Adversarial design code

The repository is code used in the paper "Bayesian experimental design without posterior calculations: an adversarial approach" (https://arxiv.org/abs/1904.05703).

Code for paper

To reproduce the results in the paper run:

python3 poisson_example.py --lr-a 1e-3 --name "poisson3"
python3 poisson_example.py --lr-a 1e-4 --name "poisson4"
python3 poisson_example.py --lr-a 1e-5 --name "poisson5"
python3 poisson_plots.py

python3 pk_example.py --gda-iterations 100000 --nsamples 100 --nparallel 100 --seed 123 --point-exchange --name "pk_gda_K100"
python3 pk_example.py --gda-iterations 100000 --nsamples 10 --nparallel 100 --seed 123 --point-exchange --name "pk_gda_K10"
python3 pk_example.py --gda-iterations 100000 --nsamples 1 --nparallel 100 --seed 123 --point-exchange --name "pk_gda_K1"
python3 pk_example.py --gda-iterations 100000 --nsamples 1 --nparallel 100 --seed 123 --sgd --point-exchange --name "pk_sgd"
python3 pk_example.py --gda-iterations 100000 --nsamples 1 --nparallel 100 --seed 123 --gaps --name "pk_gaps"
python3 pk_example.py --gda-iterations 500000 --nsamples 1 --nparallel 1 --seed 123 --multiplicative-noise --name "pk_multi"
python3 pk_example.py --gda-iterations 500000 --nsamples 1 --nparallel 1 --seed 123 --multiplicative-noise --gaps --name "pk_multi_gaps"
python3 pk_plots.py
python3 pk_posteriors.py

python3 geostats_example.py --length-scale 0.01 --name "geo1"
python3 geostats_example.py --length-scale 0.02 --name "geo2"
python3 geostats_example.py --length-scale 0.04 --name "geo3"
python3 geostats_example.py --length-scale 0.08 --name "geo4"
python3 geostats_plots.py

There is a R subdirectory to run a comparison analysis using the acebayes package. This can be run using the following command. Note you should be in the R directory when this is run. Also, this analysis takes roughly 3 days to run.

R CMD BATCH --no-save PK_SIG.R

After transferring the csv file produced to the outputs folder, plots for the paper can be produced using

python3 pk_plots_R.py

To produce plots for the methods of Foster et al (produced using code in [https://github.com/dennisprangle/pyro/tree/sgboed-reproduce]), transfer its results to the outputs folder and run

python3 pk_plots_foster.py

Finally, to produce the boxplots comparing design from all methods, run

python3 pk_design_diagnostics.py

Extra tuning analysis

This code is for an investigation of alternative tuning choices in the acebayes package. The detailed output is not included in the paper, only in the reply to reviewers.

Enter the R subdirectory and run using the following command. This takes roughly 3 days to run.

R CMD BATCH --no-save PK_SIG_tuning.R

After transferring the csv files produced to the outputs folder, summary plots can be produced using

python3 pk_plots_R_tuning.py