NNUE Evaluation

[nli33]

Results of Silverfish-dev vs Silverfish-0.1 (10+0.1, NULL, NULL, 8moves_v3.pgn):
Elo: 17.16 +/- 7.59, nElo: 30.17 +/- 13.32
LOS: 100.00 %, DrawRatio: 47.74 %, PairsRatio: 1.40
Games: 2614, Wins: 519, Losses: 390, Draws: 1705, Points: 1371.5 (52.47 %)
Ptnml(0-2): [22, 262, 624, 363, 36], WL/DD Ratio: 0.16
LLR: 2.97 (100.8%) (-2.94, 2.94) [0.00, 5.00]
--------------------------------------------------
SPRT ([0.00, 5.00]) completed - H1 was accepted

Player: Silverfish-dev
  Timeouts: 2
  Crashed: 0
Player: Silverfish-0.1
  Timeouts: 9
  Crashed: 0

Finished match
Total Time: 02:54:11 (hours:minutes:seconds)

• NNUE evaluation with (768->256)x2->1 architecture (2 feature sets, 2 accumulators)

Todo:

• Unfortunately perfts are now way slower, maybe refactor DoMove to allow not updating the NNUE
  • Related reason for this: only evaluate during quiescence since eval is only required during quiescence
• Accumulator stack / move stack for faster NNUE updates? (makes undo faster)
• 43/2614 games terminated due to illegal moves, 25 of which resulted in a win for Silverfish-dev. This is an old bug that I forgor about and likely doesn't have anything to do with move generation itself (since perfts pass)

[earthtraveller1]

Pretty amazing, not gonna lie.