Non-myopic approaches to scheduling agile sensors for multistage detection, tracking and identification

The paper addresses the problem of sensor scheduling for simultaneous target detection, tracking and identification. We consider sensors with agility in waveform and pointing direction. Scheduling decisions are made using an information based approach, where the merit of competing actions is judged by the information expected to be gained when taking the action. We focus on non-myopic scheduling, where the long-term ramifications of scheduling decisions are accounted for in decision making. Since an exact non-myopic solution is computationally prohibitive, we investigate two approximate approaches: direct approximation of Bellman's equation; reinforcement learning. We show, via simulation, that both techniques provide substantial gains over myopic scheduling.