Batch-Constrained Deep Q-Learning (BCQ)

Batch-Constrained deep Q-learning (BCQ) is the first batch deep reinforcement learning, an algorithm which aims to learn offline without interactions with the environment.

BCQ was first introduced in our ICML 2019 paper which focused on continuous action domains. A discrete-action version of BCQ was introduced in a followup Deep RL workshop NeurIPS 2019 paper. Code for each of these algorithms can be found under their corresponding folder.

Bibtex

@inproceedings{fujimoto2019off,
  title={Off-Policy Deep Reinforcement Learning without Exploration},
  author={Fujimoto, Scott and Meger, David and Precup, Doina},
  booktitle={International Conference on Machine Learning},
  pages={2052--2062},
  year={2019}
}

@article{fujimoto2019benchmarking,
  title={Benchmarking Batch Deep Reinforcement Learning Algorithms},
  author={Fujimoto, Scott and Conti, Edoardo and Ghavamzadeh, Mohammad and Pineau, Joelle},
  journal={arXiv preprint arXiv:1910.01708},
  year={2019}
}

sfujim/BCQ

Batch-Constrained Deep Q-Learning (BCQ)

Bibtex