ShakeyBot is a UCI chess engine written in C++.
It uses hand-crafted evaluation. In local testing against FabChess Version 1.8, a public engine with a listed strength around 2400 Elo, ShakeyBot scored slightly ahead. This is a local estimate, not an official rating-list result.
It uses hand-crafted evaluation. In local testing against FabChess Version 1.8, a public engine with a listed strength around 2400 Elo, ShakeyBot scored slightly ahead. This is a local estimate, not an official rating-list result.
Author: Daniel John Collins
ShakeyBot 2.0.0 what's new?
ShakeyBot v2.0.0 is the first public neural-evaluation release and a major
strength upgrade over the hand-crafted-evaluation releases. The default engine
now uses a quantized HalfKP NNUE-style evaluator with accumulator updates.
Default model:
models/halfkp_wp_h512_e15_500m_clip30_quant.txt
Default backend:
neural_halfkp_quant_accum
Strength Estimate
Local tournament against Ceibo v1.0:
Scope W-D-L Score Score % Elo diff 95% CI
Overall 651-266-202 784.0/1119 70.06% +147.7 +128.9 to +166.5
Ceibo v1.0 is listed around 2985 Elo, so this suggests ShakeyBot v2.0.0 is
provisionally around 3000 Elo under this local tournament setup.
This is a provisional estimate, not an official rating-list result. The engine
has not yet been benchmarked against a broad pool of engines and time controls.
Main Changes Since v1.1.0
Default evaluation changed from hand-crafted evaluation to HalfKP neural eval.
Default backend is now neural_halfkp_quant_accum.
Default model is the HalfKP h512 model trained on natural game-position data.
Low-material neural fallback is disabled by default for tournament play.
Accumulator checker is disabled by default for speed.
Neural eval and accumulator paths include the recent benchmarked optimization
work.
Several search/eval optimizations from the development cycle are included.
Windows Package
Included files:
ShakeyBot.exe
libgcc_s_seh-1.dll
libstdc++-6.dll
libwinpthread-1.dll
Comments
Post a Comment