e0ad7b3
This commit was signed with the committer’s verified signature.
JonathanHallstrom
Jonathan Hallström
Elo | 110.29 +- 4.75 (95%)
Conf | 40.0+0.40s Threads=4 Hash=256MB
Games | N: 3760 W: 1499 L: 344 D: 1917
Penta | [1, 53, 617, 1208, 1]
https://furybench.com/test/6704/
Elo | 111.33 +- 3.58 (95%)
Conf | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 6996 W: 2807 L: 639 D: 3550
Penta | [3, 110, 1118, 2250, 17]
https://furybench.com/test/6699/
Elo | 103.48 +- 3.21 (95%)
Conf | 10.0+0.10s Threads=1 Hash=16MB
Games | N: 10116 W: 3984 L: 1057 D: 5075
Penta | [6, 233, 1728, 3010, 81]
https://furybench.com/test/6697/
Most of the gain is from large improvements to the network on all fronts.
The network now has several hidden layers, and a much more expressive set of inputs, now including threat inputs as well as pawn-pawn inputs (my own invention). The network is fully quantised, in the vein of Stormphrax
Also various search changes, but the majority of the elo is improvements to the network.
Huge thanks to @kelseyde for his help with NEON inference in #42, @jgilchrist for fixing a fen parsing bug in #10, and @anematode for helping with compute.
Full changelog: v1.9.2...v2.0
Contributors
jgilchrist, anematode, and kelseyde