Renju

Main information

This is educational project aimed to understand and implement ideas and algorithms of deepmind AlphaGo program, but in game called Gomoku (simpler variant of renju, also called five in a row).

This repository contains game enviroment with gui and implementation of different agents:

• Human
• Dummy (makes random moves)
• Supervised Learning agent (uses policy network trained on expert moves dataset)
• BeamSearch agent (uses same policy network with beam search to improve it)
• Monte Carlo Tree Search agent (uses policy and rollout networks with mcts algorithm)

Python dependencies

You can install all python requirements in virtualenv

virtualenv -p /path/to/python3.6.5 renju_env
renju_env/bin/activate
pip install -r requirements.txt

If you are using ubuntu you also need to install tkinter

sudo apt-get update
sudo apt-get install python3-tk

Training from scratch

To train model from scratch you can use ./train script. Run ./train --help for details. If you then want to use MCTS agent you need to convert keras model to tensorflow and freeze graph with ./keras2tf script.

Building MCTS agent

The code is tested on macOS 10.12 and Ubuntu 14.04. There are two ways of building this code:

1. You can use cmake to compile project and link dynamic tensorflow library (library built for macOS is in src/libs/ (works faster).

cd src
wget https://www.dropbox.com/s/skszoznf0oih11f/libs.tar.gz
tar -xzf libs.tar.gz
rm libs.tar.gz
mkdir cmake-build && cd cmake-build
cmake .. -G"Unix Makefiles"
make bernard

2. If you don't want to use option 1. you can copy src folder inside tensorflow source directory and use bazel to build binary (this option will work slower for the first time)

git clone --recursive https://github.com/tensorflow/tensorflow /tmp/tensorflow
cp -r ./src /tmp/tensorflow/tensorflow
cd /tmp/tensorflow/
./configure # answer all the questions
cd tensorflow/src
bazel build :bernard # executable will be inside /tmp/tensorflow/bazel-bin/tensorflow/src/

Demo

To run demo simply use script ./demo --help. This will launch python GUI with chosen agents. You could also run test games between different agents (except human).