XRL-Bench: A benchmark for Explainable Reinforcement Learning
XRL-Bench is a comprehensive benchmark suite for evaluating eXplainable Reinforcement Learning (XRL) methods. It provides a standard and unified platform for researches and practitioners to develop, test, and compare their state importance-based XRL algorithms. XRL-Bench includes a wide range of environments, explainers, and evaluation methods to facilitate the development of state-of-the-art XRL techniques.
See our KDD paper for details and citations.
Key Features
- Environments: XRL-Bench includes diverse open-source environments based on tabular state form, such as Dunk City Dynasty (only supported in Python3.8), Lunar Lander, CartPole, and Flappy Bird (supported in Python3.9 or higher version), and image state form, such as Breakout, Pong, covering a wide range of reinforcement learning tasks.
- Explainers: The benchmark supports various state importance-based explainers like TabularSHAP, TabularLIME, SARFA, Perturbation Saliency, Integrated Gradient, DeepSHAP and GradientSHAP, to understand the decision-making process of the RL agents.
- Evaluation Methods: XRL-Bench provides multiple evaluation methods, including Fidelity-focused methods such as AIM, AUM, PGI and PGU, and Stability-focused methods such as RIS.
Getting Started
To get started with XRL-Bench, follow the instruction below to set up the required dependencies
git clone https://github.com/fuxiAIlab/xrl-bench.git
cd xrl-bench
pip install -r requirements.txtand run the example code
from xrlbench.explainers import Explainer
from xrlbench.evaluator import Evaluator
from xrlbench.environments import Environment
env = Environment(environment_name="LunarLander")
dataset = env.get_dataset()
actions = dataset['action']
states = dataset.drop(['action', 'reward'], axis=1)
explainer = Explainer(method="tabularShap", state=states, action=actions)
shap_values = explainer.explain()
evalutor = Evaluator(metric="AIM", environment=env)
aim = evaluator.evaluate(states, actions, shap_values, k=1)LeaderBoards
We maintain two leaderboards to compare the AUC performance of explainers for both types of state form under various evaluation metrics.
Tabular states (DunkCityDynasty)
| Explainer | AIM | AUM | PGI | PGU | RIS |
|---|---|---|---|---|---|
| TabularSHAP | 0.214 | 0.894 | 0.905 | 0.662 | 1.023 |
| DeepSHAP | 0.337 | 0.493 | 0.790 | 0.712 | 1.261 |
| GradientSHAP | 0.326 | 0.523 | 0.766 | 0.690 | 1.419 |
| Integrated Gradient | 0.323 | 0.522 | 0.808 | 0.688 | 1.465 |
| SARFA | 0.361 | 0.709 | 0.952 | 0.665 | 1.071 |
| Perturbation Saliency | 0.364 | 0.687 | 0.951 | 0.680 | 1.357 |
| TabularLIME | 0.215 | 0.779 | 0.954 | 0.274 | 1.901 |
Tabular states (LunarLander)
| Explainer | AIM | AUM | PGI | PGU | RIS |
|---|---|---|---|---|---|
| TabularSHAP | 0.116 | 0.693 | 5.258 | 4.895 | 2.646 |
| DeepSHAP | 0.188 | 0.663 | 5.988 | 4.321 | 2.623 |
| GradientSHAP | 0.203 | 0.614 | 5.963 | 4.317 | 3.134 |
| Integrated Gradient | 0.203 | 0.618 | 5.930 | 4.375 | 2.800 |
| SARFA | 0.388 | 0.363 | 4.953 | 5.169 | 6.408 |
| Perturbation Saliency | 0.382 | 0.353 | 4.847 | 5.528 | 4.878 |
| TabularLIME | 0.323 | 0.472 | 6.179 | 3.755 | 2.871 |
Image states (Breakout)
| Explainer | AIM | AUM | PGI | PGU | RIS |
|---|---|---|---|---|---|
| DeepSHAP | 0.162 | 0.630 | 1.748 | 0.347 | 0.375 |
| GradientSHAP | 0.260 | 0.655 | 1.755 | 0.384 | 0.659 |
| Integrated Gradient | 0.292 | 0.652 | 1.812 | 0.364 | 0.109 |
| SARFA | 0.253 | 0.270 | 1.225 | 0.991 | 0.653 |
| Perturbation Saliency | 0.258 | 0.387 | 1.370 | 0.621 | 0.649 |
Contributing
We welcome contirbutions to XRL-Bench! If you'd like to contirbute, please follow these guidelines:
- Fork the repository and create a new branch for your feature or bug fix.
- Add your changes and include tests to ensure your changes work as expected.
- Update the documentation as necessary.
- Submit a pull request and wait for a review from the maintainers.
License
XRL-Bench is under MIT license.