grimmer0125 / alphago-zero-tictactoe-js

A game framework based on AlphaZero/TensorFlow.js runs in browser to demonstrate tic-tac-toe AI game. Use a pre-trained model or train from scratch. Ported from suragnair/alpha-zero-general (Python)

AlphaGo Zero Tictactoe JS

Try it: https://grimmer.io/alphago-zero-tictactoe-js/. Google DeepMind AlphaGo uses enhancement learning and the algorithm is a composite of

1. Policy Network
2. Value Network
3. Monte Carlo tree search (MCTS)

Installation

npm install

Dev

npm start // build and launch its live dev web server.

After npm start, you can also use VS Code with Debugger for Chrome extension to debug.

Deployment

npm run deploy

Features and done itmes of todo list

1. Ported the algorithms from alpha-zero-general. Although its name is alpha-zero-general, it is based on AlphaGo Zero algorithm.

2. Import pretrained models from alpha-zero-general and run alphago game algorithms on Browsers. alpha-zero-general is a project to supply general game AI training frameworks. You can extend that project and add yourself game rule codes and train AI model by using Python.

TODO

1. Fix bugs to train models by this JavaScript version project. It may be a TensorFlow.js bug. Maybe waitting for native TensorFlow Node.js binding is better than WebGL solution.
2. Add UI.
3. Clean up.
4. Use service worker for cpu heavy loading part.
5. Use TypeScript instead

AlphaZero

To overcome some API limitation (Tensorflow.js export/save model/weights), so this JavaScript repo borrows one of the features of AlphaZero, always accept trained model after each iteration without comparing to previous version