Nice
“Once is happenstance, twice is coincidence, three times is enemy action.” LCZero, Stockfish and Lichess teams just posted a joint note explaining our stance on the recent Fat Fritz 2 release by ChessBase.
Interesting read about Fat Fritz 2
Moderators: hgm, Rebel, chrisw
-
- Posts: 136
- Joined: Sat Dec 04, 2010 5:31 pm
- Location: 223
Re: Interesting read about Fat Fritz 2
Judge without bias, or don't judge at all...
-
- Posts: 136
- Joined: Sat Dec 04, 2010 5:31 pm
- Location: 223
Re: Interesting read about Fat Fritz 2
Judge without bias, or don't judge at all...
-
- Posts: 136
- Joined: Sat Dec 04, 2010 5:31 pm
- Location: 223
Re: Interesting read about Fat Fritz 2
Code: Select all
Official test by SF team, at fishtest conditions, of Stockfish 13 vs
the sold version of CB/AS (ff2) net and the corresponding Stockfish player.
===========================================================================
TC 10+0.1
=========
games at: https://drive.google.com/file/d/14TVxuyOGWAVc2kJyR6K7DEUOetZfWplm/view?usp=sharing
../ordo -a 0 -p out-sf13-vs-ff2-TC10+0.1.pgn -A sf13 -s 100 -q -J -D -W
# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%)
1 sf13 : 0.0 ---- 34445.5 65536 53 100
2 ff2 : -18.1 1.3 31090.5 65536 47 ---
White advantage = 29.29 +/- 0.67
Draw rate (equal opponents) = 77.35 % +/- 0.18
TC 60+0.6
=========
games at: https://drive.google.com/file/d/10yOOoCqt12sssmKy3jSI7og-M-nNs_Xs/view?usp=sharing
../ordo -a 0 -p out-sf13-vs-ff2-TC60+0.6.pgn -A sf13 -s 100 -q -J -D -W
# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%)
1 sf13 : 0.0 ---- 33774.0 65536 52 100
2 ff2 : -10.8 1.0 31762.0 65536 48 ---
White advantage = 24.72 +/- 0.50
Draw rate (equal opponents) = 87.77 % +/- 0.14
Conclusion
==========
The ff2 net is weaker than SF13 in a proper testing setup.
Draw rates are comparable to fishtest results.
Judge without bias, or don't judge at all...
-
- Posts: 136
- Joined: Sat Dec 04, 2010 5:31 pm
- Location: 223
Re: Interesting read about Fat Fritz 2
Code: Select all
Setup and parameters
====================
Same setup as fishtest, using 8moves_v3.pgn book, run in 128 batches (command used for TC60+0.6):
../cutechess-cli -repeat -rounds 1 -games 512 -tournament round-robin -srand $RANDOM -resign movecount=3 score=400 -draw movenumber=34 movecount=8 score=20\
-concurrency 64 -openings file=../8moves_v3.pgn format=pgn order=random plies=16\
-each tc=60+0.6 option.hash=64 proto=uci -ratinginterval 128\
-engine cmd=../stockfish.sf13 name=sf13 \
-engine cmd=../stockfish.ff2 name=ff2 \
-pgnout out-sf13-vs-ff2.pgn
Corresponding to SF13 and FF2 as available Feb 21st
../stockfish.ff2 bench 2>&1 | grep "Nodes searched"
Nodes searched : 4025201
../stockfish.sf13 bench 2>&1 | grep "Nodes searched"
Nodes searched : 3766422
Compiled using exactly the same compiler and same settings
on master at SF13 commit
make clean && make -j ARCH=x86-64-avx2 profile-build && mv ./stockfish ./stockfish.sf13
on ff2 branch with the CB/AS net added to the src directory
make clean && make -j ARCH=x86-64-avx2 profile-build && mv ./stockfish ./stockfish.ff2
Judge without bias, or don't judge at all...