Bayes Risk-Based Radar Resource Management for Joint Multi-Target Track Maintenance and Classification

Abstract

Modern multifunction radar systems have to assign limited sensing resources across multiple targets and radar tasks. For mission specific scenarios, the allocation should not only improve simple performance measures such as estimation accuracy or information gain. The allocation should be focusing on the operational consequences of the decisions supported by the radar. Therefore, in this thesis a novel Bayes risk-based radar resource management framework for joint multi-target track maintenance and classification is proposed. The framework assigns resources by minimizing the expected cost of incorrect target existence and classification decisions under a shared radar time budget.

The proposed approach represents target classes and absence within a novel joint Bayesian hypothesis space. Each radar update consists of a kinematic range-azimuth measurement and a scalar class related feature measurement. These measurements are used in a recursive Bayesian update that accounts for target existence uncertainty, class uncertainty, missed detections, and clutter. The resulting belief state is then used in a Bayes risk decision formulation with a user-defined cost matrix. Mission-specific objectives can then be encoded directly, by choosing the right costs in this matrix. These could for example include different penalties for: missed targets, false tracks, class-confusion errors, high-priority target classes, or mission specific important regions.

Because the resulting expected Bayes risk cannot generally be evaluated in a closed analytical form, the conditional decision probabilities have to be approximated using Monte Carlo sampling. The sampled risk values are then converted into deterministic surrogate risk functions through smoothing, monotonicity enforcement, and finally interpolation. After this processing, the resulting surrogate functions are used in a constrained resource optimization.

The simulation results show that the proposed framework produces interpretable adaptive allocation behavior. This behavior can be related directly to the specified mission costs and the estimated belief state. The ability to encode mission-specific costs in the Bayes risk objective therefore becomes one of the key features of the framework. By changing the cost matrix, the user can determine which decision errors, target classes, or operational regions should have the strongest influence on the resource allocation.

More resources are assigned to targets for which additional sensing is expected to reduce the decision risk most strongly. This can occur because of poor measurement quality, high uncertainty, target appearance or disappearance, or increased operational cost. The framework performed particularly well in scenarios with mission-specific objectives, with cumulative risk improvements of up to approximately 30% compared with a fixed equal-budget allocation. Overall, the optimized allocation reduced the cumulative realized Bayes risk in all considered scenarios.

These results show that the proposed Bayes risk objective is suitable for joint track maintenance and classification RRM. The framework shows a direct link between sensing resource allocation and mission dependent decision costs. The observed fluctuations in the realized risk also show limitations of the myopic one step ahead implementation. This limitation motivates future work on longer horizon optimization and the joint optimization of multiple resources at the same time.