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},
}