This repo is the official codebase of our following paper:
@article{yang2021recurrent,
title={Recurrent Off-policy Baselines for Memory-based Continuous Control},
author={Yang, Zhihan and Nguyen, Hai},
journal={Deep RL Workshop, NeurIPS 2021},
year={2021}
}Paper summary: We implement and benchmark recurrent versions of DDPG, TD3 and SAC that uses full history.
This repo offers:
- DDPG, TD3 and SAC (clean PyTorch implementation and benchmarked against stable-baselines3*)
- Recurrent versions of DDPG, TD3 and SAC that use full history: RDPG, RTD3 and RSAC
- Very easy to understand and use; see our exhaustive documentation: link
*The results of benchmarking can be found in issue "Performance check against SB3" in closed Issues.
For users:
- Please feel free to ask a code question through Issues.
- When cloning this repo, please consider using shallow clone as it is large due to a large number of commits.
News:
- [2021/11/18] Noticed that I forgot to document dependencies in documentation. Added.
Paper link:
https://arxiv.org/pdf/2110.12628.pdf
Poster (click to open in new tab for better resolution):
Star history: