nebw/argos-zero

A MCTS computer go program using convolutional neural networks

Argos Zero

This is an attempt to replicate Mastering the game of Go without human knowledge. The board representation is based on the library of efficient go routines by Łukasz Lew.

Build instructions

Requirements:

• MXNet
• CMake
• Git
• Cap'n Proto

Optional requirements:

• CUDA / cuDNN
• Intel MKL
• NNPACK

How to build:

• (Install optional depedencies)
• Install MXNet (official instructions)
  • git clone --recursive --branch 1.1.0 https://github.com/apache/incubator-mxnet
  • Configure dependencies in make/config.mk file, USE_CPP_PACKAGE should be set to 1
  • make -j NUM_CPUS (replace NUM_CPUS by the number of CPUs of you computer)
• Install MXNet python package
  • Either: cd python; pip install --user -e .
  • Or: Install precompiled pypi package
• Install Cap'n Proto package
  • Either from official web page
  • Or: install with package manager apt-get install capnproto
• Download a pretrained model:
• Compile Argos Zero:
  • Clone the repository: git clone https://github.com/nebw/argos-zero.git
  • Edit the config file argos-zero/src/argos/Config.h
  • Create build folder: mkdir argos-zero-build; cd argos-zero-build
  • Run CMake: cmake -DCMAKE_BUILD_TYPE=Release -DMXNET_PATH=/your/local/mxnet/path /your/local/argos/zero/path
  • make -j NUM_CPUS (replace NUM_CPUS by the number of CPUs of you computer)

Finally, you can try the selfplay (see Command Line section for details) mode by running: argos-zero-build/src/selfplay. You can play against the program using any Go GUI that supports the GTP protocol using the argos-zero-build/src/gtp binary. A modern Go GUI is Sabaki.

Command Line

usage: selfplay networkPath [optional options]

required options:
  -n [ --networkPath ] arg         path to trained model

optional options:
  -h [ --help ]                    print this message and exit
  -c [ --config ] arg              path to config file
  -l [ --logfilePath ] arg         path where log file should be stored
                                   by default logs are stored in 
                                   `<timestamp>.log` file in the same directory
                                   as network
  -d [ --deviceType ] arg          set device type (CPU, GPU, CPU_PINNED)
  -s [ --server ] arg              set server ip, default is: 127.0.0.1
  -p [ --port ] arg                set port
  --tree-numEvaluationThreads arg  set numEvaluationThreads
  --tree-numThreads arg            set numThreads
  --tree-randomizeFirstNMoves arg  set randomizeFirstNMoves
  --tree-numLastRootNodes arg      set numLastRootNodes
  --tree-virtualPlayouts arg       set virtualPlayouts
  --tree-expandAt arg              set expandAt
  --tree-priorC arg                set priorC
  --tree-networkRollouts arg       enable|disable network rollouts
  --tree-trainingMode arg          enable|disable training mode
  --time-c arg                     set time c
  --time-maxPly arg                set maxPly
  --time-delay arg                 set delay
  --engine-totalTime arg           set totalTime in milliseconds
  --engine-resignThreshold arg     set resignThreshold