4.5 million self play games across multiple networks (v2-gen1 to v4-gen1)
{
"games_completed": 4550000,
"total_positions": 758321669,
"white_wins": 1346172,
"black_wins": 1298113,
"draws": 1905715
}
Architecture for v4 generations - (768 -> 256) x 2 -> 1
Ran a few different training configurations over different dataset combinations to try and find which works best
(0.2 / 0.3 -> Start WDL)
(0.7 / 0.8 -> End WDL)
(latest -> only v3, v4 training data)
(all -> v2,v3,v4 training data)
Rank Name Elo +/- Games Score Draw
1 rudim-nnue-latest-0.2-0.7 12 11 2000 51.7% 43.9%
2 rudim-nnue-all-0.2-0.7 11 11 2000 51.5% 44.2%
3 Rudim-3.0.4 0 11 2000 50.0% 43.4%
4 rudim-nnue-all-0.3-0.8 -3 11 2000 49.5% 45.2%
5 rudim-nnue-latest-0.3-0.8 -19 11 2000 47.3% 45.0%
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
5000 of 5000 games finished.