An agent-based modelling study and analysis of an adaptive multi-UAV virus test delivery system

Abstract

In February and March 2020, the COVID-19 disease spread rapidly across the world. It was a new disease for which many countries were not prepared. In the Netherlands there was a scarcity of virus tests, and lack of coordination on the allocation of these scarce tests. Besides, there was no proactive acquisition of
spread information, and spread predictions were not used. To address these issues, government institutions could use UAVs. The UAVs acquire virus spread information, by coordinated and proactive delivery of virus tests to people’s home. This information is important for detecting virus outbreaks early, and to respond with appropriate measures, such as social distancing measures. This research aims to develop and evaluate such a multi-UAV test delivery system, using the agent-based modelling and simulation paradigm. The system consists of two main components: (1) The system coordinates the UAVs’ region of focus, by using a Bayesian decision network. This Bayesian decision network uses the virus case predictions of a spatiotemporal generalised linear prediction model, and the observations of the UAV system to make coordination decisions. (2) The system uses neighbourhood search and exploration techniques to detect virus cases that otherwise would not be identified. The methodology has been applied to a case study of 5 municipalities in the province of Noord-Brabant in the Netherlands. These municipalities were strongly hit by the COVID-19 epidemic. The main conclusions are that increasing the number of UAVs, increases the virus detection and decreases average delivery time. Furthermore, the Bayesian decision network is effective in prioritising the allocation of test resources to regions with higher epidemic severity. It does so by allocating more time to severe regions. Additionally, neighbourhood search is an effective way to find unobserved cases. Moreover, exploration in combination with the fixed response threshold model was found to be more effective than neighbourhood search.