|Stochastic Systems Group|
Jason Williams - SSG, MIT
Sensor management problems, such as selecting the best mode of operation of a multi-mode radar system, or selecting which node or nodes to activate in a sensor network, have received increased attention in recent years due to developments in advanced sensors and information processing systems. The objective of sensor management systems is to select controls in order to minimize some measure of the expected a posteriori uncertainty. We adopt conditional entropy as our measure of uncertainty due to its behavior with multi-modal uncertainty, which commonly arises in tracking problems. The sensor management problem can be formulated as a dynamic program, which allows the single-stage methods that are commonly used to be viewed as a greedy approximation. The dynamic program can be approximated by an open loop feedback controller, which calculates the controls that maximize a lower bound of the optimal n-step rolling horizon cost function. We show how this can be used to develop a branch and bound algorithm which exploits the characteristics of mutual information to limit the required look-ahead horizon. Our initial investigations into the application of neuro-dynamic programming to the sensor management problem will be discussed briefly.
Problems with this site should be emailed to firstname.lastname@example.org