Releases: znxftw/rudim-networks
v3-gen3
3 million self play games across multiple networks (v2-gen1, v2-gen2, v3-gen1, v3-gen2)
{
"games_completed": 3050000,
"total_positions": 519371216,
"white_wins": 897206,
"black_wins": 865823,
"draws": 1286971
}
Architecture for v3 generations - (768 -> 128) x 2 -> 1
Score of v3-gen3 vs v3-gen2: 309 - 304 - 387 [0.502]
... v3-gen3 playing White: 169 - 135 - 196 [0.534] 500
... v3-gen3 playing Black: 140 - 169 - 191 [0.471] 500
... White vs Black: 338 - 275 - 387 [0.531] 1000
Elo difference: 1.7 +/- 16.8, LOS: 58.0 %, DrawRatio: 38.7 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
1000 of 1000 games finished.
(However, node counts reduced for search)
v3-gen2
2 million games of self play
1 million v2-gen1 at depth 7
500k of v2-gen2 at depth 7
500k of v3-gen1 at depth 7
{
"games_completed": 2000000,
"total_positions": 340356976,
"white_wins": 587711,
"black_wins": 568814,
"draws": 843475
}
Architecture for v3 generations - (768 -> 128) x 2 -> 1
Score of Rudim-rc vs Rudim-3.0.1-v3-gen1: 168 - 131 - 201 [0.537]
... Rudim-rc playing White: 83 - 62 - 105 [0.542] 250
... Rudim-rc playing Black: 85 - 69 - 96 [0.532] 250
... White vs Black: 152 - 147 - 201 [0.505] 500
Elo difference: 25.8 +/- 23.6, LOS: 98.4 %, DrawRatio: 40.2 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished.
v3-gen1
1 million self play games of v1-gen3 at depth 7
1 million self play games of v2-gen1 at depth 7
{
"games_completed": 2000000,
"total_positions": 338463943,
"white_wins": 608328,
"black_wins": 603581,
"draws": 788091
}
Architecture for v3 generations - (768 -> 128) x 2 -> 1
(Didn't retest vs v2-gen2 as it was same dataset, but increased HL size)
--------------------------------------------------
Results of V3-Gen1 vs V2-Gen1 (8+0.1, NULL, NULL, 8moves_v3.pgn):
Elo: 134.55 +/- 16.72, nElo: 191.40 +/- 21.53
LOS: 100.00 %, DrawRatio: 29.00 %, PairsRatio: 6.24
Games: 1000, Wins: 490, Losses: 121, Draws: 389, Points: 684.5 (68.45 %)
Ptnml(0-2): [5, 44, 145, 189, 117], WL/DD Ratio: 0.86
--------------------------------------------------
v2-gen2
1 million self play games of v1-gen3 at depth 7 + 1 million self play games of v2-gen1 at depth 7
{
"games_completed": 2000000,
"total_positions": 338463943,
"white_wins": 608328,
"black_wins": 603581,
"draws": 788091
}
Architecture for v2 generations - (768 -> 64) x 2 -> 1
Score of v2-gen2 vs v2-gen1 : 229 - 114 - 157 [0.615]
... v2-gen2 playing White: 133 - 45 - 72 [0.676] 250
... v2-gen2 playing Black: 96 - 69 - 85 [0.554] 250
... White vs Black: 202 - 141 - 157 [0.561] 500
Elo difference: 81.4 +/- 25.6, LOS: 100.0 %, DrawRatio: 31.4 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished.
v2-gen1
1 million self play games of v1-gen3 at depth 7
{
"games_completed": 1000000,
"total_positions": 167101737,
"white_wins": 321627,
"black_wins": 324699,
"draws": 353674
}
Architecture for v2 generations - (768 -> 64) x 2 -> 1
Score of Rudim-v2-gen1 vs Rudim-v1-gen3: 237 - 106 - 157 [0.631]
... Rudim-v2-gen1 playing White: 127 - 45 - 78 [0.664] 250
... Rudim-v2-gen1 playing Black: 110 - 61 - 79 [0.598] 250
... White vs Black: 188 - 155 - 157 [0.533] 500
Elo difference: 93.2 +/- 25.7, LOS: 100.0 %, DrawRatio: 31.4 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished.
v1-gen3
300,000 self play games of v1-gen2 at depth 7
----------------------------------------
Final Data Generation Summary:
Games completed : 300000
Total positions : 46797972
Total White Wins : 100415
Total Black Wins : 102928
Total Draws : 96657
Average Game Length : 155.99
----------------------------------------
Architecture for v1 generations - (768 -> 32) x 2 -> 1
Score of rudim-nnue-v1-g3 vs rudim-nnue-v1-g2: 445 - 20 - 35 [0.925]
... rudim-nnue-v1-g3 playing White: 228 - 7 - 15 [0.942] 250
... rudim-nnue-v1-g3 playing Black: 217 - 13 - 20 [0.908] 250
... White vs Black: 241 - 224 - 35 [0.517] 500
Elo difference: 436.4 +/- 51.1, LOS: 100.0 %, DrawRatio: 7.0 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished.
v1-gen2
300,000 self play games on v1-gen1 @ Depth 5
----------------------------------------
Final Data Generation Summary:
Games completed : 300000
Total positions : 39665798
Total White Wins : 126516
Total Black Wins : 123868
Total Draws : 49616
Average Game Length : 132.22
----------------------------------------
Architecture for v1 generations - (768 -> 32) x 2 -> 1
Score of rudim-nnue-v1-g2 vs rudim-nnue-v1-g1: 498 - 0 - 2 [0.998]
... rudim-nnue-v1-g2 playing White: 249 - 0 - 1 [0.998] 250
... rudim-nnue-v1-g2 playing Black: 249 - 0 - 1 [0.998] 250
... White vs Black: 249 - 249 - 2 [0.500] 500
Elo difference: 1079.2 +/- nan, LOS: 100.0 %, DrawRatio: 0.4 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished
v1-gen1
300,000 self play games from Gen 0 (Random Mover) @ depth 5
~20 million filtered positions of training
(768 -> 32) x2 -> 1
Score of rudim-nnue-v1-g1 vs rudim-nnue-v1-g0: 500 - 0 - 0 [1.000]
... rudim-nnue-v1-g1 playing White: 250 - 0 - 0 [1.000] 250
... rudim-nnue-v1-g1 playing Black: 250 - 0 - 0 [1.000] 250
... White vs Black: 250 - 250 - 0 [0.500] 500
Elo difference: inf +/- nan, LOS: 100.0 %, DrawRatio: 0.0 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
500 of 500 games finished.
Player: rudim-nnue-v1-g1
"Win: Black mates": 250
"Win: White mates": 250
Player: rudim-nnue-v1-g0
"Loss: Black mates": 250
"Loss: White mates": 250
v1-gen0
sample-network
This is a sample network trained on lichess Stockfish evals (100m)
(768 -> 1024) x 2 -> 1 (Structure might change as I try out a few more things)
This will only be used as a placeholder until v0 (Random Mover) and v1 (Gen 1) are ready from self play. Rudim will not be released with these weights live.