Major strength improvements!
Estimated ratings: 3710 CCRL Blitz, 3580 CCRL 40/15 1CPU and 3610 CCRL 40/15 4CPU
Highlights
- New NNUE architecture: Introduced king input buckets, output buckets, multi-layer topology (#136) and dual-activation (#148). Arch: (768x16hm->1280pw)x2->(16x2->32->1)x8
- New training data: Network trained purely on self-play data generated from a random initialization (#102)
- Improved time management: better and tuned time allocation strategy (#159, #161), with pv stability (#160), score stability (#163) and node scaling adjustment (#158)
- Better Search: A lot done, hard to enumerate. See Full Changelog for details
Full Changelog: v6.0.0...v7.0.0
Results against Minke 6.0.0
STC
Elo | 201.15 +- 14.46 (95%)
Conf | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 1004 W: 583 L: 59 D: 362
Penta | [0, 6, 86, 290, 120]
https://eduardomarinho.dev/test/910/
LTC
Elo | 183.98 +- 11.33 (95%)
Conf | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1002 W: 530 L: 44 D: 428
Penta | [0, 2, 80, 350, 69]
https://eduardomarinho.dev/test/911/
More tests
Against PZChessBot v7.1
LTC
Elo | 5.91 +- 11.20 (95%)
Conf | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1000 W: 265 L: 248 D: 487
Penta | [3, 114, 249, 131, 3]
https://eduardomarinho.dev/test/908/
SMP (4CPU) STC
Elo | 41.54 +- 11.10 (95%)
Conf | 8.0+0.08s Threads=4 Hash=256MB
Games | N: 1000 W: 297 L: 178 D: 525
Penta | [0, 77, 232, 186, 5]
https://eduardomarinho.dev/test/916/
Against Raphael v4.1.0
LTC
Elo | 2.78 +- 10.75 (95%)
Conf | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1000 W: 251 L: 243 D: 506
Penta | [2, 114, 258, 126, 0]
https://eduardomarinho.dev/test/915/
SMP (4CPU) STC
Elo | 3.47 +- 10.67 (95%)
Conf | 8.0+0.08s Threads=4 Hash=256MB
Games | N: 1000 W: 239 L: 229 D: 532
Penta | [4, 104, 271, 120, 1]
https://eduardomarinho.dev/test/918/
Opening book used for testing was UHO_Lichess_4852_v1.epd
Acknowledgements
I am very grateful to the chess engine community, especially the Alpha-Beta and Stockfish discord servers. Without the people in the community, Minke would be nowhere near what it is today! Thank you for all the support, ideas, and help!
Special thanks to @Ciekce and @JonathanHallstrom for all their help and shared expertise!
Huge thanks to all engine testers as well!