Tarnished - UCI chess engine
Author: Anik Patel (Bobingstern)
v2.0:
This version comes with several bugfixes, search improvements, and NNUE evaluation! Tarnished v2.0 absolutely crushes v1.0 with the power of neural networks. It can defeat Stash-30 at LTC so perhaps around ~3000 CCRL.
NNUE
Tarnished v2.0 makes use of a (768->128)x2->1x8 architecture, SCReLU activation
Starting from a random net, the NNUE was trained via self play at 5ksn and 8 random plies of opening
Networks are embedded into the executable using incbin
Inference is vectorized with the Lizard SCReLU trick for a notable speed boost
Accumulator refreshes for promotions and enpassants only (will be improved in the next release)
Search Improvements
Improving Heuristic
Internal Iterative Reductions
Improved LMR (Weiss log formula table)
Improved LMP
Improved RFP
Aspiration Windows
Mate score bugfix
Improved time management (soft time)
Comments
Post a Comment