Machines learn optimal strategies against the pandemic

Dealing with a pandemic poses major challenges for society and politics. The uncertain data situation, the countless factors to be considered and the incompatibility of different goals make it very difficult to find effective strategies to combat the pandemic.

"Usually, a few approaches developed by experts are examined and then a strategy is selected. But many potentially better strategies fall by the wayside because the number of possibilities is simply too great," says Miguel Navascués from the Institute for Quantum Optics and Quantum Information, Vienna, of the Austrian Academy of Sciences (OeAW). Together with colleagues from the OeAW and the University of Vienna, Navascués developed a mathematical framework with which optimal strategies for complex problems can be found. The work has now been published in the journal PLOS ONE.

Pandemic as an optimization problem

The physicists see the fight against the pandemic as an optimization problem. They therefore used machine learning techniques that employ systematic methods to find the best solution from an immense number of possible solutions. How this works can be illustrated with the following example: if you imagine the set of possible solutions as a landscape, computers can find the lowest point in the vicinity of a chosen starting point. "This is only a local minimum and there may always be a better solution far away, but our tests show that the best local solution is almost always very close to the global optimum," says first author Navascués from the OeAW.

The physicists' approach is very flexible and can find solutions for a wide variety of objectives. Usually, the requirements are, for example, that the maximum intensive care bed capacity is not reached and the time that a population has to spend in hard lockdowns is minimized. Then the computer finds solutions for this scenario, which can consist of lockdowns, vaccination campaigns and social distancing measures. "It can be very complicated, sometimes with a lot of short lockdowns and distancing phases, spread over a long period of time. That would probably be difficult to implement politically. But that's not a problem because we can easily add new restrictions to the model, such as a maximum number of lockdowns or the rule that lockdowns can only be introduced on Monday," Navascués explains.

Vaccination is the key

Such adjustments to the political reality lead to longer lockdown times, but the decisions are still optimal under the selected parameters, if indeed there are any solutions that can meet all expectations. "The computers find creative solutions. For example, if the number of lockdowns is limited to a maximum of five, the best strategy in this case can get away with four lockdowns," Navascués says. Machines are not immune to uncertain data, but the researchers have also come up with a solution for this. Navascués: "On the one hand, the models are equipped with buffers that allow a certain margin of deviation in the predicted development. On the other hand, regular remodeling can cushion the uncertainty."

In the end, the physicist's model is only an approximation of reality, but interesting conclusions are still possible. "It turns out, for example, that even a limited supply of vaccines can make an enormous difference that can drastically reduce lockdown times," Navascués explains. The model is not yet being used in practice, but the researchers hope that their publication will contribute to a rethink in politics. "We have a model that can give us optimal policies for complex problems. This could also be relevant for other problems in the future," Navascués says.