The XPRIZE Pandemic Response Challenge consisted of creating models to predict the course of the COVID-19 pandemic and to prescribe non-pharmaceutical interventions (NPIs) that would help mitigate it. Given past history of cases and interventions in a country or region as input, the goal in Phase 1 was to predict number of cases in the future, and in Phase 2, to prescribe intervention plans that would simultaneously minimize future cases and the stringency (i.e. cost) of the interventions. The teams that performed best in Phase 1 (Nov 2020 - Jan 2021) were selected to particate in Phase 2 (Jan - Feb 2021).
This page describes the technical details of the competition, including the competition setup, qualitative descriptions of the entires, quantitative evaluation methods, and results (see this blog post for a summary). In particular, the Phase 1 and Phase 2 leaderboards make it possible to explore the performance of the different teams interactively at different times of the pandemic, in different countries and regions, and with different weightings on the NPIs; The QE2 result document demonstrates how the Phase 2 entries were evaluated in terms of their unique and useful contributions to a combined super-prescriptor. These results are the same as those the judges used to decide the winners, however, the teams are anonymized to focus on the methodology rather than judging.
The competition was based on a research
paper
by Cognizant's Evolutionary
AI team. For more details, and for media
inquiries, contact
info@evolution.ml.
A general description of the competition goals, technical setup, timelines, evaluation, and rules for the competing teams.
Common tools for the competition, including sample predictors and prescriptors, validation, and performance evaluation.
In the near future, we plan to form an ensemble of predictor and prescriptor submissions and demonstrate their performance. The idea is to demonstrate the value of the community effort, i.e. that by encouraging diverse approaches, and working together we can create solutions that are better than the sum of their parts.
In the near future, we plan to collect the source for many of the contributions, as well as for the superpredictor and superprescriptor, and make them retrainable. In this manner, they can be utilized by the community to build applications for e.g. regional use, or to improve them further.