This is an adaptation of GCP's Leela Zero repository to chess, using Stockfish's position representation and move generation. (No heuristics or prior knowledge are carried over from Stockfish.)
The goal is to build a strong UCT chess AI following the same type of techniques as AlphaZero, as described in Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm.
We will need to do this with a distributed project, as it requires a huge amount of compute.
Please visit the LCZero forum to discuss: https://groups.google.com/forum/#!forum/lczero
The server is live at http://162.217.248.187/. Please download the client and give it a try.
Of course, we also appreciate code reviews, pull requests and Windows testers!
The weights are located at https://github.com/glinscott/lczero-weights. Currently, the best weights were obtained through supervised learning on a human dataset with elo ratings > 2000.
After compiling lczero (see below), try the following:
cd build
cp ../scripts/train.sh .
./train.sh
This should launch lczero in training mode. It will begin self-play games, using the weights from weights.txt (initial weights can be downloaded from the repo above). The training data will be written into the data subdirectory.
Once you have enough games, you can simply kill the process.
To run the training process, you need to have CUDA and Tensorflow installed. See the instructions on the Tensorflow page (I used the pip installation method into a virtual environment). NOTE: You need a GPU accelerated version of Tensorflow to train, the CPU version doesn't support the input data format that is used.
Then, make sure to set up your config, important fields to edit are the path the network is stored in, and the path to the input data.
cd training/tf
./parse.py configs/your-config.yaml
That will bring up Tensorflow and start running training. You can look at the config file in training/tf/configs/example.yaml to get an idea of all the configurable parameters. This config file is meant to be a unified configuration for all the executable pythonscripts in the training directory. After starting the above command you should see output like this:
2018-01-12 09:57:00.089784: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX TITAN X, pci bus id: 0000:02:00.0, compute capability: 5.2)
2018-01-12 09:57:13.126277: I tensorflow/core/kernels/shuffle_dataset_op.cc:110] Filling up shuffle buffer (this may take a while): 43496 of 65536
2018-01-12 09:57:18.175088: I tensorflow/core/kernels/shuffle_dataset_op.cc:121] Shuffle buffer filled.
step 100, policy loss=7.25049 mse=0.0988732 reg=0.254439 (0 pos/s)
step 200, policy loss=6.80895 mse=0.0904644 reg=0.255358 (3676.48 pos/s)
step 300, policy loss=6.33088 mse=0.0823623 reg=0.256656 (3652.74 pos/s)
step 400, policy loss=5.86768 mse=0.0748837 reg=0.258076 (3525.1 pos/s)
step 500, policy loss=5.42553 mse=0.0680195 reg=0.259414 (3537.3 pos/s)
step 600, policy loss=5.0178 mse=0.0618027 reg=0.260582 (3600.92 pos/s)
...
step 4000, training accuracy=96.9141%, mse=0.00218292
Model saved in file: /home/gary/tmp/leela-chess/training/tf/leelaz-model-4000
It saves out the new model every 4000 steps. To evaluate the model, you can play it against itself or another AI:
cd src
cp ../training/tf/leelaz-model-4000.txt ./newweights.txt
cd ../scripts
./run.sh
This runs an evaluation match using cutechess-cli.
If you have expert games you wish to train from in PGN, you can generate training data from those for the network to learn from. Run:
./lczero --supervise games.pgn
That will create a folder supervise-games, with the training data. You can
then train a network against that as usual.
It is safe to kill the training process and restart it at any time. It will automatically resume using the tensorflow checkpoint.
You can use this to adjust learning rates, etc.
- GCC, Clang or MSVC, any C++14 compiler
- boost 1.58.x or later (libboost-all-dev on Debian/Ubuntu)
- BLAS Library: OpenBLAS (libopenblas-dev) or (optionally) Intel MKL
- zlib library (zlib1g & zlib1g-dev on Debian/Ubuntu)
- Standard OpenCL C headers (opencl-headers on Debian/Ubuntu, or at https://github.com/KhronosGroup/OpenCL-Headers/tree/master/opencl22/)
- OpenCL ICD loader (ocl-icd-libopencl1 on Debian/Ubuntu, or reference implementation at https://github.com/KhronosGroup/OpenCL-ICD-Loader)
- An OpenCL capable device, preferably a very, very fast GPU, with recent
drivers is strongly recommended (OpenCL 1.2 support should be enough, even
OpenCL 1.1 might work). If you do not have a GPU, modify config.h in the
source and remove the line that says
#define USE_OPENCL. - Tensorflow 1.4 or higher (for training)
- The program has been tested on Linux.
# Install dependencies
sudo apt install libboost-all-dev libopenblas-dev opencl-headers ocl-icd-libopencl1 ocl-icd-opencl-dev zlib1g-dev
# Test for OpenCL support & compatibility
sudo apt install clinfo && clinfo
# Clone github repo
git clone git@github.com:glinscott/leela-chess.git
cd leela-chess
git submodule update --init --recursive
mkdir build && cd build
# Configure, build and run tests
cmake ..
make
./tests
See https://github.com/glinscott/leela-chess/tree/master/go/src/client/README.md.
The code is released under the GPLv3 or later, except for ThreadPool.h, cl2.hpp and the clblast_level3 subdir, which have specific licenses (compatible with GPLv3) mentioned in those files.