Deep reinforcement learning framework for controlling infectious disease outbreaks in the context of multi-jurisdictions.

In the absence of pharmaceutical interventions, social distancing and lockdown have been key options for controlling new or reemerging respiratory infectious disease outbreaks. The timely implementation of these interventions is vital for effectively controlling and safeguarding the economy.Motivated by the COVID-19 pandemic, we evaluated whether, when, and to what level lockdowns are necessary to minimize epidemic and economic burdens of new disease outbreaks. We formulated the question as a sequential decision-making Markov Decision Process and solved it using deep Q-network algorithm. We evaluated the question under two objective functions: a 2-objective function to minimize economic burden and hospital capacity violations, suitable for diseases with severe health risks but with minimal death, and a 3-objective function that additionally minimizes the number of deaths, suitable for diseases that have high risk of mortality.A key feature of the model is that we evaluated the above questions in the context of two-geographical jurisdictions that interact through travel but make autonomous and independent decisions, evaluating under cross-jurisdictional cooperation and non-cooperation. In the 2-objective function under cross-jurisdictional cooperation, the optimal policy was to aim for shutdowns at 50 and 25% per day. Though this policy avoided hospital capacity violations, the shutdowns extended until a large proportion of the population reached herd immunity. Delays in initiating this optimal policy or non-cooperation from an outside jurisdiction required shutdowns at a higher level of 75% per day, thus adding to economic burdens. In the 3-objective function, the optimal policy under cross-jurisdictional cooperation was to aim for shutdowns of up to 75% per day to prevent deaths by reducing infected cases. This optimal policy continued for the entire duration of the simulation, suggesting that, until pharmaceutical interventions such as treatment or vaccines become available, contact reductions through physical distancing would be necessary to minimize deaths. Deviating from this policy increased the number of shutdowns and led to several deaths.In summary, we present a decision-analytic methodology for identifying optimal lockdown strategy under the context of interactions between jurisdictions that make autonomous and independent decisions. The numerical analysis outcomes are intuitive and, as expected, serve as proof of the feasibility of such a model. Our sensitivity analysis demonstrates that the optimal policy exhibits robustness to minor alterations in the transmission rate, yet shows sensitivity to more substantial deviations. This finding underscores the dynamic nature of epidemic parameters, thereby emphasizing the necessity for models trained across a diverse range of values to ensure effective policy-making.