README.md

Active Exploration for Inverse Reinforcement Learning (AceIRL)

This repository contains code for the paper "Active Exploration for Inverse Reinforcement Learning". Here, we describe how to reproduce the experiments presented in the paper.

Citation

David Lindner, Andreas Krause, and Giorgia Ramponi. Active Exploration for Inverse Reinforcement Learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022.

@inproceedings{lindner2022active,
    title={Active Exploration for Inverse Reinforcement Learning},
    author={Lindner, David and Krause, Andreas and Ramponi, Giorgia},
    booktitle={Conference on Neural Information Processing Systems (NeurIPS)},
    year={2022},
}

Setup

We recommend to use Anaconda for setting up an environment with the required dependencies. After installing Anaconda, you can create an environment using:

conda create -n "aceirl" python=3.9
conda activate aceirl

Then, install dependencies by running

pip install -e . 

in the root folder of this repository.

Reproducing experiments in the main paper

All experiments can be run using the scripts/experiments/experiments.py script. We use sacred for keeping track of experiment parameters.

Running active IRL experiments

Run the following commands to reproduce the experiments in the main paper:

python scripts/experiments/experiment.py with four_paths n_ep_per_iter=50 results_file="result_aceirl_four_paths_50ep_50runs.csv"
python scripts/experiments/experiment.py with four_paths n_ep_per_iter=100 results_file="result_aceirl_four_paths_100ep_50runs.csv"
python scripts/experiments/experiment.py with four_paths n_ep_per_iter=200 results_file="result_aceirl_four_paths_200ep_50runs.csv"

python scripts/experiments/experiment.py with double_chain n_ep_per_iter=50 results_file="result_aceirl_double_chain_50ep_50runs.csv"
python scripts/experiments/experiment.py with double_chain n_ep_per_iter=100 results_file="result_aceirl_double_chain_100ep_50runs.csv"
python scripts/experiments/experiment.py with double_chain n_ep_per_iter=200 results_file="result_aceirl_double_chain_200ep_50runs.csv"

python scripts/experiments/experiment.py with random_env results_file="result_aceirl_random_mdp_50runs.csv"
python scripts/experiments/experiment.py with chain results_file="result_aceirl_chain_mdp_50runs.csv"
python scripts/experiments/experiment.py with gridworld results_file="result_aceirl_gridworld_50runs.csv"

To parallelize experiments, you can additionally pass the n_jobs parameter, for example:

python scripts/experiments/experiment.py with four_paths n_ep_per_iter=200 n_jobs=50

Running reward-free exploration experiments in the appendix

Run the following commands to reproduce the reward-free exploration experiments:

python scripts/experiments/experiment.py with double_chain_rfe n_ep_per_iter=1000
python scripts/experiments/experiment.py with double_chain_rfe n_ep_per_iter=3000
python scripts/experiments/experiment.py with double_chain_rfe n_ep_per_iter=5000

Creating plots

The results will be saved in results/. To plot the results and produce the results from Table 1 in the paper, use the scripts/experiments/plot_results.ipynb notebook.