Weighted likelihood filter

Citation

@inproceedings{duran2024outlier,
  title={Outlier-robust Kalman filtering through generalised Bayes},
  author={Duran-Martin, Gerardo and Altamirano, Matias and Shestopaloff, Alexander Y and S{\'a}nchez-Betancourt, Leandro and Knoblauch, Jeremias and Jones, Matt and Briol, Fran{\c{c}}ois-Xavier and Murphy, Kevin},
  booktitle={Proceedings of the 41st International Conference on Machine Learning},
  pages={12138--12171},
  year={2024}
}

Installation

To run the experiments, make sure to have installed jax>=0.4.2, rebayes-mini, flax, and the BayesianOptimization package:

pip install git+https://github.com/gerdm/rebayes-mini.git
pip install flax
pip install bayesian-optimization

Script workflow (2D tracking outliers)

The 2D tracking experiment is now available as standalone scripts:

• Run full experiment and store results:

cd experiments
python 2d_tracking_outlier_run.py --kind both --n-steps 1000 --n-samples 500 --init-points 20 --n-iter 30

This writes result files in experiments/results:

• 2d-ssm-outlier-mean.pkl

• 2d-ssm-outlier-covariance.pkl

• Generate plots from saved results:

cd experiments
python 2d_tracking_outlier_plot.py --kinds mean covariance

This writes figures in experiments/figures:

• 2d-ssm-comparison-outlier-mean.png
• 2d-ssm-comparison-outlier-covariance.png
• 2d-ssm-comparison-single-run-mean.png
• 2d-ssm-comparison-single-run-covariance.png
• 2d-ssm-comparison-outlier-both.png

Script workflow (all converted notebook experiments)

Each experiment now follows the same pattern:

1. Run script writes a .pkl result file to experiments/results.
2. Plot script reads the .pkl and writes figures to experiments/figures.

Gamma process

cd experiments
python gamma_process_run.py --n-steps 2000
python gamma_process_plot.py

Intro figure

cd experiments
python intro_fig_run.py --n-steps 20 --grid-size 150
python intro_fig_plot.py

Online MLP training

cd experiments
python online_mlp_training_run.py --n-obs 500
python online_mlp_training_plot.py

UCI regression with outliers

cd experiments
python uci_regression_outliers_run.py --dataset kin8nm --noise-type target --n-runs 100 --p-error 0.10
python uci_regression_outliers_plot.py

Run all full experiments end to end

cd experiments
python gamma_process_run.py --n-steps 2000
python gamma_process_plot.py

python intro_fig_run.py --n-steps 20 --grid-size 150
python intro_fig_plot.py

python online_mlp_training_run.py --n-obs 500
python online_mlp_training_plot.py

python uci_regression_outliers_run.py --dataset kin8nm --noise-type target --n-runs 100 --p-error 0.10
python uci_regression_outliers_plot.py

python 2d_tracking_outlier_run.py --kind both --n-steps 1000 --n-samples 500 --init-points 20 --n-iter 30
python 2d_tracking_outlier_plot.py --kinds mean covariance