https://arxiv.org/abs/2405.16158
The repository contains the implementation of the BRO algorithm (NeurIPS 2024 spotlight) that can be used to reproduce our results. The codebase is heavily inspired by JaxRL and Parallel JaxRL.
To run the BRO algorithm:
python3 train_parallel.py --benchmark=dmc --env_name=dog-run --num_seeds=10 --updates_per_step=10
To run the BRO (fast) version simply reduce the replay ratio to 2:
python3 train_parallel.py --benchmark=dmc --env_name=dog-run --num_seeds=10 --updates_per_step=2
To install the dependencies for the DMC experiments, run 'pip install -r jaxreqs.txt'. Due to incompatibilities, MetaWorld and MyoSuite has to be installed in separate environments.
- NewMujoco branch - we migrated BRO to new mujoco versions, as well as moved from gym to gymnasium
- BiggerRegularizedOptimistic Torch - we plan to release a minimal implementation of BRO in torch (https://github.com/naumix/BiggerRegularizedOtimistic_Torch). Work in progress
If you find this repository useful, feel free to cite our paper using the following bibtex.
@inproceedings{
nauman2024bigger,
title={Bigger, Regularized, Optimistic: scaling for compute and sample-efficient continuous control},
author={Michal Nauman and Mateusz Ostaszewski and Krzysztof Jankowski and Piotr Miłoś and Marek Cygan},
booktitle={Advances in Neural Information Processing Systems},
year={2024},
url={https://arxiv.org/pdf/2405.16158},
}