sweetice / MEPE

Official implementation of MEPE

Code for paper: A Minimalist Ensemble Policy Evaluation Operator for Deep Reinforcement Learning

Install

Optionally: Create a new python virtual environment

We recommend readers use the anaconda tool.

conda create -n mepe python=3.7.4

And then activate the created environment

conda activate mepe

Firstly, install the requirements.txt

pip install -r requirements.txt

The required python package is as follows:

torch==1.7.1+cu101
numpy==1.18.1
gym==0.12.1
pybullet==2.7.1
roboschool==1.0.48
pandas==1.0.1
tqdm
scikit-image
tensorboard
torch
torchvision
gym==0.25.1
patchelf
termcolor
seaborn==0.9.0
opencv-python
kornia

Run

For running ME-DDPG:

cd Bullet
python main_me_ddpg.py --env HopperBulletEnv-v0  --seed 0 --dropout_p 0.1

(mepe) python main_me_ddpg.py --env HopperBulletEnv-v0  --seed 0 --dropout_p 0.1

---------------------------------------
Policy: ME-DDPG, Env: HopperBulletEnv-v0, Seed: 0
---------------------------------------
3%|██▍                              | 28308/1000000 [00:28<59:30, 272.17it/s]

For running ME-SAC:

cd Bullet
python main_me_sac.py --env HopperBulletEnv-v0  --seed 0 --dropout_p 0.1

(mepe) python main_me_sac.py --env HopperBulletEnv-v0  --seed 0 --dropout_p 0.1
---------------------------------------
Policy: ME-SAC, Env: HopperBulletEnv-v0, Seed: 0
---------------------------------------
3%|██▌                          | 30202/1000000 [00:47<1:36:08, 168.12it/s]

For running ME-CURL

cd Atari
python main.py --game=ms_pacman

(mepe) python main.py --game=ms_pacman
Algorithm name:  mepe
2%|█▍                      | 1703/100000 [00:11<1:22:28, 19.86it/s]