Code for the experiments in Generalized Posteriors in Approximate Bayesian Computation .
misc{schmon2020generalized,
title={Generalized Posteriors in Approximate Bayesian Computation},
author={Sebastian M Schmon and Patrick W Cannon and Jeremias Knoblauch},
year={2020}, eprint={2011.08644}, archivePrefix={arXiv}, primaryClass={stat.ME}
}
The required dependencies can be installed using pip install -r requirements.txt.
The code comprises:
- a collection of kernels
kernels.py, - an ABC-MCMC class and algorithm
gb_alg.py, - a series of three experiments
gb_experiments.py, and - plotting the output of the experiments
gb_plots.py
Three experiments are described in the paper and can be recreated here (up to stochasticity in the ABC algorithms).
The first is a straightforward demonstration of the output of the algorithms (the posterior samples) as a function of the weight w chosen. Find the code for this experiment in initial_experiment from gb_experiments.py, and plotted by initial_experiment_plot from gb_plots.py.
The second experiment, detailing the effective sample size of the ABC-MCMC samples is performed by loss_average_experiment from gb_experiments.py, and plotted by loss_average_plot from gb_plots.py.
Finally the experiment on robustness to misspecification is performed by misspecification_experiment in gb_alg.py, and plotted by misspecification_plot from gb_plots.py.
