Facebear-ljx/RGM

The official implementation of "Mind the Gap: Offline Policy Optimization for Imperfect Rewards" (ICLR2023)

Mind the Gap: Offline Policy Optimization for Imperfect Rewards (ICLR 2023)

This is the official implementation of RGM (Reward Gap Minimization) (https://openreview.net/forum?id=WumysvcMvV6). RGM can be perceived as a hybrid offline RL and offline IL method that can handle diverse types of imperfect rewards include but not limited to partially correct reward, sparse reward, multi-task datasharing setting and completely incorrect rewards.

RGM formalizes offline policy optimization for imperfect rewards as a bilevel optimization problem, where the upper layer optimizes a reward correction term that performs visitation distribution matching w.r.t. some expert data and the lower layer solves a pessimistic RL problem with the corrected rewards.

Usage

To install the dependencies, use

    pip install -r requirements.txt

If you want conduct experiments on Robomimic datasets. You need to install Robomimic according to the instructions in Robomimic

Benchmark experiments

You can reproduce the Mujoco tasks and Robomimic tasks like so:

    bash run d4rl.sh

    bash run_robomimic.sh

For the experiments on multi-task datasharing setting, we'll release soon.

Visulization of Learning curves

You can resort to wandb to login your personal account via export your own wandb api key.

export WANDB_API_KEY=YOUR_WANDB_API_KEY

and run

wandb online

to turn on the online syncronization.

If you find our code and paper can help, please cite our paper as:

Bibtex

@inproceedings{
li2023mind,
title={Mind the Gap: Offline Policy Optimization for Imperfect Rewards},
author={Jianxiong Li and Xiao Hu and Haoran Xu and Jingjing Liu and Xianyuan Zhan and Qing-Shan Jia and Ya-Qin Zhang},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=WumysvcMvV6}
}