Graph Neural Networks for Resource Allocation in Multi-Channel Wireless Networks

Changyang She; Jamie Evans; Jingge Zhu; Lili Chen

REVIEW 4 major objections 7 minor 16 references

Graph Neural Networks for Resource Allocation in Multi-Channel Wireless Networks

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A graph neural network trained on a Lagrangian loss achieves higher sum rate than iterative WMMSE for multi-channel wireless networks, at a fraction of the inference time.

desk verdict A real but incremental multi-channel GNN extension that deserves review, with the performance claim currently unverified due to an underspecified baseline and an unmeasured post-processing projection. read the letter →

arxiv 2506.03813 v1 pith:ZN3D63PR submitted 2025-06-04 cs.LG eess.SP

Lili Chen , Changyang She , Jingge Zhu , Jamie Evans This is my paper

classification cs.LGeess.SP

keywords jointresourcemanagementgraphneuralnetworkswirelesscommunicationmultiplechannelspowerallocationchannelLagrangianprimal-duallearningWMMSE

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 tries to show that a graph neural network can replace an iterative optimizer for joint channel and power allocation in dense multi-channel wireless networks. The proposed JCPGNN-M treats each channel as a subgraph of interfering transceiver pairs and learns, in an unsupervised way, how much power each user should send on every channel. The training signal is the Lagrangian of the sum-rate maximization, so the per-user total power constraint is enforced through Lagrange multipliers instead of through labeled examples. If the reported simulations hold, a single forward pass of the network gives better data rates than the enhanced WMMSE baseline while taking about one percent of its running time, and it transfers to larger networks without retraining.

What carries the argument

The central object is the Lagrangian reformulation in Eq. (12), where the GNN weights are the primal variables and the per-user total-power constraints are dualized with Lagrange multipliers. Training alternates stochastic gradient descent on the network weights with stochastic gradient ascent on the multipliers, following Eq. (13), so that the power constraint is enforced systematically rather than by supervised labels. The GNN architecture itself is a message-passing network over M separate complete subgraphs, one per channel: each vertex carries the direct channel gain as its node feature, each interference edge carries the two cross gains as edge features, and the output layer applies a sigmoid to keep powers in a valid range. A post-processing step then normalizes a user's total power to the maximum only when the raw output would exceed it.

What would settle it

Run JCPGNN-M on a small instance (for example, D=4 transceiver pairs and M=2 channels) where the global optimum can be found by exhaustive search over channel assignments and power levels; if the raw network output before post-processing violates a user's total power constraint on many test samples, or the post-processed sum rate falls below eWMMSE, the central claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that the joint channel and power allocation problem in a multi-channel wireless network can be solved by a graph neural network, JCPGNN-M, that is trained without labels by minimizing the Lagrangian of the sum-rate maximization. The network takes one subgraph per channel, with node features equal to the direct channel gain and edge features equal to the mutual interference terms, and outputs a power for every user on every channel. Because every user is allowed to transmit on all channels at once, the method realizes spectrum reuse that single-channel GNN schemes cannot. On the simulated scenarios the learned policy reports higher average sum rate than the enhanced WMMSE algorithm, around one percent of its running time, and retains about 98-99 percent of its own per-setup performance when tested on networks with more users or more channels.

Load-bearing premise

The load-bearing premise is that alternately updating the GNN weights and the power-constraint penalties converges to a good local optimum of the non-convex sum-rate problem, so that the learned allocations genuinely respect the per-user total-power limit.

Editorial extensions

If this is right

Operators could deploy JCPGNN-M for near-real-time resource allocation in dense networks, since inference is much faster than iterative eWMMSE.
The support for simultaneous multi-channel access means spectrum can be reused by many transceiver pairs, with the reported rate advantage over single-channel GNN baselines growing as the number of channels increases.
The unsupervised training recipe avoids generating expensive optimal labels, so the method can be retrained for new deployments using only channel measurements and the sum-rate objective.
The generalization results suggest a model trained on a small network (ten pairs, two channels) can be applied to roughly five times more channels and three times more users while keeping near its full-rate performance.

Reading between the lines

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

A direct extension of the paper's logic is that the same Lagrangian primal-dual training could handle other non-convex allocation problems with coupled per-agent constraints, such as energy harvesting or latency constraints, by swapping the constraint terms.
One point the paper leaves implicit is that the reported comparison is against eWMMSE's local optima; whether JCPGNN-M also approaches the global optimum of the non-convex problem is untested, so a small-scale exhaustive-search benchmark would clarify the ceiling.
The raw network output is only normalized when a user exceeds Pmax, which suggests feasibility is not guaranteed by the architecture itself; an architecture that projects onto the power simplex could remove that post-processing and strengthen the claim.
Because the generalization tables are normalized by JCPGNN-M's own performance at each size, they measure relative stability rather than absolute optimality; testing on out-of-distribution channel statistics (different path loss, shadowing, or user density) would be a stronger transfer test.

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

Short version: this is a real but incremental extension of GNN-based resource allocation to the multi-channel case where each user can use several channels at once. The architecture (M subgraphs, per-user Lagrange multipliers) is a natural next step from [9] and [10], and the authors don't oversell it as a paradigm shift. If the central performance claim holds up, it's a useful engineering result for dense networks. But the paper as submitted does not give me enough to verify that claim, and one methodological gap needs explicitly addressing.

What's new and good: the binary-to-continuous reformulation is exact for nonnegative powers, the WMMSE extension is standard and correctly imported from [1], and the unsupervised primal-dual training follows [14] and [15] without fitting test data. The generalization to larger networks is tested on fresh realizations, and the self-normalized tables are at least honest about what they show. The M-subgraph message passing with per-user power constraints is a clean way to share channel information while keeping the total power constraint. The citation pattern is appropriate, with no circular self-reliance.

Soft spots, in order of importance. First, the 'outperforms eWMMSE' claim rests on a baseline that is underspecified: no iteration count, stopping criterion, or lambda update rule for eWMMSE, no error bars, no parameter values, and no code or data. As it stands, Figure 4 could be comparing against a partially converged optimizer. That is not an accusation, just a missing verification. Second, the Lagrangian max-min training is justified by citing [14] and [15], but the multi-channel case with per-user total power constraints is not analyzed for convergence or constraint satisfaction. The post-processing normalization in Section IV.C.3 is ad hoc; the paper never reports how often it fires. If it fires often, the evaluated policy is not the trained policy, and the comparison could reflect the projection rather than the learned allocation. Both issues are empirically fixable and do not constitute a demonstrated contradiction. The generalization tables are self-normalized by the algorithm's own performance, so they prove consistency rather than absolute quality.

Bottom line: the paper is coherent, the extension is genuinely new in the cited context, and the math anchors to external results without circularity. It deserves a serious referee, but not acceptance in its current form. I would send it to review with a request for full simulation specs, error bars, code or data release, and a report of how often the normalization fires (or better, a projection that respects constraints by construction).

Referee Report

4 major / 7 minor

Summary. The paper studies joint channel and power allocation (JCPA) in multi-channel wireless networks where each transceiver pair may use all channels simultaneously. It first proposes eWMMSE, an extension of the classical WMMSE algorithm, and then introduces JCPGNN-M, a graph neural network trained unsupervised on a Lagrangian objective to output per-user per-channel powers. The binary channel assignment of the original problem is replaced by continuous nonnegative powers, the sum-power constraints are handled through Lagrange multipliers updated by a primal-dual scheme, and a post-processing step enforces feasibility. The central empirical claims are that JCPGNN-M achieves higher average sum rate than eWMMSE, uses about 1% of eWMMSE's running time, and generalizes to larger networks and more channels.

Significance. If the central claims are correct, JCPGNN-M would be a practically valuable result: a single learned model replaces an iterative optimizer for a non-convex multi-channel resource allocation problem, with near-real-time inference and transfer to larger network sizes. The paper has several genuine strengths: the reformulation from Problem (4) to Problem (5) is exact for nonnegative powers; the GNN is trained on the Lagrangian objective without labeled optimal solutions; the graph construction over independent per-channel subgraphs is natural for the problem; and the generalization experiments are a useful step beyond many GNN resource-allocation papers. However, the main performance claims rest on simulation details that are currently not reported, in particular the implementation of the eWMMSE baseline and the behavior of the feasibility post-processing step. These omissions make the headline claims plausible but not yet verified.

major comments (4)

[IV.C.3 / Fig. 4] The post-processing layer is not part of the training computation graph, and the paper never reports how often the normalization is actually applied on the 1000 test samples. If the normalization fires frequently, the transmitted power vector is not the vector whose Lagrangian loss was minimized, and the comparison in Figure 4 may reflect the projection rule rather than the learned allocation. The authors should report the fraction of users/samples for which the sum power exceeds Pmax before normalization, report the resulting constraint violations, and either include the projection in the forward model (with a differentiable approximation) or add a feasibility penalty to the training loss so that the trained policy and the evaluated policy coincide.
[V.B] The eWMMSE baseline is underspecified: no iteration count, stopping criterion, initialization, or lambda_i update procedure is given, and Equation (9) does not state how lambda_i is found (e.g., bisection, subgradient, or fixed-point). Without these details, the claim that JCPGNN-M 'achieves better data rate compared to eWMMSE' in all scenarios of Figure 4 is not reproducible, and the comparison could be against a partially converged optimizer. The authors should provide the full eWMMSE implementation details, report its convergence behavior (e.g., number of iterations used, final constraint violation), and, if possible, show that the reported eWMMSE points are at a consistent convergence criterion across all D and M values.
[IV.A] The paper asserts that solving the max-min Lagrangian problem (12) by the primal-dual updates (13) yields a locally optimal solution of the non-convex Problem (5), citing [15]. This is not established for the present setting: Problem (5) has no convexity, and for non-convex problems the dual value need not equal the primal value, so a saddle point of the Lagrangian is not automatically a local optimum of the original problem. The authors should either state and verify the precise conditions from [15] under which convergence applies to this multi-channel sum-rate problem with per-user total power constraints, or weaken the claim to an empirical one backed by constraint-violation and objective-value measurements. Without this, the optimality interpretation of the trained policy is unsupported.
[V.C] The generalization claim depends on a normalization whose denominator is not fully defined: the text says performance is 'normalized by the sum rate that the algorithm can achieve under the same set-up,' but it is not clear whether the denominator is the rate of JCPGNN-M retrained at the target size, the rate of a baseline, or the rate of a larger architecture. The reported values of 98-99% are only meaningful with absolute sum rates and with the same random channel realizations across the compared settings. Please specify the exact normalization, report absolute rates for at least one row/column, and add confidence intervals or standard deviations over the random network topologies.

minor comments (7)

[Abstract] The abstract contains a typo: 'the inference time of JCPGNN-M is much lower than eWMMS' should read 'eWMMSE'.
[Eqs. (6)-(7)] The MSE variable e_i^m is used in the objective but never explicitly defined, and the noise variance notation changes from sigma_i^2 in Section II to (sigma_i^m)^2 in Equation (7a). Please define e_i^m and standardize the noise variance notation.
[IV.C.1] The update rule for the hidden state x_i^{m(s)} is not given; Equation (14) uses x_i^{m(s-1)} and Equation (15) outputs \hat p_i^{m(s)}, but there is no equation showing how x_i^{m(s)} is computed from n_i^{m(s)}. Please specify the full message-passing recurrence.
[V.A] The simulation setup omits several hyperparameters needed for reproducibility: the learning rates phi_omega and phi_lambda, the number of training iterations/epochs, batch size, optimizer, activation functions, and how the Lagrange multipliers are initialized. Please add these values.
[Figure 4] Figure 4 shows no error bars or confidence intervals even though the channel realizations are random; without them, the visual claim that JCPGNN-M 'outperforms eWMMSE in all scenarios' cannot be assessed statistically.
[V.D] The running-time comparison in Figure 5 would be more informative with error bars and with a description of the hardware/software environment and the number of repeated runs; the y-axis label also should state the exact time unit.
[Introduction] The sentence 'WMMSE imposes a restrictive constraint... each transceiver pair can access only one subchannel at a time' may be misleading, because the classical WMMSE of [1] is formulated for a MIMO interfering broadcast channel with beamforming vectors, not specifically for a one-subchannel restriction. Please rephrase to avoid conflating the original WMMSE model with the single-channel constraint adopted by later GNN-based papers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GNN is trained unsupervised on a Lagrangian loss and tested on fresh channel realizations; the central claim is anchored to external results rather than reduced to its inputs by construction.

full rationale

The derivation chain runs from Problem (5) to the Lagrangian (10), then to the GNN policy G_p(omega_p; H), then to the max-min formulation (12) solved by the primal-dual updates (13). The training loss is the unsupervised Lagrangian (10), and the reported sum rates are computed on 1,000 held-out channel realizations (Section V.A), so the data-rate claim is not fitted to the test set. The eWMMSE baseline is an iterative algorithm run on the same channels rather than a fitted parameter, so the Figure 4 comparison is not forced by construction. The load-bearing external anchors are [1] (WMMSE equivalence), [13] (universal approximation of graph functions), and [15] (primal-dual convergence to a local optimum of the original problem). Reference [14] is co-authored by C. She, but it is a supplementary pointer to the same max-min/convergence argument and is not the sole support for the central claim; the self-cited JCPGNN [9] appears only as a baseline. The generalization tables compare transferred models against per-setup trained models, so the near-99% numbers are not tautological. The post-processing normalization in Section IV.C.3 is a standard feasibility projection and does not rename a fitted coefficient as a prediction. No equation in the paper reduces to its own inputs by construction; the concerns about non-convex convergence and eWMMSE specification are correctness risks, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or mediators. Its load-bearing structure is standard optimization machinery (WMMSE equivalence, GNN approximation, primal-dual updates) plus domain assumptions about perfect CSI, flat fading, and complete interference graphs. The trained GNN weights and unreported hyperparameters are the main data-fitted quantities. The most fragile assumption is that the unconstrained GNN parameterization, combined with the ad hoc post-processing projection, reliably tracks the per-user total power constraint.

free parameters (3)

GNN trainable weights omega_p (message and update MLPs) = not reported (trained on 10,000 unlabeled samples)
The performance of JCPGNN-M depends on these learned weights; no trained model, initialization, or random seed is released.
Lagrange multiplier learning rates phi_omega_p and phi_lambda_i = not stated
Chosen by hand for the primal-dual updates in Eq. (13); values are not reported, so convergence and final performance cannot be reproduced.
Power scaling and normalization rule in post-processing = Pmax (exact scale not stated)
The sigmoid output lies in (0,1), and the conversion to watts is unstated. Section IV.C.3 only says to normalize to Pmax when the constraint is violated.

assumptions (5)

standard math Theorem 1 of [1]: sum-rate maximization is equivalent to weighted MSE minimization.
Used in Section III.B to justify Eq. (6) through Eq. (8) and the eWMMSE update in Eq. (9).
standard math Universal invariant and equivariant GNN approximation theorem [13].
Used in Section IV.A to justify representing the power allocation policy by a GNN with uniform approximation capability.
standard math Primal-dual method converges to a locally optimal solution of Problem (5) [14], [15].
Invoked after Eq. (13) to support the claim that SGD/SGA updates on the Lagrangian solve the non-convex problem.
domain assumption Perfect CSI H is available as a block-diagonal matrix and each resource block exhibits flat fading.
System model in Section II.A; no channel estimation error or feedback overhead is modeled.
domain assumption Dropping binary c_i^m by identifying activity with p_i^m > 0 preserves Problem (4).
Section III.A reformulates (4) to (5); this is exact for p >= 0 but suppresses the integer nature of channel assignment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph Neural Networks for Resource Allocation in Multi-Channel Wireless Networks." pith.science (2026). https://pith.science/paper/ZN3D63PR

@misc{pith2026250603813,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Networks for Resource Allocation in Multi-Channel Wireless Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZN3D63PR}},
  note         = {Machine review of arXiv:2506.03813}
}

read the original abstract

As the number of mobile devices continues to grow, interference has become a major bottleneck in improving data rates in wireless networks. Efficient joint channel and power allocation (JCPA) is crucial for managing interference. In this paper, we first propose an enhanced WMMSE (eWMMSE) algorithm to solve the JCPA problem in multi-channel wireless networks. To reduce the computational complexity of iterative optimization, we further introduce JCPGNN-M, a graph neural network-based solution that enables simultaneous multi-channel allocation for each user. We reformulate the problem as a Lagrangian function, which allows us to enforce the total power constraints systematically. Our solution involves combining this Lagrangian framework with GNNs and iteratively updating the Lagrange multipliers and resource allocation scheme. Unlike existing GNN-based methods that limit each user to a single channel, JCPGNN-M supports efficient spectrum reuse and scales well in dense network scenarios. Simulation results show that JCPGNN-M achieves better data rate compared to eWMMSE. Meanwhile, the inference time of JCPGNN-M is much lower than eWMMS, and it can generalize well to larger networks.

Figures

Figures reproduced from arXiv: 2506.03813 by the authors.

**Figure 2.** Graph representation of joint resource allocation in multi-channel networks with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The structure of proposed JCPGNN-M algorithm. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Average sum rate comparison for different number of channels with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Average running time of different algorithms when [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 9 canonical work pages

[9]

GNN-based joint channel and power allo- cation in heterogeneous wireless networks,

L. Chen, J. Zhu, and J. Evans, “GNN-based joint channel and power allo- cation in heterogeneous wireless networks,” in2024 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2024, pp. 233–238

work page 2024
[10]

Graph neural networks approach for joint wireless power control and spectrum allocation,

M. Marwani and G. Kaddoum, “Graph neural networks approach for joint wireless power control and spectrum allocation,”IEEE Transactions on Machine Learning in Communications and Networking, 2024

2024
[1]

An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,

Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,”IEEE Transactions on Signal Processing, vol. 59, no. 9, pp. 4331–4340, 2011

2011
[14]

Unsupervised deep learning for opti- mizing wireless systems with instantaneous and statistic constraints,

C. Sun, C. She, and C. Yang, “Unsupervised deep learning for opti- mizing wireless systems with instantaneous and statistic constraints,” Ultra-Reliable and Low-Latency Communications (URLLC) Theory and Practice: Advances in 5G and Beyond, pp. 85–117, 2023

work page 2023
[15]

Learning optimal resource allocations in wireless systems,

M. Eisen, C. Zhang, L. F. Chamon, D. D. Lee, and A. Ribeiro, “Learning optimal resource allocations in wireless systems,”IEEE Transactions on Signal Processing, vol. 67, no. 10, pp. 2775–2790, 2019

2019
[2]

Underlay device-to-device communications on multiple channels,

M. Elnourani, M. Hamid, D. Romero, and B. Beferull-Lozano, “Underlay device-to-device communications on multiple channels,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 3684–3688

2018
[3]

A joint channel and power al- location scheme for device-to-device communications underlaying uplink cellular networks,

F. Hajiaghajani, R. Davoudi, and M. Rasti, “A joint channel and power al- location scheme for device-to-device communications underlaying uplink cellular networks,” in2016 IEEE Conference on Computer Communica- tions Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 768–773

work page 2016
[4]

Resource allocation for D2D communication with multiple D2D pairs reusing multiple channels,

P. Mach, Z. Becvar, and M. Najla, “Resource allocation for D2D communication with multiple D2D pairs reusing multiple channels,” IEEE Wireless Communications Letters, vol. 8, no. 4, pp. 1008–1011, 2019

work page 2019

Show all 16 references

[5]

Learning power allocation for cell-free massive mimo system with graph neural networks,

Y . Peng, T. Liu, and C. Yang, “Learning power allocation for cell-free massive mimo system with graph neural networks,” inGLOBECOM 2024-2024 IEEE Global Communications Conference. IEEE, 2024, pp. 2653–2658

2024
[6]

Graph neural networks for power allocation in wireless networks with full duplex nodes,

L. Chen, J. Zhu, and J. Evans, “Graph neural networks for power allocation in wireless networks with full duplex nodes,” in2023 IEEE In- ternational Conference on Communications Workshops (ICC Workshops), 2023, pp. 277–282

2023
[7]

Heterogeneous graph neural network for power allocation in multicarrier-division du- plex cell-free massive mimo systems,

B. Li, L.-L. Yang, R. G. Maunder, S. Sun, and P. Xiao, “Heterogeneous graph neural network for power allocation in multicarrier-division du- plex cell-free massive mimo systems,”IEEE Transactions on Wireless Communications, vol. 23, no. 2, pp. 962–977, 2024

2024
[8]

A GNN- based supervised learning framework for resource allocation in wireless iot networks,

T. Chen, X. Zhang, M. You, G. Zheng, and S. Lambotharan, “A GNN- based supervised learning framework for resource allocation in wireless iot networks,”IEEE Internet of Things Journal, vol. 9, no. 3, pp. 1712– 1724, 2021

2021
[11]

Deep reinforcement learning-based channel allocation for wireless LANs with graph convolutional networks,

K. Nakashima, S. Kamiya, K. Ohtsu, K. Yamamoto, T. Nishio, and M. Morikura, “Deep reinforcement learning-based channel allocation for wireless LANs with graph convolutional networks,”IEEE Access, vol. 8, pp. 31 823–31 834, 2020

2020
[12]

Learning to optimize: Training deep neural networks for interference management,

H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos, “Learning to optimize: Training deep neural networks for interference management,”IEEE Transactions on Signal Processing, vol. 66, no. 20, pp. 5438–5453, 2018

2018
[13]

Universal invariant and equivariant graph neural networks,

N. Keriven and G. Peyr ´e, “Universal invariant and equivariant graph neural networks,”Advances in Neural Information Processing Systems, vol. 32, 2019

2019
[16]

Graph neural networks for wireless communications: From theory to practice,

Y . Shen, J. Zhang, S. Song, and K. B. Letaief, “Graph neural networks for wireless communications: From theory to practice,”IEEE Transactions on Wireless Communications, vol. 22, no. 5, pp. 3554–3569, 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [9]

GNN-based joint channel and power allo- cation in heterogeneous wireless networks,

L. Chen, J. Zhu, and J. Evans, “GNN-based joint channel and power allo- cation in heterogeneous wireless networks,” in2024 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2024, pp. 233–238

work page 2024

[2] [10]

Graph neural networks approach for joint wireless power control and spectrum allocation,

M. Marwani and G. Kaddoum, “Graph neural networks approach for joint wireless power control and spectrum allocation,”IEEE Transactions on Machine Learning in Communications and Networking, 2024

2024

[3] [1]

An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,

Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,”IEEE Transactions on Signal Processing, vol. 59, no. 9, pp. 4331–4340, 2011

2011

[4] [14]

Unsupervised deep learning for opti- mizing wireless systems with instantaneous and statistic constraints,

C. Sun, C. She, and C. Yang, “Unsupervised deep learning for opti- mizing wireless systems with instantaneous and statistic constraints,” Ultra-Reliable and Low-Latency Communications (URLLC) Theory and Practice: Advances in 5G and Beyond, pp. 85–117, 2023

work page 2023

[5] [15]

Learning optimal resource allocations in wireless systems,

M. Eisen, C. Zhang, L. F. Chamon, D. D. Lee, and A. Ribeiro, “Learning optimal resource allocations in wireless systems,”IEEE Transactions on Signal Processing, vol. 67, no. 10, pp. 2775–2790, 2019

2019

[6] [2]

Underlay device-to-device communications on multiple channels,

M. Elnourani, M. Hamid, D. Romero, and B. Beferull-Lozano, “Underlay device-to-device communications on multiple channels,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 3684–3688

2018

[7] [3]

A joint channel and power al- location scheme for device-to-device communications underlaying uplink cellular networks,

F. Hajiaghajani, R. Davoudi, and M. Rasti, “A joint channel and power al- location scheme for device-to-device communications underlaying uplink cellular networks,” in2016 IEEE Conference on Computer Communica- tions Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 768–773

work page 2016

[8] [4]

Resource allocation for D2D communication with multiple D2D pairs reusing multiple channels,

P. Mach, Z. Becvar, and M. Najla, “Resource allocation for D2D communication with multiple D2D pairs reusing multiple channels,” IEEE Wireless Communications Letters, vol. 8, no. 4, pp. 1008–1011, 2019

work page 2019