Using Ecosystem Dynamics to Evolve Skillful Ensemble Forecasts

Evolutionary programming (EP) has been shown to be a means to produce skillful ensemble forecasts for temperature and other weather variables. Here, we present a further development of this technique in which we employ ecosystem dynamics to drive more diversity in spatial ensemble forecasts, thus addressing the underdispersive nature of forecast ensembles. Specifically, co-evolution occurs in a predator-prey context with two underlying EP architectures, where the selection pressure on each species is a function of algorithm performance (root mean square error and Brier skill score). The use of input information by the “prey” EP algorithm varies across the development domain and, in combination with the clustering effect of predation, results in diverse algorithmic groups. Additionally, within-species learning occurs such that “good” solutions propagate faster through the population than is possible with only generation-to-generation progress. In this talk, we demonstrate the method in the context of forecasts for temperature and convective occurrence on the North American domain.