Adaptive Observation Strategies for Forecast Error Minimization

Using a scenario of multiple mobile observing platforms (UAVs) measuring weather variables in distributed regions of the Pacific, we are developing algorithms that will lead to improved forecasting of high-impact weather events. We combine technologies from the nonlinear weather prediction and planning/control communities to create a close link between model predictions and observed measurements, choosing future measurements that minimize the expected forecast error under time-varying conditions. We have approached the problem on three fronts. We have developed an information-theoretic algorithm for selecting environment measurements in a computationally effective way. This algorithm determines the best discrete locations and times to take additional measurement for reducing the forecast uncertainty in the region of interest while considering the mobility of the sensor platforms. Our second algorithm learns to use past experience in predicting good routes to travel between measurements. Experiments show that these approaches work well on idealized models of weather patterns.