Frozenight

NNUE chess engine written in Rust. Play against it on lichess.org

The current minimum supported Rust version for the UCI binary is 1.59.0.

Rating Lists

CCRL 40/15:

• Frozenight 6.0: 3226
• Frozenight 5.1: 3027
• Frozenight 5.0: 3000
• Frozenight 4.0: 2964
• Frozenight 3.0: 2842
• Frozenight 2.1: 2680
• Frozenight 2.0: 2614

CCRL Blitz:

• Frozenight 6.0: 3331
• Frozenight 5.1: 3098
• Frozenight 5.0: 3085
• Frozenight 4.0: 3005
• Frozenight 3.0: 2891
• Frozenight 2.1: 2677
• Frozenight 1.0: 2448

CCRL 40/2 FRC:

• Frozenight 6.0: 3493
• Frozenight 5.1: 3148
• Frozenight 5.0: 3103
• Frozenight 4.0: 2994
• Frozenight 3.0: 2761

Features

• cozy-chess for move generation
• Principal Variation Search
• Aspiration windows
• LazySMP multithreading
• NNUE evaluation
  • 768 -> 384x2 (-> 1)x16
  • Network bucketing based on modified piece material values (Queen = 8)
    • This is based on game phase tuning in Koivisto done by Luecx
  • Training data generated through self-play, originally starting with a random network
  • Trained using (a modified version of) Pali's marlinflow
• Quiescense search
  • SEE ordering & pruning with MVV-LVA for ties
  • Check Evasions
• Check Extensions
• PV Extensions
• Late move reductions
• Late move pruning
• Null move pruning
• Reverse futility pruning, except using qsearch instead of static eval
• Internal Iterated Deepening
• Transposition Table
  • Depth-preferred with aging
• Move ordering
  • Hash move
  • SEE captures, losing captures last, with MVV-LVA for ties
  • Killer heuristic (ordered near pawn captures pawn)
  • Relative history heuristic (side-by-side piece-tosq and fromsq-tosq tables)
  • Underpromotions last
• Time management
  • Uses at least 2% remaining + increment/2
  • Avoids stopping in the middle of an iteration

License

Frozenight is dual-licensed under the MIT License and Apache License (Version 2.0) licesnses.

Thanks

• Analog (Tantabus), for cozy-chess and helping me understand search techniques
• Pali (Black Marlin), for marlinflow and helping me understand NN training and search techniques
• Authors of the chess programming wiki, for its wealth of knowledge