sonnestolz/codes-for-rl-per

A novel DDPG method with prioritized experience replay (IEEE SMC 2017)

Codes for conference paper "A novel DDPG method with prioritized experience replay"

Demo videos

The following videos record the performance of our trained model running on five tasks in the OpenAI gym:

Requirements

• Tensorflow 1.4.0
• MuJoCo
• Gym 0.7.4

Install necessary components

conda create -n tensorflow_gpu pip python=2.7
source activate tensorflow_gpu
pip install --upgrade tensorflow-gpu==1.4
pip install gym==0.7.4
pip install mujoco-py==0.5.5

Run the code

source activate tensorflow_gpu
cd PER-in-RL
CUDA_VISIBLE_DEVICES=0 python run_ddpg_mujoco.py

Notes

export MUJOCO_PY_MJKEY_PATH=/path/to/mjpro131/bin/mjkey.txt
export MUJOCO_PY_MJPRO_PATH=/path/to/mjpro131

You need to have the above mujoco key file in your path. Now, you can reproduce the results in our paper.

Cite RL-PER

@inproceedings{hou2017novel,
  title={A novel DDPG method with prioritized experience replay},
  author={Hou, Yuenan and Liu, Lifeng and Wei, Qing and Xu, Xudong and Chen, Chunlin},
  booktitle={Systems, Man, and Cybernetics (SMC), 2017 IEEE International Conference on},
  pages={316--321},
  year={2017},
  organization={IEEE}
}

Acknowledgement

This repo is built upon Tensorflow-Reinforce and prioritized-experience-replay

Contact

If you have any problems in reproducing the results, just raise an issue in this repo.