alphago-lite

A lightweight implementation of Alphago style training pipeline based on the book

• Deep Learning and the game of Go, and the paper
• Mastering the game of Go with Deep Neural Networks and Tree Search

implementation

• fast policy network
• strong policy network
• self-play loop -q learning, policy learning
  • policy gradient learning
  • Q-learning
• Bot vs Bot gameplay
• web interface(next.js frontend)
• train strong policy network
• alphago style self-play
• benchmark against known bots

implementation overview

The project roughly follows the pipeline described in Deep Learning and the Game of Go.

1. supervised learning

Train a policy network to imitate human moves from professional Go games.

2. policy improvement

Improve the policy using reinforcement learning through self-play.

3. evaluation

Evaluate trained agents by running automated matches between models.

dataset

This project uses the Computer Go Dataset:

https://github.com/yenw/computer-go-dataset.git

references

• Deep Learning and the Game of Go
• Silver et al., Mastering the Game of Go with Deep Neural Networks and Tree Search https://www.nature.com/articles/nature16961