Phase II: Prescriptor Development
Phase II of the XPRIZE Pandemic Response Challenge is about proposing interventions for all the countries in the world in order to reduce the number of cases without having a huge negative effect in the country (stringency). Teams have to implement prescriptors (up to 10 different) that propose intervention plans for each country in a window of time, taking into account the previous interventions applied, and the weight each intervention type (closing schools, restricting mobility, forcing masks, etc.) may have in the country’s stringency. A standard predictor is provided by the organization in order to estimate the impact the intervention plan will have in the number of COVID-19 new cases.
The following plots shows the performance of the 10 prescriptor algorithms developed by the CoronaSurveys team in the Phase II of the XPRIZE Pandemic Response Challenge. The results are shown for the next 90 days and are updated daily.
The metrics presented are
- PredictedDailyNewCases: The number of new cases the standard predictor forecasts for each day using:
- The real intervention plans used by the countries, for the period until the current date.
- The intervention plan proposed by the prescriptor algorithm for the country/region selected, for 90 days starting from the current date.
- PredictedDailyNewDeaths: The number of daily fatalities, derived from the forecasted number of cases.
- PredictedDailyNewHospital: The number of daily new hospital cases, derived from the forecasted number of cases.
- PredictedDailyNewICU: The number of daily new hospital cases, derived from the forecasted number of cases.
- Intervention Stringency: The daily stringency (or cost) of the interventions implemented in the 90 proposed days. The cost depends on the weights vector selected (fixed_equal_costs or uniform_random_costs). The intervention types are described
in the Codebook for the Oxford Covid-19 Government Response Tracker.
The 10 prescriptor algorithms proposed by the CoronaSurveys team are the following:
- P0-Dance: This algorithm proposes to use no intervention for the whole period.
- P1-LSTM1: This approach uses Long Short-Term Memory(LSTM) Recurrent Neural Networks (RNN) that are trained trying to optimize the number of cases and the cost simultaneously.
- P2-Greedy 10: This approach creates the intervention plan in periods of 15 days, all with the same interventions, from the past to the future. A goodness function that combines normalized cases and cost is used to select the optimal interventions for a period before moving to the next. In this version the number of cases is normalized around an incidence of 10 cases pero 100,000 people.
- P3-HammerDance1: This algorithm proposes an intervention plan with the highest level of interventions for 45 days, and no intervention afterwards. This algorithm, and the other HammerDance algorithms below, are inspired by Tomás Pueyo’s proposal “The Hammer and the Dance”.
- P4-Hammer: This algorithm proposes an intervention plan with highest level of interventions for the whole period.
- P5-HammerDance2: This algorithm proposes the fastest intervention plan in reducing the number of cases (breaking ties by the cost) for 45 days, and no intervention afterwards.
- P6-HammerDance3: This algorithm proposes the fastest intervention plan in reducing the number of cases (breaking ties by the cost) for 45 days. After that, the cheapest intervention plan that does not increase the number of cases is proposed.
- P7-Greedy50: This algorithm is like P2-Greedy 10 but in the goodness function the number of cases is normalized around an incidence of 50 cases pero 100,000 people.
- P8-LSTM2: This approach uses LSTM RNN that are trained trying to optimize the number of cases and the cost alternating between both objectives.
- P9-LSTM3: This approach uses LSTM RNN as P8-LSTM2 but applies a heuristic post-processing to reduce the cost of the intervention plan proposed.
These algorithms are compared with
- PR (Real Interventions): These are the predictions of the evolution of number of cases and other parameters with the (closest in time) real interventions deployed by authorities.