Sunday, November 8, 2026

ChessKISS 2.3, yes! this time with tuning

Oh well, some tunings were tough and also weird, but I manage to somehow have better values and thus improve the ELO (not sure how much...)

What's new?
  • SPSA (search) and Texel (evaluation) tuning done, my room has never been so warm
  • All evaluation tables moved to the parameters, so the tuners can play with them
  • Many parameters were not symmetric (a-file vs h-file, white vs black): chess is symmetric, my tables weren't (shame on me, again)
  • Mobility tables are now symmetric
  • Material imbalance scored for both sides, before its value depended on who was to move (oops)
  • Threat detection was not symmetric between sides
  • SEE treats knights and bishops as equals
  • The weak f-pawn penalty no longer depends on the castling history
  • Safer pruning in quiescence: delta and futility use the biggest of the middlegame and endgame values
  • Smarter skipping of silly captures in quiescence
  • New ProbCut phase in the move picker, and a faster ProbCut
  • A mate on the hundredth half-move is a mate, not a fifty-move draw (FIDE agrees)
  • Data reordered for better memory locality
  • The board can write its own FEN
I also took the opportunity to clean and improve the readme.txt which was to say the least too lazy...

please download it at ChessKISS 2.3

No comments:

Post a Comment