Pea - a UCI compliant chess engine written in Rust with a NNUE trained from scratch.
Author: WGCodings
Pea 9.1 what's new?
What's Changed
corrhist-majors-and-kings in #2
wdl-normalization in #3
added mate detection to q search and changed return value to static eval if static eval > beta in #4
eval-to-tt by in #5
history-tables-refactor in #6
also do lmr in pv nodes in #7
history tables passed to engine state after search in #8
Pass pawn corrhist table on to engine state after search. in #9
tt-best-move-storage in #10
pruning-conditions-cutnode in #11
simd-in-evaluate-nnue in #12
Implemented fuse updates in #13
lazy-acc-updates in #14
Full Changelog: v9.0...v9.1
There are some more features added that can be found on my openbench instance (before I started using pull requests) : https://peaob.pythonanywhere.com/greens/.
These include correction histories for pawns and material, mate distance pruning, smarter aspiration window choice, quiet history pruning, better time management and some bug fixes concerning draw and mate detection.
STC (+195 Elo)
Results of PEA-9.1 vs Pea-9.0 (8+0.08, 1t, 3MB, UHO_Lichess_4852_v1.epd):
Elo: 194.57 +/- 14.67, nElo: 349.26 +/- 21.51
LOS: 100.00 %, DrawRatio: 18.16 %, PairsRatio: 36.27
Games: 1002, Wins: 556, Losses: 47, Draws: 399, Points: 755.5 (75.40 %)
Ptnml(0-2): [0, 11, 91, 278, 121], WL/DD Ratio: 0.65
LTC (+159 Elo)
Results of PEA-9.1 vs Pea-9.0 (40+0.4, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: 158.51 +/- 12.43, nElo: 314.62 +/- 21.53
LOS: 100.00 %, DrawRatio: 24.00 %, PairsRatio: 33.55
Games: 1000, Wins: 478, Losses: 51, Draws: 471, Points: 713.5 (71.35 %)
Ptnml(0-2): [0, 11, 120, 300, 69], WL/DD Ratio: 0.50
Recommended to download the avx2 version, which is around 40% faster.
Comments
Post a Comment