| --- |
| title: "Faïence" |
| emoji: "🀄" |
| colorFrom: blue |
| colorTo: yellow |
| sdk: static |
| pinned: true |
| thumbnail: https://remifabre-faience.static.hf.space/social.png |
| short_description: "Azul rules vs a self-play neural net, all in your browser" |
| --- |
| |
| <!-- The card for the Faïence Space. This file lives in the ludometer repo at |
| web/player/space/README.md; scripts/deploy_player.sh copies it to the |
| Space root at deploy time (the space/ directory itself is never shipped |
| to visitors). --> |
| |
| Faïence is a free, open-source implementation of the rules of *Azul*, the |
| tile-laying game by Michael Kiesling: a fan project for research, with its own |
| code and artwork, not affiliated with or endorsed by the game's publishers. |
|
|
| Your opponent is a neural network that learned the game from scratch by |
| self-play. The net and its tree search run entirely in your browser: your |
| browser downloads the model once and everything after that is local. No |
| server plays for it, no account, no ads, nothing to buy. |
|
|
| It began as a machine-learning experiment: Ludometer, a research framework |
| that measures how good a board game is from the shape of an AI's learning |
| curve. The games people play here are the research material: when a game |
| ends, the page sends an anonymous record (moves, deals, net, score; sharing |
| can be switched off in Settings) to the public dataset |
| [RemiFabre/faience-games](https://huggingface.co/datasets/RemiFabre/faience-games), |
| where it becomes training data. |
|
|
| Code, training logs and methodology: |
| [RemiFabre/ludometer](https://github.com/RemiFabre/ludometer) (MIT). |
|
|
