UnrealGo is a experimental implementation of alphagoZero in C++. It's developed with MCTS and deep reinforcement learning. It can be trained from scratch.
Boost - 1.64+
http://www.boost.org/
# build boost from source
$ ./bootstrap.sh
$ ./b2
# set LIB_INSTALL_DIR
$ cp -rf stage/lib/ ${LIB_INSTALL_DIR}
# edit and add lib directory to file /etc/ld.so.conf
$ sudo ldconfig
ZMQ - 4.2.2
$ git clone https://github.com/zeromq/libzmq.git
$ sudo apt-get install autogen autoconf libtool
$ ./autogen.sh
$ ./configure
$ ./make
$ ./make install
tensorflow cmake interface
Assume $PROJ_WORKSPACE is your project home directory
# clone tensorflow
$ cd $PROJ_WORKSPACE
$ git clone https://github.com/tensorflow/tensorflow.git
# git checkout to desired version(e.g. 1.6) if you like
# build interface
$ cd $PROJ_WORKSPACE
$ git clone https://github.com/unrealgo/tensorflow_cmake_interface.git
$ cd tensorflow_cmake_interface
$ mkdir build
$ cd build
$ cmake -DTENSORFLOW_SHARED=ON -DTENSORFLOW_SOURCE_DIR="$PROJ_WORKSPACE/tensorflow" ..
$ make
$ make install
After finishing build, the interface is installed at TensorflowCC="$PROJ_WORKSPACE/tensorflow_cmake_interface/build/lib/cmake/TensorflowCC"
Build UnrealGo with tensorflow cmake interface:
$ git clone https://github.com/unrealgo/unrealgo.git
$ mkdir build
$ cd build
$ cmake -DTensorflowCC="$PROJ_WORKSPACE/tensorflow_cmake_interface/build/lib/cmake/TensorflowCC" ../
$ make -j4
The binary is installed at $PROJ_WORKSPACE/unrealgo/bin. To run unrealgo, the network checkpoint(weights) is needed. A bootstrap checkpoint of 20 residual block is provided, decompress and copy to bin/
# first copy config/dlcfg-19 to bin/
$ cd $PROJ_WORKSPACE/
$ cp config/dlcfg-19 bin/
$ tar xvf bootstrap.tar.gz
$ mv bootstrap-ckpt.* bin/
# start unrealgo, type selfplay to start self play.
$ ./unrealgo
selfplay
meta_graph: path to metagraph
checkpoint: prefix of checkpoint files
nn_input: input name of the model inputs
nn_outputs: policy:value name pairs of the model outputs
value_transform: transform formula for network value output
0: v = v
1: v = -v
2: v = 1-v
reuse_searchtree: whether to reuse search tree
$ git clone https://github.com/SabakiHQ/Sabaki.git
$ cd Sabaki/dist/linux-unpacked
$ ./sabaki
# after sabaki is started, select [Engines]->[Attach] to attach unrealgo binary.
The hard part is neural network training. UnrealGo run in client mode by default, allowing user to run command 'selfplay' or 'clienteval'
- command 'selfplay' executes mcts search by single player with identical network weight
- command 'clienteval' executes mcts search by two players with different network weights.
The generated selfplay data or evaluation result will be synced to server for training/update.
UnrealGo is licensed under GPLv3 or later.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent