Prototype-Wrapper Network (PW-Net) - ICLR-2023

This is the repository for the paper "Towards Interpretable Deep Reinforcement Learning with Human-Friendly Prototypes"¹.

The paper was published with a spotlight presentation at ICLR 2023, Kigali, Rwanda, see paper here.

The paper proposes a "wrapper" component which can be added to any pre-trained agent to make it interpretable-by-design. Specifically, the original agent $f_{enc}$ is used as an encoder, which produces an output $z$. This representation is then transformed by going through multiple different MLP networks $h_{i,j}$ which then assess its similarity to pre-defined human-interpretable prototypes $p_{i,j}$ for common concepts used in the domain. For example, in the figure above, common concepts used in driving are "right turn", "left turn", "accelerate", and "brake". So, we defined these a-priori alongside the output weight matrix $W'$, which together form an interpretable design for the agent.

In the above figure for example, the agent sees its current state $S$ as being most similar to the prototypes for "turning left" and and "accelerating", so the similarity function outputs a high value for these, which then produces the final output action.

This repo uses a lot of the code from Jain (2022)², from which we use the pre-trained agent for the Car Racing domain.

---

Car Racing (updated Dec 2024: You must use Numpy < 2)

To reproduce these results for the Car Racing Domain Simply copy the whole repo and run these commands in the terminal whilst in the "Car Racing" directory:

python3 -m venv pwnet
source pwnet/bin/activate
pip3 install --upgrade pip
pip install toml
pip install numpy
pip3 install torch torchvision torchaudio
pip install gym'[box2d]'
pip install tqdm
pip install gym==0.24.0
pip install gym-notices==0.0.7
pip install scikit-learn
pip install numpy==1.26.4
pip install matplotlib

Then run

Python collect_data.py

When you have the collected data you can now train the various wrappers. Simply run...

Python run_kmeans.py
Python run_pwnet*.py
Python run_pwnet.py

And the terminal will print off the results, reproducing the results for Car Racing from the paper.

---

Atari Pong

For the Atari Pong repeat the above process for car racing, but this time install this environment:

python3 -m venv pwnet
source pwnet/bin/activate
pip3 install --upgrade pip
pip install gym==0.21.0  
pip install opencv-python
pip3 install torch torchvision torchaudio
pip install atari-py==0.2.9
pip install gym'[atari]'
pip install gym'[accept-rom-license]'
pip install scikit-learn
pip install toml

The code in this directory is heavily based upon Şentürk (2020)³

---

To-do List

1. Make package for easier usage.

For any questions or comments, please email ekenny@mit.edu.

---

Bibtex

@inproceedings{
kenny2023towards,
title={Towards Interpretable Deep Reinforcement Learning with Human-Friendly Prototypes},
author={Eoin M. Kenny and Mycal Tucker and Julie Shah},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=hWwY_Jq0xsN}
}

Footnotes

1. Kenny, E.M., Tucker, M. and Shah, J., Towards Interpretable Deep Reinforcement Learning with Human-Friendly Prototypes. In The Eleventh International Conference on Learning Representations. Kigali, Rwanda, 2023. (Spotlight, notable paper) ↩

2. Jinay Jain. Jinayjain/deep-racing: Self-driving racecar using reinforcement learning (proxi- mal policy optimization) in pytorch, 2022. URL https://github.com/JinayJain/ deep-racing. ↩

3. Behc¸et Şentürk. https://github.com/bhctsntrk/openaipong-dqn, 2022. URL ¨ https://github. com/nikhilbarhate99/Actor-Critic-PyTorch. ↩