Best
Frank

Jouni · Post by **Jouni** » Mon Oct 30, 2023 5:55 pm

There was Brainfish book project. But finished July 24, 2020 as useless.