sir_pomdp
Matlab Toolbox for disease control using partially observable decision process framework for action selection
Why?
During the Corona pandemic in 2020 with no vaccine available several measures where proposed to deal with the epidemic.
Which of these measures/actions or combination thereof is most efficient?
Having worked in epidemiology decades ago I knew how to model a disease. Having worked with partially observable Markov chains also decades ago I also knew how to find a best set of actions, given that the state of the world is not well known. So I decided to combine both to find an answer.
How?
The disease is modeled in compartments
- Susceptible
- Infectious
- Recovered
In order to model actions as well compartments like
- vaccinated
- isolated being infectious
- isolated being susceptible
- intensive care
- dead
- ...
are added. The transitions of changes between these compartments, i.e. states, are modeled using transition matrices.
Possible observations
- infectious
- isolated
- intensive care
- recovered
- vaccinated
- dead
- ...
are also modeled.
The model uses a POMDP approach to model the uncertainty in observations (Partially Observable Markov Decision Process). I recommend to read the article "planning and acting in partially observable stochastic domains" by L.P. Kaebling et al. 1997.
The possible actions are
- do nothing
- try to find out if someone is infectious
- isolate an infectious
- isolate a susceptible
- vaccinate
Given a set of "rewards" for each state and action combination the model tries to find the optimal "greedy" action. I.e. the current version of the code implements just a 1-step policy of action selection.