meatball

This is a hobbyist's attempt at machine learning by creating a chess engine in C++.

Features

Consider reading the wiki for more information on the theory behind some of the engine's components.

• Board Representation
  • Bitboards
• Move Generation
  • Magic bitboards
• Search
  • Monte Carlo Tree Search (MCTS)
• Evaluation
  • Deep neural network

This code is written with readability in mind and is not meant to be fully optimized.

Installation

Dependencies

• C++17 (minimum)

Clone the directory and build the executable

git clone https://github.com/hgducharme/meatball.git
cd meatball/
make meatball

If successful, you should see the executable /bin/meatball.

Run make help for more information about what targets can be ran.

Tests

Running tests using Docker

1. Make sure the docker daemon is already running (see: starting docker daemon)
2. Run the test suite

cd /path/to/meatball/
make docker-tests

Running tests locally

Dependencies

• googletest 1.13.0
• cmake for building googletest

To install googletest-1.13.0:

cd /tmp
wget https://github.com/google/googletest/archive/refs/tags/v1.13.0.tar.gz
tar xf v1.13.0.tar.gz
cd googletest-1.13.0
mkdir install
cd install
cmake ..
make
make install
cd ../../
rm -rf googletest-1.13.0 && v1.13.0.tar.gz

Next, edit the GOOGLETEST variable inside meatball/Makefile to point to where googletest was installed. Replace /usr/local/lib with the directory that houses your libgtest.a and libgtest_main.a files.

GOOGLETEST := -L/usr/local/lib -lgtest -lgtest_main

Build and run the tests:

make tests
./bin/tests

To build and run only the unit tests:

make unit_tests
./bin/unit_tests

To build (and run) only the inegration tests:

make integration_tests
./bin/integration_tests

Test coverage

This only works locally. Make sure you have gcov and gcovr installed on your system:

• gcov is included with gcc and clang
• gcovr can be installed with pip and brew

make coverage

The test_coverage/ directory will contain .gcov files generated by gcov and html files generated by gcovr.

To see the HTML coverage report just open the file meatball/test_coverage/html/coverage.html in your browser.

Debugging move generation using the perft

It can be difficult to know where to start debugging the move generation when a test case fails because the number of calculated nodes don't match up with the expected number of nodes. To achieve a more specific error output we can do a perft divide. For each top level move, the perft divide function will show the number of child nodes. So in order to debug move generation in a practical way, the following steps are recommended:

1. Go to your stockfish repo and make sure the executable has been built.
2. Start the executable ./stockfish.
3. Run the commands in the shell: position startpos and then go perft [depth]. You should see the number of nodes that fall under each top level move. You can recurse down the tree to get more and more precise results.
4. In test_movegeneration.cpp, make sure the showDivideOutput argument in the perft() calls are set to true. This will print the divide output to the console.
5. Compare the results.

Docker Support

If you have docker installed then you can do any of the following things within a docker container.

NOTE: The docker daemon must be already running

1. Build the source code (BUILD_FLAGS is optional)

make docker-build BUILD_FLAGS="CXX=g++ DEBUG_LEVEL=0 OPTIMIZATION_LEVEL=3"

3. Run a linter

make docker-linter

5. Run the test suite

make docker-tests