Conference AAMAS: COVID-19 boosts agent-based simulations

How do you get an epidemic under control? Which measures can most effectively contain the number of infections while minimizing undesirable social consequences? Agent-based simulations can provide decision support. Two studies have now been presented at the AAMAS conference (International Conference on Autonomous Agents and Multi-Agent Systems).

Veronika Kurchyna from the University of Trier worked with the simulation model SoSAD (Social Simulation for Analysis of Infectious Disease Control) examined the effects of different strategies for organizing school lessons on the infection process. She reports on this in the workshop Multi-Agent Based Simulation (MABS).

SoSAD was developed based on the social structure of Kaiserslautern. The model consists of 102,798 agents, of which 15,888 represent children and young people, 67,169 adults and 19,741 pensioners. These are spread across 56,663 households, can choose from 175 leisure activities and visit 175 workplaces and 33 schools with a total of 400 classes.

In the context of project Ascore The study conducted at the German Research Center for Artificial Intelligence (DFKI) simulated three different measures: continuation of school operations as usual, supplemented by hygiene precautions such as respiratory masks and regular ventilation of the classrooms; complete closure of schools; and division of the school classes into two halves, which are taught alternately on site and at home via video conference on a weekly basis.

It was shown that the principle of divided classes reduced the infection rates in the general population the most: after 60 days, 30,147 people were infected with the omicron variant of the corona virus in schools that were completely open, 23,512 in closed schools and 23,512 in schools split classes only 22,401.

According to Kurchyna, this surprising result can be explained by the fact that the students in the divided classes continue to be tested regularly for COVID-19, so that infected people can go into quarantine at an early stage. With closed schools, on the other hand, this is not guaranteed. In general, according to the conclusion of the study, the superiority of agent-based simulations over purely statistical models has been shown. The former could better capture individual activities and household structures.

Also inspired by the COVID-19 pandemic is research presented by Christian Kammler from the Swedish University of Umeå. He was concerned with modeling the effects of restrictions in the catering industry: How do restaurant operators react to the reduction in the number of guests they are allowed to serve? How will visitors adjust their behavior?

According to Kammler, in order to be able to model such effects, the agents in the simulation would have to show human behavior as realistically as possible. This behavior is in turn influenced by norms that can both restrict and motivate – and are occasionally deliberately disregarded. For example, an innkeeper can decide to allow a few more guests than is actually allowed instead of turning away an entire group.

Restaurant visitors can also deal with the applicable standards in very different ways, depending on their individual needs. For example, being with friends in a pleasant atmosphere may be more important to some people than the quality of the food, so they continue to visit the restaurant even if the owner uses cheaper ingredients to reduce costs.

To take all of this into account, Kammler and his research colleagues have developed an agent architecture designed to model the complex interplay of long-term and short-term goals, personal values, and available actions. Unlike Kurchyna, who was able to point out that her simulation reproduced infection patterns that were also observed in reality, Kammler did not carry out this empirical quality test for the time being.

Last year, researchers presented simulations of the acceptance of COVID-19 measures at AAMAS 2021.


To home page

Leave a Comment