How AI Makes Intervention Recommendations

This demo illustrates the non-pharmaceutical interventions (NPIs) that the AI generates for different countries and regions over time, their predicted effect. You first select country/region from the map at right, and then select the degree to which you'd like to minimize the number of cases vs. minimize the stringency of NPIs from the slider on the left. Among the ~101240 possible NPI schedules for the next 180 days, the AI will then recommend one that implements that tradeoff with as few cases and NPIs as possible. The predicted number of cases and the recommended NPIs will be shown over time in the charts below. See Optimizing COVID-19 Interventions for the general context, in-depth paper for the technical details, and ESP Introduction for the Evolutionary AI technology and its other applications. See also the XPRIZE Pandemic Response Challenge (and its technical details) for a recent machine learning competition inspired by this demo.

What is going on?

The ESP system learned to make recommendations based on historical data, i.e. the publicly available COVID-19 data on number of cases and NPIs over time provided by Oxford University (Hale, Webster, Petherick, Phillips, Kira, 2020, Oxford COVID-19 Government Response Tracker, Blavatnik School of Government, Oxford University). As described in this detailed technical paper, ESP consists of two learning tasks: (1) learning to predict, and (2) using the predictor as a surrogate for the real world, learning to prescribe actions that lead to desirable outcomes.

The Predictor was first trained with historical data. It is an LSTM recurrent neural network. Given the number of cases and NPIs in the past three weeks, its task is to predict the cases the next day. This prediction and any changes to NPIs are then fed back into the LSTM as input data to predict the next day, cascading predictions upto 180 days into the future.

Using the Predictor as a surrogate, a Prescriptor neural network was then created through population-based search, i.e. evolutionary optimization. Initially a population of random Prescriptor neural networks was created. During each generation, each network was evaluated in turn, based on how well their NPI recommendations fared in two dimensions: the number of cases, and the sum of NPI stringency levels, both averaged over the following 180 days and over the 20 countries with the most deaths. As usual in evolution, those networks that performed well along one or both objectives were crossed over and mutated to form offspring networks, replacing those that performed poorly. The population thus discovered a Pareto front of tradeoffs between these two objectives (for an animation of this process, see Demo 2.2).

Twenty different Prescriptors from Demo 2.2 are used in this demo. The selected one is initialized with data for a selected country, and it recommends NPIs over time. The same initial sequence and the recommended NPIs are given to the Predictor, and it generates the plot of cases over time (you can zoom in with the mouse by dragging). Uncertainty in the predictions is estimated through RIO, a Gaussian process approach for estimating residuals (RIO). The gray area represents the upper and lower quartiles of the distribution for each day. It is formed by rolling out uncertainties stochastically day by day for 180 days, thus resulting in jagged boundaries.

Explore especially the intermediate prescriptors that try to balance the two objectives by combining and alternating various NPI recommendations. It is possible to see certain informative trends to emerge, and they change as the pandemic changes. For example in Spring 2020, schools and workplace restrictions, i.e. those where people spend a lot of time with the same people mostly indoors, were often favored by the model. Also, the model sometimes suggested creative ways to implement these restrictions gradually, for instance by alternating them on and off over time. In June there was a marked shift towards more stringency, presumably because the NPIs were no longer as effective as they had been (possibly because people were eager to emerge from lockdowns). Also, as schools were out for the summer in many countries, travel restrictions become more important. In the Fall, with much of the world adapting schools, work, and travel to the pandemic, stay-at-home requirements became important. As the world is entering vaccination phase, other types of presriptions may emerge. The data is updated almost daily, and the Predictors and Prescriptors trained with it, and thus the system is continuously adjusting to the changing face of the pandemic.