Plan-To-Predict

This code accompanies the paper "Plan To Predict: Learning an Uncertainty-Foreseeing Model for Model-Based Reinforcement Learning".

Installation

1. Install MuJoCo 1.50 at ~/.mujoco/mjpro150 and copy your license key to ~/.mujoco/mjkey.txt
2. Clone P2P

git clone https://github.com/ZifanWu/Plan-to-Predict.git

3. Create a conda environment and install Plan-to-Predict

cd src/Plan-to-Predict
conda env create -f environment/gpu-env.yml
conda activate p2p
pip install -e .

Usage

python src/main_p2p.py --num_epoch 150

Optimal parameters

The optimal parameters are contained in .src/configs/ folder.

Reference

@article{wu2022plan,
  title={Plan To Predict: Learning an Uncertainty-Foreseeing Model For Model-Based Reinforcement Learning},
  author={Wu, Zifan and Yu, Chao and Chen, Chen and Hao, Jianye and Zhuo, Hankz Hankui},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={15849--15861},
  year={2022}
}