[2024] github/dogefromage/Stalemater2000/README.md

Stalemater2000 - a C++ chess engine from scratch with NNUE

Stalemater2000 searches for the best chess move in a position using iterative deepening framework like most popular chess engines. The engine uses a simple two layer NNUE (efficiently updatable neural network) which predict the evaluation of the leaf nodes in the search space given the board squares as input features. The engine is written in C++ with no dependencies and runs in the command line featuring the UCI protocol for inputs and communication with other software. The NNUE model is trained using torch and uses the lichess.org evaluation dataset.

Skill level

It beats stockfish lvl 4 on lichess but I don't know if that's something to be proud of :/

Usage

Input position and display board using FEN:

position fen rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq - 0 1
d

  -------------------
8 | r n b q k b n r |
7 | p p p p p p p p |
6 | . . . . . . . . |
5 | . . . . . . . . |
4 | . . . . P . . . |
3 | . . . . . . . . |
2 | P P P P . P P P |
1 | R N B Q K B N R |
  -------------------
    a b c d e f g h  

Halfmove clock: 0
Fullmove number: 1

FEN: rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq-
Hash: 2824af1b2a2d7300

Is legal: 1
Checks:

Do a perft analysis to validate the engine by counting legal subpositions:

go perft 5

b8a6 -> 398321
b8c6 -> 475842
g8f6 -> 474457
g8h6 -> 400579
a7a6 -> 370351
b7b6 -> 430756
c7c6 -> 456786
d7d6 -> 639932
e7e6 -> 822484
f7f6 -> 342008
g7g6 -> 442824
h7h6 -> 363576
a7a5 -> 442101
b7b5 -> 411896
c7c5 -> 482325
d7d5 -> 788468
e7e5 -> 728887
f7f5 -> 427835
g7g5 -> 426723
h7h5 -> 445481
Total: 9771632

Find the best move in the current position up to a depth of 7 halfmoves:

go depth 7

info depth 1 score cp 0 nodes 21 nps 512195 pv b1a3
info depth 2 score cp 0 nodes 81 nps 1538461 pv b1a3 b8a6
info depth 3 score cp 0 nodes 667 nps 1693641 pv b1a3 b8a6 g1f3
info depth 4 score cp 0 nodes 2162 nps 1752637 pv b1a3 b8a6 g1f3 a6b4
info depth 5 score cp 100 nodes 26955 nps 1876835 pv b2b3 b8a6 c1b2 a6b4 b2g7
info depth 6 score cp -100 nodes 201096 nps 1342872 pv b1a3 b7b6 g1f3 c8a6 a3b1 a6e2
info depth 7 score cp 100 nodes 998975 nps 1180470 pv b1c3 b8a6 g1f3 a6b4 c3b5 b4d5 b5a7
bestmove b1c3

Useful links

Everything you'd ever would want to know about chess programming can be found on the chess programming wiki. It has lots of pseudocode and details about both historic and leading-edge approaches.

The bitboard logic I used was really well explained in this Youtube series by Logic Crazy Chess.

TODO: Unfinished work

  • Quantize NNUE model such that it uses integer arithmetic and SIMD instead of lousy float32
  • Do profiling to get performance up (I'm sure there are a lot of redundant copies in the search algorithm)
  • Find better dataset with less chaotic positions for better model (loss on current model is really high, but performs alright)

Future ideas