First of all, this is NOT a submodule. This is a separate area for python scripts related to training. We want to keep a record of some important training scripts but they are not decent enough to be merged into the mainn library. This folder exists only for documenting purposes. Please proceed with caution.
The best method so far is to use JAX to implement MCTS+neural network like AlphaZero for our game, because we have the luxury of encoding the game into our policy. The general training script is this one (in practice, one can tweak it all you like).
The most straightforward and perhaps lazy way to solve an RL problem is to plug into a general-purpose RL trainer and let it find a policy for the players. It wraps the environment into OpenAI Gym, and there are some amazing implementations that connects to these (stable-baselines3, etc).
This script uses our gym wrapper gym_env to train using stable_baselines3.
The JAX training script. It runs and is our best effort so far (with the right hyper-parameters of course).