New chess engine: Molybdenum 3.0
v.3.1:
This release features an Elo gain of around 150 mostly coming from NNUE improvements:
STC:
Elo | 142.76 +- 18.12 (95%)
Conf | 8.0+0.08s Threads=1 Hash=8MB
Games | N: 1002 W: 517 L: 127 D: 358
Penta | [1, 33, 153, 203, 111]
http://aytchell.eu.pythonanywhere.com/test/347/
LTC:
Elo | 152.60 +- 18.14 (95%)
Conf | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 1000 W: 526 L: 113 D: 361
Penta | [3, 21, 130, 252, 94]
http://aytchell.eu.pythonanywhere.com/test/346/
Elo gains:
Search Improvements:
While this release I mostly focused on improving the NNUE there are a few search patches:
Adjust the singular extensions margin based on the rootmovecount of the current branch: #66
(Note that the commit message may seem confusing, that's because I intended exactly the opposite effect of what this currently does, I tried to "fix" it, but so far unsuccessfully)
Don't singular extend the rootnode: #67
Movecount prune more when not improving #68
NNUE improvements:
Increased hidden layer size of 128 alongside an additional 40m fens: #69
New nets trained on additional data: #71, #72, #74, #78, #79
SCReLU instead of CReLU activation function: #76
Further increase in hidden layer size to 256: #77
All networks mentioned in this release were trained using the bullet trainer by @jw1912, without bullet this release would not have been possible, so thanks a lot for providing this awesome trainer
The provided binaries use avx2, bmi2... (v3) / sse3, popcnt.. (v2) optimization respectively. The avx2bmi2 binary is faster so it should be preferred if your CPU supports these instructions sets. Note that even ryzens before zen 3 should be able to use the bmi2 binaries without slowdown, as they contain no pext / pdep instructions
Comments
Post a Comment