Abstract
This paper presents a scalable, non-myopic beam scheduling policy for tracking multiple reactive "smart" targets within phased array radar networks. By formulating the scheduling challenge as a Restless Multi-Armed Bandit Problem (RMABP) and employing Whittle indices, the proposed approach accounts for long-term target maneuverability. Numerical simulations demonstrate that this method significantly outperforms traditional myopic policies, offering superior tracking accuracy and computational efficiency in large-scale scenarios.
Introduction
Modern phased array radars can electronically steer beams to track multiple targets simultaneously. However, radar resources are finite, necessitating efficient scheduling. The problem is further complicated by "smart targets"—reactive entities that detect tracking and intentionally maneuver (e.g., switching from a steady flight path to a high-acceleration turn) to evade detection or increase tracking error.
Traditional scheduling methods are often myopic, focusing only on immediate error reduction. This can be a "trap" because tracking a smart target now might trigger a maneuver, making it much harder to track in the future. This paper proposes a non-myopic approach to avoid this pitfall, addressing the "curse of dimensionality" that typically plagues optimal scheduling for many targets.
Methodology
The system is modeled as a Restless Multi-Armed Bandit Problem (RMABP). In this framework:
- Bandits: Each target is treated as a "bandit" with a state defined by its Tracking Error Covariance (TEC)—a measure of positional and velocity uncertainty.
- Restlessness: Unlike standard bandits, these "restless" bandits change state even when they are not being tracked.
- Whittle Index: The core solution is the Whittle index, a heuristic that assigns a numerical priority to each target based on the "marginal productivity" of tracking it. It calculates the long-term benefit of tracking a target now versus later.
Key innovations include applying the Whittle index to complex multi-dimensional Kalman filter matrices and providing numerical evidence for "indexability"—a critical mathematical property for the index's effectiveness. The resulting algorithm achieves linear computational complexity ($O(N\tau)$) relative to the number of targets ($N$) and the time horizon ($\tau$), ensuring real-world scalability.
Experimental Setup
The study relies on extensive numerical simulations rather than static datasets. The researchers designed scenarios involving:
- Dynamics Models: Constant Velocity (CV) and Constant Turn (CT) models.
- Target Types: "Reckless" targets (high maneuver probability when tracked) and "Cautious" targets (lower probability).
- Radar Configurations: Networks with varying numbers of radars ($K$) and targets ($N$), maintaining a fixed ratio to test scalability across different network sizes.
Results and Discussion
The Whittle index policy acts as a "prudent" tracker. It sometimes chooses not to track a target with a high current error if doing so would likely trigger a maneuver that degrades performance in the long run.
Key findings include:
- Superior Accuracy: In all simulated scenarios, the Whittle index policy achieved lower total discounted tracking errors than both "TEC-only" (greedy) and standard myopic policies.
- Reduced Suboptimality: The "suboptimality gap" (the distance from the theoretical best possible performance) was significantly smaller for the Whittle policy—often around 3%, compared to 8-13% for baseline methods.
- Stability: As the number of radars and targets increased, the performance of the Whittle index policy remained stable and efficient, confirming its suitability for large-scale radar networks.
Conclusion and Future Work
While the proposed method shows strong numerical performance and scalability, formal mathematical proofs for indexability in multi-dimensional states remain an open challenge. Future work will explore integrating Reinforcement Learning to allow radars to learn target maneuver patterns on the fly and expanding the model to include environmental constraints like clutter or jamming.
Summary updated