Pith. sign in

REVIEW 3 major objections 4 minor 59 references

Distributed Link Sparsification for Scalable Scheduling Using Graph Neural Networks (Journal Version)

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read GNN-adjusted thresholds let dense wireless networks drop scheduling overhead without losing utility.

desk verdict A credible engineering advance with a clean overhead gradient and broad experiments, but the utility constraint rides on unvalidated surrogate gradients—treat the Pareto claim as conditional, not proven. read the letter →

arxiv 2509.05447 v1 pith:33ZXUXRR submitted 2025-09-05 cs.NI cs.DMcs.LGeess.SP

classification cs.NIcs.DMcs.LGeess.SP
keywords graphneuralnetworkslinksparsificationdistributedschedulingthresholdpolicyconstrainedunsupervisedlearningwirelessmulti-hopCSMAMaxWeight
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 dense wireless network can shed much of its link-scheduling overhead without losing capacity by letting each link decide locally whether to enter the contention for the current time slot. The decision is a per-link threshold on link utility, with the threshold scaled by a multiplier produced by a graph neural network that sees only local conflict topology and traffic statistics. The authors claim this GNN-based sparsification dominates a statistical global-threshold baseline: it keeps expected total utility at least as high while reducing the expected number of conflict edges, which in turn lowers average backlog, collision rates, and radio footprint in simulated networks with up to 500 links. Training is offline and constrained, so the trained GCN can be deployed in a fully distributed, asynchronous way across networks of different sizes and even across different scheduling protocols.

What carries the argument

The central object is the per-link threshold policy hv(u(v); z) = u(v)H(u(v) − z(v)u(η)), a Heaviside gate that mutes a link when its utility falls below an individually scaled version of a global quantile threshold. The multiplier z(v) is generated by a graph convolutional network z = ΨG(1;ω) defined on the conflict graph, whose normalized-Laplacian message passing lets each link compute its threshold from local neighborhood information. The argument is carried by Alt-SGD, which alternates between a gradient that reduces the expected number of sparsified edges (available analytically) and a surrogate gradient that raises expected utility, switching according to whether the utility constrain

What would settle it

On a small conflict graph of roughly 20 links where all schedules can be enumerated, compute the true expected utility of the sparse scheduler for many random threshold multipliers and compare it with the two analytical proxies used in training. If the proxies' gradients frequently point in the opposite direction of the true utility gradient, the training signal underlying Alt-SGD is not reliable.

Watch

Extended reading notes

Core claim

The central claim is that the threshold-sparsification policy hv(u(v); z) = u(v)H(u(v) − z(v)u(η))—where η is a global utility quantile and z(v) is a link-specific multiplier generated by a graph convolutional network on the conflict graph—can be optimized to be Pareto superior to the statistical baseline of a single global threshold. The optimization, called Alt-SGD, minimizes the expected number of edges in the sparsified conflict graph under the constraint that expected total utility is no smaller than the baseline's, alternating between utility-increasing updates when the constraint is violated and overhead-decreasing updates when it is satisfied. Because the scheduler is non-differentia

Load-bearing premise

The method depends on the assumption that the two approximate formulas used during training to estimate how sparsification changes expected utility are accurate enough to steer the learned policy into the region where it keeps at least the baseline utility; if they are not, the claimed Pareto improvement may not appear beyond the training setup.

Editorial extensions

If this is right

  • In dense wireless networks, MaxWeight-like schedulers can run with substantially lower point-to-point message complexity without sacrificing total utility.
  • A single offline-trained GCN can be reused on networks of different sizes and topologies, and even carried from one scheduler family (LGS) to another (CSMA), without retraining.
  • A one-layer GCN with very few trainable parameters is enough to beat the global-threshold baseline in dense regimes, which keeps the deployment cost low for resource-constrained devices.
  • The threshold computations are decoupled from real-time scheduling and can be amortized over many time slots, so the sparsification step does not reintroduce the signaling overhead it removes.
  • With a hybrid policy that applies thresholds only to high-degree links, the network keeps low latency at low densities while gaining sparsification benefits at high densities, easing integration into existing protocols.

Reading between the lines

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

  • Beyond the paper, the same muting mechanism could be applied to any distributed protocol whose cost grows with the number of active participants—consensus, gossip, or neighbor discovery—whenever each participant can compute a local utility for joining.
  • The paper's objective counts conflict-graph edges as the overhead proxy; an extension would replace this with the expected number of actual control messages in the scheduler, making the cost model tighter and more directly tied to protocol behavior.
  • Because the GCN input is featureless and the threshold multiplier is a scalar per link, the trained policy could be evaluated on trace-driven topologies with measured interference rather than synthetic random graphs, testing whether the Pareto improvement survives realistic physical-layer effects.
  • The paper notes that the global utility quantile would need periodic refinement under network dynamics but does not analyze the required refresh rate; a testable extension is to bound how quickly η must be re-estimated before the utility constraint is violated.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes a GNN-based per-link threshold policy for link sparsification in wireless multi-hop networks. For each link v, a GCN produces a multiplier z(v) that scales a global utility quantile threshold u(η); links with u(v) ≤ z(v)u(η) are muted before contention. The authors formulate a constrained stochastic program (Problem 2) that minimizes the expected number of conflict-graph edges subject to the expected scheduled utility being at least that of a global-threshold baseline. To train the GCN through non-differentiable schedulers, they derive an exact gradient for the overhead objective under i.i.d. utilities and heuristic surrogate gradients for the utility constraint, and propose an alternating SGD algorithm (Algorithm 2). Simulations compare GCN policies with global-threshold and dense baselines under LGS and CSMA schedulers on ER and BA graph models, reporting lower average backlog and post-sparsification conflict degree in dense networks. The paper claims generalization across topologies and, in some cases, across schedulers.

Significance. The topic is timely and the approach is practical: a featureless GCN that learns local contention thresholds offline and can be applied asynchronously is an appealing way to reduce scheduling overhead in dense networks. The exact closed-form overhead gradient in (11), derived under the i.i.d. utility assumption, is a useful analytic contribution. The evaluation covers four distributed schedulers, multiple graph models, and network sizes up to 500 links, and the source code is released. If the constrained-utility claim can be validated under the deployment distribution, this would be a solid contribution to scalable scheduling. The main weakness is that the utility arm of the training algorithm relies on unvalidated heuristic gradients, and the paper does not directly verify the central constraint (8h) in the deployment experiments.

major comments (3)
  1. [Sec. V, Eqs. (12)-(14); Algorithm 2, line 7; Figs. 5-7] The central feasibility claim (8h) is enforced only by the utility-update branch, whose gradient is computed from the analytical proxies (13a)-(13b). These replace E[u_v hat_v^s_v | G] with a1[1-p(v)] or a2[1-p(v)][1-a3 d_s(v)]; a1, a2, a3 are not reported, and the proxies are not derived from c(.) or from how retaining v displaces conflicting links. Fig. 3(a) validates total utility only on ER test graphs with the same utility eCDF used in training. In the deployment experiments (Sec. VI-C), the dense BA graphs and CSMA schedulers are deliberately outside the training distribution, but only backlog and post-sparsification degree are reported, not E[u(hat_v_s)-u(tilde_v_s)] or the fraction of instances violating (8h). Since u=q*r, lower backlog can be accompanied by lower scheduled utility, so Figs. 5-7 cannot certify the constraint. Please report the constants, validate the surrogate gr
  2. [Sec. V, Algorithm 2] The paper states that convergence of Algorithm 2 is not theoretically guaranteed and uses a fixed 25-epoch stop criterion. The alternating updates switch between the overhead gradient and the utility proxy based on a single-sample comparison, with updates applied after mini-batch aggregation. As a result, there is no certificate that the final iterate satisfies (8h), even on the training distribution. I request empirical convergence diagnostics: e.g., the evolution of E[u(hat_v_s)-u(tilde_v_s)] and E[|E^s|] over epochs, and the final constraint violation on held-out graphs. This is especially important because the utility branch is a heuristic (see previous comment).
  3. [Sec. V, Eq. (11); Sec. VI-A] The exact overhead gradient (11) is derived under the i.i.d. per-link utility assumption, and it requires the PDF f'_u, which is obtained by fitting an MLP to the empirical CDF. In the deployment simulations, utilities are generated by queueing dynamics and are therefore correlated across links and over time; Fig. 4(b) also shows that the eCDF changes across schedulers. The paper does not quantify how these departures from the i.i.d. assumption affect the accuracy of (11) and hence the learned policy. I am not asking for a full theoretical treatment, but a sensitivity analysis, e.g., training with correlated utilities or measuring the bias of the closed-form gradient on simulated states, would substantially strengthen the claim.
minor comments (4)
  1. [Sec. V, Algorithm 2 line 12] The text says 'global norm being clipped', but the update normalizes the gradient by its L2 norm and multiplies by N; this is normalization/scaling, not clipping. Please reword.
  2. [Table I, BA training set] The BA training set lists m in {2, 5, 7.5, 10, 12.5}. In the Barabasi-Albert model, m is the integer number of edges added per new node. Please clarify how non-integer m values were realized, or whether these entries denote average degrees.
  3. [Figs. 5-7] Each plotted point is the average of 60-100 test instances, but no error bars or confidence intervals are provided. For close comparisons, such as 'GCN(2) is slightly better than GCN(1) and GCN(3)' in Fig. 5(a), statistical significance cannot be assessed.
  4. [Sec. V, final paragraph] The sentence 'the policy can always converge after a prescribed number of epochs' is misleading given the preceding admission that convergence is not theoretically guaranteed; 'terminate' would be more accurate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the utility floor is a training constraint, and overhead/latency gains are measured against held-out baselines and schedulers.

full rationale

The paper's central claim is that a GCN-based threshold policy trained by Alt-SGD reduces scheduling overhead while maintaining at least the baseline utility. The utility floor in constraint (8h) is indeed defined as the baseline policy's expected utility, so Fig. 3(a)'s alignment between the GCN and the baseline on total utility is a check that the training constraint is satisfied on the test set, not an independent empirical prediction. However, the paper's substantive claims—reduced scheduling overhead, lower backlog, and lower post-sparsification conflict degree—are not encoded in the constraint and are measured independently on held-out ER and BA test sets and across four distinct schedulers (LGS fixed-deadline, LGS flexible-overhead, Q-CSMA, unweighted CSMA). The overhead objective (8a) and its analytic gradient (11) are derived from the model in Appendix A, not from the baseline. The surrogate utility gradients (13a)-(13b) are explicitly stated as analytical proxies for a quantity with no closed form, and the paper admits that convergence of Algorithm 2 is not theoretically guaranteed; any failure of these proxies would be a correctness or robustness issue, not circularity. Self-citations to [1], [11], [12], [22] provide background and prior GNN-scheduler machinery but are not used to justify the constrained-learning result, which is benchmarked against ZOO and statistical baselines. Overall, the derivation chain is self-contained and the central findings do not reduce to the training inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard modeling assumptions for conflict graphs and scheduling plus several hand-set training constants. No new physical or mathematical entities are introduced. The most fragile inputs are the utility-gradient proxies and the representativeness of the empirical utility CDF.

free parameters (5)
  • a1, a2, a3 = not reported
    Constants in the surrogate utility gradients (13a) and (13b). No fitting procedure or sensitivity analysis is given.
  • global threshold quantile eta and u(eta) = eta in {0.61, 0.66, 0.75} per protocol
    Set by quasi-concave search (Algorithm 3) or manual choice; determines the baseline and all threshold policies.
  • GCN weights omega = trained via Alt-SGD on ER/BA datasets
    Neural network parameters learned from simulated data. For a 1-layer GCN there are only two parameters, but they are still fitted.
  • hybrid policy degree threshold D = 25
    Chosen in (7) so that D tau < 30 tau; no optimization or sensitivity analysis is reported.
  • learning hyperparameters alpha, beta, N, batch size = alpha=1e-4, beta=0.996, N=0.05, batch=100
    Training schedule parameters from Section VI-A; they affect convergence but are not claimed as physical quantities.
assumptions (5)
  • domain assumption The conflict graph G is known or can be estimated from local exchange.
    Invoked in Section III and Section IV-B. The entire sparsification and scheduling pipeline depends on an accurate conflict graph.
  • domain assumption Per-link utilities u(v) are i.i.d. draws from a distribution Omega_u.
    Problem 2 states this explicitly; the analytical expectations in (9) to (14) rely on exchangeability of link utilities.
  • domain assumption Signaling overhead is accurately represented by the expected number of edges |E_s| in the sparsified conflict graph.
    The objective in (3a) equates message complexity with edge count. Actual overhead also involves protocol constants such as frame fraction tau and window size W.
  • ad hoc to paper The optimal global cut-off quantile eta* can be found by quasi-concave search.
    Appendix B uses a peak search for a quasi-concave function; quasi-concavity is only observed empirically in Fig. 4(c), not proven.
  • domain assumption The empirical CDF of per-link utility under dense scheduling is representative of utility under sparsified scheduling.
    Both the GCN training and the global threshold u(eta) rely on the eCDF in Fig. 1(a). If sparsification shifts the utility distribution, the threshold and training distribution become miscalibrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distributed Link Sparsification for Scalable Scheduling Using Graph Neural Networks (Journal Version)." pith.science (2026). https://pith.science/paper/33ZXUXRR

@misc{pith2026250905447,
  author       = {Pith},
  title        = {Pith review of: Distributed Link Sparsification for Scalable Scheduling Using Graph Neural Networks (Journal Version)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/33ZXUXRR}},
  note         = {Machine review of arXiv:2509.05447}
}
read the original abstract

In wireless networks characterized by dense connectivity, the significant signaling overhead generated by distributed link scheduling algorithms can exacerbate issues like congestion, energy consumption, and radio footprint expansion. To mitigate these challenges, we propose a distributed link sparsification scheme employing graph neural networks (GNNs) to reduce scheduling overhead for delay-tolerant traffic while maintaining network capacity. A GNN module is trained to adjust contention thresholds for individual links based on traffic statistics and network topology, enabling links to withdraw from scheduling contention when they are unlikely to succeed. Our approach is facilitated by a novel offline constrained {unsupervised} learning algorithm capable of balancing two competing objectives: minimizing scheduling overhead while ensuring that total utility meets the required level. In simulated wireless multi-hop networks with up to 500 links, our link sparsification technique effectively alleviates network congestion and reduces radio footprints across four distinct distributed link scheduling protocols.

Figures

Figures reproduced from arXiv: 2509.05447 by the authors.

Figure 1
Figure 1. Architecture of our GCN-based distributed sparse scheduler [1]. (a) Select [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Random instances of wireless multi-hop networks generated [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The performance of different threshold policies normalized by that of zero-threshold policy with identical input network state [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: a) Illustration of two LGS-based protocols: With fixed scheduling deadline, a link will be scheduled only if the signaling meets the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The performance of threshold policies vs average conflict [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: The performance of threshold policies vs average conflict [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 57 canonical work pages

  1. [1]

    Distributed link sparsification for scalable scheduling using graph neural networks,

    Z. Zhao, A. Swami, and S. Segarra, “Distributed link sparsification for scalable scheduling using graph neural networks,” in IEEE Intl. Conf. Acoustics, Speech and Signal Process. (ICASSP) , pp. 5308–5312, 2022

  2. [2]

    Cisco annual internet report (2018–2023),

    “Cisco annual internet report (2018–2023),” white paper, Cisco Systems, Inc., Mar. 2020. [Online]. Available: https: //www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/ annual-internet-report/white-paper-c11-741490.html

  3. [3]

    The internet of battle things,

    A. Kott, A. Swami, and B. J. West, “The internet of battle things,” Computer, vol. 49, no. 12, pp. 70–75, 2016

  4. [4]

    6G and beyond: The future of wireless communications systems,

    I. F. Akyildiz, A. Kak, and S. Nie, “6G and beyond: The future of wireless communications systems,” IEEE Access , vol. 8, pp. 133995– 134030, 2020

  5. [5]

    Massive access for 5G and beyond,

    X. Chen, D. W. K. Ng, W. Yu, E. G. Larsson, N. Al-Dhahir, and R. Schober, “Massive access for 5G and beyond,” IEEE J. Sel. Areas Commun., vol. 39, no. 3, pp. 615–637, 2021

  6. [6]

    Toward massive machine type commu- nications in ultra-dense cellular IoT networks: Current issues and ma- chine learning-assisted solutions,

    S. K. Sharma and X. Wang, “Toward massive machine type commu- nications in ultra-dense cellular IoT networks: Current issues and ma- chine learning-assisted solutions,” IEEE Commun. Surveys & Tutorials , vol. 22, no. 1, pp. 426–471, 2020

  7. [7]

    A tutorial on cross-layer opti- mization in wireless networks,

    X. Lin, N. B. Shroff, and R. Srikant, “A tutorial on cross-layer opti- mization in wireless networks,” IEEE J. Sel. Areas Commun. , vol. 24, no. 8, pp. 1452–1463, 2006

  8. [8]

    S. K. Sarkar, T. G. Basavaraju, and C. Puttamadappa, Ad Hoc Mobile Wireless Networks: Principles, Protocols and Applications. Boca Raton, FL, USA: CRC Press, Taylor & Francis Group, 2nd ed., 2013

Show all 59 references
  1. [9]

    Performance evaluation of backpressure routing in integrated satellite-terrestrial backhaul for PPDR networks,

    N. Patriciello, C. A. Grazia, J. Núñez-Martínez, J. Baranda, J. Mangues- Bafalluy, and M. Casoni, “Performance evaluation of backpressure routing in integrated satellite-terrestrial backhaul for PPDR networks,” in Proc. IEEE Intl. Conf. Wireless Mobile Comput., Netw., Commun. ...

  2. [10]

    Integrated access and backhaul: A key enabler for 5G millimeter-wave deployments,

    M. Cudak, A. Ghosh, A. Ghosh, and J. Andrews, “Integrated access and backhaul: A key enabler for 5G millimeter-wave deployments,” IEEE Commun. Mag., vol. 59, no. 4, pp. 88–94, 2021

  3. [11]

    Graph-based deterministic policy gradient for repetitive combinatorial optimization problems,

    Z. Zhao, A. Swami, and S. Segarra, “Graph-based deterministic policy gradient for repetitive combinatorial optimization problems,” in Intl. Conf. Learn. Repres. (ICLR) , pp. 1–21, 2023

  4. [12]

    Link scheduling using graph neural networks,

    Z. Zhao, G. Verma, C. Rao, A. Swami, and S. Segarra, “Link scheduling using graph neural networks,” IEEE Trans. Wireless Commun., vol. 22, no. 6, pp. 3997–4012, 2023

  5. [13]

    Understanding the capacity region of the greedy maximal scheduling algorithm in multihop wireless networks,

    C. Joo, X. Lin, and N. B. Shroff, “Understanding the capacity region of the greedy maximal scheduling algorithm in multihop wireless networks,” IEEE/ACM Trans. Netw., vol. 17, no. 4, pp. 1132–1145, 2009

  6. [14]

    Optimal cross-layer design of wireless fading multi-hop networks,

    A. G. Marques, N. Gatsis, and G. B. Giannakis, “Optimal cross-layer design of wireless fading multi-hop networks,” in Cross Layer Designs in WLAN Systems (N. Zorba, C. Skianis, and C. Verikoukis, eds.), pp. 1– 44, Leicester, U.K.: Troubador Publishing Ltd., 2011

  7. [15]

    Finding a maximal weighted independent set in wireless networks,

    S. Basagni, “Finding a maximal weighted independent set in wireless networks,” Telecommun. Systems, vol. 18, no. 1-3, pp. 155–168, 2001

  8. [16]

    The complexity of channel scheduling in multi-radio multi-channel wireless networks,

    W. Cheng, X. Cheng, T. Znati, X. Lu, and Z. Lu, “The complexity of channel scheduling in multi-radio multi-channel wireless networks,” in IEEE Intl. Conf. Computer Comms. (INFOCOM) , pp. 1512–1520, 2009

  9. [17]

    Local greedy approximation for scheduling in multihop wireless networks,

    C. Joo and N. B. Shroff, “Local greedy approximation for scheduling in multihop wireless networks,” IEEE Trans. Mobile Computing , vol. 11, no. 3, pp. 414–426, 2012

  10. [18]

    Distributed greedy approxima- tion to maximum weighted independent set for scheduling with fading channels,

    C. Joo, X. Lin, J. Ryu, and N. B. Shroff, “Distributed greedy approxima- tion to maximum weighted independent set for scheduling with fading channels,” IEEE/ACM Trans. Netw., vol. 24, no. 3, pp. 1476–1488, 2015

  11. [19]

    A new distributed approximation algorithm for the maximum weight independent set problem,

    P. Du and Y . Zhang, “A new distributed approximation algorithm for the maximum weight independent set problem,” Mathematical Problems in Engineering, vol. 2016, no. 1, p. 9790629, 2016

  12. [20]

    A message-passing algorithm for wireless network scheduling,

    I. C. Paschalidis, F. Huang, and W. Lai, “A message-passing algorithm for wireless network scheduling,” IEEE/ACM Trans. Netw. , vol. 23, pp. 1528–1541, Oct. 2015

  13. [21]

    On the complexity of scheduling in wireless networks,

    C. Joo, G. Sharma, N. B. Shroff, and R. R. Mazumdar, “On the complexity of scheduling in wireless networks,” EURASIP J. Wireless Commun. Netw., vol. 2010, no. 1, p. 418934, 2010

  14. [22]

    Distributed scheduling using graph neural networks,

    Z. Zhao, G. Verma, C. Rao, A. Swami, and S. Segarra, “Distributed scheduling using graph neural networks,” in IEEE Intl. Conf. Acoustics, Speech and Signal Process. (ICASSP) , pp. 4720–4724, 2021

  15. [23]

    Q-CSMA: Queue-length-based CSMA/CA algorithms for achieving maximum throughput and low delay in wireless networks,

    J. Ni, B. Tan, and R. Srikant, “Q-CSMA: Queue-length-based CSMA/CA algorithms for achieving maximum throughput and low delay in wireless networks,” IEEE/ACM Trans. Netw., vol. 20, no. 3, pp. 825– 836, 2012

  16. [24]

    A distributed CSMA algorithm for throughput and utility maximization in wireless networks,

    L. Jiang and J. Walrand, “A distributed CSMA algorithm for throughput and utility maximization in wireless networks,” IEEE/ACM Trans. Netw., vol. 18, no. 3, pp. 960–972, 2010

  17. [25]

    Constant-time distributed scheduling policies for ad hoc wireless networks,

    X. Lin and S. B. Rasool, “Constant-time distributed scheduling policies for ad hoc wireless networks,” IEEE Trans. Auto. Control, vol. 54, no. 2, pp. 231–242, 2009

  18. [26]

    A simple parallel algorithm for the maximal independent set problem,

    M. Luby, “A simple parallel algorithm for the maximal independent set problem,” in Proc. 17th annual ACM Symp. Theory of computing , pp. 1–10, 1985

  19. [27]

    Blind wireless network topology inference,

    E. Testi and A. Giorgetti, “Blind wireless network topology inference,” IEEE Trans. Commun. , vol. 69, no. 2, pp. 1109–1120, 2021

  20. [28]

    Medium access control with coordinated adaptive sleeping for wireless sensor networks,

    W. Ye, J. Heidemann, and D. Estrin, “Medium access control with coordinated adaptive sleeping for wireless sensor networks,” IEEE/ACM Trans. Netw., vol. 12, no. 3, pp. 493–506, 2004

  21. [29]

    Topology control in wireless ad hoc and sensor networks,

    P. Santi, “Topology control in wireless ad hoc and sensor networks,” ACM Comput. Surv., vol. 37, pp. 164–194, June 2005

  22. [30]

    Antenna beamforming and power control for ad hoc networks,

    R. Ramanathan, “Antenna beamforming and power control for ad hoc networks,” in Mobile Ad Hoc Networking (S. Basagni, M. Conti, S. Giordano, and I. Stojmenovic, eds.), pp. 139–173, Hoboken, NJ, USA: John Wiley & Sons, Ltd, 2004

  23. [31]

    A hybrid energy efficient protocol for mobile ad hoc networks,

    N. K. Ray and A. K. Turuk, “A hybrid energy efficient protocol for mobile ad hoc networks,” J. Comput. Netw. Commun. , vol. 2016, no. 1, p. 2861904, 2016

  24. [32]

    Green wave sleep scheduling: Optimizing latency and throughput in duty cycling wireless networks,

    S. Guha, P. B. Basu, C.-K. C. Chau, and R. Gibbens, “Green wave sleep scheduling: Optimizing latency and throughput in duty cycling wireless networks,” IEEE J. Sel. Areas Commun. , vol. 29, no. 8, pp. 1595–1604, 2011

  25. [33]

    Collaborative duty cycling strategies in energy harvesting sensor networks,

    J. Long and O. Büyüköztürk, “Collaborative duty cycling strategies in energy harvesting sensor networks,” Computer-Aided Civil and Infras- tructure Engineering, vol. 35, pp. 534–548, May 2020

  26. [34]

    Low-complexity and distributed energy minimization in multihop wireless networks,

    L. Lin, X. Lin, and N. B. Shroff, “Low-complexity and distributed energy minimization in multihop wireless networks,” IEEE/ACM Trans. Netw., vol. 18, no. 2, pp. 501–514, 2010

  27. [35]

    Energy-efficient link selection and transmission scheduling in mobile cloud computing,

    X. Xiang, C. Lin, and X. Chen, “Energy-efficient link selection and transmission scheduling in mobile cloud computing,” IEEE Commun. Lett., vol. 3, no. 2, pp. 153–156, 2014

  28. [36]

    Energy-efficient link scheduling in time-variant dual-hop 60GHz wireless networks,

    W. Wu, W. Xu, Z. Chen, and M. Yang, “Energy-efficient link scheduling in time-variant dual-hop 60GHz wireless networks,” Concurrency and Computation: Practice and Experience , vol. 32, no. 23, p. e5903, 2020

  29. [37]

    Network information flow,

    R. Ahlswede, N. Cai, S.-Y . Li, and R. Yeung, “Network information flow,” IEEE Trans. Info. Theory , vol. 46, no. 4, pp. 1204–1216, 2000

  30. [38]

    Delay-optimal back- pressure routing algorithm for multihop wireless networks,

    L. Hai, Q. Gao, J. Wang, H. Zhuang, and P. Wang, “Delay-optimal back- pressure routing algorithm for multihop wireless networks,” IEEE Trans. Vehicular Tech., vol. 67, no. 3, pp. 2617–2630, 2018

  31. [39]

    Generalization of graph neural networks is robust to model mismatch,

    Z. Wang, J. Cerviño, and A. Ribeiro, “Generalization of graph neural networks is robust to model mismatch,” in Proc. AAAI Conf. Artificial Intell. (AAAI) , (Philadelphia, PA, USA), pp. 2387–2395, AAAI Press, 2025. ZHAO et al.: DISTRIBUTED LINK SPARSIFICATION FOR SCALABLE SCHED...

  32. [40]

    Learning decentralized wireless resource allocations with graph neural networks,

    Z. Wang, M. Eisen, and A. Ribeiro, “Learning decentralized wireless resource allocations with graph neural networks,” IEEE Trans. Signal Process., vol. 70, pp. 1850–1863, 2022

  33. [41]

    A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications,

    S. Liu, P.-Y . Chen, B. Kailkhura, G. Zhang, A. O. Hero III, and P. K. Varshney, “A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications,” IEEE Signal Process. Mag. , vol. 37, no. 5, pp. 43–54, 2020

  34. [42]

    A low-complexity congestion control and scheduling algorithm for multihop wireless networks with order-optimal per-flow delay,

    P.-K. Huang, X. Lin, and C.-C. Wang, “A low-complexity congestion control and scheduling algorithm for multihop wireless networks with order-optimal per-flow delay,” IEEE/ACM Trans. Netw., vol. 21, no. 2, pp. 495–508, 2012

  35. [43]

    Low-complexity distributed schedul- ing algorithms for wireless networks,

    A. Gupta, X. Lin, and R. Srikant, “Low-complexity distributed schedul- ing algorithms for wireless networks,” IEEE/ACM Trans. Netw., vol. 17, no. 6, pp. 1846–1859, 2009

  36. [44]

    Timescale decoupled routing and rate control in intermittently connected networks,

    J. Ryu, L. Ying, and S. Shakkottai, “Timescale decoupled routing and rate control in intermittently connected networks,” IEEE/ACM Trans. Netw., vol. 20, no. 4, pp. 1138–1151, 2012

  37. [45]

    Queueing systems with a removable service station,

    M. Yadin and P. Naor, “Queueing systems with a removable service station,” J. Operational Research Soc. , vol. 14, pp. 393–405, 1963

  38. [46]

    The threshold policy in the M/G/1 queue with server vacations,

    O. Kella, “The threshold policy in the M/G/1 queue with server vacations,” Naval Research Logistics, vol. 36, no. 1, pp. 111–123, 1989

  39. [47]

    The n-policy of a discrete time Geo/G/1 queue with disasters and its application to wireless sensor networks,

    D. H. Lee and W. S. Yang, “The n-policy of a discrete time Geo/G/1 queue with disasters and its application to wireless sensor networks,” Applied Mathematical Modelling, vol. 37, no. 23, pp. 9722–9731, 2013

  40. [48]

    Traffic-aware two-stage queueing communication networks: Queue analysis and energy saving,

    N. Qi, N. I. Miridakis, M. Xiao, T. A. Tsiftsis, R. Yao, and S. Jin, “Traffic-aware two-stage queueing communication networks: Queue analysis and energy saving,” IEEE Trans. Commun. , vol. 68, no. 8, pp. 4919–4932, 2020

  41. [49]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 32, no. 1, pp. 4–24, 2021

  42. [50]

    A comprehensive survey of multiagent reinforcement learning,

    L. Busoniu, R. Babuska, and B. De Schutter, “A comprehensive survey of multiagent reinforcement learning,” IEEE Trans. Systems, Man, Cybernetics, Part C (Appl. Rev.) , vol. 38, no. 2, pp. 156–172, 2008

  43. [51]

    Giselsson and A

    P. Giselsson and A. Rantzer, Large-scale and Distributed Optimization , vol. 2227. Cham, Switzerland: Springer, 2018

  44. [52]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc. 14th Intl. Conf. Artificial Intelligence and Statistics , pp. 627–635, JMLR Wrksp. & Conf. Proc., 2011

  45. [53]

    Q-learning,

    C. J. Watkins and P. Dayan, “Q-learning,” Machine learning , vol. 8, no. 3-4, pp. 279–292, 1992

  46. [54]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V . Panneershelvam, M. Lanctot, et al., “Mastering the game of go with deep neural networks and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016

  47. [55]

    Deterministic policy gradient algorithms,

    D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Ried- miller, “Deterministic policy gradient algorithms,” inIntl. Conf. Machine Learning (ICML), pp. 387–395, PMLR, 2014

  48. [56]

    Distributed graph coloring: Fundamentals and recent developments,

    L. Barenboim and M. Elkin, “Distributed graph coloring: Fundamentals and recent developments,” Synthesis Lectures on Distributed Computing Theory, vol. 4, no. 1, pp. 1–171, 2013

  49. [57]

    LTE physical layer: Performance analysis and evaluation,

    H. Mousavi, I. S. Amiri, M. Mostafavi, and C. Choon, “LTE physical layer: Performance analysis and evaluation,” Applied Computing and Informatics, vol. 15, no. 1, pp. 34 – 44, 2019

  50. [58]

    On random graphs I,

    P. Erd ˝os and A. Rényi, “On random graphs I,” Publ. Math. Debrecen 6 , pp. 290–297, 1959

  51. [59]

    Barabási and M

    A.-L. Barabási and M. Pósfai, Network Science . Cambridge, U.K.: Cambridge Univ. Press, 2016. Zhongyuan Zhao (Member, IEEE) received his B.Sc. and M.S. degrees in Electronic Engineering from the University of Electronic Science and Tech- nology of China, Chengdu, China, in 200...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.