search

liamt19 / Lizard

Chess engine written in C#
MIT License
31 stars 2 forks source link

New network with selfplay data #65

Closed liamt19 closed 3 months ago

liamt19 commented 3 months ago

Networks are now trained entirely on selfplay data instead of Leela data. The current arch is (768 -> 1536)x2 -> 8, horizontally mirrored.

Decent elo hits in both STC and LTC:

Elo   | -33.04 +- 6.83 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3006 W: 661 L: 946 D: 1399
Penta | [34, 496, 710, 247, 16]
http://somelizard.pythonanywhere.com/test/1440/
Elo   | -36.45 +- 9.13 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1502 W: 288 L: 445 D: 769
Penta | [10, 257, 369, 110, 5]
http://somelizard.pythonanywhere.com/test/1442/

But a significant gain in (D)FRC due to the larger proportion of DFRC to standard data included in this data:

Elo   | 112.14 +- 11.09 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 3000 W: 1479 L: 543 D: 978
Penta | [46, 201, 461, 355, 437]
http://somelizard.pythonanywhere.com/test/1446/

Correction history was also added to make up for the slightly weaker evaluation, which passed relatively easily:

Elo   | 4.84 +- 2.90 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=32MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 3.00]
Games | N: 15646 W: 3995 L: 3777 D: 7874
Penta | [88, 1748, 3948, 1936, 103]
http://somelizard.pythonanywhere.com/test/1427/