Contrails: Crowd-Sourced Learning of Human Models in an Aircraft Landing Game

Models of human agents play an important role in the design of automation. However, creating human models is a difficult and complex process. Cognitive processes are not observable and can only be inferred based on carefully constructed experiments and interviews, while actual data on human actions and intent is difficult to obtain. In order to enable the learning of quantitative models directly on data, we developed the Contrails game, an aircraft landing game released to the public on Android phones. Having players perform an abstracted aircraft routing task in the game has allowed us to gather datasets several orders of magnitude larger than typical experiments in the field. Using this data, we have been able to learn a probabilistic model of human control action when presented with an airspace, and to derive quantitative results supporting qualitative observations made in the literature. These results are useful for understanding the simplifications that humans make when presented with complex scenarios involving many moving objects.