HIDIL

Offline Imitation Learning with a Misspecified Simulator

This repository is code for the paper

Shengyi Jiang, Jing-Cheng Pang, Yang Yu. Offline imitation learning with a misspecified simulator. In: Advances in Neural Information Processing Systems 33 (NeurIPS'20), Virtual Conference, 2020.

Requirement

• python 3.6
• Anaconda
• gym
• tmux
• tqdm
• tmuxp
• numpy
• pyyaml
• mujoco_py
• matplotlib
• tensorflow==1.13.1

Instructions

For automatic running this code in parallel, you may firstly install tmux and Anaconda. You have access to pretrain expert models on different dynamics mismatch from Google Drive.

# Preparations:
sudo apt-get install tmux
conda create -n your_venv python=3.6
source activate your_venv
pip install -r requirements.txt

# For a minimum running, you can run the following command: 
python main.py --process_num 5 --env_list "Walker2d-v2" --variety_list "0.5" --transfer_type "gravity"

# For customized running, you can modify generate_tmux_config.py and run:
python generate_tmux_config.py --conda_name your_venv
tmuxp load run_all.yaml

After one experiment has been done, the summary data will be saved in ./result/summary, which can be loaded by:

import numpy as np
summary = np.load("file_name.npy", allow_pickle=True).item()

where summary is a basic type [dict].