AlphaGo Zero based RL agent

Made during 'AI Weekend' in Stockholm.

Structure

├── README.md
├── agz.py              # MCTS logic. File can also visualise etc
├── goboard.py          # Go implementation code
├── scoring.py          # Go implementation code
├── gostate.py          # Go environment wrapping goboard, scoring
├── gostate_pachi.py    # Go environment wrapping the fast pachi implementation
├── resnet.py           # Neural network for evaluating board positions
├── policyvalue.py      # Predictor class wrapping the resnet CNN
└── training.py         # Training loop performing self play 

Installation

Requires pachi-py.

pip install numpy
pip install keras
pip install tensorflow

python agz.py

Todo

• Cleanup code structure with folders etc
• Implement random reflections of board
• Tune how much time is spend exploring / training (c.f. AGZ paper)
• Parallelize training and simulation.
• Use code from agz.play_game to create MCTSAgent class
• Use same model on other environments
• Learn the transition dynamics of step(state, action)
• Refactor MCTSAgent to implement .update_state and .decision methods