gamebase used to create the current net

Skip to first unread message


Apr 28, 2021, 8:39:05 AMApr 28
to FishCooking
Hello, does anyone knows, if we can reach and obtain the games used to create the current main net?
My theory is that by starting from that gamebase as core, and removing some games and adding some other games, using some specific criteria, we might be able to create a stronger net

Joost VandeVondele

May 1, 2021, 12:02:19 PMMay 1
to FishCooking
Sergio's latest set of games are but we have never been able to recreate a strong net from them.


May 16, 2021, 12:13:04 PMMay 16
to FishCooking
If you're interesting in tuning the set of positions used for training, you could start with a very large collection of positions, randomly sample from it to train a bunch of small/quick networks for a reasonable number of iterations, run them in tournaments against each other, and gather statistics about which positions were most likely to be in the training sets of the best networks. (Where best could be Elo, some measure of variance or exploitability, a combination of multiple measures, etc.)

Another thought is to co-train the usual network with a GAN whose task is to generate chesslike positions that the network evaluates poorly (i.e. maximize deviation between depth 1 and depth 8 search or whatever), and have the GAN randomly generate some portion of the positions used each training step.
Reply all
Reply to author
0 new messages