Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Unsupervised Learning-Based Joint Resource Allocation and Beamforming Design for RIS-Assisted MISO-OFDMA Systems

T0 review · 4 major / 6 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 two-stage unsupervised network sets RIS phases, resource blocks, and beamforming at 99.93% of the SCA sum rate and 0.036% of its runtime.

desk verdict A useful, well-executed unsupervised-learning joint allocation paper whose headline near-optimality claim is anchored to a weaker quantized-SCA baseline than it appears. read the letter →

arxiv 2506.22448 v1 pith:H2WE5MXB submitted 2025-06-12 eess.SP cs.AIcs.ITmath.IT

Yu Ma , Xingyu Zhou , Xiao Li , Le Liang , Shi Jin This is my paper · ORCID
classification eess.SPcs.AIcs.ITmath.IT
keywords reconfigurableintelligentsurfaceMISO-OFDMAunsupervisedlearningresourceallocationbeamformingdesignGumbel-softmaxquality-of-servicesum-ratemaximization
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 aims to establish that a label-free, two-stage neural network can jointly solve the resource-allocation problem in an RIS-assisted MISO-OFDMA downlink—choosing the RIS reflection phases, the resource-block assignments, and the base-station beamforming—at a fraction of the computing cost of the standard successive convex approximation (SCA) method. A reconfigurable intelligent surface is a panel of programmable passive reflection elements; the system is a multi-antenna base station serving single-antenna users over OFDMA subcarriers and timeslots. The first network, BeamNet, predicts per-timeslot RIS phases from channel state information, and the second, AllocationNet, assigns resource blocks based on the equivalent channels those phases create; the active beamforming is then computed in closed form with maximum ratio transmission and water-filling. Trained without labels using a loss equal to negative sum rate plus a quality-of-service penalty, the discrete version reaches 99.93% of the SCA baseline sum rate at 0.036% of its runtime in the main configuration, and the continuous-phase version reaches 113.60% of that baseline. If the claim holds, RIS control in frequency-selective multiuser systems can run in a single forward pass, making real-time optimization practical at a small, quantified performance cost.

What carries the argument

The load-bearing mechanism is the two-stage network pair plus closed-form active beamforming. BeamNet's quantization layer approximates the non-differentiable 1-bit step function by $f(\phi)=\pi\,\mathrm{sigmoid}(\beta(\phi-\pi))$, so discrete RIS phases can receive gradients. AllocationNet uses the Gumbel-softmax reparameterization $\alpha_{n,k,q} = \exp((P(n,k,q)+g_{n,k,q})/\tau)/\sum_{k'} \exp((P(n,k',q)+g_{n,k',q})/\tau)$, which makes the discrete resource-block assignment differentiable and enforces the constraint that each subcarrier goes to at most one user. Active beamforming is then fixed by maximum ratio transmission with water-filling power allocation, taking the beamforming variables out of the trainable parameter set. A phased training schedule trains each sub-network separately before joint fine-tuning, which the paper says avoids the poor convergence of training the large two-network model from scratch.

What would settle it

Train or test the exact same networks on CSI corrupted by estimation noise (say, 10 dB channel-estimation SNR) or by one timeslot of feedback delay, and compare the resulting sum rate with SCA run on the same imperfect CSI; a large drop in the learned method's relative performance would show the near-SCA result depends on perfect CSI.

Watch

Extended reading notes

Core claim

The central claim is that iterative optimization can be replaced by one forward pass through two coupled networks without a labeled training set. BeamNet takes the direct and cascaded CSI and outputs 1-bit RIS phase shifts per timeslot through a differentiable approximation of the step function; AllocationNet then receives the resulting equivalent BS-user channels and outputs near-one-hot resource-block decisions via the Gumbel-softmax trick. With phases and allocations fixed, the base station uses maximum ratio transmission with water-filling power allocation, so beamforming is not learned at all. The networks are trained in phases—BeamNet alone, AllocationNet alone, then jointly—against a loss that is the negative system sum rate plus a QoS penalty. The paper reports that at 64 RIS elements the discrete algorithm reaches 99.93% of the discrete SCA sum rate in 18.21 ms versus 49,642.81 ms for the baseline, and that the continuous-phase variant reaches 113.60% of the discrete SCA baseline.

Load-bearing premise

The entire pipeline assumes perfect channel state information at the input: BeamNet and AllocationNet are trained and evaluated on exact direct and cascaded channels, with no estimation error, feedback delay, or CSI mismatch; if real CSI is imperfect, the reported near-SCA gains may not persist.

Editorial extensions

If this is right

  • One forward pass through the two networks produces RIS phases, RB assignments, and beamforming, cutting per-decision runtime from about 49.6 seconds to about 18 ms at 64 RIS elements.
  • No labeled optimal solutions are needed for training, so the costly step of generating SCA labels is avoided.
  • The learned dynamic allocation often assigns a timeslot to a single user, which concentrates the RIS passive beamforming gain and explains part of the performance.
  • The networks transfer to unseen delay-tap settings (90.55% / 94.7% of the SCA benchmarks), shifted user positions (94.33% / 97.44%), and higher Rician factors, but degrade when the Rician factor falls below 0 dB.
  • Retraining is required when the number of subcarriers or RIS elements changes, because the network input and output dimensions are fixed.

Reading between the lines

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

  • The paper does not test imperfect CSI; a natural next experiment is to inject channel estimation noise or feedback delay during training and see whether the same unsupervised loss learns a robust phase policy.
  • Since the continuous variant already beats the discrete SCA baseline by 13.6 percentage points, intermediate phase resolutions such as 2-bit quantizers might close the small remaining gap to the discrete baseline while keeping hardware simple.
  • The allocation patterns show the learned policy often dedicates a timeslot to a single user; reading AllocationNet's soft probabilities as a user-scheduling prior could scale the idea to systems with many more users without retraining the full network.
  • Because retraining is tied to fixed N and M, transfer to new system sizes would likely require meta-learning or an input-encoding scheme, which the paper lists as future work.
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

4 major / 6 minor

Summary. This paper studies downlink RIS-assisted MISO-OFDMA sum-rate maximization with 1-bit RIS phase constraints, RB allocation, and BS beamforming. It proposes a two-stage unsupervised learning framework: BeamNet outputs RIS phases through a differentiable quantization approximation, AllocationNet outputs RB allocations via Gumbel-softmax, and active beamforming is computed by MRT plus water-filling. A phased training procedure and a hinge-penalty loss for QoS are introduced. Simulations compare against continuous/discrete SCA baselines, random allocation, random RIS, and no-RIS, reporting 99.93% of the discrete SCA sum rate at M=64 with only 0.036% of its runtime.

Significance. The contribution is potentially useful: if the results hold, it provides a low-complexity, label-free learning solution to a nontrivial mixed-integer resource allocation problem in RIS-OFDMA. The paper's strengths include external SCA and random baselines, separate ablation studies for the penalty factor, learning rate, and phased training, and robustness checks across delay taps, user distributions, and Rician factors. The main caveats are that the discrete SCA baseline is only a quantized heuristic and the QoS constraint is handled by a soft penalty, so several headline statements are stronger than the evidence.

major comments (4)
  1. [Section IV-A, Table III] The 'Discrete SCA' baseline is defined as continuous SCA whose RIS phases are quantized a posteriori; it is a feasible heuristic, not a solution to the discrete-constrained problem in (7). Therefore the headline '99.93% of the SCA baseline' only establishes closeness to this heuristic, and the near-optimality claim for the actual discrete problem is unsupported. The continuous variant's 113.60% of the same baseline corresponds to roughly 96.3% of continuous SCA, a non-negligible gap. Please add a stronger discrete benchmark (e.g., local search over discrete phase profiles, or exact optimization for small M) or explicitly relabel the claim as matching quantized-continuous SCA.
  2. [Section III-D, Fig. 7] QoS is enforced by the hinge penalty L_QoS = lambda_1 * sum_k (R_QoS - R_k,0)_+ in (17), not by a hard constraint. The abstract and Section IV-C state that the method 'satisfies QoS constraints,' but the evidence is that the 5th percentile rate exceeds R_QoS for lambda_1 = 5; this does not guarantee that every user meets (7e). Please report the fraction of users satisfying the constraint, or the minimum per-user rate, and describe the soft-constraint nature in the claims.
  3. [Section IV-G, Table III] The runtime comparison is not hardware-matched: the SCA benchmarks run on an Intel i7-12700 CPU while the proposed network inference uses an RTX 3060 GPU. The reported 0.036% runtime ratio therefore mixes algorithmic speedup with hardware acceleration. Please provide CPU-only inference times, or FLOPs and iteration counts, and compare on the same platform.
  4. [Section III-B, Section IV-F] The networks are trained and tested with perfect CSI; the input to BeamNet is described as CSI in Section III-B, and no channel estimation error or feedback delay appears in Section IV. The robustness experiments change delay taps, user positions, and Rician factors, but they do not address CSI mismatch. Since practical deployment would rely on imperfect CSI, please add a mismatch analysis (e.g., Gaussian CSI error or outdated CSI) or explicitly state this limitation.
minor comments (6)
  1. [Table II] The description of lambda_1 reads 'Softmax temperature for gumbel softmax,' but lambda_1 is the QoS penalty factor defined in Section III-D; the entries for lambda_1 and tau appear to be swapped.
  2. [Fig. 3 caption and text] The text says the SE-Res block is shown in Fig. 3(e) and the layer symbols in Fig. 3(d), while the captions label (d) as the SE-Res block and (e) as the symbols; please reconcile this discrepancy.
  3. [Eqs. (14)-(15)] Equation (14) uses g_n but the Gumbel noise is later defined in (15) as g_n = -ln(-ln(u)); the notation should use the same indices as P(n,k,q) to avoid confusion.
  4. [Section IV-A] The 'Without RIS' baseline with M=0 is not accompanied by a description of how the fixed-dimensional networks are adapted or retrained for M=0; please clarify.
  5. [Section III-B, Eq. (12)] The quantization function in (12) is described as producing discrete 0/pi phases, but it is a smooth approximation; for beta=100 the outputs are close to but not exactly {0, pi}. A sentence acknowledging this residual approximation would help.
  6. [Section IV-G] The complexity expressions for BeamNet and AllocationNet are given without derivation; please add a short derivation or a reference for the counting argument.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the unsupervised training objective equals the evaluation objective by design, the SCA benchmarks are external, and the few self-citations are not load-bearing.

full rationale

The derivation chain is self-contained. BeamNet and AllocationNet are trained by minimizing the negative sum-rate loss in Eqs. (16)-(18), which is the same metric used for evaluation; that is the intended mechanism of unsupervised learning, not a hidden circular dependency, since no label or fitted target is reused as a prediction. Active beamforming via MRT and water-filling (Eqs. (8)-(11)) and the equivalent channel construction (Eq. (13)) are explicit physical formulas independent of the learned parameters. The benchmarks are external: continuous SCA is the algorithm of [35] (Yang, Zhang, and Zhang), and discrete SCA is its phase-quantized variant defined in Section IV-A. The 99.93% and 0.036% figures therefore compare against an external baseline, even though the baseline is only a quantized heuristic and the runtime comparison mixes CPU (i7-12700) and GPU (RTX 3060); those are benchmarking-quality concerns, not circularity. The only author-overlapping citations, [36] for the delay-tap formula L = max(L0, L1+L2-1) and [40] for multi-agent RL background, are not load-bearing: the delay formula is a standalone parameter-free identity, and the RL remark is motivational. No uniqueness theorem, ansatz, or fitted parameter is smuggled in via self-citation. The stated retraining limitation and the low-Rician robustness gap are acknowledged limitations that do not make the derivation circular.

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

The central numerical claims rest on hyperparameters selected by simulation sweeps and on standard domain assumptions such as perfect CSI and 1-bit quantized phases. The two neural networks are computational components, not new physical entities, so no invented entities are introduced. The strongest unstated load-bearing assumption is perfect CSI availability.

free parameters (5)
  • QoS penalty factor lambda_1 = 5
    Chosen by sweeping values from 0 to 10 in Fig. 7 to balance 5th percentile rate against sum rate; not derived from first principles.
  • Gumbel softmax temperature tau = 0.5
    Set by hand as a trade-off between near-one-hot outputs and gradient stability; no optimality criterion given.
  • Quantization steepness beta = 100
    Hyperparameter in Eq. (12) controlling how closely the sigmoid approximates the step function; chosen manually.
  • Weight decay lambda_2 = 5e-5
    Regularization term in the loss (17); small value chosen by convention, not by an explicit selection procedure.
  • Learning rates mu_1, mu_2, mu_3 = 0.001, 0.001, 0.0005
    Selected based on loss comparison in Fig. 8(a); influence convergence but are standard tuning choices.
assumptions (5)
  • domain assumption Perfect channel state information is available to the networks at training and inference
    The input to BeamNet and AllocationNet is described as CSI in Section III-B/C, with no channel estimation error model; performance under imperfect CSI is not tested.
  • domain assumption Quasi-static block fading with cyclic prefix longer than the maximum delay spread
    Section II assumes N_CP >= L to remove inter-symbol interference, a standard but restrictive condition for delay spread.
  • domain assumption RIS reflection phases are 1-bit quantized (0 or pi) and identical over all subcarriers in a timeslot
    Constraint (7d) and the system model set 1-bit phase resolution, which shapes the quantization layer design and the reported gains.
  • standard math Each subcarrier is allocated to at most one user, eliminating multi-user interference
    Constraint (7b) yields orthogonal access, so the SNR expression in Eq. (5) contains no interference terms.
  • domain assumption The SCA algorithm from [35] is a valid near-optimal baseline
    The headline 99.93% and 0.036% numbers are measured relative to this baseline; any inaccuracy in the SCA implementation propagates into the claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Learning-Based Joint Resource Allocation and Beamforming Design for RIS-Assisted MISO-OFDMA Systems." pith.science (2026). https://pith.science/paper/H2WE5MXB

@misc{pith2026250622448,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Learning-Based Joint Resource Allocation and Beamforming Design for RIS-Assisted MISO-OFDMA Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2WE5MXB}},
  note         = {Machine review of arXiv:2506.22448}
}
read the original abstract

Reconfigurable intelligent surfaces (RIS) are key enablers for 6G wireless systems. This paper studies downlink transmission in an RIS-assisted MISO-OFDMA system, addressing resource allocation challenges. A two-stage unsupervised learning-based framework is proposed to jointly design RIS phase shifts, BS beamforming, and resource block (RB) allocation. The framework includes BeamNet, which predicts RIS phase shifts from CSI, and AllocationNet, which allocates RBs using equivalent CSI derived from BeamNet outputs. Active beamforming is implemented via maximum ratio transmission and water-filling. To handle discrete constraints while ensuring differentiability, quantization and the Gumbel-softmax trick are adopted. A customized loss and phased training enhance performance under QoS constraints. Simulations show the method achieves 99.93% of the sum rate of the SCA baseline with only 0.036% of its runtime, and it remains robust across varying channel and user conditions.

Figures

Figures reproduced from arXiv: 2506.22448 by the authors.

Figure 1
Figure 1. RIS-assisted downlink communication system. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of fixed and dynamic allocation schemes: different fill [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Network architecture. B. BeamNet Design To predict the RIS reflection phase shift, BeamNet is proposed, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The step function and the approximating function. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Sum rate versus transmit power. • Proposed discrete algorithm: The proposed Algo￾rithm 1. • Random allocation: The RB allocation decisions are ran￾domly set, while active beamforming and RIS reflection phase shift are optimized using the proposed algorithm. • Random RI…
Figure 7
Figure 7. Figure 7: The impact of different penalty factors on the proposed algorithm. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: The impact of different learning rates and training methods on the proposed algorithm. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: The impact of the number of reflection elements on the system [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Comparison of various RB allocation schemes: the RB colored in [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: The sum rate versus transmit power evaluated with [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Sum rate versus transmit power, tested at [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 42 canonical work pages

  1. [1]

    6G wireless networks: Vision, requirements, architec- ture, and key technologies,

    Z. Zhanget al., “6G wireless networks: Vision, requirements, architec- ture, and key technologies,”IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 28–41, 2nd Quart. 2019

    work page 2019

  2. [2]

    Terahertz-band joint ultra-massive MIMO radar-communications: Model-based and model- free hybrid beamforming,

    A. M. Elbir, K. V . Mishra, and S. Chatzinotas, “Terahertz-band joint ultra-massive MIMO radar-communications: Model-based and model- free hybrid beamforming,”IEEE J. Sel. Top. Signal Process, vol. 15, no. 6, pp. 1468–1483, Nov. 2021

    work page 2021

  3. [3]

    Terahertz band communication: An old problem revisited and research directions for the next decade,

    I. F. Akyildiz, C. Han, Z. Hu, S. Nie, and J. M. Jornet, “Terahertz band communication: An old problem revisited and research directions for the next decade,”IEEE Trans. Commun., vol. 70, no. 6, pp. 4250–4285, Jun. 2022

    work page 2022

  4. [4]

    Integrated sensing and communication for RIS-assisted backscatter systems,

    X. Wang, Z. Fei, and Q. Wu, “Integrated sensing and communication for RIS-assisted backscatter systems,”IEEE Internet Things J., vol. 10, no. 15, pp. 13 716–13 726, Aug. 2023

    work page 2023

  5. [5]

    RIS- assisted integrated sensing and backscatter communications for future IoT networks,

    N. Wu, X. Wang, Z. Fei, F. Xia, J. Huang, and A. Nallanathan, “RIS- assisted integrated sensing and backscatter communications for future IoT networks,”IEEE Internet of Things Mag., vol. 7, no. 4, pp. 44–50, Jul. 2024

    work page 2024

  6. [6]

    Stacked Intelligent Metasurfaces for Wireless Communications: Applications and Challenges

    H. Liuet al., “Stacked intelligent metasurfaces for wireless sens- ing and communication: Applications and challenges,”arXiv preprint arXiv:2407.03566, 2024

    work page Pith review arXiv 2024

  7. [7]

    Multi-user MISO with stacked intelligent metasurfaces: A DRL-based sum-rate optimization approach,

    H. Liu, J. An, G. C. Alexandropoulos, D. W. K. Ng, C. Yuen, and L. Gan, “Multi-user MISO with stacked intelligent metasurfaces: A DRL-based sum-rate optimization approach,”IEEE Trans. Cognit. Commun. Net- working, pp. 1–1, 2025

    work page 2025

  8. [8]

    Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,

    M. Di Renzoet al., “Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, Nov. 2020

    work page 2020

Show all 43 references
  1. [9]

    Reconfigurable intelligent surfaces for 6G systems: Principles, applications, and research directions,

    C. Panet al., “Reconfigurable intelligent surfaces for 6G systems: Principles, applications, and research directions,”IEEE Commun. Mag., vol. 59, no. 6, pp. 14–20, Jun. 2021. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2021

  2. [10]

    Coverage enhancement by deploying ris in 5G commer- cial mobile networks: Field trials,

    J. Sanget al., “Coverage enhancement by deploying ris in 5G commer- cial mobile networks: Field trials,”IEEE Wireless Commun., vol. 31, no. 1, pp. 172–180, Feb. 2024

    2024

  3. [11]

    Reconfigurable intelligent surfaces: Principles and op- portunities,

    Y . Liuet al., “Reconfigurable intelligent surfaces: Principles and op- portunities,”IEEE Commun. Surv., vol. 23, no. 3, pp. 1546–1577, 3rd Quart. 2021

    2021

  4. [13]

    Intelligent reflecting surface enhanced wireless network via joint active and passive beamforming,

    ——, “Intelligent reflecting surface enhanced wireless network via joint active and passive beamforming,”IEEE Trans. Wireless Commun., vol. 18, no. 11, pp. 5394–5409, Nov. 2019

    2019

  5. [14]

    Weighted sum- rate maximization for intelligent reflecting surface enhanced wireless networks,

    H. Guo, Y .-C. Liang, J. Chen, and E. G. Larsson, “Weighted sum- rate maximization for intelligent reflecting surface enhanced wireless networks,” inProc. IEEE GLOBECOM, Dec. 2019, pp. 1–6

    2019

  6. [15]

    Intelligent reflecting surface enhanced wireless network: Joint active and passive beamforming design,

    Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless network: Joint active and passive beamforming design,” inProc. IEEE GLOBECOM, Dec. 2018, pp. 1–6

    2018

  7. [17]

    Robust beamforming for RIS-assisted wireless communications with discrete phase shifts,

    H. Gao, K. Cui, C. Huang, and C. Yuen, “Robust beamforming for RIS-assisted wireless communications with discrete phase shifts,”IEEE Wireless Commun. Lett., vol. 10, no. 12, pp. 2619–2623, Dec. 2021

    2021

  8. [18]

    Quantized phase alignment by discrete phase shifts for reconfigurable intelligent surface-assisted communication systems,

    J. Sanget al., “Quantized phase alignment by discrete phase shifts for reconfigurable intelligent surface-assisted communication systems,” IEEE Trans. Veh. Technol., vol. 73, no. 4, pp. 5259–5275, Apr. 2024

    2024

  9. [19]

    Rate optimization and power allocation in RIS- assisted multi-user OFDM communication,

    S. Hassouna, M. A. Jamshed, M. Ur-Rehman, K. Arshad, M. A. Imran, and Q. H. Abbasi, “Rate optimization and power allocation in RIS- assisted multi-user OFDM communication,” inProc. IEEE WCNC, 2024, pp. 01–05

    2024

  10. [20]

    Joint beamforming optimization for reconfigurable intelligent surface-enabled MISO-OFDM systems,

    K. Feng, X. Li, Y . Han, and Y . Chen, “Joint beamforming optimization for reconfigurable intelligent surface-enabled MISO-OFDM systems,” China Commun., vol. 18, no. 3, pp. 63–79, Mar 2021

    2021

  11. [21]

    Harmony search-based optimization for multi-RISs MU-MISO OFDMA systems,

    J. Lee, J. Choi, and J. Kang, “Harmony search-based optimization for multi-RISs MU-MISO OFDMA systems,”IEEE Wireless Commun. Lett., vol. 12, no. 2, pp. 257–261, Feb. 2023

    2023

  12. [22]

    Sum-rate maximization for IRS-assisted UA V OFDMA communication systems,

    Z. Wei, Y . Cai, Z. Sun, D. W. K. Ng, J. Yuan, M. Zhou, and L. Sun, “Sum-rate maximization for IRS-assisted UA V OFDMA communication systems,”IEEE Trans. Wireless Commun., vol. 20, no. 4, pp. 2530–2550, Apr. 2021

    2021

  13. [23]

    AI empowered wireless communications: From bits to semantics,

    Z. Qin, L. Liang, Z. Wang, S. Jin, X. Tao, W. Tong, and G. Y . Li, “AI empowered wireless communications: From bits to semantics,”Proc. IEEE, vol. 112, no. 7, pp. 621–652, 2024

    2024

  14. [24]

    RIS-assisted mmwave channel estimation using convolutional neural networks,

    E. Shtaiwi, H. Zhang, A. Abdelhadi, and Z. Han, “RIS-assisted mmwave channel estimation using convolutional neural networks,” inProc. IEEE WCNCW, 2021, pp. 1–6

    2021

  15. [25]

    Deep learning based multi-user power allocation and hybrid precoding in massive MIMO systems,

    A. Koc, M. Wang, and T. Le-Ngoc, “Deep learning based multi-user power allocation and hybrid precoding in massive MIMO systems,” in Proc. IEEE ICC, 2022, pp. 5487–5492

    2022

  16. [26]

    Deep channel learning for large intelligent surfaces aided mm-Wave massive MIMO systems,

    A. M. Elbir, A. Papazafeiropoulos, P. Kourtessis, and S. Chatzinotas, “Deep channel learning for large intelligent surfaces aided mm-Wave massive MIMO systems,”IEEE Wireless Commun. Lett., vol. 9, no. 9, pp. 1447–1451, Sep. 2020

    2020

  17. [27]

    Enabling large intelligent surfaces with compressive sensing and deep learning,

    A. Taha, M. Alrabeiah, and A. Alkhateeb, “Enabling large intelligent surfaces with compressive sensing and deep learning,”IEEE Access, vol. 9, pp. 44 304–44 321, Mar. 2021

    2021

  18. [28]

    Deep learning for physical-layer 5G wireless tech- niques: Opportunities, challenges and solutions,

    H. Huanget al., “Deep learning for physical-layer 5G wireless tech- niques: Opportunities, challenges and solutions,”IEEE Wireless Com- mun., vol. 27, no. 1, pp. 214–222, Feb. 2020

    2020

  19. [29]

    Deep unsupervised learning for joint antenna selection and hybrid beamforming,

    Z. Liu, Y . Yang, F. Gao, T. Zhou, and H. Ma, “Deep unsupervised learning for joint antenna selection and hybrid beamforming,”IEEE Trans. Commun., vol. 70, no. 3, pp. 1697–1710, Mar. 2022

    2022

  20. [30]

    Semi-supervised learning via cross- prediction-powered inference for wireless systems,

    H. Sifaou and O. Simeone, “Semi-supervised learning via cross- prediction-powered inference for wireless systems,”IEEE trans. mach. learn. commun. netw., vol. 3, pp. 30–44, Nov. 2024

    2024

  21. [31]

    Channel quality prediction for TSCH blacklisting in highly dynamic networks: A self-supervised deep learning approach,

    M. Farahmandand and M. Nabi, “Channel quality prediction for TSCH blacklisting in highly dynamic networks: A self-supervised deep learning approach,”IEEE Sens. J., vol. 21, no. 18, pp. 21 059–21 068, Sep. 2021

    2021

  22. [32]

    Unsupervised learning for passive beamforming,

    J. Gao, C. Zhong, X. Chen, H. Lin, and Z. Zhang, “Unsupervised learning for passive beamforming,”IEEE Commun. Lett., vol. 24, no. 5, pp. 1052–1056, May 2020

    2020

  23. [33]

    Unsupervised learning- based joint active and passive beamforming design for reconfigurable intelligent surfaces aided wireless networks,

    H. Song, M. Zhang, J. Gao, and C. Zhong, “Unsupervised learning- based joint active and passive beamforming design for reconfigurable intelligent surfaces aided wireless networks,”IEEE Commun. Lett., vol. 25, no. 3, pp. 892–896, Mar. 2021

    2021

  24. [34]

    Unsupervised learning for joint beamforming design in RIS-aided ISAC systems,

    J. Ye, L. Huang, Z. Chen, P. Zhang, and M. Rihan, “Unsupervised learning for joint beamforming design in RIS-aided ISAC systems,” IEEE Wireless Commun. Lett., vol. 13, no. 8, pp. 2100–2104, 2024

    2024

  25. [35]

    IRS-enhanced OFDMA: Joint re- source allocation and passive beamforming optimization,

    Y . Yang, S. Zhang, and R. Zhang, “IRS-enhanced OFDMA: Joint re- source allocation and passive beamforming optimization,”IEEE Wireless Commun. Lett., vol. 9, no. 6, pp. 760–764, Jun. 2020

    2020

  26. [36]

    Deep reinforcement learning based power minimization for RIS-assisted MISO-OFDM systems,

    P. Chen, W. Huang, X. Li, and S. Jin, “Deep reinforcement learning based power minimization for RIS-assisted MISO-OFDM systems,” China Commun., vol. 20, no. 4, pp. 259–269, Apr. 2023

    2023

  27. [37]

    IRS-enhanced OFDM: Power allocation and passive array optimization,

    Y . Yang, S. Zhang, and R. Zhang, “IRS-enhanced OFDM: Power allocation and passive array optimization,” inProc. IEEE GLOBECOM, 2019, pp. 1–6

    2019

  28. [38]

    Intelligent reflecting surface meets OFDM: Protocol design and rate maximization,

    Y . Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting surface meets OFDM: Protocol design and rate maximization,”IEEE Trans. Commun., vol. 68, no. 7, pp. 4522–4535, Jul. 2020

    2020

  29. [39]

    SOQ: Structural reinforcement learning for constrained delay minimization with channel state information,

    Y . Zhao, Y . Kim, and J. Lee, “SOQ: Structural reinforcement learning for constrained delay minimization with channel state information,”IEEE Internet of Things Journal, vol. 11, no. 3, pp. 4628–4644, Feb. 2024

    2024

  30. [40]

    Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,

    L. Liang, H. Ye, and G. Y . Li, “Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,”IEEE J. Sel. Areas Com- mun., vol. 37, no. 10, pp. 2282–2292, Oct. 2019

    2019

  31. [41]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,”arXiv preprint arXiv:1611.01144, 2016

    2016 arXiv

  32. [42]

    Towards optimal power control via ensembling deep neural networks,

    F. Liang, C. Shen, W. Yu, and F. Wu, “Towards optimal power control via ensembling deep neural networks,”IEEE Trans. Commun., vol. 68, no. 3, pp. 1760–1776, Mar. 2020

    2020

  33. [43]

    Wireless communications through reconfigurable intelligent surfaces,

    E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and R. Zhang, “Wireless communications through reconfigurable intelligent surfaces,”IEEE Access, vol. 7, pp. 116 753–116 773, Aug. 2019

    2019

  34. [44]

    Hybrid beamforming for reconfigurable intelligent surface based multi-user communications: Achievable rates with limited discrete phase shifts,

    B. Di, H. Zhang, L. Song, Y . Li, Z. Han, and H. V . Poor, “Hybrid beamforming for reconfigurable intelligent surface based multi-user communications: Achievable rates with limited discrete phase shifts,” IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1809–1822, Aug. 2020

    2020

  35. [45]

    Learning resilient radio resource management policies with graph neural networks,

    N. NaderiAlizadeh, M. Eisen, and A. Ribeiro, “Learning resilient radio resource management policies with graph neural networks,”IEEE Trans. Signal Process., vol. 71, pp. 995–1009, 2023

    2023

Pith tools

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