This project was a final project for the deep learning module of ESPCI Paris - PSL. It is an implementation of the Hive Board game with a reinforcement learning algorithm inspired by AlphaZero and with a connection-based representation of the game.
Here is a GIF of the generation of a game.
If you have computational power and want to train it even further, look at the
train.py file which will show you all you need. Basically you have 2
methods for the trainer, the start_play which will make you play
against the currently loaded algorithm and the train one which will train and
save the model after each iteration of the training process.
You can also load a previous model thanks to the load_history_and_net
method. There's already the latest weights and game histories of the pretraining
and training phase described in the PDF