This is an unusually fast second release, but we have a few good reasons to do it:
- two solid algorithm improvements (new history gravity formula and scaling NNUE eval against remaining material)
- discovering that a NNUE adjusted by training on Anand's games is better than default
- considerably reducing memory footprint
- adding UCI_Elo option, so that you can play against a weaker versions of personalities
- fixing NNUE access for personalities
- adding user manual
Rodent gained about 70 Elo in a week, and should be at 3470 CCRL at short time control.
Test against Rodent 1.0:
--------------------------------------------------
Results of RodentTest vs RodentBase (40+0.4, 1t, 32MB, UHO_Lichess_4852_v1.epd):
Elo: 78.45 +/- 15.33, nElo: 144.16 +/- 27.22
LOS: 100.00 %, DrawRatio: 42.17 %, PairsRatio: 4.84
Games: 626, Wins: 231, Losses: 92, Draws: 303, Points: 382.5 (61.10 %)
Ptnml(0-2): [0, 31, 132, 130, 20], WL/DD Ratio: 0.86
LLR: 2.95 (100.1%) (-2.94, 2.94) [0.00, 5.00]
--------------------------------------------------
SPRT ([0.00, 5.00]) completed - H1 was accepted
Rodent Asylum - Test of NNUE personalities
Rank Name Elo +/- Games Score Draw
1 Rodent V 1.1 Anand 100 9 3000 64.1% 44.0%
2 Rodent V 1.1 Nakamura -26 9 3000 46.3% 48.3%
3 Rodent V 1.1 -35 9 3000 45.1% 49.2%
4 Rodent V 1.1 Tal -38 9 3000 44.6% 47.9%