Lc0: Large nets are still best for RTX
Posted: Fri Sep 24, 2021 8:36 pm
Although the T75 (small 15x192) net has been improving and did well at https://www.computerchess.org.uk/ccrl/404/ ,
The official https://github.com/LeelaChessZero/lc0/w ... ts-for-Lc0 still has:
Below: lastest Stockfish in a gauntlet vs.:
T60 run 1 (30x384) net 610069 - Sep 16, 2021
T75 run 3 (15x192) net 752187 - Aug 18, 2021 - same net as at CCRL blitz (although newer might be better)
Both Lc0's (separate copies) defaults with multi_stream=true.
The links go to a page with the game in a viewer. ('open in a new tab' is an idea.)
Also the game #'s and round numbers are given so they can be found in the .pgn.
( Not shown: Game #49, Round 13 - dubious white gambit, Sf 1-0 Lc75.
In spite of the opening Sf had +0.72 then Lc0's 34...a2?? lost the ending. )
Game #50, Round 13 - same dubious white gambit, Lc75 0-1 Sf.
Although the gambit pawn dawn Lc0 sac'ed another with 41.d6!?.
Stockfish refuted that with an exchange sac. 42...Qc6! and went on to win with black B+B+2 pawns vs. white R+N.
http://chess.tuxtown.net/game-replayer. ... de66bcd14b
( Not shown: Games #51 and #52, Round 13 - same dubious white gambit, Sf 1/2-1/2 Lc61 and Lc61 1/2-1/2 Sf.
With the same opening the large net was able to draw with both colors. )
Games #261-#264, Round 66 - KGA white advantage, Sf 1-0 Lc75, Lc75 1/2-1/2 Sf ; Sf 1-0 Lc61 (only 21 moves, diagram), Lc61 1-0 Sf (78 moves)
Games #221-#224, Round 56 - Urusov Gambit, Sf won with white and drew with black vs. both Lc0's.
#223, Sf 1-0 Lc61, was only 15 moves, one of the best games that I have seen
5.Bxg6!! B sac.; 6.Rxe6!! exchange sac.; 7.Qxd7 recovers a N; 9.Ne5!! N offer won although declined:
http://chess.tuxtown.net/game-replayer. ... f991059486
Games .pgn: https://www.mediafire.com/file/jstgxunq ... 3.pgn/file
The official https://github.com/LeelaChessZero/lc0/w ... ts-for-Lc0 still has:
Below: lastest Stockfish in a gauntlet vs.:
T60 run 1 (30x384) net 610069 - Sep 16, 2021
T75 run 3 (15x192) net 752187 - Aug 18, 2021 - same net as at CCRL blitz (although newer might be better)
Both Lc0's (separate copies) defaults with multi_stream=true.
Code: Select all
G/75 secs. + 0.8 secs., 2Gh Sy.6, R7 2700X 16ths | RTX 2060
Each all 100 DOTS3 reversed.
Rank Name Elo +/- Games Points Score Draws
0 Stockfish_21092119 130 25 400 271.5 67.9% 47.3%
1 Lc0-0.28.0_610069 -106 35 200 70.5 35.3% 48.5%
2 Lc0-0.28.0--752187 -156 35 200 58.0 29.0% 46.0%
400 games, 100 rounds @ 1 round per opening.Also the game #'s and round numbers are given so they can be found in the .pgn.
( Not shown: Game #49, Round 13 - dubious white gambit, Sf 1-0 Lc75.
In spite of the opening Sf had +0.72 then Lc0's 34...a2?? lost the ending. )
Game #50, Round 13 - same dubious white gambit, Lc75 0-1 Sf.
Although the gambit pawn dawn Lc0 sac'ed another with 41.d6!?.
Stockfish refuted that with an exchange sac. 42...Qc6! and went on to win with black B+B+2 pawns vs. white R+N.
http://chess.tuxtown.net/game-replayer. ... de66bcd14b
( Not shown: Games #51 and #52, Round 13 - same dubious white gambit, Sf 1/2-1/2 Lc61 and Lc61 1/2-1/2 Sf.
With the same opening the large net was able to draw with both colors. )
Games #261-#264, Round 66 - KGA white advantage, Sf 1-0 Lc75, Lc75 1/2-1/2 Sf ; Sf 1-0 Lc61 (only 21 moves, diagram), Lc61 1-0 Sf (78 moves)
Games #221-#224, Round 56 - Urusov Gambit, Sf won with white and drew with black vs. both Lc0's.
#223, Sf 1-0 Lc61, was only 15 moves, one of the best games that I have seen
5.Bxg6!! B sac.; 6.Rxe6!! exchange sac.; 7.Qxd7 recovers a N; 9.Ne5!! N offer won although declined:
http://chess.tuxtown.net/game-replayer. ... f991059486
Games .pgn: https://www.mediafire.com/file/jstgxunq ... 3.pgn/file