Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems

Bin Guo; Bo Yang; Chau Yuen; Chenyang Yan; George C. Alexandropoulos; Merouane Debbah; Xuelin Cao; Xueyao Zhang; Zhiwen Yu

REVIEW 4 major objections 5 minor 15 references

Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read In covert underwater pursuit, AUVs should transmit only when the sensed information's task value survives acoustic and covert constraints, not merely when a link exists.

desk verdict A well-motivated integration of task-oriented information value with realistic acoustic and covert constraints, but the 20% gain over IC3Net is not pinned on the value estimator—missing ablations and reproducibility details keep this conditional. read the letter →

arxiv 2607.13880 v1 pith:RML3SNGG submitted 2026-07-15 cs.LG

Xueyao Zhang , Chenyang Yan , Bo Yang , Xuelin Cao , Zhiwen Yu , Bin Guo , George C. Alexandropoulos , Merouane Debbah

show 1 more author

Chau Yuen

This is my paper · ORCID

classification cs.LG

keywords autonomousunderwatervehiclesmulti-agentreinforcementlearningcovertcommunicationtask-orientedbearing-onlylocalizationacousticchannelsvalueofinformationpursuit-evasion

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

This paper argues that in covert multi-AUV cooperation, communication decisions should be driven by the task-level value of the sensed information after it has been degraded by real underwater acoustic channels and covert-detection constraints. It proposes SVR-MARL, a multi-agent reinforcement learning framework that decomposes each AUV's policy into a communication phase and a movement phase, and defines a 'value of sensed information' that is reshaped by propagation delay, interference, decoding reliability, and eavesdropper detection risk into a 'practical information value.' In a simulated bearing-only pursuit scenario, the paper claims this approach achieves a task efficiency of about 1.2 × 10^-2, roughly 20% higher than a learned-gating baseline and better than always-transmit or no-communication schemes. If correct, the work demonstrates that task-oriented covert communication is an information-selection problem, not just a link-optimization problem.

What carries the argument

The key machinery is the conversion of 'value of sensed information' (the potential collaboration benefit of a local message under ideal sharing) into 'practical information value' (the benefit after passing through a physics-grounded communication environment). This environment computes received power from the Thorp path-loss model and fading, models packet arrival windows and overlap interference, checks decoding via SINR, and enforces covertness through a KL-divergence detection constraint. The policy uses separate communication and task heads, a GRU-based temporal context encoder, a learned sensed-information value estimator, and PPO updates with Lagrangian penalty terms.

What would settle it

A field trial or high-fidelity re-simulation using measured underwater acoustic propagation statistics and a realistic passive eavesdropper detector, comparing the value-driven policy against always-transmit and learned-gating baselines, would settle whether the reported 20% task-efficiency gain persists. Specifically, if the real eavesdropper's detection probability as a function of received power deviates strongly from the KL-divergence model, or if packet-overlap interference behaves differently in practice, the policy's transmission schedule could become either too aggressive or too conser

Watch

Extended reading notes

Core claim

The central claim is that a multi-AUV system can improve covert collaborative localization and pursuit by learning when to transmit and at what power, based on an estimate of how much the local sensing message will actually help teammates after passing through realistic acoustic links and covert constraints. The framework introduces a two-stage policy: a communication actor decides whether and how to transmit, while a task actor decides movement actions. The environment simulates Thorp-model path loss, large- and small-scale fading, propagation delays, packet-overlap interference, decoding outcomes, and a KL-divergence-based covert detection constraint. The communication reward captures coun

Load-bearing premise

The entire evaluation rests on the assumption that the custom simulator's underwater acoustic channel model (Thorp path loss, fading, packet-overlap interference) and the KL-divergence covert detection model faithfully represent real covert underwater acoustic links; if they do not, the learned policy's selective-transmission advantage is an artifact of the simulation.

Editorial extensions

If this is right

Task-aware, physics-aware communication scheduling can outperform both always-transmit and learned-gating baselines in covert multi-AUV pursuit, with roughly a 20% task-efficiency gain.
Covert communication and task movement are coupled: task geometry affects exposure risk, so communication and movement policies should be learned jointly rather than designed separately.
Under realistic acoustic links, indiscriminate full communication is suboptimal because channel contention and decoding failures degrade efficiency.
Adaptive transmit power allocation uses the covert budget more effectively, increasing transmission investment when target uncertainty is high and reducing it when localization is confident.
The framework produces fully distributed policies, where each AUV decides based on local observations and received messages, avoiding reliance on a central scheduler.

Reading between the lines

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

The same value-realization logic could generalize to other resource-constrained cooperative sensing domains, such as radio-frequency or optical networks, where transmission carries detection or energy costs.
A practical extension would be to test sensitivity to the eavesdropper's detection threshold: if Eve is weaker or stronger than modeled, the learned selective-transmission advantage may shrink or grow, suggesting an operating envelope for the policy.
The paper's framing implies that optimal communication in such settings is event-triggered rather than periodic, driven by spikes in estimation uncertainty or task-critical states.
The counterfactual reward design points toward a general principle for task-oriented communication: schedule transmissions by expected reduction in teammate uncertainty per unit exposure risk, rather than by information content or link quality alone.

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

The paper to know about: it extends MARL communication to realistic underwater acoustic channels and covert constraints by adding a learned value-of-sensed-information estimate that shapes both gating and transmit power. The problem is real, and the authors deserve credit for assembling the pieces—Thorp path loss, delay, packet overlap, decoding, and KL-divergence-based covertness—into a single framework that is clearly described. Table I is honest about prior work, and the case study is a sensible testbed for bearing-only pursuit.

The soft spot is exactly what the stress-test note says. SVR-MARL differs from IC3Net in two major ways: it adds the value estimator, but also swaps binary gating with fixed power for a continuous Beta-distributed power action plus Lagrangian penalty terms for covertness and reliability. The reported 20% improvement over IC3Net could come entirely from the extra control knob and the constraint penalties, not from the "value-driven" mechanism the title and abstract foreground. There is no ablation isolating SV-Est, and the paper does not report hyperparameters, reward weights, or statistical variance. That is load-bearing, since the central claim is that value-aware scheduling is what helps.

I also share the reader's circularity concern: "practical information value" is defined through the same task-efficiency reward used for training and evaluation, so it is closer to reward shaping than an independently predicted benefit. That is not disqualifying for an engineering paper, but it should be said plainly, and the authors should at least discuss what would distinguish a genuine value estimator from a shaped reward.

The math and modeling are internally consistent, and I see no contradiction that would sink it. The citation pattern looks fair; related work is represented accurately. The paper just does not yet provide enough evidence for the headline claim.

Who gets value from this? Researchers working on underwater multi-robot systems or MARL with physically grounded communication will appreciate the framework and the clearly identified gap, even if the empirical support is thin. It deserves a serious referee but needs major revision: add ablations that separate the value estimator from adaptive power/constraint penalties, release code and seed data, report error bars, and provide a full specification of the SV-Est objective and reward terms.

My recommendation: send it to peer review, but with a strong request for those ablations and reproducibility details before acceptance.

Referee Report

4 major / 5 minor

Summary. This paper addresses the gap between MARL with ideal communication and link-level communication optimization in covert multi-AUV collaboration. The authors propose SVR-MARL, a framework that decomposes each agent's policy into a communication phase and a task-execution phase, with a GRU-based encoder, a 'sensed information value estimator' (SV-Est), and a physics-grounded communication environment that models Thorp path loss, fading, propagation delay, overlap interference, decoding reliability, and KL-divergence-based covert detection. The framework is instantiated on a case study of four AUVs, one target, and one eavesdropper in a 3D pursuit task with bearing-only passive sonar, EKF state estimation, and PPO training. The paper reports that SVR-MARL reaches a task efficiency of about 1.2×10^-2, outperforming IC3Net by about 20%, and claims that value-driven scheduling improves localization and task completion efficiency while reducing unnecessary transmissions.

Significance. If the claimed result survives scrutiny, the paper makes a useful conceptual contribution by placing communication decision-making inside the loop of task value, physical-layer realism, and covert constraints—a combination rarely considered together in MARL. The framework is general and the case study is well motivated. However, the current evidence is too weak to establish the central claim: the evaluation is a single simulation with no error bars, no ablations, and no sensitivity analysis. The manuscript also leaves the notion of 'information value' under-specified. With a rigorous empirical evaluation (ablations, hyperparameters, seeds, sensitivity sweeps) and a formal definition of the value estimator's objective, the paper could become an interesting contribution to task-oriented communication and multi-agent learning.

major comments (4)

[§IV.B and §III.C.3/5] The reported ~20% gain over IC3Net is the central quantitative claim, but SVR-MARL changes two things relative to IC3Net: it adds the SV-Est value estimator and replaces the binary gate/fixed-power communication with a continuous Beta-distributed power control plus Lagrangian constraint penalties. No ablation isolates the value estimator. Without variants such as (a) IC3Net-style gate+fixed power augmented with SV-Est and (b) SVR-MARL with the value-estimator input removed, the result is consistent with the explanation that adaptive power and explicit constraint shaping alone account for the improvement. The title/abstract foreground 'value-driven communication scheduling,' so this missing ablation is load-bearing.
[§IV.B] Task efficiency is reported as a single point estimate (≈1.2×10^-2; 'outperforming IC3Net by about 20%'). No standard deviation, number of seeds, training curves, or statistical test is provided, and no sensitivity analysis is given for the reward weights, Lagrangian penalty coefficients, EKF covariances, PPO hyperparameters, or the covertness KL threshold. The evaluation therefore does not establish that the improvement is robust, and the reader cannot reproduce or assess the significance of the claim. Provide at least mean±std over seeds and a table of hyperparameters.
[§III.C.3/5] The central concepts of the paper, 'value of sensed information (VSI)' and 'practical information value (PIV),' are described verbally but never given a formal definition or an equation. Specifically, the text says SV-Est supervision is 'based on changes in the receiver's state or task-related loss,' but the actual learning objective, the counterfactual benefit computation, and the communication reward function r_comm are unspecified. As presented, PIV appears to be the communication reward shaped from the same success signal (capture rate/task steps) used for evaluation, making it a reward-shaping term rather than an independently measured utility. This lack of formalization makes the framework's internal mechanism unfalsifiable and needs to be fixed.
[§III.C.4] All performance results are generated inside the custom communication environment (Thorp path loss, fading, overlap interference, KL-divergence covert detection). No parameter sensitivity analysis or known-reference validation is provided for any of the acoustic or covertness parameters, including the detection threshold ε, transmit power bounds, noise levels, or interference model. Since the advantage of learned selective communication depends on the realism of Eve's detection statistics and the channel, the 20% improvement may be a simulator artifact. Add sensitivity sweeps over the key physical-layer parameters and compare at least one aggregate metric with an analytical benchmark.

minor comments (5)

[Abstract] The abstract uses 'practical information' while the body consistently uses 'practical information value'; standardize terminology throughout.
[Fig. 3] The equations for f(γ_i) and D(Q_i^e||Q_i) appear only inside the figure; move them into the text and refer explicitly to them, since they are central to the covertness and reliability constraints.
[Reference [14]] Reference [14] appears to have an author error ('Y. L. L. Brekhovskikh'); verify the correct author name and bibliographic details of the monograph.
[§IV.A] The passive sonar observation model is described qualitatively ('angular observation error increases with target distance'); state the exact sensor model and noise covariance used in the EKF.
[§IV.B] It is unclear whether the 'True Obs' scheme uses the same movement policy as the other schemes or an oracle policy; clarify to interpret this upper bound correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claim is benchmarked against external baselines and no equation in the paper reduces its result to its own inputs.

full rationale

The paper's central claim is an empirical comparison: an SVR-MARL policy trained with a communication reward that includes counterfactual task benefit, physical-layer success, and covert-cost penalties is evaluated on task efficiency (capture rate / task steps) against True Obs, Full Comm, No Comm, and IC3Net (Section IV.B). The evaluation metric is not defined in terms of the trained communication reward, so the reported ~20% gain over IC3Net is not forced by construction. The 'practical information value' concept is operationalized as a reward signal, but using the task objective as reward is standard RL design, not a fitted parameter being renamed as a prediction. The SV-Est estimator is learned from receiver-state changes and used as an input to the communication policy; no equation identifies that estimate with the performance metric. Reference [13] is a self-citation, but it is used only as a Table I example of communication-centric optimization and carries no load-bearing premise, uniqueness theorem, or ansatz. The lack of an ablation isolating SV-Est from adaptive transmit power is a confound/correctness concern, not circularity by construction. No circular step can be exhibited from the paper's own equations, so the score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central claim rests on several unvalidated domain models (covert detection, acoustic propagation, bearing-only observability) and on learned value constructs that are defined through the same reward used to measure success. Nearly all numerical settings are undisclosed, so the framework as described imports a large set of unmeasured assumptions and hidden tuning parameters.

free parameters (6)

Covertness threshold (KL divergence bound)
The covert constraint is written as D(Q_e || Q_0) <= epsilon in Figure 3, but the epsilon value is not reported and no sensitivity analysis is given.
Communication reward weights and Lagrangian penalty coefficients
The reward r_comm balances counterfactual benefit against communication and covert costs; the balance weights are never listed, so the learned behavior depends on undisclosed tuning.
EKF process and measurement noise covariances
Passive sonar angular errors increase with distance, but the noise parameters are not reported, even though they directly affect belief quality and the value estimates.
Transmit power bounds, bandwidth, and access-slot parameters
Communication actions include transmit power, bandwidth, frequency, and access slots; their bounds are essential for decoding and covertness but are not specified.
PPO hyperparameters and network sizes
The paper says it uses a two-stage PPO structure but reports no learning rates, batch sizes, horizons, or network architectures, making the result non-reproducible.
Target motion and initial uncertainty parameters
Mission area and episode length are given, but target dynamics, initial AUV-target geometry, and initial belief uncertainties are not, all of which affect localization difficulty.

assumptions (4)

domain assumption Eavesdropper detection is a binary hypothesis test and covertness is governed by a KL-divergence bound D(Q_e || Q_0) <= epsilon.
Section III.B.1 assumes Eve's observation statistics and noise are known; if this model is wrong, the covertness coupling between task and communication is inaccurate.
domain assumption The Thorp model and standard fading/overlap interference accurately capture underwater acoustic propagation in this scenario.
Section III.C.4 builds the communication environment on Thorp path loss, channel fading, and packet-overlap interference without field validation.
domain assumption Bearing-only passive sonar with EKF gives an unobservable range for a single AUV, while multi-AUV fusion restores trackability.
Section IV.A relies on this observability argument to justify the value of communication; no simulated or real observability analysis is supplied.
domain assumption PPO with reward shaping and Lagrangian penalties converges to a good policy in this partially observable Markov game.
The paper assumes the two-stage PPO training finds useful communication and task policies; convergence in POMGs is not guaranteed and is not analyzed.

invented entities (2)

Value of sensed information (VSI)
purpose: Quantify the potential collaborative benefit of a local observation before physical transmission.
VSI is defined internally and learned from the same task-related losses used in the reward; no external metric or falsifiable prediction is provided (Section III.C.3).
Practical information value (PIV)
purpose: Serve as the communication reward and selection criterion, reshaping VSI by link, delay, interference, and covert constraints.
PIV is implemented as a reward-shaping term in the simulator; there is no independent measurement or prediction outside the paper, so the claimed improvement partly depends on this invented construct (Section III.C.5).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems." pith.science (2026). https://pith.science/paper/RML3SNGG

@misc{pith2026260713880,
  author       = {Pith},
  title        = {Pith review of: Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RML3SNGG}},
  note         = {Machine review of arXiv:2607.13880}
}

read the original abstract

In underwater covert cooperative missions, autonomous underwater vehicles (AUVs) often cannot rely on active sonar to continuously obtain complete information, since active sensing and frequent communications increase the risk of exposure. As a result, AUVs primarily rely on passive observation, an approach that yields incomplete local perception and limited task efficiency. Although underwater acoustic communications can mitigate this limitation through information sharing, they are simultaneously constrained by long delays, severe interference, low reliability, and the risk of covert exposure. Existing communications-oriented multi-agent reinforcement learning (MARL) studies often model communication as an ideal information flow, whereas traditional communication optimization primarily focuses on link-level performance. However, both are insufficient to characterize the actual contribution of perceptual information to cooperative tasks under realistic conditions of covert physical communications. This paper proposes a Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL) framework that leverages practical information to characterize the utility of information for cooperative tasks and learns distributed cooperative policies under realistic communication and covert constraints. Through a case study of covert multi-AUV cooperative localization and tracking, the potential of the proposed framework to improve collaborative task efficiency while reducing unnecessary communication and exposure risks is demonstrated.

Figures

Figures reproduced from arXiv: 2607.13880 by the authors.

**Figure 2.** A perspective on realizing collaborative value under realistic commu [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** A general framework for covert multi-AUV collaboration. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Capture rate and task steps under different communication strategies. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Trajectory-based communication behaviors, including communication [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references

[1]

Underwater Internet of Things in smart ocean: System architecture and open issues,

T. Qiu, Z. Zhao, T. Zhang, C. Chen, and C. L. P. Chen, “Underwater Internet of Things in smart ocean: System architecture and open issues,” IEEE Transactions on Industrial Informatics, vol. 16, no. 7, pp. 4297– 4307, Jul. 2020

2020
[2]

A survey of autonomous underwater vehicle formation: Performance, formation control, and communication capability,

Y . Yang, Y . Xiao, and T. Li, “A survey of autonomous underwater vehicle formation: Performance, formation control, and communication capability,”IEEE Communications Surveys & Tutorials, vol. 23, no. 2, pp. 815–841, Secondquarter 2021

2021
[3]

A Survey of Underwater Multi-Robot Systems,

Z. Zhou, J. Liu and J. Yu, “A Survey of Underwater Multi-Robot Systems,”IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 1, pp. 1-18, January 2022

2022
[4]

Learning when to communicate at scale in multiagent cooperative and competitive tasks,

A. Singh, T. Jain, and S. Sukhbaatar, “Learning when to communicate at scale in multiagent cooperative and competitive tasks,”Proc. Int. Conf. Learn. Represent. (ICLR), 2019

2019
[5]

Low probability detec- tion constrained underwater acoustic communication: A comprehensive review,

S. Liu, M. A. Khan, M. Bilal, and H. H. Zuberi, “Low probability detec- tion constrained underwater acoustic communication: A comprehensive review,”IEEE Communications Magazine, vol. 63, no. 2, pp. 21–30, Feb. 2025

2025
[6]

Learning multiagent commu- nication with backpropagation,

S. Sukhbaatar, A. Szlam, and R. Fergus, “Learning multiagent commu- nication with backpropagation,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pp. 2244–2252, 2016

2016
[7]

TarMAC: Targeted multi-agent communication,

A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat, and J. Pineau, “TarMAC: Targeted multi-agent communication,”Proc. Int. Conf. Mach. Learn. (ICML), pp. 1538–1546, 2019

2019
[8]

On Reliable Data Transfer in Underwater Acoustic Networks: A Survey From Networking Perspective,

S. Jiang, “On Reliable Data Transfer in Underwater Acoustic Networks: A Survey From Networking Perspective,”IEEE Communications Sur- veys & Tutorials, vol. 20, no. 2, pp. 1036-1055, Secondquarter 2018

2018

Show all 15 references

[9]

Joint Bayesian Channel Estimation and Data Detection for Underwater Acoustic Com- munications,

Y . Liang, H. Yu, L. Xu, H. Zhao, F. Ji and S. Yan, “Joint Bayesian Channel Estimation and Data Detection for Underwater Acoustic Com- munications,”IEEE Transactions on Communications, vol. 72, no. 9, pp. 5868-5883, Sept. 2024

2024
[10]

Joint Link Scheduling and Power Allocation in Imperfect and Energy-Constrained Underwater Wireless Sensor Networks,

T. Zhang, Y . Gou, J. Liu, S. Song, T. Yang and J. -H. Cui, “Joint Link Scheduling and Power Allocation in Imperfect and Energy-Constrained Underwater Wireless Sensor Networks,”IEEE Transactions on Mobile Computing, vol. 23, no. 10, pp. 9863-9880, Oct. 2024

2024
[11]

Joint Autonomous Underwater Vehicle Trajectory and Energy Optimization for Underwater Covert Communications,

J. Chen, J. Wang, Z. Wei, Y . Ren, C. Masouros and Z. Han, “Joint Autonomous Underwater Vehicle Trajectory and Energy Optimization for Underwater Covert Communications,”IEEE Transactions on Com- munications, vol. 72, no. 11, pp. 7327-7341, Nov. 2024

2024
[12]

Multi-Agent Graph- Attention Communication and Teaming,

Y . Niu, R. R. Paleja, and M. C. Gombolay, “Multi-Agent Graph- Attention Communication and Teaming,” inProc. Int. Conf. Auton. Agents Multiagent Syst. (AAMAS), pp. 964–973, 2021

2021
[13]

Cooperative target detection with AUVs: A dual-timescale hierarchical MADRL approach,

X. Zhang, B. Yang, Z. Yu, X. Cao, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Cooperative target detection with AUVs: A dual-timescale hierarchical MADRL approach,”arXiv preprint arXiv:2509.13381, 2025

2025
[14]

Y . L. L. Brekhovskikh,Fundamentals of Ocean Acoustics. Springer, New York, 1990

1990
[15]

A new method of target tracking by EKF using bearing and elevation measurements for underwater environment,

N. Modalavalasa, G. S. B. Rao, K. S. Prasad, L. Ganesh, and M. N. V . S. S. Kumar, “A new method of target tracking by EKF using bearing and elevation measurements for underwater environment,”Robotics and Autonomous Systems, vol. 74, pp. 221–228, Dec. 2015

2015

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Underwater Internet of Things in smart ocean: System architecture and open issues,

T. Qiu, Z. Zhao, T. Zhang, C. Chen, and C. L. P. Chen, “Underwater Internet of Things in smart ocean: System architecture and open issues,” IEEE Transactions on Industrial Informatics, vol. 16, no. 7, pp. 4297– 4307, Jul. 2020

2020

[2] [2]

A survey of autonomous underwater vehicle formation: Performance, formation control, and communication capability,

Y . Yang, Y . Xiao, and T. Li, “A survey of autonomous underwater vehicle formation: Performance, formation control, and communication capability,”IEEE Communications Surveys & Tutorials, vol. 23, no. 2, pp. 815–841, Secondquarter 2021

2021

[3] [3]

A Survey of Underwater Multi-Robot Systems,

Z. Zhou, J. Liu and J. Yu, “A Survey of Underwater Multi-Robot Systems,”IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 1, pp. 1-18, January 2022

2022

[4] [4]

Learning when to communicate at scale in multiagent cooperative and competitive tasks,

A. Singh, T. Jain, and S. Sukhbaatar, “Learning when to communicate at scale in multiagent cooperative and competitive tasks,”Proc. Int. Conf. Learn. Represent. (ICLR), 2019

2019

[5] [5]

Low probability detec- tion constrained underwater acoustic communication: A comprehensive review,

S. Liu, M. A. Khan, M. Bilal, and H. H. Zuberi, “Low probability detec- tion constrained underwater acoustic communication: A comprehensive review,”IEEE Communications Magazine, vol. 63, no. 2, pp. 21–30, Feb. 2025

2025

[6] [6]

Learning multiagent commu- nication with backpropagation,

S. Sukhbaatar, A. Szlam, and R. Fergus, “Learning multiagent commu- nication with backpropagation,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pp. 2244–2252, 2016

2016

[7] [7]

TarMAC: Targeted multi-agent communication,

A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat, and J. Pineau, “TarMAC: Targeted multi-agent communication,”Proc. Int. Conf. Mach. Learn. (ICML), pp. 1538–1546, 2019

2019

[8] [8]

On Reliable Data Transfer in Underwater Acoustic Networks: A Survey From Networking Perspective,

S. Jiang, “On Reliable Data Transfer in Underwater Acoustic Networks: A Survey From Networking Perspective,”IEEE Communications Sur- veys & Tutorials, vol. 20, no. 2, pp. 1036-1055, Secondquarter 2018

2018