This release brings a network trained with the new, highly-advanced rewrite of the Bullet neural network trainer.
Upgrades, in chronological order:
- Significantly improve SIMD code quality (#185)
- Add binpack subsampling (#187)
- New network, ID
lila(#188) - Use FT_SHIFT of 9 instead of 10 (#189)
- Make inference robust across architectures (#190)
- Fused NNZ Computation (#191)
- Propagate L1 two-at-a-time (#192)
- Unroll feature transformer two-at-a-time (#193)
- Only cut via SNMP under some conditions (#194)
- Use a linear history bonus (#196)
- Add a range of additional correction histories (#197)
- Speculatively force accumulator materialisation for some positions (#198)
- Use 2-back
improvingwith guarding for checks (#199) - Simplify the movepicker (#202)
- Use history to modify SEE pruning margin (#203)
- Remove the guard on history LMR (#204)
- New network, ID
voyager(#205) - Use threefold repetition instead of twofold (#206)
Elo comparison against v14.0.0
STC:
Elo | 35.28 +- 6.86 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 3004 W: 917 L: 613 D: 1474
Penta | [11, 251, 704, 495, 41]
https://chess.swehosting.se/test/8737/
LTC:
Elo | 43.51 +- 6.42 (95%)
Conf | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 3002 W: 939 L: 565 D: 1498
Penta | [3, 206, 737, 524, 31]
https://chess.swehosting.se/test/8736/
(D)FRC:
Elo | 33.88 +- 5.97 (95%)
Conf | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 3004 W: 624 L: 332 D: 2048
Penta | [2, 181, 876, 409, 34]
https://chess.swehosting.se/test/8738/
Thanks as always to everyone who helps me out with viri's development, especially everyone on https://chess.swehosting.se/, you're a great bunch.