Skip to content

Code for Latent Action Space for Offline Reinforcement Learning [CoRL 2020]

License

Notifications You must be signed in to change notification settings

Wenxuan-Zhou/PLAS

Repository files navigation

PLAS: Latent Action Space for Offline Reinforcement Learning

This is the repository for our paper "PLAS: Latent Action Space for Offline Reinforcement Learning" in CoRL 2020. Please visit our website for more information.

This repository is built on top of BCQ. The logger is from BEAR.

Requirements

You may install the above packages following the instructions in their repositories, or run the following command:

pip3 install -r requirements.txt

Note that the latest d4rl repository has some problem loading the mujoco dataset. We recommend the users to install this commit version.

Instructions

To train the Latent Policy for the d4rl datasets:

python main.py --env_name walker2d-medium-expert-v0 --algo_name Latent --max_latent_action 2

To train the Latent Policy with the perturbation layer:

python main.py --env_name walker2d-medium-expert-v0 --algo_name LatentPerturbation --max_latent_action 2 --phi 0.05

By default, the algorithm trains a VAE before the policy to model the behavior policy of the dataset. You may also load a pre-trained vae and then train policy.

python main.py --env_name walker2d-medium-expert-v0 --algo_name Latent --vae_mode v6

This command will load the vae models under the "models/vae_v6" folder according to the name of the dataset and the random seed automatically.

The results will be saved under the "results" folder. You may use viskit to visualize the curves.

Citation

@inproceedings{PLAS_corl2020,
 title={PLAS: Latent Action Space for Offline Reinforcement Learning},
 author={Zhou, Wenxuan and Bajracharya, Sujay and Held, David},
 booktitle={Conference on Robot Learning},
 year={2020}
}

About

Code for Latent Action Space for Offline Reinforcement Learning [CoRL 2020]

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages