Adaptive Resource Management in Cognitive Radar via Deep Deterministic Policy Gradient

Chilukuri K. Mohan; M. Cenk Gursoy; Pramod K. Varshney; Ziyang Lu

REVIEW 4 major objections 5 minor 18 references

Adaptive Resource Management in Cognitive Radar via Deep Deterministic Policy Gradient

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A cognitive radar can learn to allocate its limited dwell time between scanning and multi-target tracking under a hard time budget, using a constrained DDPG agent that updates a Lagrangian dual variable.

desk verdict Useful incremental DDPG radar resource management paper with a correct central idea, but a sign inconsistency in the pseudocode and an unjustified dual update make the constraint-satisfaction claim unverifiable as written. read the letter →

arxiv 2507.04195 v1 pith:2KLVDENN submitted 2025-07-06 eess.SP

Ziyang Lu , M. Cenk Gursoy , Chilukuri K. Mohan , Pramod K. Varshney This is my paper · ORCID

classification eess.SP

keywords cognitiveradarresourcemanagementdeepdeterministicpolicygradientconstrainedreinforcementlearningmulti-targettrackingscanningtimeallocationLagrangiandual

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a cognitive radar can learn, with no hand-crafted allocation rule, how to divide a fixed time budget between scanning for new targets and tracking several known targets, while never exceeding the budget. The authors formulate this as a constrained optimization problem and solve it with a DDPG-based reinforcement learning agent that simultaneously updates its policy networks and a Lagrangian dual variable. If the approach works as claimed, a radar could adapt in real time to changing target counts, distances, and maneuvers, outperforming fixed-allocation baselines in both tracking cost and overall utility.

What carries the argument

The carrying mechanism is the constrained DDPG loop: a continuous action vector $a_t = \{\tau_t^n\}$ of dwell times for all targets, a state vector containing the previous tracking costs, previous dwell times, and the dual variable $\lambda_{t-1}$, and a reward $r_t = U_t - \lambda_t \left(\sum_n \tau_t^n / T_0 - \Theta_{\max}\right)$. The dual variable is updated by $\lambda_{t+1} = \max\left(0, \lambda_t + \alpha\left(\sum_n \tau_t^n / T_0 - \Theta_{\max}\right)\right)$, a simplification of the discounted Lagrangian gradient that uses only the current slot's budget violation. The extended Kalman filter tracking cost and the uniform circular array scanning model, with detection probability mapped via Shnidman's equation, determine the utility.

What would settle it

Run the trained CDRL policy over many test episodes with different target spawn patterns and measure the discounted sum of $\left(\sum_n \tau_t^n / T_0 - \Theta_{\max}\right)$; if this sum is systematically positive (budget violated) or the utility is no better than a well-chosen fixed allocation as $\beta$ varies, the claim of solving Eq. (19) would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that the proposed CDRL algorithm solves the constrained problem in Eq. (19) by alternating between DDPG updates that maximize a Lagrangian reward and a dual-variable update (Eq. (25)) that raises the penalty when the instantaneous tracking-time budget is exceeded. Numerical simulations with five maneuvering targets show that the learned policy keeps total dwell time at or below the budget, spends more time on distant or hard-to-track targets, and achieves better tracking cost and utility than fixed-percentage allocation schemes. The scanning phase uses the remaining time, and the tradeoff between tracking and scanning is set by a coefficient $\beta$.

Load-bearing premise

The result rests on the unproved simplification that updating the dual variable with only the current time slot's budget violation, instead of the discounted sum of violations, is enough to enforce the long-run budget constraint.

Editorial extensions

If this is right

The CDRL policy keeps the total tracking time at or below the budget in most time slots, showing that a learned allocation can respect a hard resource limit.
The learned strategy adapts to target distance and count, allocating more dwell time to distant targets and more time to scanning when tracking is easy.
The continuous action space allows all targets' dwell times to be decided simultaneously, avoiding the discretization and sequential-decision limitations of prior DQN-based approaches.
Scanning performance can be shifted toward faster new-target detection by increasing the tradeoff coefficient $\beta$, giving an operator a tunable knob.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The simplified dual update substitutes the instantaneous budget violation for the discounted sum required by the Lagrangian objective; a convergence or optimality proof for this substitution is not given, so the claim that Eq. (19) is truly solved remains open.
A testable extension would replace the hand-picked step size $\alpha$ and initial $\lambda_0$ with an adaptive scheme and measure whether the budget constraint is satisfied more consistently across different target spawn patterns.
Comparing CDRL against a constrained policy optimization baseline that estimates the constraint value function could isolate whether the dual simplification or the DDPG training is responsible for the reported gains.
The same constrained DDPG structure could apply to other radar resources such as transmit power or bandwidth, where the per-slot budget constraint has the same Lagrangian form.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, quick read of arXiv:2507.04195. It's a legitimate incremental extension of the authors' DQN work [11] to DDPG with continuous simultaneous dwell-time allocation and a GNN track initiation model. The radar model is standard EKF tracking plus scanning with Shnidman detection, and the utility tradeoff is sensible. What's new is applying Lagrangian-dual DDPG (known from [7],[10]) to this radar resource management problem; that's a fair contribution, not a breakthrough.

The paper does a decent job explaining the model and the algorithm. The numerical comparison against fixed-allocation baselines shows the DDPG policy learns to prioritize distant/high-cost targets and generally stays within budget. That part is believable.

But there are soft spots the authors need to fix. The biggest is an internal inconsistency in the dual update. Eq. (25) writes λ_{t+1}=max(0, λ_t + α g_t) with g_t = Σ τ/T0 − Θmax, which is the correct subgradient direction (increase λ when violating). Algorithm 1 line 9 prints λ_{t+1}=max(0, λ_t − α g_t). With the printed update, exceeding the budget lowers λ, weakening the penalty; the pseudocode trains the agent to violate the constraint. Either the implementation follows Eq. (25) and the pseudocode has a sign typo, or the implementation follows the pseudocode and the central claim isn't true. Neither case is acceptable in a preprint. Also, the move from Eq. (24)'s discounted sum to Eq. (25)'s instantaneous term is presented as a simplification with zero convergence or optimality argument. The authors call it a simplification, but it's a heuristic that changes the problem; the claim 'solves the constrained optimization problem in (19)' needs at least a discussion or an Abelian argument.

Minor issues: results are single-run with no error bars; the reward uses N_miss, the true number of undetected targets, which the radar doesn't actually know, so the trained policy may rely on unobservable information. And the comparison baselines are only fixed allocations; a smarter heuristic (e.g., water-filling) would be a stronger yardstick.

Who's this for? Researchers in radar resource management who want a concrete DDPG+Lagrangian recipe. It's not a theoretical contribution. I'd send it to peer review because the application is timely and the model is clearly presented; with the sign fixed and some added analysis or at least experimentation it could be acceptable. As written, I wouldn't cite it.

Recommendation: engage with it, but ask for a corrected pseudocode, error bars, and an honest discussion of the instantaneous dual update before any acceptance.

Referee Report

4 major / 5 minor

Summary. The paper considers a cognitive radar that performs both scanning for new targets and tracking of multiple maneuvering targets, and it frames the resulting dwell-time allocation as a constrained optimization problem with a discounted budget constraint on the fraction of time spent tracking. The authors propose a constrained deep reinforcement learning (CDRL) algorithm based on DDPG, in which the actor and critic networks are trained with a Lagrangian-style reward that includes a penalty controlled by a dual variable, and the dual variable is updated simultaneously with the neural networks. Numerical results compare the learned policy against fixed-allocation baselines and report that CDRL achieves lower tracking cost and higher utility while respecting the time budget. The paper claims that CDRL can autonomously allocate time between scanning and tracking without exceeding the predefined time constraint.

Significance. If the central claims are correct, the paper makes a useful contribution to radar resource management by demonstrating a continuous-action DRL solution to a joint scanning/tracking allocation problem, going beyond the discretized DQN approach of the authors' prior work. The system model is reasonably detailed: an EKF tracking model, a UCA scanning model with Shnidman's detection equation, a GNN-based track initialization mechanism, and a time-budget constraint are all explicitly formulated. The proposed algorithm is an original combination of DDPG with a Lagrangian dual update rather than a fitting of constants to a known result, and the reward function is tied to the radar-specific performance metrics. However, the central claim that CDRL solves the constrained problem in Eq. (19) is undermined by an internal sign inconsistency between Eq. (25) and Algorithm 1, by an unproved simplification of the dual update, and by the absence of statistical validation. These issues are fixable but need to be addressed before the experimental conclusions can be accepted.

major comments (4)

[Section V-C (Eq. (25) and Algorithm 1, line 9)] There is a sign inconsistency in the dual update. Eq. (24) derives λ_{t+1}=max(0, λ_t + α Σ_{m=0}^∞ γ^m (Σ_{n=1}^N τ^n_{t+m}/T_0 − Θ_max)), and Eq. (25) keeps a plus sign for the instantaneous term: λ_{t+1}=max(0, λ_t + α(Σ_{n=1}^N τ^n_t/T_0 − Θ_max)). Algorithm 1, line 9, however, prints λ_{t+1}=max(0, λ_t − α(Σ_{n=1}^N τ^n_t/T_0 − Θ_max)). For a positive violation g_t = Σ_n τ^n_t/T_0 − Θ_max > 0, the printed update decreases λ, which weakens the penalty term in the reward (22) and therefore trains the agent toward violating the budget constraint in (19); for g_t < 0 it increases λ and punishes compliance. Thus the pseudocode and the equation cannot both describe the algorithm actually used. The paper must either correct the sign in Algorithm 1 or explicitly state which version was implemented, since the claim that CDRL satisfies the constraint depends directly on this update.
[Section V-C (Eqs. (24)-(25))] The replacement of the discounted-sum gradient in Eq. (24) with the instantaneous violation in Eq. (25) is presented as a simplification, but no convergence or optimality argument is provided. The constraint in Eq. (19) is a discounted-sum constraint, so updating λ_t from the current g_t alone need not drive the discounted constraint violation to zero; a policy that alternately exceeds and then satisfies the budget can keep λ_t near zero while the discounted violation remains positive. To support the claim that CDRL solves the constrained problem, the authors should either give an analysis of the dual-ascent update (e.g., under a CMDP framework or with a bound on the approximation error) or, failing that, report empirical constraint-violation statistics over complete evaluation episodes across multiple random seeds, rather than only showing a single budget line in Fig. 3.
[Section VI-B (Figs. 4-6)] The numerical comparisons against fixed allocation are based on what appears to be a single simulation run, with no error bars, confidence intervals, or multiple independent seeds. The target spawning and initial conditions are random (Section VI-A1), and the DDPG training itself involves stochastic mini-batch sampling, so the reported differences in tracking cost, waiting time, and utility may be artifacts of a single trajectory or a single training run. The claim that CDRL 'always' outperforms fixed allocation requires repeated independent evaluations, a statement of how many seeds were used, and a report of the variance of each metric.
[Section IV-A and VI-B (Eq. (18), Figs. 4-5)] The utility function in Eq. (18) and the scanning-performance evaluation use N_miss, the true number of undetected targets. In a real radar system, this quantity is not directly observable; what is observable is the track-confirmation process described in Section III-C. The paper should clarify whether N_miss is assumed to be available to the controller during training and evaluation, or whether it is an oracle used only for the numerical study. If the latter, the authors should explain how the policy would be learned or evaluated with an estimate of N_miss, since this directly affects the practical validity of the reported utility values and of the learned allocation strategy.

minor comments (5)

[Section III-C] There is a typo in 'We futher note' (should be 'further'), and the track-initialization rule could be clearer: the text says a slot accumulates K=3 measurements, but also says slots are cleared when no associated measurement is received; please state whether the K detections must be consecutive.
[Section V-B (Eqs. (21)-(22))] The state in Eq. (21) includes λ_{t-1}, while the reward in Eq. (22) uses λ_t, and Algorithm 1 updates λ after the DDPG training step; the paper should specify the exact ordering of reward computation, experience storage, and the dual-variable update so that the time indices are unambiguous.
[Section VI-B (Fig. 5)] The metric 'Waiting Time before Track Confirmation' is not defined in the text; please define how it is computed and whether it is averaged over confirmed tracks, over all spawned targets, or over the simulation horizon.
[Section VI-B] The fixed-allocation baselines are described as arbitrary fixed percentages with dwell times evenly distributed among currently tracked targets, but no details are given on how the total tracking time is capped (e.g., whether 90% means 90% of T_0 even when the budget allows more) or how these baselines are matched in total radar energy; please clarify these implementation details.
[Introduction/Contributions] Although the paper compares CDRL with fixed allocation, it does not compare against the DQN-based approach of the authors' prior work [11] or against the CMDP baselines cited in [7]-[10]; such a comparison, even if only in the numerical section, would help position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CDRL is a standard DDPG plus dual-variable optimization scheme, self-contained and optimized against its own reward by construction rather than masquerading as an external prediction.

full rationale

The paper's central claim is that a DDPG-based constrained reinforcement learning framework can learn dwell-time allocations that maximize a defined utility while approximately respecting a time budget. This is an optimization/learning claim evaluated in simulation, not a first-principles prediction derived from fitted inputs. The utility function in Eq. (18), the Lagrangian relaxation in Eq. (20), the reward in Eq. (22), and the dual-variable updates in Eqs. (24)-(25) are all internal components of the proposed algorithm; using the same utility in training and evaluation is standard for RL and does not constitute circularity. The comparison against fixed-allocation baselines in Figs. 4-6 demonstrates optimization ability, not a hidden equivalence between inputs and outputs. The only self-citation, [11], is used to contrast the prior DQN approach with the present DDPG extension and is not load-bearing for the derivation. The unproved simplification from the discounted violation sum in Eq. (24) to the instantaneous violation in Eq. (25), and the sign inconsistency between Eq. (25) and Algorithm 1 line 9, are correctness/rigor concerns, not circularity. No fitted parameter is renamed as a prediction, no known result is repackaged, and no uniqueness theorem is imported from prior work by the same authors. The derivation chain is therefore self-contained rather than circular; the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard tracking assumptions (constant velocity, Gaussian noise, EKF), a scaling law between SNR and measurement variance, an oracle reward that knows N_miss, and a simplified dual update that is not proven to converge. The paper contributes no new physical entities.

free parameters (6)

beta (tracking/scanning tradeoff coefficient) = 2e4
Hand-chosen in Table I. It sets the relative value of tracking error versus missed targets in Eq. (18), so it directly determines what behavior counts as optimal; no sensitivity analysis is given.
lambda_0 (initial dual variable) = 5000
Hand-chosen in Table I. The Lagrange multiplier starts large, making the budget penalty dominant from the first step; convergence and final feasibility depend on this choice.
alpha (dual variable step size) = 5000
Hand-chosen in Table I. Controls how quickly the constraint penalty grows or shrinks in Eq. (25); no principled tuning or adaptivity is provided.
DRL hyperparameters (actor/critic sizes, learning rates, batch size, replay buffer, discount) = actor 256/128, critic 100/100, lr 0.001, batch 128, replay 1e6, gamma 0.9
Set in Section VI-A without sensitivity analysis. Different choices can change both constraint satisfaction and utility.
K (required consecutive detections for track confirmation) = 3
Set in Section III-C. Changes how quickly a new target becomes tracked and therefore influences scanning performance.
Theta_max (time budget limit) = 0.9
Constraint limit chosen for the scenario; the algorithm is not tested under different budgets.

assumptions (4)

domain assumption Measurement noise variances scale inversely with SNR (Eq. 6)
This relation, attributed to [14], is the mechanism by which dwell time improves tracking accuracy; if the scaling is different, the learned allocation would change.
domain assumption Constant-velocity target motion with Gaussian process noise (Eqs. 1-3)
Standard tracking model but restricts validity to targets that do not maneuver strongly within a revisit interval.
domain assumption The simulator knows N_miss, the number of undetected targets, and uses it in the reward (Eq. 18)
A real radar cannot directly observe how many targets it has failed to detect, so the reward is an oracle signal available only in simulation.
ad hoc to paper Instantaneous constraint violation is a valid stand-in for the discounted-sum gradient in the dual update (Eq. 25)
The paper explicitly calls this a simplification; no convergence proof is provided, and the constrained problem (19) is stated with a discounted sum.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Resource Management in Cognitive Radar via Deep Deterministic Policy Gradient." pith.science (2026). https://pith.science/paper/2KLVDENN

@misc{pith2026250704195,
  author       = {Pith},
  title        = {Pith review of: Adaptive Resource Management in Cognitive Radar via Deep Deterministic Policy Gradient},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KLVDENN}},
  note         = {Machine review of arXiv:2507.04195}
}

read the original abstract

In this paper, scanning for target detection, and multi-target tracking in a cognitive radar system are considered, and adaptive radar resource management is investigated. In particular, time management for radar scanning and tracking of multiple maneuvering targets subject to budget constraints is studied with the goal to jointly maximize the tracking and scanning performances of a cognitive radar. We tackle the constrained optimization problem of allocating the dwell time to track individual targets by employing a deep deterministic policy gradient (DDPG) based reinforcement learning approach. We propose a constrained deep reinforcement learning (CDRL) algorithm that updates the DDPG neural networks and dual variables simultaneously. Numerical results show that the radar can autonomously allocate time appropriately so as to maximize the reward function without exceeding the time constraint.

Figures

Figures reproduced from arXiv: 2507.04195 by the authors.

**Figure 1.** Trajectories of the Targets 010002000300040005000600070008000900010000 Time Slot 0 0.5 1 1.5 2 Target Distance to Radar(m) ×104 Target1 Target2 Target3 Target4 Target5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

**Figure 2.** Distances of the Targets to the Radar [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Comparison of Tracking Performances 10%30%50%70%90% Tracking Time Percentage 20 40 60 80 100 120 140 160 180 Waiting Time before Track Confirmation (s) Fixed Allocation CDRL [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Comparison of Scanning Performances 10%30%50%70%90% Tracking Time Percentage -3 -2.5 -2 -1.5 -1 -0.5 0 ×105 Utility Function (Ut ) Fixed Allocation CDRL [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Comparison of Utility Functions time it takes for track confirmation for a new target) that when the percentage of tracking time is below 90%, it takes less time for the radar to detect a target with the fixed allocation algorithm. This is expected since fixed allocati…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 12 canonical work pages

[11]

Learning- based cognitive radar resource management for scanning and multi- target tracking,

Z. Lu, M. C. Gursoy, C. K. Mohan, and P. K. Varshney, “Learning- based cognitive radar resource management for scanning and multi- target tracking,” inICC 2024 - IEEE International Conference on Communications, 2024, pp. 2785–2790

work page 2024
[7]

Reward constrained policy optimization,

C. Tessler, D. J. Mankowitz, and S. Mannor, “Reward constrained policy optimization,”arXiv preprint arXiv:1805.11074, 2018

arXiv 2018
[10]

A Lagrangian dual framework for deep neural networks with constraints,

F. Fioretto, T. Mak, F. Baldo, M. Lombardi, and P. Van Hentenryck, “A Lagrangian dual framework for deep neural networks with constraints,” arXiv preprint arXiv:2001.09394, 2020

arXiv 2001
[1]

Cognitive radar: a way of the future,

S. Haykin, “Cognitive radar: a way of the future,”IEEE Signal Process- ing Magazine, vol. 23, no. 1, pp. 30–40, 2006

work page 2006
[2]

The devel- opment from adaptive to cognitive radar resource management,

A. Charlish, F. Hoffmann, C. Degen, and I. Schlangen, “The devel- opment from adaptive to cognitive radar resource management,”IEEE Aerospace and Electronic Systems Magazine, vol. 35, no. 6, pp. 8–19, 2020

work page 2020
[3]

Human-level control through deep reinforcement learning,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015

2015
[4]

Deep reinforcement learning control for radar detection and tracking in congested spectral environments,

C. E. Thornton, M. A. Kozy, R. M. Buehrer, A. F. Martone, and K. D. Sherbondy, “Deep reinforcement learning control for radar detection and tracking in congested spectral environments,”IEEE Transactions on Cognitive Communications and Networking, vol. 6, no. 4, pp. 1335– 1349, 2020

work page 2020
[5]

Reinforcement learning for adaptable bandwidth tracking radars,

E. Selvi, R. M. Buehrer, A. Martone, and K. Sherbondy, “Reinforcement learning for adaptable bandwidth tracking radars,”IEEE Transactions on Aerospace Electronic Systems, vol. 56, no. 5, pp. 3904–3921, 2020

work page 2020

Show all 18 references

[6]

Deep reinforcement learning-based radar network target assignment,

F. Meng, K. Tian, and C. Wu, “Deep reinforcement learning-based radar network target assignment,”IEEE Sensors Journal, vol. 21, no. 14, pp. 16 315–16 327, 2021

2021
[8]

Optlayer-practical con- strained optimization for deep reinforcement learning in the real world,

T.-H. Pham, G. De Magistris, and R. Tachibana, “Optlayer-practical con- strained optimization for deep reinforcement learning in the real world,” in2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 6236–6243

2018
[9]

Batch policy learning under con- straints,

H. Le, C. V oloshin, and Y . Yue, “Batch policy learning under con- straints,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 3703–3712

2019
[12]

A constrained POMDP formulation and algorithmic solution for radar resource management in multi-target tracking,

M. Sch ¨ope, H. Driessen, and A. Yarovoy, “A constrained POMDP formulation and algorithmic solution for radar resource management in multi-target tracking,”ISIF Journal of Advances in Information Fusion, vol. 16, no. 1, p. 31, 2021

2021
[13]

Adaptive parameter control for phased-array tracking,

W. Koch, “Adaptive parameter control for phased-array tracking,” in Signal and Data Processing of Small Targets 1999, vol. 3809. SPIE, 1999, pp. 444–455

1999
[14]

Meikle,Modern Radar Systems

H. Meikle,Modern Radar Systems. Artech House, 2008

2008
[15]

An introduction to the Kalman filter,

G. Welch, G. Bishopet al., “An introduction to the Kalman filter,” 1995

1995
[16]

Determination of required SNR values [radar detection],

D. Shnidman, “Determination of required SNR values [radar detection],” IEEE Transactions on Aerospace and Electronic Systems, vol. 38, no. 3, pp. 1059–1064, 2002

2002
[17]

Altman,Constrained Markov Decision Processes

E. Altman,Constrained Markov Decision Processes. CRC Press, 1999, vol. 7

1999
[18]

Continuous control with deep reinforcement learning,

T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,”arXiv preprint arXiv:1509.02971, 2015

2015 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [11]

Learning- based cognitive radar resource management for scanning and multi- target tracking,

Z. Lu, M. C. Gursoy, C. K. Mohan, and P. K. Varshney, “Learning- based cognitive radar resource management for scanning and multi- target tracking,” inICC 2024 - IEEE International Conference on Communications, 2024, pp. 2785–2790

work page 2024

[2] [7]

Reward constrained policy optimization,

C. Tessler, D. J. Mankowitz, and S. Mannor, “Reward constrained policy optimization,”arXiv preprint arXiv:1805.11074, 2018

arXiv 2018

[3] [10]

A Lagrangian dual framework for deep neural networks with constraints,

F. Fioretto, T. Mak, F. Baldo, M. Lombardi, and P. Van Hentenryck, “A Lagrangian dual framework for deep neural networks with constraints,” arXiv preprint arXiv:2001.09394, 2020

arXiv 2001

[4] [1]

Cognitive radar: a way of the future,

S. Haykin, “Cognitive radar: a way of the future,”IEEE Signal Process- ing Magazine, vol. 23, no. 1, pp. 30–40, 2006

work page 2006

[5] [2]

The devel- opment from adaptive to cognitive radar resource management,

A. Charlish, F. Hoffmann, C. Degen, and I. Schlangen, “The devel- opment from adaptive to cognitive radar resource management,”IEEE Aerospace and Electronic Systems Magazine, vol. 35, no. 6, pp. 8–19, 2020

work page 2020

[6] [3]

Human-level control through deep reinforcement learning,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015

2015

[7] [4]

Deep reinforcement learning control for radar detection and tracking in congested spectral environments,

C. E. Thornton, M. A. Kozy, R. M. Buehrer, A. F. Martone, and K. D. Sherbondy, “Deep reinforcement learning control for radar detection and tracking in congested spectral environments,”IEEE Transactions on Cognitive Communications and Networking, vol. 6, no. 4, pp. 1335– 1349, 2020

work page 2020

[8] [5]

Reinforcement learning for adaptable bandwidth tracking radars,

E. Selvi, R. M. Buehrer, A. Martone, and K. Sherbondy, “Reinforcement learning for adaptable bandwidth tracking radars,”IEEE Transactions on Aerospace Electronic Systems, vol. 56, no. 5, pp. 3904–3921, 2020

work page 2020