Suppose that LC0 were trained not from self-play games but from data extracted from its games on TCEC, CCRL, and CEGT?
Since those contests include lots of arbitrary openings, many of which are imbalanced, I wonder if it would help the problem.
Fixing the LC0 imbalance defect
Moderators: hgm, Rebel, chrisw
-
Dann Corbit
- Posts: 12540
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Fixing the LC0 imbalance defect
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
frankp
- Posts: 228
- Joined: Sun Mar 12, 2006 3:11 pm
Re: Fixing the LC0 imbalance defect
Superficially it seems that including the relevant opening positions (endpoints) in lc0 training should improve its performance in such positions, but then it would not be lc_zero.
The alphaGo match against lee sedol demonstrated the potential for holes ('delusions', I think the team called them) resulting from self-play training - when positions had not be explored sufficiently. I have in mind the game lee sedol won, when alphaGo was basically clueless.
The alphaGo match against lee sedol demonstrated the potential for holes ('delusions', I think the team called them) resulting from self-play training - when positions had not be explored sufficiently. I have in mind the game lee sedol won, when alphaGo was basically clueless.
-
dkappe
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: Fixing the LC0 imbalance defect
1. Too little data.Dann Corbit wrote: ↑Thu Oct 15, 2020 3:24 pm Suppose that LC0 were trained not from self-play games but from data extracted from its games on TCEC, CCRL, and CEGT?
Since those contests include lots of arbitrary openings, many of which are imbalanced, I wonder if it would help the problem.
2. There have been a number of experiments using openings and start positions, most recently t72 and t60. The opening books didn’t seem to help. More info in the discord.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".