Taming MAML: Efficient unbiased meta-reinforcement learning

Reference Tensorflow implementation of Taming MAML: Efficient unbiased meta-reinforcement learning. We will release Pytorch version later.

Getting started

You can use Dockerfile to build an image with conda environment called tmaml included, activating this conda env:

conda activate tmaml

you can also use tmaml.yml to create a conda env called tmaml.

conda env create -f tmaml.yml

then activate this conda env

conda activate tmaml

Usage

You can use the tmaml_run_mujoco.py , vpg_run_mujoco.py and dice_vpg_run_mujoco.py scripts in order to run reinforcement learning experiments with different algorithm. MAML:

python vpg_run_mujoco.py --env HalfCheetahRandDirecEnv

MAML + DICE:

python dice_vpg_run_mujoco.py --env HalfCheetahRandDirecEnv

TMAML:

python tmaml_run_mujoco.py --env HalfCheetahRandDirecEnv

References

To cite TMAML please use

@InProceedings{pmlr-v97-liu19g,
  title = 	 {Taming {MAML}: Efficient unbiased meta-reinforcement learning},
  author = 	 {Liu, Hao and Socher, Richard and Xiong, Caiming},
  booktitle = 	 {Proceedings of the 36th International Conference on Machine Learning},
  pages = 	 {4061--4071},
  year = 	 {2019},
  editor = 	 {Chaudhuri, Kamalika and Salakhutdinov, Ruslan},
  volume = 	 {97},
  series = 	 {Proceedings of Machine Learning Research},
  address = 	 {Long Beach, California, USA},
  month = 	 {09--15 Jun},
  publisher = 	 {PMLR},
}

TODOs

•  Adding TMAML

•  Adding MAML

•  Adding DICE

•  Benchmarking

•  Pytorch version

Acknowledgements

This repository is based on ProMP repo.