The award-winning paper describes the design, deployment, and impact of a reinforcement learning system for targeted COVID-19 testing of all visitors to Greece during summer 2020. This system, nicknamed “Eva,” allowed Greece to simultaneously: 1) optimize the allocation of its scarce testing resources and 2) develop reliable, real-time estimates of COVID-19 prevalence from different origin countries to inform national-level travel protocols. Eva operated continuously from Aug. 6 to Oct. 30, 2020, processing travelers from more than 40,000 households each day, and represents the first national-scale reinforcement learning system deployed to combat the COVID-19 pandemic.
The interdisciplinary team behind Eva included operations researchers, epidemiologists, software developers, and government officials. This team first designed a COVID-19 testing supply chain consisting of 300 medical and emergency personnel to collect biological samples at 40 distinct points of entry, 32 private and public testing labs to process samples, and logistics teams to transport samples from these points of entry to the central labs twice daily. The robustness and speed of this supply chain was crucial to ensuring that real-time information from border testing could be used to quickly adapt testing protocols. Next, the team deployed a novel reinforcement learning algorithm to simultaneously: 1) allocate tests to passenger subpopulations with high-prevalence and 2) learn COVID-19 prevalence across all passenger subpopulations. Although developing this algorithm required new techniques to circumvent a number of practical challenges, the most critical design elements of Eva were shaped by the practical realities of earning trust and “buy-in” from a largely nontechnical set of decision-makers in a period of crisis. Hence, transparent reasoning that enabled human-in-the-loop decision-making was paramount.
Over the course of its operation, Eva identified twice as many asymptomatic infected individuals as more traditional surveillance testing, effectively doubling Greece’s testing capacity at the border. Decision-support tools built around Eva’s prevalence estimates informed within-country pandemic operations, including allocation of mobile testing units and social distancing guidelines. Finally, these real-time estimates of prevalence were also used by the Greek government to adapt national travel protocols and shared with the European Union. Overall, Eva represents a blueprint for the design of future high-stakes algorithmic decision-making systems in public policy and highlights the importance of transparency and interpretability in system design.