Pea - a UCI compliant chess engine written in Rust with a NNUE trained from scratch.
Author: WGCodings
Pea 8.0 what's new?
Training was done in 6 sessions of genning and training.
Training parameters for these runs :
Positions : 40M Net 0, 160m Net 1
5k nodes per position
Initial LR : 0.0001-0.0005
Final LR : 0.0001
~2 epochs
start WDL : 0.0
end WDL : 0.15
Results vs previous generation
Net 0
Results of GEN8_NET0 vs GEN7_NET0 (8+0.08, 1t, 256MB, book.epd):
Elo: 199.96 +/- 30.48, nElo: 273.26 +/- 33.55
Games: 412, Wins: 267, Losses: 53, Draws: 92, Points: 313.0 (75.97 %)
Net 1
Results of GEN8_NET1 vs GEN7_NET1 (8+0.08, 1t, 256MB, book.epd):
Elo: 122.91 +/- 25.10, nElo: 151.48 +/- 28.42
Games: 574, Wins: 312, Losses: 117, Draws: 145, Points: 384.5 (66.99 %)
Elo estimation done vs Stash v26 (~3000 Elo) in a 8+0.08 TC.
Net 0 ~2993 Elo
Results of GEN8_NET0 vs STASH26 (8+0.08, 1t, 256MB, book.epd):
Elo: -7.04 +/- 7.10, nElo: -8.34 +/- 8.41
Games: 6562, Wins: 2539, Losses: 2672, Draws: 1351, Points: 3214.5 (48.99 %)
Net 1 ~2963 Elo
Results of GEN8_NET1 vs STASH26 (8+0.08, 1t, 256MB, book.epd):
Elo: -36.66 +/- 14.71, nElo: -42.79 +/- 17.03
Games: 1598, Wins: 548, Losses: 716, Draws: 334, Points: 715.0 (44.74 %)
Net 1 embedded in binaries.
Comments
Post a Comment