dlxj / alphago-zero-tictactoe-js Goto Github PK

View Code? Open in Web Editor NEW

A JavaScript + TensorFlow.js Porting of https://github.com/suragnair/alpha-zero-general (Python). 1. Import Keras pretrained model and run it successfully 2. self-train from the beginning also works on the browser. Try it: https://grimmer.io/alphago-zero-tictactoe-js/, use "npm run deploy" to upload. 2 output of the network: policy and value.

License: MIT License

JavaScript 93.62% HTML 4.59% CSS 1.80%

alphago-zero-tictactoe-js's Introduction

alphago-zero-tictactoe-js

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.

Features and done itmes of todo list

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

~~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.~~
~~Add UI.~~
Clean up.
Use service worker for cpu heavy loading part.

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

Recommend Projects

dlxj / alphago-zero-tictactoe-js Goto Github PK