Flattening the curve through reinforcement learning driven test and trace policies

Abstract

An effective way of limiting the diffusion of viruses when vaccines are unavailable or insufficiently potent to eradicate them is through running widespread “test and trace” programmes. Although these have been instrumental during the COVID-19 pandemic, they also lead to significant increases in public spending and societal disruptions caused by the numerous isolation requirements. What is more, after the health measures were relaxed across the world, these programmes were unable to prevent substantial upsurges in infections. Here we propose an alternative approach to conducting pathogen testing and contact tracing that is adaptable to the budgeting requirements and risk tolerances of regional policy makers, while still breaking the high risk transmission chains. To that end, we propose several agents that rank individuals based on the role they possess in their interaction network and the epidemic state over which this diffuses, showing that testing or isolating just the top ranked can achieve adequate levels of containment without incurring the costs associated with standard strategies. Additionally, we extensively compare all the policies we derive, and show that a reinforcement learning actor based on graph neural networks outcompetes the more competitive heuristics by up to 15% in the containment rate, while far surpassing the standard random samplers by margins of 50% or more. Finally, we clearly demonstrate the versatility of the learned policies by appraising the decisions taken by the deep learning agent in different contexts using a diverse set of prediction explanation and state visualization techniques.

More information

Identifiers

ORCID for Andrei Rusu: orcid.org/0000-0002-6053-1685

ORCID for Kate Farrahi: orcid.org/0000-0001-6775-127X

ORCID for Mahesan Niranjan: orcid.org/0000-0001-7021-140X

Catalogue record

Export record

Altmetrics