kristian-georgiev/SGMRL

Codebase for SG-MRL (On the Convergence Theory of Debiased Model-Agnostic Meta-Reinforcement Learning)

Code for On the Convergence Theory of Debiased Model-Agnostic Meta-Reinforcement Learning

• NeurIPS 2021 version
• ArXiv version

(a previous version of the manuscript was entitled Provably Convergent Policy Gradient Methods for Model-Agnostic Meta-Reinforcement Learning)

This codebase is based off of the ProMP repository.

To run the setup from the numerical experiments section of our paper, please use sgmrl_experiments/vpg_run.py. Our modified implementation is located in maml_zoo/meta_algos/vpg_sgmrl.py.

Cite as

@article{fallah2021convergence,
  title={On the convergence theory of debiased model-agnostic meta-reinforcement learning},
  author={Fallah, Alireza and Georgiev, Kristian and Mokhtari, Aryan and Ozdaglar, Asuman},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  pages={3096--3107},
  year={2021}
}