Discovery of Nonmyopic Value Functions for Sensors Management

New sensors, harder-to-detect platforms, and accurate target identification are increasing the need to effectively and rapidly collect and exploit sensor data so that appropriate actions can be taken. However, the complexities of real world sensor data and scenarios has challenged efforts to automatically manage sensors. In particular, although several theoretical approaches (decision-theoretic, information-theoretic, automatic control theory, heuristic) have been proposed, all of these approaches have been myopic - they make decisions with respect to very short time horizons. This clearly in not optimal, and a very important challenge for sensor management is to develop methods that can develop plans and schedules which take into account a substantial time horizon.We propose t use off-line processing to discover the relationship among computable attributes of a situation (e.g., a marginal value function), possible actions, and the long term (i.e, the nonmyopic value of actions taken in that context). We propose reinforcement learning techniques to operationalize this process. We propose to embed these techniques in a far-sighted decision-theoretic architecture (FS-DTSM) for sensor management. The FS_DTSM architecture is based on the decision-theoretic sensor management (DTSM) architecture for sensor management developed at Wright Laboratory.