Codes for conference paper "A novel DDPG method with prioritized experience replay"
The following videos record the performance of our trained model running on five tasks in the OpenAI gym:
- Tensorflow 1.4.0
- MuJoCo
- Gym 0.7.4
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
source activate tensorflow_gpu
cd PER-in-RL
CUDA_VISIBLE_DEVICES=0 python run_ddpg_mujoco.py
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.
@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}
}
This repo is built upon Tensorflow-Reinforce and prioritized-experience-replay
If you have any problems in reproducing the results, just raise an issue in this repo.