Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Update default Network to [ECF0F596]
I was unable to reproduce a Network from scratch which rivaled master. As a result, this Network is built by taking the previous best (A258642E), un-quantizing it, and converting it back into the format for training. This meant losing some precision from the original network; losing the factorized split of inputs; losing the ADAM optimizer state. It was then trained for 74 iterations of 1024x1024x128 steps, using 100% WDL and 0% Eval. The learning rate was 0.000025, or a typical 1/40th drop in LR as is done in previous network series. The data is as follows. No validation data was used. S44 8,384,993,675 S51 349,252,596 S52 698,714,372 S53 781,050,465 S54 365,505,860 S55 146,908,394 S56 769,986,626 S44 is self-play, the rest are against Stockfish, but WITHOUT using any evaluations or positions from Stockfish. IE, Ethereal is still strictly original data. Elo | 42.67 +- 11.49 (95%) SPRT | 10.0+0.10s Threads=1 Hash=8MB LLR | 2.95 (-2.94, 2.94) [0.00, 3.00] Games | N: 1604 W: 463 L: 267 D: 874 Penta | [4, 118, 375, 288, 17] http://chess.grantnet.us/test/36386/ Elo | 35.63 +- 8.10 (95%) Conf | 60.0+0.60s Threads=1 Hash=64MB Games | N: 3356 W: 968 L: 625 D: 1763 Penta | [4, 247, 837, 582, 8] http://chess.grantnet.us/test/36388/ BENCH : 3,139,722
- Loading branch information