Sensor Management with Uncertain Performance Characteristics

Previous work has presented an information-theoretic sensor management framework for the detection of static targets. This framework is based on the expected discrimination gain maximization technique of Kastella. The sensor manager searches for N targets within a grid of cells using M sensors, which may be thought of as a reconfigurable sensor array. Sensor probabilities of detection and false alarm are used in the mathematical structure of the sensor manager, and these probabilities have previously been assumed to be certain. Realistic problems, however, will inevitably involve uncertainty. This paper introduces uncertain sensor P_d and P_f values into the mathematical framework and allows for their incorporation into the sequential structure of the sensor manager. The performance of the presented sensor management technique is then compared to direct search, where the sensors sweep through the grid in a predefined sampling pattern. The sensor manager is found to be superior to direct search when the uncertainty present in the problem is properly modeled. When uncertainty is present but not modeled, the performance of the sensor manager is severely degraded. This result indicates that uncertainty modeling will be important and necessary for the successful application of the presented sensor manager to real-world problems. Additional simulations examine the robustness of the sensor manager to errors in the assumed densities for P_d and P_f, and the performance of the sensor manager is found to remain strong even when the assumed P_d and P_f densities differ from the true densities