Repository navigation
Total Elo gain estimate: ~300 from playing against v1.1, though in my testing this version beat Pirarucu 3.3.5 and played roughly equally against Stash 30, and this would put it closer to a gain of about 400 Elo
- changed activation function of net to SCReLU
- removed psqt neurons
- 768x8 for NNUE inputs instead of halfKA_hm feature set
- 3 more NNUE generations
- stopped doing TT cutoffs in PV nodes
- adjusting LMR reduction value based on history
- avoided object creation in many places, reducing GC load and object creation overhead
- allowed LMR in PV nodes
- added
go nodes <n>option for limiting search - log based formula for LMR
- TT cutoffs in quiescence search
- counting quiescence search nodes to the total searched nodes
- added late move pruning
- improving flag and using it to adjust RFP and NMP
- storing static eval in transposition table
- added SEE pruning to main search
- added counter move history (1-ply cont hist)
- added recapture extensions
- now sorting bad quiet moves after bad captures
- replaced debug option with debug UCI command
- reworked UCI command parsing