During the pandemic I rediscovered chess and played a lot of games with my friends.
Then I started to program my first engine python-smallbrain in python, with the help of python-chess.

I quickly realized how slow python is for chess engine programming, so I started to learn C++.
My first try was cppsmallbrain, though after some time I found the code very buggy and ugly.
So I started Smallbrain from scratch, during that time, I also joined Stockfish development.

After some time I began implementing a NNUE into Smallbrain, with the help of Luecx from Koivisto.
As of now Smallbrain has a NNUE trained on 1 billion depth 9 + depth 7 + 150 million depth 9 DFRC fens generated with the built in data generator and using CudAD trainer to ultimately train the network.

The latest development versions support FRC/DFRC.

Compile it using the Makefile

make -j
.\smallbrain.exe bench

compare the Bench with the Bench in the commit messages, they should be the same.

or download the latest the latest executable directly over Github.
At the bottom you should be able to find multiple different compiles, choose one that doesnt crash.

Ordered by performance you should try x86-64-avx2 first then x86-64-modern and at last x86-64. If you want maximum performance you should compile Smallbrain yourself.

• Hash The size of the hash table in MB.
• Threads The number of threads used for search.
• EvalFile The neural net used for the evaluation, currently only default.nnue exist.
• SyzygyPath Path to the syzygy files.
• UCI_ShowWDL Shows the WDL score in the UCI info.
• UCI_Chess960 Enables Chess960 support.

• go perft <depth> calculates perft from a set position up to depth.
• print prints the current board
• eval prints the evaluation of the board.

• bench Starts the bench.
• perft fen=<fen> depth=<depth> fen and depth are optional.
• -eval fen=<fen>
• -version/--version/--v/-v Prints the version.
• -see Calculates the static exchange evaluation of the current position.
• -generate Starts the data generation.
• -tests Starts the tests.

• Evaluation
  • As of v6.0 the NNUE training dataset was regenerated using depth 9 selfplay games + random 8 piece combinations.

• Starts the data generation.

  -generate
  

• Specify the number of threads to use. default: 1

  threads=<int>
  

• If you want to start from a book instead of using random playout. default: ""

  book=<path/to/book>
  

• Path to TB, only used for adjudication. default: ""

  tb=<path/to/tb>
  

• Analysis depth, values between 7-9 are good. default: 7

  depth=<int>
  

• Analysis nodes, values between 2500-10000 are good. default: 0

  nodes=<int>
  

• The amount of hash in MB. This gets multiplied by the number of threads. default: 16

  hash=<int>
  

• Example:

.\smallbrain.exe -generate threads=30 book=E:\Github\Smallbrain\src\data\DFRC_openings.epd tb=E:/Chess/345

.\smallbrain.exe -generate threads=30 depth=7 hash=256 tb=F:\syzygy_5\3-4-5
.\smallbrain.exe -generate threads=30 depth=9 tb=H:/Chess/345
.\smallbrain.exe -generate threads=30 nodes=5000 tb=H:/Chess/345

I'd also like to thank the following people for their help and support:

• A big thanks to Luecx for his amazing CudAd trainer and his help with the NNUE implementation.
• Andrew Grant for the OpenBench platform https://github.com/AndyGrant/OpenBench
• Morgan Houppin, author of Stash https://github.com/mhouppin/stash-bot for his debug sessions.
• Various other people from Stockfish discord for their help.
• Chess.com for their Smallbrain inclusion in the Computer Chess Championship (CCC)
• TCEC for their Smallbrain invitation.

The following engines have taught me a lot about chess programming and I'd like to thank their authors for their work:

• Stockfish
• Koivisto
• Weiss

Included: The following parts of the code are from other projects, I'd like to thank their authors for their work and their respective licenses remain the same:

• Fathom MIT
• Incbin UNLICENSE

External:

• OpenBench GPL-3.0