HyperDQN: A Randomized Exploration Method for Deep Reinforcement Learning

HyperDQN is a randomized exploration based on Deep Q-Network (DQN). The paper can be found here.

Requirement

Experiments are based on Python 3.6. Packages can be installed by the following cmd:

pip install -r requirement.txt

Note that our implementation highly relies on tianshou==0.4.1.

Usage

bash scripts/hyper_dqn/run_atari.sh

Training results can be found in the results folder.

Bibtex

@inproceedings{
    li2022hyperdqn,
    title={Hyper{DQN}: A Randomized Exploration Method for Deep Reinforcement Learning},
    author={Ziniu Li and Yingru Li and Yushun Zhang and Tong Zhang and Zhi-Quan Luo},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=X0nrKAXu7g-}
}

Acknowledgment

Our codebase is based on the Tianshou framework.