Power-Aware Cognitive Radar

• The paper demonstrates that power-aware cognitive radar frameworks couple estimation accuracy, track continuity, latency, and robustness with resource constraints for improved performance.
• Methodologies involve fixed-total-power multiband waveform design, online bandwidth control via reinforcement learning, and distributed update scheduling to optimize sensing and tracking.
• Networked and spatial deployment strategies, including anti-jamming placement and adaptive mode switching, advance the framework by balancing power, spectrum, and communication resources.

Power-aware cognitive radar framework denotes a family of adaptive radar and cognitive radar network formulations in which sensing actions are selected under explicit or implicit resource constraints. In the recent literature, power-awareness appears as fixed-total-power multi-band waveform design, online bandwidth selection, anti-jamming deployment under SINR-based detection models, active-versus-passive mode switching, communication-budgeted update scheduling, power-aware beampattern design in massive MIMO, and constrained scan/track allocation under a time budget (Mishra et al., 2017, Tholeti et al., 2024, Cai et al., 5 Dec 2025, Bouhou et al., 23 Jul 2025, Lu et al., 4 Jun 2026). Several works explicitly state that the objective is not always direct power minimization; instead, the framework couples estimation accuracy, track continuity, latency, interceptability, communication load, and adversarial robustness to resource-aware sensing decisions.

1. System concept and architectural scope

A standard cognitive-radar loop is stated as: transmit waveform with chosen parameters, receive measurement, track target state using a filter, and use feedback to adapt the next waveform (Tholeti et al., 2024). In this formulation, adaptation occurs before each transmission, and the transmitter uses prior measurements and earlier waveform choices, written as $$\mathbf{I}_k = \left(\mathbf{z}^{k-1},\boldsymbol{\theta}^{k-1}\right)$$, to choose the next control vector $\boldsymbol{\theta}_k$. The same paper makes the waveform-control vector explicit as $\boldsymbol{\theta}=[\lambda,b]^T$, where $\lambda$ is PRF and $b$ is bandwidth.

In networked settings, the architecture expands to multiple radar or radar/ESM nodes and a fusion center. One line of work models a cognitive radar network as a communication-constrained tracking system in which nodes observe multiple maneuvering targets and decide which updates should reach the fusion center under an average update-rate constraint (Howard et al., 2023). A related line considers nodes that choose between active radar mode and passive signal estimation mode, with the fusion center performing class formation over time and broadcasting class information back to the nodes (Howard et al., 2023). The centralized architecture is described as more informed but suffering from network latency, whereas distributed control reduces latency sensitivity.

This scope implies that “power-aware” is broader than transmit-amplitude control alone. It includes decisions over occupied spectrum, bandwidth, dwell time, node-update opportunities, sensing mode, deployment geometry, and hardware actuation. This suggests that the operative abstraction is a constrained adaptive sensing system rather than a single waveform-optimization module.

2. Fixed-power waveform adaptation and estimation efficiency

A foundational waveform-level formulation studies a single-target, single-antenna cognitive radar that adapts its transmitted spectrum while keeping the total radiated power fixed (Mishra et al., 2017). The cognitive waveform occupies only $N_b$ disjoint frequency bands,

$$\mathcal B_i=\left[\omega_{C_i}-\frac{B_i}{2},\,\omega_{C_i}+\frac{B_i}{2}\right], \qquad i=1,\dots,N_b,$$

with total-power preservation enforced by

$$\int_{B_h} |H(\omega)|^2\,d\omega = \sum_{i=1}^{N_b}\int_{B_i} |\widetilde H(\omega)|^2\,d\omega = \sum_{i=1}^{N_b}P_i = P.$$

The paper explicitly emphasizes that the cognitive radar has the same total power as the conventional waveform, but more power per occupied Hz. Its delay-estimation analysis gives

$$CRLB_R(\hat\tau_0)=\frac{1}{\mathrm{SNR}\cdot \overline F^{\,2}}, \qquad CRLB_{CR}(\hat\tau_0) = \frac{1}{\sum_{i=1}^{N_b}\mathrm{SNR}_i\cdot \overline F_i^{\,2}},$$

and proves that, for flat spectra, equal-width subbands are always at least as good as the conventional radar in CRLB terms. The extended Ziv–Zakai analysis further states that the cognitive radar performs well in low-SNR regions because the effective occupied-band SNR is increased under fixed total power.

The same paper treats band selection and power allocation as coupled design variables. Given a radar environment map, subband selection is formulated as a block-sparse approximation problem, while power is redistributed to maximize the Fisher-information term $\sum_i \mathrm{SNR}_i\overline F_i^2$. The stated design trade-off is explicit: less occupied bandwidth gives higher power per Hz and better robustness in noise, but less bandwidth can reduce intrinsic delay sensitivity unless power is allocated carefully.

A distinct online formulation studies bandwidth selection for ballistic-target tracking (Tholeti et al., 2024). Its trend study concludes that PRF variation had little effect on range error, whereas bandwidth strongly affects both range error and track continuity. Track loss is defined as no correlation between predicted and measured range for 5 consecutive transmissions. The paper proposes three adaptive methods—bandwidth scaling, tabular Q-learning, and $\boldsymbol{\theta}_k$0-step lookahead Q-learning—with action set

$\boldsymbol{\theta}_k$1

and reward

$\boldsymbol{\theta}_k$2

The paper explicitly states that it is not a power-minimization study, but it also states that bandwidth is a radar resource-control variable with operational implications for power-aware systems.

Taken together, these works define a waveform-level power-aware framework in which cognition redistributes a fixed power budget across spectrum, or adapts bandwidth online, to balance estimation accuracy against uncertainty, interference, and continuity-of-track. A plausible implication is that fixed-power spectral agility and feedback-driven bandwidth control are complementary rather than competing interpretations of power-aware radar.

3. Spatial deployment under jamming and physically grounded resource models

Power-aware cognition also appears in spatial deployment. FARDA, the Fast Anti-Jamming Radar Deployment Algorithm, models the placement of radars in an enemy-jammed region so that final detection coverage is maximized (Cai et al., 5 Dec 2025). The surveillance area is discretized into a grid $\boldsymbol{\theta}_k$3, the detected set is

$\boldsymbol{\theta}_k$4

and the optimization objective is

$\boldsymbol{\theta}_k$5

The task is cast as an MDP

$\boldsymbol{\theta}_k$6

with state

$\boldsymbol{\theta}_k$7

continuous action $\boldsymbol{\theta}_k$8, reward $\boldsymbol{\theta}_k$9, and $\boldsymbol{\theta}=[\lambda,b]^T$0. Before PPO training, the method applies dimension reduction to the deployable boundary

$\boldsymbol{\theta}=[\lambda,b]^T$1

continuous relaxation on a continuous boundary $\boldsymbol{\theta}=[\lambda,b]^T$2, and environment simplification through a reduced heatmap and coarser sampling.

The perception module combines a DCNN and an LSTM. A thresholded channel is added to the heatmap, convolution and pooling produce $\boldsymbol{\theta}=[\lambda,b]^T$3, the LSTM yields $\boldsymbol{\theta}=[\lambda,b]^T$4, and the final policy input is

$\boldsymbol{\theta}=[\lambda,b]^T$5

The actor predicts $\boldsymbol{\theta}=[\lambda,b]^T$6 for a Gaussian action distribution, and the critic predicts $\boldsymbol{\theta}=[\lambda,b]^T$7. Empirically, FARDA achieves coverage comparable to or better than GA1D and PSO1D while deploying radars approximately 7,000 times faster; the paper also reports a custom efficiency improvement from about 0.13 to 4.21.

The same work explicitly states that it is not an energy-minimization paper in the strict sense, but it identifies several indirect power-aware aspects. The radar model uses transmit powers, antenna gains, wavelength, bandwidth, and noise power in the SINR and detection-probability calculations; the simplifications reduce computational load; and neural inference replaces repeated population evaluation. This suggests a broader interpretation of power-awareness in which deployment latency and compute burden become system-level resource variables alongside physical transmit parameters.

4. Networked update scheduling, mode control, and class-aware sensing

In cognitive radar networks, power-aware control often appears as communication-aware updating. One framework uses Age of Information and Age of Incorrect Information metrics to decide when nodes should update a fusion center under a shared spectrum budget (Howard et al., 2023). If there are $\boldsymbol{\theta}=[\lambda,b]^T$8 resource blocks per CPI and only a fraction $\boldsymbol{\theta}=[\lambda,b]^T$9 is available, the network update capacity is $\lambda$0, and the average update rate available per node is

$\lambda$1

The centralized policy chooses exactly $\lambda$2 nodes per CPI according to

$\lambda$3

with node score

$\lambda$4

The distributed policy assigns each target to the closest tracker and uses an AoII threshold rule. The paper concludes that the distributed AoII policy performs best overall and achieves nearly double the probability of less than 100 meter accuracy while using the same spectrum budget.

A complementary CRN framework explicitly trades off tracking quality against energy and radiated-power usage by allowing each node to choose between active radar mode and passive signal estimation mode (Howard et al., 2023). Passive ESM is described as reducing the node’s power usage, especially radiated power, lowering observability to adversaries, and providing extra information about target class. Targets are modeled by motion and emission DTMCs, classes are associated through the $\lambda$5-Wasserstein metric, and mode selection is posed as a two-arm bandit problem. The paper reports that centralized and distributed class-aware policies use radar only about 80% of the time, lower effective radiated power relative to radar-only operation, and that centralized control with zero latency achieves the best tracking performance, while increased latency causes roughly an order-of-magnitude increase in tracking error.

Another class-based formulation uses a single central coordinator and one bandit algorithm per node rather than a joint $\lambda$6 action space (Howard et al., 2023). The reward is based on normalized Shannon entropy of the estimated motion and signal distributions,

$\lambda$7

and the system clusters targets over many tracks to learn class distributions that become prior information for future sensing decisions. Its simulations report 70–80% active radar, a random policy that performs 10–15% worse than the bandit policy, and a final-epoch tracking advantage of about 5% over radar-only.

These networked results place power-aware cognition at the intersection of sensing, communication, and classification. They also show that lower radiated-power usage need not be achieved by degrading information quality; passive observations and age-aware scheduling can be used to preserve or improve tracking if the network exploits class information and freshness structure.

5. Planning engines, constrained reinforcement learning, and scalable implementations

Recent power-aware frameworks differ sharply in planning mechanism. A massive MIMO formulation for joint multi-target detection and tracking extends a single-target POMCP algorithm by assigning each target an independent POMCP tree (Bouhou et al., 23 Jul 2025). For target $\lambda$8, the action is

$\lambda$9

so the planner predicts both angle and a power-related coefficient derived from estimated range. Waveform design then solves

$b$0

subject to fixed total transmit power and a fairness constraint relative to the uniform beampattern baseline. The reward is modified so that it is 1 only when both the predicted angle bin and predicted power coefficient match the true future values. With $b$1 targets, $b$2 virtual channels, and $b$3, the power-aware waveform improves detection probability for the weakest target and gives slightly better position RMSE than uniform allocation.

A constrained deep reinforcement learning framework treats track-while-scan time management as a CMDP and learns both policy parameters and a dual variable simultaneously (Lu et al., 4 Jun 2026). The utility is

$b$4

with discounted constraint

$b$5

The training reward is the Lagrangian-relaxed form

$b$6

and the dual variable is updated by

$b$7

The paper reports track initiation times of 6 s for CDRL-DDPG, 7.83 s for CDRL-DQL, 10.50 s for equal allocation, 14.67 s for distance-based allocation, and 20.33 s for optimization-based allocation, and states that online decision-making takes about 1.9 ms for CDRL versus about 5.4 ms for the optimization-based approach.

A hardware-efficient implementation replaces a conventional cognitive MIMO transmitter with a single-antenna transmitter plus a transmissive reconfigurable intelligent surface whose diagonal phase matrix is

$b$8

and whose effective transmission vector is $b$9 (Umra et al., 17 Sep 2025). Multi-target detection is posed through a CFAR adaptive matched filter, while a SARSA agent learns target-bin selection and phase-shift control. The reward is

$N_b$0

In simulations at 28 GHz with 400 spatial bins and 500 Monte Carlo trials, the TRIS-aided RL radar matches MIMO and, with $N_b$1, surpasses MIMO for the lowest-SNR target when the number of RIS elements exceeds 169. The paper attributes the gain to passive beamforming with far fewer RF chains and reduced cost and energy requirements.

Across these implementations, the common structure is not a single learning algorithm but a recurring control template: a belief or state abstraction, a resource-sensitive action space, a reward or utility tied to detection/tracking quality, and either a fairness constraint or a budget constraint. This suggests that power-aware cognitive radar is defined more by constraint coupling than by any particular planner.

6. Adversarial observability, coordination secrecy, and interpretive boundaries

Power-aware cognition has also been treated as a secrecy and counter-countermeasure problem. One framework models the radar as a constrained utility maximizer,

$N_b$2

and studies how an adversary can infer the radar’s utility or constraint through revealed preference and inverse reinforcement learning (Pattanayak et al., 2022). The proposed cognition-masking scheme deliberately uses slightly sub-optimal responses so that the adversary’s Neyman–Pearson detector becomes unreliable. Theoretical guarantees are given in terms of reduced feasibility margin and a Type-I error condition $N_b$3. The paper explicitly frames this as a system-level ECCM objective: small controlled performance sacrifice buys strategic secrecy.

A related inverse multi-objective optimization framework asks whether intercepted radar emissions are consistent with coordinated Pareto-optimal behavior under a shared total power budget (Snow et al., 2023). Coordination is defined by

$N_b$4

or equivalently by a weighted-sum scalarization with positive weights. The paper then derives necessary and sufficient linear feasibility conditions for whether noisy observed emissions are rationalizable by such a coordinated optimization and proves a false-alarm bound for the detector. In this view, power-aware radar coordination itself becomes an observable statistical signature.

The literature also places clear boundaries on the term “power-aware.” FARDA states that it is not explicitly an energy-minimization work in the strict sense, though it is relevant through physically grounded power modeling, deployment efficiency, and lower runtime cost (Cai et al., 5 Dec 2025). The online waveform-selection paper explicitly states that it does not formulate a power-minimization objective, but that adaptive bandwidth control can support power-aware operation (Tholeti et al., 2024). The AoI/AoII scheduling paper states that it is not itself a power-allocation paper, even though reduced communications imply lower signaling overhead and likely lower energy usage (Howard et al., 2023). The constrained DRL time-allocation paper likewise manages time rather than explicitly power, while presenting itself as a generic constrained resource-allocation framework that could be reinterpreted for power-aware control (Lu et al., 4 Jun 2026).

This suggests that, in current usage, a power-aware cognitive radar framework is best understood as a constrained adaptive sensing architecture in which power, time, bandwidth, communication opportunities, deployment choices, secrecy margins, or hardware degrees of freedom are coupled to estimation, detection, tracking, and survivability objectives. The contemporary research direction is therefore not a single canonical framework, but a technically coherent family of formulations linked by resource-aware cognition.