MPPCA Importance Sampling

This repository contains code and experiments for the paper:

Kruse, L. A., Schlichting, M. R., & Kochenderfer, M. J., Scalable Importance Sampling in High Dimensions with Low-Rank Mixture Proposals, CoDIT 2025.

Dependencies

See environment.yaml for required packages, or use this to create a Conda environment with all dependencies:

conda env create -f environment.yaml

This repository was tested with Python 3.10 and PyTorch 2.7.

End-to-end Experiments

Run the experiments.sh file to reproduce all experimental results. This script kicks off the following steps:

1. Generate Monte Carlo datasets for reference data and metric calculations.
2. Perform importance sampling for all five problems (branches with 40 and 60 dimensions, oscillator with 100 and 200 dimensions, and f16gcas.
3. Generate a results table and figure using the logged metrics and trajectories.

Individual steps are next discussed in more detail:

Data

The experiments require Monte Carlo datasets for reference failure probabilities and coverage/NDB metric calculations. Datasets can be created using the following syntax:

python3 utils/mc_data_generation.py --problem Branches --d 40
python3 utils/mc_data_generation.py --problem Branches --d 60
python3 utils/mc_data_generation.py --problem Oscillator --d 100
python3 utils/mc_data_generation.py --problem Oscillator --d 200
python3 utils/mc_data_generation.py --problem F16GCAS

Experiments

Example commands to reproduce the f16gcas results are shown below. To change the problem, replace f16gcas with either branches or oscillator and update the --n_features flag with the corresponding problem dimensionality.

python3 experiments.py --problem "f16gcas" --is_method "CE" --proposal "MPPCA" --n_features 202 --n_components 8 --n_factors 8
python3 experiments.py --problem "f16gcas" --is_method "SIS" --proposal "MPPCA" --n_features 202 --n_components 8 --n_factors 8
python3 experiments.py --problem "f16gcas" --is_method "CE" --proposal "GMM" --n_features 202 --n_components 8
python3 experiments.py --problem "f16gcas" --is_method "SIS" --proposal "GMM" --n_features 202 --n_components 8

Results

All saved metrics can be viewed in tabular form by running the following command:

python3 utils/results_table.py

Furthermore, code to reproduce the results plot in the paper is provided:

python3 utils/results_fig.py

How to Cite

If you find this code useful in your research, please cite the following publication:

@inproceedings{kruse2025scalable,
  title={Scalable Importance Sampling in High Dimensions with Low-Rank Mixture Proposals},
  author={Kruse, Liam A and Schlichting, Marc R and Kochenderfer, Mykel J},
  booktitle={2025 11th International Conference on Control, Decision and Information Technologies (CoDIT)},
  year={2025}
}