FRED

FRED simulator

Getting started

Singularity images

We provide a singularity image that can be used for running the experiments. It has all the dependencies (pyprob, pyprob_cpp, PPX, and FRED simulator) installed. To create the singularity image from the recipe files, clone this repo and cd to FRED/singularity. Then,

sudo singularity build covid_base.sif covid_base.def
sudo singularity build covid.sif covid.def

Once the images are successfully created, start up the image:

singularity shell <path-to-covid.sif>

To do a quick test on your Singularity image, you can run [FRED/tests/pyprob_cpp](this test). To do so, run the following:

singularity shell -B <path-to-FRED/>:/workdir <path-to-covid.sif>
cd /workdir/tests/pyprob_cpp
cmake . && cmake --build .
python inference.py

It runs a test on a Gaussian with unknown mean model is in C++. It connects the C++ model to pyprob, samples from its prior and prints the empirical mean and standard deviation of the samples along with the analytically computed values for them.

Dependencies

The the core ingredients of this experiment are the following:

• FRED simulator: An agent-based simulator of epidemiological dynamics in C++.
• pyprob: A probabilistic programming library in python that can be coupled to the existing simulators in different platforms.
• pyprob_cpp A C++ library of pyprob interface.
• PPX: The protocol over which pyprob and pyprob_cpp communicate.

Running the experiments

Once you have the repo cloned and the singularity image created,

singularity shell -B <path-to-FRED/>:/workdir
cd /workdir
python main.py

The following list is the most important configurations for this experiment:

• num_traces: Number of simulator runs for the inference task (default: 10)
• constraint_threshold: Imposed threshold on the ratio of infectious population (default: 0.1)
• days: Number of days to simulate in each run (default: 150) For a list of configurations, look at main.py file.

Configuration files

FRED simulator has a large number of configurations. They should be stored in a text file. The parameters we used for our experiments can be found here. However, the five parameters we impose priors on will be overwritten be a value sampled from their prior. For more information, refer to the paper.

Visualizing the results

Once the experiments are ran, it generates a zip file containing the results (simulator outputs, log files, inference-related files, etc.) The plotting script can then be used to plot an array of 2D histograms of marginal posterior distribution over the policy parameters, like this.

python plot_histogram.py with exp_dir=<results_directory> county=<experiment_name>

where <results_directory>/<experiment_name> is the path of the directory containing the zip files.