LoryPack / GenBayes_LikelihoodFree_ScoringRules

Code for the paper "Generalized Bayesian Likelihood-Free Inference Using Scoring Rules Estimators"

Generalized Bayesian Likelihood-Free Inference Using Scoring Rules Estimators

Code for the paper: Generalized Bayesian Likelihood-Free Inference Using Scoring Rules Estimators, which can be found here.

Instructions

All experiments and plots from the paper can be reproduced using the code provided here.

The content of this repository is as follows:

• src contains source code for the models, scoring rules and various utilities
• scripts contains Python and bash scripts to run the different experiments.
• mg1_code contains Matlab code to perform exact MCMC inference for the M/G/1 model, as described in Shestopaloff and Neal, "On bayesian inference for the M/G/1 queue with efficient MCMC sampling", 2014. This was kindly provided by the authors, except for the file mg1_code/do_runs_mine.sh which was added by me.

Reproducing the experiments

We provide bash scripts calling the Python scripts with the same parameters used in the paper. All figures in the paper can be generated by calling those scripts.

For instance, the MA2 experiment may be run by calling:

./scripts/MA2.sh

Additionally, running:

./scripts/run.sh

reproduces all experiments by calling the different bash scripts.

The scripts work on a single core. However, runnning everything on one single core can be slow; parallelization can be used to run different inferences at the same time and thus reducing computing time.

Requirements

The inference routines with the Scoring Rules are implemented using the ABCpy Python package.

All requirements are listed in the requirements.txt file. You can install those with:

pip install -r requirements.txt

Tests

We provide some tests for our source code. To run them, do:

python -m unittest src/tests.py

Citation

Please use the following .bib entry:

@misc{pacchiardi2021generalized,
      title={Generalized Bayesian Likelihood-Free Inference Using Scoring Rules Estimators}, 
      author={Lorenzo Pacchiardi and Ritabrata Dutta},
      year={2021},
      eprint={2104.03889},
      archivePrefix={arXiv},
      primaryClass={stat.ME}
}