Dynamic Resource Management for Radar Networks

Abstract

Distributed radar networks improve tracking and detection performance through spatial diversity, yet the complexity of shared resource management remains a key barrier to achieving those gains. In particular, such networks require increasingly large spectrum occupancy to meet their tracking requirements. This demand is not unique to radar as communication and sensing systems face similar pressure, making the electromagnetic spectrum a shared, finite resource under growing congestion. We therefore consider scenarios in which a frequency-agile, phased-array radar network must operate in the presence of other transmitters acting as sources of interference. In this thesis, we address radar spectrum management challenges through two scenarios. First, we consider a scenario in which a decentralized radar network is tracking a target in an environment where interference conditions are statistically non-stationary and communications are not guaranteed. The proposed framework aims to maximize the probability of target detection while minimizing the number of sub-band changes across frequency to improve stability of the shared spectral band using only sensed information. A multi-agent multi-armed bandit (M-MAB) learning framework is utilized, where a standard discounted upper-confidence-bound (DUCB) algorithm is extended to include spatial weighting and multi-objective reward fusion. Results demonstrate that the proposed policy reduces excessive frequency switching while achieving similar target detection performance compared to other learning techniques and improved target detection performance compared to rules-based techniques. Second, we consider a scenario in which a radar network with centralized control is tasked with tracking multiple targets simultaneously. We propose a solution where a central controller jointly allocates node--target assignment and frequency assignment with a McCormick relaxation and subsequent rounding and greedy dwell time allocation technique. Furthermore, the inputs to this optimization algorithm require knowledge on detection and estimation performance of each radar node--target pair. We consider the realistic case where we do not have prior knowledge of the spectrum and target strength, which additionally changes over time due to the movement of the targets. Thus, we propose a spectrum mapping solution that estimates target and spectrum states from measured statistics and a learning technique that prioritizes exploration more when map uncertainty is high. Results show that the optimization technique significantly improves tracking performance compared to a nearest-neighbor, fixed spectrum approach, with additional gains over time when exploration is intelligently incorporated.