Pingu is a command-line program - in order to interact with other engines or people it is recommended to use a suitable GUI e.g. cutechess.

When running Pingu in the command line, type 'help' for a list of available commands.

Pingu accepts many of the usual UCI commands (go/stop/position etc.) and it has some additional custom commands.

More information on CCRL.

• Staged move generation
• Bitboard representation
• Plain magic bitboards

• Root search
  1. PV move
  2. Others ordered by descending subtree size
• Main search
  1. Hash move
  2. Good captures + promotions
  3. Killers
  4. Bad captures
  5. Quiet moves
• Quiescence search
  1. Good captures + promotions
  2. Check evasions
• Winning captures ordered by MVV/LVA
• Losing captures ordered by static exchange evaluation
• Quiet moves ordered by history, counter moves history and piece square tables

• Root search
  • Iterative deepening
  • Aspiration windows
  • Principal variation search
• Main search
  • Fail-soft alpha-beta
  • Principal variation search
  • Reverse futility pruning
  • Null move pruning
  • Internal iterative reduction
  • Late move reductions
  • Futility pruning
  • Late move pruning
  • Repetition detection
  • Transposition tables
    1. Always replace
    2. Depth-preferred
• Quiescence search
  • Fail-soft alpha-beta
  • Stand-pat
  • Forward prune SEE < 0
• Basic time management
• Parallel search with Lazy SMP algorithm

NNUE with HalfKA feature set and subnetworks indexed by piece count.

Architecture: 2 x (45056 -> 32) -> 1

Trained with PyTorch and a custom dataloader written in C++ (source code in /nnue).

Final model uses quantized weights and AVX2 intrinsics for optimal performance.

All NNUE versions of Pingu were trained with the engine's own self-play data. The process is described in nnue.md, and more recently in nnue_2.md and nnue_3.md.

• Pingu 4.0.0 dataset (new)
• Pingu 4.0.0 dataset (old)
• Pingu 3.0.0 dataset
• Pingu 2.0.0 dataset

Early HCE versions of Pingu were Texel-tuned using self-play data from Ethereal, obtained from this forum post. The process is described in evaluation_tuning.md.

Chess Programming Wiki for its useful resources.

CCRL for testing Pingu.