This repository contains code for the paper "Learning Safety Constraints from Demonstrations with Unknown Rewards". Here, we describe how to reproduce the experiments presented in the paper.

David Lindner, Xin Chen, Sebastian Tschiatschek, Katja Hofmann, Andreas Krause. Learning Safety Constraints from Demonstrations with Unknown Rewards. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.

@inproceedings{lindner2024learning,
    title={Learning Safety Constraints from Demonstrations with Unknown Rewards},
    author={Lindner, David and Chen, Xin and Tschiatschek, Sebastian and Hofmann, Katja and Krause, Andreas},
    booktitle={International Conference on Artificial Intelligence and Statistics (AISTATS)},
    year={2024},
}

We recommend using Anaconda to set up an environment with the dependencies of this repository. Run the following commands from this repository to set up the environment:

conda create -n cocorl python=3.9
conda activate cocorl
pip install -e .

This sets up a Anaconda environment with the required dependencies and activates it.

To run the experiments with synthetic single-state CMDP instances, use

python src/constraint_learning/linear/synthetic_experiment.py

Parameters and experiment tracking are handled via sacred. Results are stored in results/, and arguments can by passed using with. A sweep of experiments is defined in experiment_configs/synthetic/synthetic.json, and it can be run with

python run_sacred_experiments.py --config experiment_configs/synthetic/synthetic.json --num_jobs 1

The --num_jobs flag can be used to parallelize the runs over multiple CPUs.

To run the Gridworld experiments, use

python src/constraint_learning/linear/gridworld_experiment.py

In the paper, we present 3 different experiments to test reward transfer. These can be run with the following three parameter sweeps:

python run_sacred_experiments.py --config experiment_configs/gridworld/gridworld_exp1_no_transfer.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/gridworld/gridworld_exp2_reward_transfer.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/gridworld/gridworld_exp3_env_transfer.json --num_jobs 1

Note that these sweeps run a lot of random seeds by default. You'll likely want to increase the number of parallel jobs using --num_jobs.

To run experiments in highway-env, we first need to precompute potential demonstrations for different reward functions (in the standard environment, as well as environments with aggressive and defensive drivers). To do that, run

python run_sacred_experiments.py --config experiment_configs/highway_ce/ce_demonstrations.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/highway_ce/ce_demonstrations_aggressive.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/highway_ce/ce_demonstrations_defensive.json --num_jobs 1

For convenience, we provide a set of precomputed demonstrations. To use them, unpack the archive demonstrations.zip into the folder demonstrations/.

Now, experiments can be run using the experiment script:

python src/constraint_learning/linear/highway_experiment.py

To run the three experiments presented in the paper, execute the following commands:

python run_sacred_experiments.py --config experiment_configs/highway_ce/exp1_no_transfer.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/highway_ce/exp2_goal_transfer.json --num_jobs 1
python run_sacred_experiments.py --config experiment_configs/highway_ce/exp3_env_transfer.json --num_jobs 1

By default the results of all experiments are stored in results/. Each result is stored in an individual sub-folder labelled with a timestamp. To aggregate and evaluate the experiment results, we provide a script aggregate_results.py. To use it, execute

python scripts/aggregate_results.py --results_folder results/ --out_file aggregated.csv

This produces a csv file aggregated.csv containing the results from all experiments in results/. This csv file can be used to reproduce the plots shown in the paper.

We use pytest for unit tests. Run all tests with pytest .

We use black, isort, flake8, mypy, darglint, and jsonlint-php. Run all linters with lint.sh.