ziq-ao / NFEE

Entropy estimation via normalizing flow

UMestimator

Code for reproducing the numerical experiments in the paper: "Entropy estimation via normalizing flow".

How to run the numerical experiments

The files are summarized as follows:

File	Experiments
demo_mvn.py	Multivariate normal distribution, RMSE vs dimension
demo_mvn_cr.py	Multivariate normal distribution, RMSE vs sample size
demo_hr.py	Multivariate Rosenbrock distribution, RMSE vs dimension
demo_hr_cr.py	Multivariate Rosenbrock distribution, RMSE vs sample size
demo_lv_max_ent_ed.py	Experimental design for the Lotka-Volterra model
lv_val.py	Compute “reference value” of the entropy for the optimal observation time placements

To run the experiment of multivariate normal distribution in Fig. 2 (RMSE vs dimension), you can issue the command:

python demo_mvn.py

To run the experiment of multivariate normal distribution in Fig. 3 (RMSE vs sample size), you can issue the command:

python demo_mvn_cr.py

To run the experiments of multivariate Rosenbrock distribution, you can issue the command:

python demo_hr.py <mdl_name> <n_trials> 

for the figure of RMSE vs dimension, and issue the command:

python demo_hr_cr.py <mdl_name> <n_trials> 

for the figure of RMSE vs sample size. Here <mdl_name> is the name of distribution tested, and can be hybrid_rosenbrock or even_rosenbrock. <n_trials> is the number of repeated trials and can be any positive integer.

To run the experiment of the optimal experimental design, you can issue the command:

python demo_lv_max_ent_ed.py <method> <n_samples>

Here, <method> is the entropy estimator used, and can be any of kl, ksg, umtkl , umtksg and nf. <n_samples> is the sample size, and can be any positive even number.

To compute the “reference value” of the optimal observation time placements obtained in this paper, you can issue the command:

python lv_val.py