AI Arena Python Environment
To get started with our python environment you can run the training.py
file.
This file shows you how to do a few things in our environment:
- Initialize a new model
- Import a pretrained model
- Set up the game environment
- Run training with one-sided and selfplay reinforcement learning
- Save your model in the format that works with our researcher platform
We have set you up with a starter model in the starter_model
directory. This is a simple Policy Gradient that implements a version of the REINFORCE algorithm. We encourage you to replace this with your own models!
Additionally, we set up some basic training loops in the simulation_methods.py
file. Feel free to change these up and make them your own!
NOTE: There are two variables in the training.py
file which you should not change because our game requires these to be constant:
-
n_features
: This is the dimensionality of the state -
n_actions
: This is the dimensionality of the policy
Lastly, we have included the rules-based agent agent_sihing.py
(the researcher platform benchmark) in case you want to train specifically against it. But be careful about overfitting because we will introduce more benchmarks which require generalization...