https://github.com/uduse/moozi

https://github.com/uduse/moozi

Last synced: about 1 year ago
JSON representation

• Host: GitHub
• URL: https://github.com/uduse/moozi
• Owner: uduse
• Created: 2021-05-18T16:47:24.000Z (about 5 years ago)
• Default Branch: master
• Last Pushed: 2024-02-05T18:15:20.000Z (over 2 years ago)
• Last Synced: 2025-04-10T02:16:53.591Z (about 1 year ago)
• Language: Python
• Size: 6.16 MB
• Stars: 8
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# MooZi: A General-Game-Playing System

*MooZi* is a high-performance general-game-playing system that plans with a learned model.

The *MooZi* core learning algorithm is based on [*MuZero*](https://www.deepmind.com/blog/muzero-mastering-go-chess-shogi-and-atari-without-rules) with a few extensions.

*MooZi* uses [JAX eco-system](https://www.deepmind.com/blog/using-jax-to-accelerate-our-research) and [Ray](https://github.com/ray-project/ray) for the best performance.

## Supported Games

*MooZi* current supports all [MinAtar](https://github.com/kenjyoung/MinAtar) games and some two-players games from [OpenSpiel](https://github.com/deepmind/open_spiel).

## Architecture

*MooZi* deploys a [hierarchial control](https://arxiv.org/abs/1712.09381) paradigm to parallizes its workers.

![*MooZi* Architecture](assets/moozi_architecture.png)

![*MooZi* Training Step](assets/training_step.png)

## Tools

*MooZi* is shipped with tools to help you montior understand the system.

### Tensorboard

*MooZi* Tensorboard hows all important metrics regarding the behavior and performance of the system.

![*MooZi* Tensorboard](assets/tb.png) 

### Hidden Space Visualizer

*MooZi* hidden space visualizer collects hidden states during acting and visualize them either in 3D or 2D.

![*MooZi* t-SNE 2D Projector](assets/tsne.png)