Uncertainty Quantification and Propagation in Surrogate-based Bayesian Inference

Authors: Philipp Reiser, Javier Enrique Aguilar, Anneli Guthke, Paul-Christian Bürkner

---

Overview

This repository contains code to replicate key experiments from our paper Uncertainty Quantification and Propagation in Surrogate-based Bayesian Inference.

---

Replicating

In this section, we will detail how to run both the baseline approach (Point), as well as E-Post, E-Lik, and E-Log-Lik. We will showcase this on the two Case studies for a linear and a logistic simulator.

Case Study 1: Linear Model

To replicate the uncertainty propagation (UP) via analytic and numeric integration for the linear model, use the script

src/eval/run_eval_two_step_linear.R

The results are then stored in plots/analytic_case/linear_intercept/.

Case Study 2: Logistic Model

To replicate the UP via MCMC for the logistic model and plot respective I-posteriors, use the script

src/eval/run_eval_two_step_logistic.R

The results are then stored in plots/logistic_case_densities/true_model.

To replicate the UP via MCMC for the polynomial surrogate model (PCE) and plot respective I-posteriors, use the script

src/eval/run_eval_two_step_logistic.R

and set

config_file_name <- "pce_config.yml"

The results are then stored in plots/logistic_case_densities/pce.

To replicate the calibration results for the logistic model, use

src/eval/run_sbc_logistic.R
src/eval/eval_sbc_logistic.R

The results are then stored in plots/logistic_case_sbc/.

---

Related repositories

• Bayesian-sparse-PCE

• SBC

• sbc_test_quantities_paper

---

Citation

@misc{reiser2023uncertainty,
      title={Uncertainty Quantification and Propagation in Surrogate-based Bayesian Inference}, 
      author={Philipp Reiser and Javier Enrique Aguilar and Anneli Guthke and Paul-Christian Bürkner},
      year={2023},
      eprint={2312.05153},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}