The Meta-Uncertainty Framework

This repository contains the code for the paper Meta-Uncertainty in Bayesian Model Comparison: https://arxiv.org/abs/2210.07278

Note that the R code is structured as a package, thus requiring a local installation with subsequent loading via library(MetaUncertaintyPaper).

Installation Instructions

The {ggsimplex} plot package

The current paper code uses a highly experimental version of the ggsimplex R package. Install it from GitHub via

devtools::install_github('marvinschmitt/ggsimplex')

R environment

The R environment is captured with renv. Install the renv package and load the environment with

renv::restore()

Python environment

The package requirements of the Python environment (except BayesFlow, see below) are captured in the requirements.txt file. Recreate the environment using

pip install -r requirements.txt

The amortized model comparison network (BayesFlow) in Experiment 3 uses BayesFlow at commit c4208418ad19b6648be216cfe013c8f5317a652c: https://github.com/stefanradev93/BayesFlow/tree/c4208418ad19b6648be216cfe013c8f5317a652c.

Should you fail to install this BayesFlow version or encounter unexpected errors, you can load the trained neural networks’ weights from the folder python/checkpoints_exp3/ and avoid re-training the network.