COVID-19 agent-based simulation using C++

After seeing this article and checking out this cool project it occurred to me that there is room for improvement both in terms of agent-based modeling to simulate pandemics more realistically and to use them to come up with sound scientific intervention policies to bring the disease under control quickly without suffocating the economy. I created an agent based model with a lot of features in terms of code and execution optimization so that we can use exhaustive search and optimization algorithms to find those policies that hit that sweet-spot between controlling the pandemic and preserving jobs and economic activity.

I used the Eigen linear algebra library to efficiently manipulate and store matrices that describe the state of the agents. I coupled the model with a lightweight Qt user interface and a scientific data visualization tool qcustomplot to visualize and control the model in realtime. I used CUDA to accelerate the computation of the many different interactions that occur between the agents in hopes of scaling this model to include millions of agents in the future.

I provide Python scripts to call the model in batch mode and run Monte-Carlo simulations to understand how different intervention measures affect the spread of the disease.

Setup

Dependencies

• SGTELIB (optional) for sensitivity analysis
• Qt 5.15.1 for user-interface
• (optional) CUDA 11.2.0 for GPU acceleration
• (optional) Python 3.8.3 or newer for statistics and post-processing
  • pyDOE
  • python-csv
• Microsoft Visual Studio 2017 or newer
  • Microsoft Visual Studio Qt Plugin

Setting environment variables (only if using CUDA)

Install CUDA. Check that the %CUDA% environment variable has been set correctly by:

• Right click This PC, click properties
• Goto Advanced system properties
• Goto Environment variables
• Under System variables check that %CUDA% is set correctly to

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\libnvvp;
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\bin;

• Add %CUDA% to the PATH environment variable

Compiling executable files (CUDA-capabale systems only)

Build the project executable using COVID_SIM_UI_CUDA.sln. This will dump the executable and all relevant dll's in the build directory.

IMPORTANT: Check that you have a CUDA-capable GPU before building the project.

Compiling executable files

First disable GPU acceleration during compilation by going to header/Defines.h and commenting the line.

#define GPU_ACC // comment to disable GPU acceleration

Build the project executable using COVID_SIM_UI.sln.

Example usage

To run the model simply execute COVID_SIM_UI.exe. The user-interface shown at the beginning will be shown.

Click the run button to start the simulation. Click the pause button to pause the simulation. Use the sliders to adjust the different intervention measures shown

• Infection chance
• Social distancing
• Number of tests/day

Other model parameters can be configured in the file configuration_debug.ini. Simply open the file with a text editor such as Notepad++ and change the included paramters. Not all paramters are defined in the configuration file and removing any value from the configuration_debug.ini file will reset it to its default value. The list of parameters, their meaning, and their default values are given in the table in Configuration_list.md.

Example usage in batch mode

The system command used to call the executable is

COVID_SIM_UI <arg1> <arg2> <arg3> <arg4>

• <arg1> is the run index
• <arg2> is the number of essential workers (social_distance_violation configuration parameter)
• <arg3> is the number of social distancing factor (social_distance_factor configuration parameter)
• <arg4> is the number of tests/day (number_of_tests configuration parameter)
• <arg5> is the output directory for objective (mobility) and constraint (maximum number of infections)

Copy the executable COVID_SIM_UI.exe and all dll files in build to build/COVID_DOE.

• Launch the python file MCS_model_points.py
• Wait for execution to finish
• Copy the resulting .pkl files in build/COVID_DOE/data/ to build/COVID_visualize/population/
• Launch the python file plot_profiles_stochastic.py
• Copy the resulting .pkl files in build/COVID_DOE/data/ to build/COVID_post/data/opts/
• Launch the python file MCS_model_points_post.py
• The following plots are generated (may slightly differ due to randomness of the model)

infections	fatalities

recoveries	reproductive number

mobility	Monte-Carlo (mobility)

Monte-Carlo (fatalities)	Monte-Carlo (infections)

Example for performing a sensitivity study

After completing the previous example, follow the following procedure to perform a latin hypercube search across all the different possible combinations of interventions:

• Open the python file MCS_model_LHS.py
• adjust n_samples_LH and n_samples to reduce the number of simulations needed to obtain the data and launch
• Wait for execution to finish (can take a long time possibly hours, much faster with CUDA)
• Copy the resulting .pkl files in build/COVID_DOE/data/ to build/COVID_post/data/LHS/
• Open the python file MCS_model_post.py
• adjust n_samples_LH according to what was selected in MCS_model_LHS.py and launch MCS_model_post.py
• Open the python file main.py
• adjust n_samples_LH according to what was selected in MCS_model_LHS.py and launch main.py
• The following plots are generated (may slightly differ due to randomness of the model)

Expected value of mobility

Red hatched regions show which combinations lead to infections exceeding healthcare capacity with a probability of 90 %

Standard deviation of infections

Further reading

Please check the following article for more information related to optimization of interventions using agent-based models

• Optimization of Infectious Disease Prevention and Control Policies Using Agent-based Modeling 1