TY - THES
AU - Miehling, Erik J.
PY - 2011
TI - Stochastic target scheduling for radar resource management : threat assessment and optimal threshold policies
KW - Thesis/Dissertation
LA - eng
M3 - Text
AB - This thesis formulates a stochastic scheduler for use in adaptive resource management of a single Ground Moving Target Indicator (GMTI) radar when faced with tracking multiple, weakly maneuvering targets. The general problem involves first determining a priority allocation of the L targets, then determining the optimal time to spend using the specified allocation. We present a framework for computing the threat estimate and associated priority of each target in a surveillance environment, termed the radar macro-manager. The threat level of each target is computed based on its heading and proximity relative to a set of user-specified static assets. We present a weight assignment algorithm based on the geography of the surveillance region and use eigenvector centrality to assign vulnerability weights to each asset. The error in the threat level is computed based on the error-covariance matrix of each target, provided by a Kalman filter. Both the threat level and threat error are used to compute the respective priority rank distributions. From the priority distributions of each target we specify a queue of tasks to maximize a reward function associated with processing the queue. The queue is determined with the aid of structural results from the field of optimal issuing which involves ordering the priority rank distributions with respect to the monotone likelihood ratio. In particular, we compute an optimal queue which specifies the order in which we observe individual targets. The length of each target observation is controlled by an external stochastic process, termed the radar micro-manager. The problem of determining this optimal stopping time is formulated as a sequential decision process, a type of Markov decision process. We provide conditions such that the optimal policy is characterized by a monotone threshold policy on the partially ordered set of positive definite error covariance matrices of each target. Given that the optimal policy is monotone, we can efficiently approximate its form with an affine hyperplane using a hybrid random search - Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm. Detailed numerical simulations evaluate the performance of both the radar macro-manager and radar micro-manager.
N2 - This thesis formulates a stochastic scheduler for use in adaptive resource management of a single Ground Moving Target Indicator (GMTI) radar when faced with tracking multiple, weakly maneuvering targets. The general problem involves first determining a priority allocation of the L targets, then determining the optimal time to spend using the specified allocation. We present a framework for computing the threat estimate and associated priority of each target in a surveillance environment, termed the radar macro-manager. The threat level of each target is computed based on its heading and proximity relative to a set of user-specified static assets. We present a weight assignment algorithm based on the geography of the surveillance region and use eigenvector centrality to assign vulnerability weights to each asset. The error in the threat level is computed based on the error-covariance matrix of each target, provided by a Kalman filter. Both the threat level and threat error are used to compute the respective priority rank distributions. From the priority distributions of each target we specify a queue of tasks to maximize a reward function associated with processing the queue. The queue is determined with the aid of structural results from the field of optimal issuing which involves ordering the priority rank distributions with respect to the monotone likelihood ratio. In particular, we compute an optimal queue which specifies the order in which we observe individual targets. The length of each target observation is controlled by an external stochastic process, termed the radar micro-manager. The problem of determining this optimal stopping time is formulated as a sequential decision process, a type of Markov decision process. We provide conditions such that the optimal policy is characterized by a monotone threshold policy on the partially ordered set of positive definite error covariance matrices of each target. Given that the optimal policy is monotone, we can efficiently approximate its form with an affine hyperplane using a hybrid random search - Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm. Detailed numerical simulations evaluate the performance of both the radar macro-manager and radar micro-manager.
UR - https://open.library.ubc.ca/collections/24/items/1.0072219
ER - End of Reference