Prune

this is a hobby project made by two passionate about informatic.

currently, the engine includes:

• Evaluation:

  • nnue
    • (60144+768x4hm -> 384-pw)x2 -> (16 -> 32 -> 1)x8
    • 13 iterations from random net
  • trained on self-gen data using bullet (if you want the data used to train it, you can directly ask one of us)
  • correction history :
    • pawn
    • previous move
    • continuation move
    • minor+king pieces
  • material scaling

• Search:

  • iterative Deepening
  • quiescence search
  • aspiration window
  • enhanced forward pruning
  • Reduction:
    • late move reduction :
      • log base
      • history
      • tactical moves
  • pruning:
    • null move pruning
    • reverse futility pruning (improving and not improving)
    • razoring at depth 1
    • late move pruning
    • history pruning
    • capture history pruning
    • futility pruning
    • SEE pruning in QS
  • transposition table
    • buckets of 3 entries
  • move ordering:
    • killer move
    • PV move
    • SEE
    • MVV-LVA
    • history heuristic
    • capthist
  • singular extension :
    • singular
    • double
    • negative
    • cutnode negext
  • IIR

• most of the search parameters has been tuned over 2k iterations of 16 40+0.4 games

• movegen

  • use bitboards
  • fully legal movegen

• UCI protocol:

  • fancy commands:
    • version (give the commit's hash when it was compiled)
    • arch (give the arch used to compile)
    • bench (test of 48 positions, which give the node count/depth sum to compare search, also give average nps)
    • runQ (run quiescnce search on the given position)
    • print (print the current position [after the moves])
    • position kiwipete (set kiwipete's position)

• option :

  • SyzygyPath: path to Syzygy tablebase files
  • SyzygyProbeDepth: minimum search depth to probe tablebases
  • SyzygyProbeLimit: maximum piece count to probe (0–7)
  • TB scoring: WIN = 20000 - ply, LOSS = -20000 + ply, DRAW = 0; ply adjustment prefers faster wins and slower losses; cursed win/blessed loss are treated as draws since the 50-move rule makes them unachievable in practice
  • the other ones should be known, refer to the stockfish readme

• syzygy probing is done thanks to a copy of Fathom from : https://github.com/jdart1/Fathom

• support multithreading (lazy smp)

• don't support yet frc, planed for v4

to compile, just run

make prune

in the core/ directory, and you will have the executable prune, which contains the engine (so it is moveable)

note

for the version naming, we currently use the following :

• 3rd digit is only for bug fix versions
• 2nd digit is for elo improvement (for now, we target a +190 elo per release)
• 1st digit is for new features (as smp for v3, or planed frc for v4)

thanks

• stockfish discord :
  1. TM improvement (took some constant from heimdall for init)
  2. corrhist help (took some constant from sirius for update)
  3. speedup of the nnue (lizard updates)
• chessprogramming wiki
  • warn for new dev, can be outdated, so don't forget to sprt every changes
• especially to :
  • swedishchef (the dev from pawnocchio and vine) for helping me debugging the viriformat writer, and giving me some search ideas
  • Ciekce, who helped me a lot making working the threat net (by giving me his bullet config, and helping me debugging it)
  • Yoshie, who helped me to know some speedup techniques for the ti, directly or by his code