Interpretable Reinforcement Learning using SAE discovered Feature Steering

Done for the Reprogramming AI hackathon by Apart Research. Currently the code is very messy and I'm not certain when I will clean it up.

There are two main ways to run the training code after getting the requirements:

1. Run the main.py file after modifying the if name == "main" block to run the desired experiment.

2. Use the main.ipynb notebook to run the code in a more interactive way which demonstrates how the code works and various experiment configurations.

Aside from the training code, there is a also a visualize.py script which can be used to visualize the steering of features for a given state. If you clone the repo, you can run the file directly. Otherwise, you will need to download the features.pkl file. The visualization UI is quite intuitive to use and you can cycle through labels in a cell by continuously clicking. Not all states will have feature steering vectors, so you can use the next board state button to cycle through states.