AI-powered cognitive radar uses DRL to autonomously balance tracking and scanning

A team from Syracuse University has introduced a novel constrained deep reinforcement learning (CDRL) framework to solve one of radar's toughest problems: how to optimally split a fixed time budget between tracking multiple maneuvering targets and scanning for new threats. The paper, published on arXiv (2606.05526), addresses the track-while-scan mode where a single radar must simultaneously maintain tracks on several targets and sweep the environment. The authors formulate this as a constrained optimization problem, then train a CDRL agent to learn the optimal time allocation policy. The agent uses the Deep Deterministic Policy Gradient (DDPG) algorithm, which handles continuous action spaces better than discrete methods like deep Q-learning. Numerical experiments show the CDRL framework autonomously allocates more time to high-priority tracking tasks—those requiring greater attention—while still freeing up time for scanning and staying under the total time budget. The approach outperformed heuristic strategies and a traditional optimization baseline in both tracking accuracy and scanning coverage.

The implications extend beyond radar. This work demonstrates how reinforcement learning can solve real-time resource allocation problems with hard constraints—a challenge common in defense systems, autonomous vehicles, and communication networks. By jointly learning the neural network weights and the dual variable for the constraint, the CDRL framework avoids manual tuning and adapts to changing conditions (e.g., target maneuvering dynamics). The use of DDPG is particularly relevant for any domain where actions are continuous (like power levels, bandwidth, or time slices). While the paper focuses on cognitive radar, the methodology is transferable to other constrained scheduling problems. The authors note that future work could extend the framework to handle multiple radars or dynamic priority settings. For now, this marks a practical step toward fully autonomous radar systems that can maintain situational awareness without human intervention.