- new NNUE architecture (768x4hm -> 768)x2 -> 1x8 (768 hidden neurons, horizontal mirrorring, 4 king buckets, 8 output buckets)
- improved and rewritten Static Exchange Evaluator
- shrinking hash entry
- LMR tweaks and search reorganisation
- some speed improvements
- solving problems with memory consumption
- DFRC Support and +200 ELO gains (Est. 3700 on FRC list)
- retrained on 2.1b dataset (1b 512hl standard + 1.1b 512hl dfrc)
- trained tal personality on 200M filtered dataset from above dataset for higher EAS gameplay approx around same as Patricia (600k EAS) but around at 3300 cELO
NOTE: testers versions had multithreading disabled, please redownload if you downloaded on 30.08
Results of RodentTest vs RodentBase (8+0.08, 1t - NULL, 16MB, UHO_Lichess_4852_v1.epd):
Elo: 66.19 +/- 15.61, nElo: 113.49 +/- 26.11
LOS: 100.00 %, DrawRatio: 44.12 %, PairsRatio: 3.22
Games: 680, Wins: 257, Losses: 129, Draws: 294, Points: 404.0 (59.41 %)
Ptnml(0-2): [0, 45, 150, 117, 28], WL/DD Ratio: 1.27
LLR: 2.95 (100.1%) (-2.94, 2.94) [0.00, 5.00]