Contains PyTorch implementation of the AdVIL, AdRIL, and DAeQuIL algorithms.

To install dependencies, run:

conda env create -f environment.yml

To train an expert, run:

python experts/train.py -e env_name

To train a learner, run:

python learners/train.py -a algo_name -e env_name -n num_runs

This package supports training via:

• Behavioral Cloning
• AdVIL
• SQIL
• GAIL
• AdRIL

We also support a comparison of the following algorithms that require an interactive expert on both OpenAI Gym and a custom environment:

• DAgger
• DAeQuIL

To explore these algorithms and environments, run:

jupyter notebook

and open up learners/imm.ipynb.

Run:

jupyter notebook

and open up vis.ipynb.

@article{swamy2021moments, author = {Gokul Swamy and Sanjiban Choudhury and J. Andrew Bagnell and Zhiwei Steven Wu}, title = {Of Moments and Matching: A Game-Theoretic Framework for Closing the Imitation Gap}, conference = {Proceedings of the 38th International Conference on Machine Learning}, url = {https://arxiv.org/abs/2103.03236}, }