Code accompanying the paper "Off-Policy Primal-Dual Safe Reinforcement Learning".
cd ./env/safety-gym/
pip install -e .
We follow the environment implementation in the CVPO repo to accelerate the training process. All the compared baselines in the paper are also evaluated on this environment. For further description about the environment implementation, please refer to Appendix B.2 in the CVPO paper.
Refer to https://github.com/openai/mujoco-py.
Configurations for experiments, environments, and algorithmic components as well as hyperparameters can be found in /arguments.py
.
For Safety-Gym tasks:
python main.py --env_name Safexp-PointButton1-v0 --num_epoch 500
For MuJoCo tasks:
python main.py --env_name Ant-v3 --num_epoch 300 --c 100 --qc_ens_size 8
We adopt the same hyperparameter setting across all Safety-Gym tasks tested in our work (PointButton1, PointButton2, CarButton1, CarButton2, PointPush1), which is the default setting in /arguments.py
.
The configurations different from the default setting are as follows:
-
The conservatism parameter
$k$ (--k
in/arguments.py
) is 0. for Humanoid. -
The convexity parameter
$c$ (--c
) is 100 for Ant, and 1000 for HalfCheetah and Humanoid. -
The replay ratio (
--num_train_repeat
) is 20 for HalfCheetah. -
The ensemble size
$E$ of the safety critic (--qc_ens_size
) is 8 for all MuJoCo tasks (may be smaller, like 4, for Hopper and Humanoid).In my test runs, thanks to the batch matrix multiplication function provided by PyTorch, the size of the ensemble does not significantly affect the running speed.
-
The option
--intrgt_max
is True for Humanoid.While in CAL conservatism is originally incorporated in policy optimization, for the Humanoid task we found it more effective to instead incorporate conservatism into
$Q_c$ learning.
The codebase contains wandb as a visualization tool for experimental management. The user can initiate a wandb experiment by adding --use_wandb
in the command above and specifying the wandb user account by --user_name [your account]
.
@article{wu2024off,
title={Off-Policy Primal-Dual Safe Reinforcement Learning},
author={Wu, Zifan and Tang, Bo and Lin, Qian and Yu, Chao and Mao, Shangqin and Xie, Qianlong and Wang, Xingxing and Wang, Dong},
journal={arXiv preprint arXiv:2401.14758},
year={2024}
}