search

glinscott / leela-chess

**MOVED TO https://github.com/LeelaChessZero/leela-chess ** A chess adaption of GCP's Leela Zero
http://lczero.org
GNU General Public License v3.0
759 stars 298 forks source link

LC0 Games database #482

Open GuardianRM opened 6 years ago

GuardianRM commented 6 years ago

Good day, it seems LC0 reaches enough strength and lots of games for a quality database. Are there any plans to create an online database of the analogue Fritz Lets check !? Thanks.

Akababa commented 6 years ago

Interesting idea, but you might notice that all of the self-play games are partially randomized (to avoid duplicates), so the games in such a database wouldn't necessarily be of high quality.

GuardianRM commented 6 years ago

I drew attention to this, because there are quite interesting theoretical questions. As far as we know, a large volume of statistical sampling allows us to make fairly accurate statistical conclusions. If the level of lc0 is now approximately equal to the level of the top GM, then with the presence of about 50 million games, the level of its base can surpass the level of any chess base, including correspondence. It may be worthwhile to take as a starting point the later versions, but it seems, since version 227 it can already start working.

GuardianRM commented 6 years ago

It is quite possible in the future it can be useful when choosing manually the definite debut direction for further training or vice versa for the expansion of statistics for the rare first moves. In any case, a fairly extensive theory does not yet exist, so that we can not with certainty say what is better, for example, Najdorf Sicilian or Nimzowich Defense, or there may be a move 1) c2-c4 is the best choice. I think it's can be interesting to everyone.

Akababa commented 6 years ago

If you would use it that way why wouldn't you just put the moves into lc0 and see what happens?

GuardianRM commented 6 years ago

Statistics save time for practitioners. To approach, for example, to the current level of 2300 games by correspondence, it is necessary to use the engine for about an hour for a move on good hardware. The latest versions of the top engines, including Stockfish, can not guarantee anything now, for example in the Old Indian defense. Even after several hours of deliberation. The statistics gives some answers.

GuardianRM commented 6 years ago

I also wanted to ask such a question - is it possible to assemble the Network on the basis of any database and how much it will be strong!? And is it possible, on the basis of any database, to derive the optimal network ratio and the speed of the search? Can, for example, a network based on the database of masters ICCF and higher be a stronger network, than the already trained !?