Pea - a UCI compliant chess engine written in Rust with a NNUE trained from scratch.
Author: WGCodings
Pea 6.0 what's new?
Turned the knobs on the hyperparameters a lot here. These ones seemed to give the best performance.
Trained both Net 0 and Net 1 with the same data and parameters.
Dategen done with 10k nodes per position as an experiment. (turned out well)
Training parameters for this run :
Positions : 20M Net 0
10k nodes per position
Initial LR : 0.001
Final LR : 0.0001 * 0.3f32.powi(5)
~2 epochs
start WDL : 0.0
end WDL : 0.25
Results vs previous generation
Net 0
Results of GEN6_NET0 vs GEN5_NET0 (8+0.08, 1t, 256MB, book.epd):
Elo: 350.17 +/- 47.62, nElo: 508.27 +/- 37.15
Games: 336, Wins: 275, Losses: 18, Draws: 43, Points: 296.5 (88.24 %)
Net 1
Results of GEN6_NET1 vs GEN5_NET1 (8+0.08, 1t, 256MB, book.epd):
Elo: 471.62 +/- 69.61, nElo: 752.72 +/- 38.43
Games: 314, Wins: 286, Losses: 11, Draws: 17, Points: 294.5 (93.79 %)
Elo estimated around 2742 for Net 0 and 2811 for Net 1 (8+0.08s).
Net 1 (64HL) included in binary.
Comments
Post a Comment