New LC0 option test: CpuctRoofOffset

Discussion of computer chess matches and engine tournaments.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
mmt
Posts: 216
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

New LC0 option test: CpuctRoofOffset

Post by mmt » Wed Feb 26, 2020 12:07 am

Match: 384x30-t60-2800 kiudee bonus with CpuctRootOffset=2 vs 384x30-t60-2800 kiudee bonus
LC0 version (both): 2/20 master Win build with double instead of float for updates
LC0 options (both): cudnn-fp16, 3 threads, backend=multiplexing
Hardware: 2080Ti + Titan V
Time control: 10s+0.5s
Book: Random from Silver Large with colors switched
Tablebases: important 7-piece TBs, all 6-piece TBs, WDL only
Adjudication: 5-man TB, 50 moves, 35 moves at 0.00, <-900 centipawns
Software: Arena with a clean-up of duplicate games
Comments: CpuctRoofOffset=2 looks better than the default but within the margin of error. The error should be 6.4 using the pentanomial model.

Code: Select all

   # PLAYER                       :  RATING  ERROR  POINTS  PLAYED   (%)
   1 lc0-kuidee bonus wide root   :  2301.5    7.6   513.5    1018    50
   2 lc0-mine sergio kuidee 2     :  2298.5    7.6   504.5    1018    50
   

https://drive.google.com/open?id=1pTbWI ... E-cfL8M4Gi

While analyzing the 1. g4 game, several times I thought that LC0 focuses too much on the move it finds early instead of searching wider. This match suggests but not proves that it might indeed be the case. Additional testing is needed.

Post Reply