Batch Reinforcement Learning with Hyperparameter Gradients

This repository is the official implementation of Batch Reinforcement Learning with Hyperparameter Gradients.

Requirements

To install requirements:

conda env create -f environment.yml
conda activate batchrl

To download the batch trajectories used in the paper, please run the following:

python download_dataset.py

Finite MDP experiments

To run the finite MDP experiments in the paper, run this command:

python finite_run.py

Continuous control experiments

To run the Mujoco continuous control experiments in the paper, run this command:

python cont_run.py

References

If this repository helps you in your academic research, you are encouraged to cite our paper. Here is an example bibtex:

@inproceedings{lee2020batch,
	title={Batch Reinforcement Learning with Hyperparameter Gradients},
	author={Byung-Jun Lee* and Jongmin Lee* and Peter Vrancx and Dongho Kim and Kee-Eung Kim},
	booktitle={Proceedings of the 37th International Conference on Machine Learning},
	year={2020}
}