Search Applications for Dynamic Targeting of Satellite Observations

In the Dynamic Targeting (DT) planning problem, a satellite uses a look-ahead sensor to gain information on the upcoming environment. This information can be used to intelligently plan observations for the future as to maximize utility. The DT problem is an exciting application of automated planning to further Earth science missions. We build on previous work to analyze the performance of well-known search algorithms on realistic DT problem instances derived from data sets and operational constraints such as power and slewing. We also examine the relevant, varying properties of DT problems and how different search algorithms and strategies may be better suited for these variations. We evaluate Monte-Carlo tree searches, beam searches, partitioned depth first searches, and a greedy approach on three DT problems. Empirical results and analysis highlights the effectiveness in deeper tree searches that exploit the look-ahead view with little exploration.