1. Title of the paper From Prediction to Prescription: AI-Based Optimization of Non-Pharmaceutical Interventions for the COVID-19 Pandemic 2. Authors Risto Miikkulainen, Olivier Francon, Elliot Meyerson, Xin Qiu, Elisa Canzani, Babak Hodjat. Cognizant Technology Solutions, 649 Front St., Suite 300, San Francisco, CA 94111, +1 415-471-4324, firstname.lastname@cognizant.com Risto Miikkulainen also: The University of Texas at Austin, Austin TX 78712, +1 512 471 9571, risto@cs.utexas.edu 3. Corresponding author Babak Hodjat; babak@cognizant.com 4. Abstract Several models have been developed to predict how the COVID-19 pandemic spreads, and how it could be contained with non-pharmaceutical interventions (NPIs) such as social distancing restrictions and school and business closures. This paper demonstrates how evolutionary AI could be used to facilitate the next step, i.e., determining most effective intervention strategies automatically. Through evolutionary surrogate-assisted prescription (ESP), it is possible to generate a large number of candidate strategies and evaluate them with predictive models. In principle, strategies can be customized for different countries and locales, and balance the need to contain the pandemic and the need to minimize their economic impact. While still limited by available data, early experiments suggest that workplace and school restrictions are the most important and need to be designed carefully. It also demonstrates that results of lifting restrictions can be unreliable, and suggests creative ways in which restrictions can be implemented softly, e.g.\ by alternating them over time. As more data becomes available, the approach can be increasingly useful in dealing with COVID-19 as well as possible future pandemics 5. Criteria satisfied D, E, F, G 6. Justification of criteria satisfied The COVID-19 crisis is unprecedented in modern times, and caught the world largely unprepared. The decisions as to what non-pharmaceutical interventions (NPIs) to put in place, and when, are based on epidemiological models that include many assumptions that may or may not hold true with this as-yet unknown virus. Furthermore, the manner by which people in different regions adhere to different policies is unknown. Most importantly, it is very difficult for a decision-maker to know which NPIs to put in place in order to achieve a desired balance of containing the spread of the virus, while minimizing the economic cost. In short, these human decisions have huge implications and are very difficult to make. Since there is little experience and guidance, authorities have been responding in a variety of ways. Many different NPIs have been implemented at different stages of the pandemic and in different contexts. The NPIs put in place at the same point of the spread of the disease in South Korea, China, and Germany, for instance, are quite different than those in Brazil, Iran, and Sweden, resulting in very different trajectories of the disease cases and deaths, as well as economic cost. This paper describes how Evolutionary Surrogate-assisted Prescription system can be used to discover effective NPI strategies. An LSTM neural network is first trained with the sequence of cases and NPIs so far to predict how the pandemic spreads in the future in a particular country. A feedforward neural network is then evolved to recommend NPIs to be enacted for each day in the future. The predictive network is used as a surrogate during evolution to evaluate the prescriptive network candidates. Evolution results in a Pareto front of prescriptions from which the decision maker can choose one that represents a desirable balance between number of cases and the stringency (i.e. cost) of the NPIs. To compare the evolved prescriptors with human performance, a number of tests were run by asking what they would have done if they had been applied to various countries at various points in the past. The trained predictors are highly accurate especially in the early stages of the pandemic, so they can be used reliable to evaluate the outcomes---even when they are counterfactual. Indeed in these counterfactual tests, prescriptors often discover policy recommendations that would have likely reduced the number of cases and reduced cost. For instance, they would have recommended a lockdown in the US on February 25th, reducing the number of cases, and later allowing for a more rapid lifting of the NPIs, thus minimizing cost. (The benefit of early interventions have recently been suggested through other modeling efforts as well, e.g. Pei, Kandula, and Shaman 2020.) Remarkably, in some cases the prescriptors could have possibly avoided full lockdowns. For instance in the UK on March 16th, the NPIs actually in effect were the mild 'recommend work from home' and 'recommend cancel public events', which would have resulted in a high number of predicted cases. A lockdown was eventually implemented on March 24th, and the actual case numbers were significantly smaller. However, evolution discovered a Prescriptor that would have required closing schools and recommended work-from home already on March 16th, while avoiding other restrictions. As a result, the predicted cases could have been much fewer---even without an extensive lockdown. In terms of the criteria: (D) ESP produces a schedule of placing and replacing NPIs in order to control a regional outbreak given historical data on the outbreak's progression so far, and information on the region in question. This is in itself a publishable result in epidemiological publications. (E) The human designed NPI strategy based on epidemiological assumptions has been devised and implemented throughout the COVID-19 pandemic. Counterfactual studies show the strategies automatically generated by ESP are better. (F) Human decisions based on trial and error against epidemiological models with many untested assumptions are the best the experts can create. (G) Devising effective NPI strategies in the face of an unfolding pandemic is very difficult and there is no consensus strategy, as evidenced by how different countries and regions have dealt with it so far. 7. Citation Miikkulainen, R., Francon, O., Meyerson, E., Qiu, X., Canzani, E., and Hodjat, B. (2020). From Prediction to Prescription: AI-Based Optimization of Non-Pharmaceutical Interventions for the COVID-19 Pandemic. arXiv 2005:13766. 8. Prize money, if any, will be divided equally among the co-authors. 9. Why this entry is the best It solves an important problem in the real world: Policy decision making to contain the spread of an as-yet unknown disease is very difficult and, up until now, at the sole discretion of human decision makers. Where possible, ESP creatively suggests policies to achieve the best balance of containing the spread and cost. Across several countries, for example, at different stages of the pandemic, a consistent pattern emerges: in order to keep the number of cases flat, other NPIs can be lifted gradually, but workplace and school restrictions need to be in effect much longer. Indeed these are the two activities where people spend a lot of time with other people indoors, where it is possible to be exposed to significant amounts of the virus. In other activities, such as gatherings and travel, they may come to contact with many people briefly and often outdoors, mitigating the risk. Therefore, the main conclusion that can already be drawn from these preliminary prescription experiments is that it is not the casual contacts but the extended contacts that matter. Consequently, when planning for lifting NPIs, attention should be paid in particular to how workplaces and schools can be opened safely. Even though yet unproven, recent epidemiological studies are consistent with this point (Kay 2020; Park et al. 2020; Lu et al. 2020). Another interesting conclusion can be drawn from some of the prescribed policies: Alternating between weeks of opening workplaces and partially closing them may be an effective way to lessen the impact on the economy while reducing cases. This solution is interesting because it shows how evolution can be creative and find surprising and unusual solutions that are nevertheless effective. While on/off alternation of school and workplace closings may sound unwieldy, it is a real possibility (Chowdhury et al. 2020). Note also that it is the only creative solution available to the Prescriptor: there are no options in its output for e.g., alternating remote and in-person work, extending school to wider hours, improving ventilation, moving classes outside, or other ways of possibly reducing exposure. How to best implement such distancing at school and workplace is left for human decision makers at this point; the model, however, makes a suggestion that coming up with such solutions may make it possible to lift the NPIs gradually, and thereby avoid secondary waves of cases. Thus, in the early stages, the ESP approach will suggest how to ``flatten the curve'', i.e., what NPIs should be implemented in order to slow down the spread of the disease. At later stages, it may recommend how the NPIs can be lifted and the economy restarted safely. 10. Type of EC used Genetic Algorithms 11. Date of Publication May 28th, 2020. References Chowdhury, R., Heng, K., Shawon, M. S. R., Goh, G., Okonofua, D., Ochoa-Rosales, C., Gonzalez-Jaramillo, V., Bhuiya, A., Reidpath, D., Prathapan, S., Shahzad, S., Althaus, C. L., Gonzalez-Jaramillo, N., Franco, O. H., and COVID-19 Collaborative Group, T. G. D. I. S. for Dynamic interventions to control COVID-19 pandemic: a multivariate prediction modelling study comparing 16 worldwide countries. European Journal of Epidemiology, 2020.10.1007/s10654-020-00649-w. Kay, J. COVID-19 superspreader events in 28 countries: Critical patterns and lessons. Quillette, April 23 2020. Lu, J., Gu, J., Li, K., Xu, C., Su, W., and Lai, Z. . e. a. COVID-19 outbreak associated with air conditioning in restaur ant, guangzhou, china, 2020. Emerging Infectious Diseases, 2020. Park, S. Y., Kim, Y. M., Yi, S., Lee, S., Na, B. J., and Kim, C. e. a. Coronavirus disease outbreak in call center, south korea. Emerging Infectious Diseases, 2020. Pei, S., Kandula, S., and Shaman, J. Differential effects of intervention timing on COVID-19 spread in the united states. medRxiv, 2020.