Skip to content

Latest commit

 

History

History
82 lines (47 loc) · 3.34 KB

README.md

File metadata and controls

82 lines (47 loc) · 3.34 KB

Queen Bee

hiveGo

Hive game, backed by an AlphaZero based AI.

Ant

Warning

🚧🛠 Currently the AI is broken 🚧🛠
This is a 2017/2018 project built with TensorFlow, that I'm bringing back to life, this time using GoMLX.

Command line version

Simplest version to run. Hotseat mode (no AI) is already working:

    go run github/janpfeifer/hiveGo/cmd/hive

image

Gnome Version

Easy to build in XWindows -- well, more or less, depending on the dependencies working out fine. Probably one could compile it in windows as well ...

TODO: use Fyne to make it portable.

    go run github/janpfeifer/hiveGo/cmd/gnome-hive

image

Web Version

Spider

The Gnome version works nicely ... but asking anyone to install it is cruel. And I wouldn't want to distribute a binary -- then I would have to try to compile everything staticly.

So instead, why not a web version ... duh, I should have done this first, but I wanted to try out coding for Gnome.

Anyway, GopherJS, a compiler of Go to js, made this very easy. SVG + JQuery libraries for GopherJS are not well documented ... so it took a little while to learn the ropes ... but the nice thing is: it all work on the web. Unfortunately too slow ... this game speed (or slowness) is dominated by generating the list of moves (updating connectivity of the graph and moves for Ants), and that can't be pushed to tensorflow.js. So the fully javascript solution is not very viable for higher levels of play.

And since the Go code that runs the game can also run in compiled javascript, the whole game runs on the machine -> TODO: provide a link to a static version.

TODO: Convert the TensorFlow model to TensorFlow.js, so we can get good models also running on the client browser.

Trainer

Allows playing games among AI's automatically.

  • Play new games. Optionally save the games.
  • Load and rescore old games.
  • Compare AIs
  • Train models while playing the game.
  • Can train TF models.
  • Can distill from previous models: very handy to quickly ramp up a new model to a moderate quality model.

Hexagonal Convolutional model

The latest, more fancy model, uses the full board as input to the NN, and runs a bunch of layers of convolution on that, with residual connections.

The convolutions for hexagonal mappings require some extra care, since the neighbourhood kernel is different depending if one are on odd/even columns.

See experiments in this colab

Thanks

  • Thanks for Florence Poirel for the awesome drawings!
  • The project started from earlier version in Python, but it was greatly improved since.

Background pictures: