Pith. sign in

REVIEW 3 major objections 6 minor 32 references

A soft actor-critic agent can jointly allocate subchannels, power, and RIS phases in shared-spectrum local networks, reaching about 96% of an exhaustive-search benchmark's utility at far lower computational cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

SAC-based deep reinforcement learning jointly allocates subchannels, power, and RIS phases in a multi-operator spectrum-sharing setting, outperforming DDPG and approaching an exhaustive-search benchmark in simulation.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Useful engineering framework with a genuinely new problem formulation, but the headline 'near-optimal' claim rests on a benchmark that never optimizes RIS phases, so the 96% number is not evidence of near-optimality. the 3 major comments →

arxiv 2603.25332 v2 pith:PAF5FKRM submitted 2026-03-26 eess.SY cs.SY

DRL-Based Spectrum Sharing for RIS-Aided Local High-Quality Wireless Networks

Hamid Reza Hashempour , Mina Khadem , Eduard A. Jorswieck , Hien Quoc Ngo This is my paper
classification eess.SY cs.SY
keywords spectrum sharingreconfigurable intelligent surfacedeep reinforcement learningsoft actor-criticlicensed shared accessresource allocationutility maximizationlocal high-quality wireless networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 establish that a deep reinforcement-learning approach—specifically soft actor-critic (SAC)—can solve a joint spectrum, power, and reflective-surface allocation problem in near-real time for networks where multiple vertical service providers share an operator's spectrum. The problem, a mixed-integer non-linear program, is reformulated as a Markov decision process with a hybrid continuous-discrete action space, and the learned policy is compared against an exhaustive discrete search with successive convex approximation. The central quantitative claim is that SAC reaches about 96 percent of the benchmark's utility while reducing online decision cost from exponential enumeration to a forward pass of a neural network. If true, this would make dynamic licensed shared access with reconfigurable intelligent surfaces deployable in practice, and would show that such surfaces, often seen as interference sources, can be turned into utility-enhancing assets.

Core claim

The paper's central claim is that the mixed-integer utility-maximization problem for RIS-aided multi-VSP spectrum sharing—maximizing VSP revenue minus spectrum, RIS, and power costs under QoS constraints—can be solved nearly optimally by a SAC-based agent. The agent observes full CSI, achieved rates, and the previous action; its actor outputs a continuous vector that is projected onto the feasible action set by thresholding and clipping, covering subchannel assignment, BS association, transmit powers, and RIS phases. In simulations with two VSPs, up to twelve users, and a sixteen-element RIS, SAC attains approximately 96 percent of the EDS-SCA benchmark reward and outperforms DDPG in converg

What carries the argument

The carrying mechanism is the MDP formulation plus the SAC learning rule. The state stacks instantaneous CSI for direct, base-station-to-RIS, and RIS-to-user links, achieved user rates, and the previous feasible action to preserve the Markov property. The action space is hybrid; the actor outputs continuous relaxations that a deterministic projection maps into feasible binary subchannel and association decisions and bounded power and phase values. SAC's entropy-regularized objective with twin critics and an automatically tuned temperature provides stable exploration over the combinatorial decisions, while the EDS-SCA benchmark—exhaustive enumeration of discrete variables followed by successi

Load-bearing premise

The load-bearing premise is that the agent can observe perfect instantaneous channel state information for every direct, base-station-to-RIS, and RIS-to-user link at each time step, so the learned policy's near-optimal performance has not been shown to survive channel-estimation error, delayed feedback, or mobility.

What would settle it

Train the identical SAC agent on the same simulated network, but replace the perfect CSI in the state with noisy estimates (for example, independent Gaussian errors calibrated to a realistic pilot-based estimator) or delay the state by one time slot. If the attained utility falls substantially below the claimed 96 percent of the EDS benchmark—say, below 80 percent—the central deployability claim is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • eLSA-style dynamic spectrum sharing among vertical service providers could run at runtime, since online decisions reduce to neural-network inference rather than combinatorial search.
  • RIS leasing becomes economically attractive: with phase configuration optimized by the same policy, VSP utility improves enough to offset the leasing cost.
  • Spectrum partitioning policy is a first-order lever: dedicated subchannels can raise utility several-fold over heavily shared ones, giving operators guidance on reuse factors.
  • The MDP-and-projection recipe is a template for larger, multi-antenna or multi-RIS deployments where exhaustive benchmarks are infeasible.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The EDS benchmark fixes RIS phases during enumeration, so the 96 percent figure likely understates the joint phase-optimization benefit; a benchmark that also searches phases might shift the comparison.
  • The perfect-CSI state assumption is the main gap between simulation and deployment; training the same agent under channel-estimation noise and measuring how the 96 percent degrades would test deployability.
  • Because RIS-BS association is fixed by proximity rather than optimized, the reported gains may not capture the full value of RIS placement; making association a decision variable could close part of that gap.
  • The same MDP-and-projection approach could extend to uplink/downlink joint scheduling or multi-cell coordination, which share the hybrid action-space structure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a deep reinforcement learning (DRL) framework for joint subchannel assignment, transmit power control, and RIS phase configuration in an eLSA-based multi-VSP spectrum-sharing system. The resource allocation problem is formulated as a mixed-integer nonlinear program (MINLP) with QoS constraints and economic costs, then modeled as a Markov decision process. Two DRL algorithms, DDPG and SAC, are tailored to the hybrid discrete-continuous action space via a feasibility projection. Numerical results compare the DRL policies against an exhaustive discrete search (EDS) with SCA-based power refinement, reporting that SAC reaches roughly 96% of the benchmark reward for M1=16 and outperforms DDPG in stability and convergence. The paper concludes that the SAC-based solution is near-optimal and suitable for real-time deployment.

Significance. If the central near-optimality claim were supported, the paper would make a useful contribution by addressing a timely problem: scalable DRL-based resource management for RIS-assisted spectrum sharing in local high-quality networks. The paper has several strengths: a detailed system and MDP formulation, explicit treatment of the hybrid discrete-continuous action space through projection, a constrained EDS-SCA benchmark, and ablation studies on learning-rate and mini-batch sensitivity. However, the main quantitative claim, that SAC attains approximately 96% of the near-optimal benchmark, is currently not supported because the benchmark never optimizes the RIS phase variables that are part of the original problem and of the DRL action space. This is a load-bearing issue that affects the abstract, the numerical results, and the conclusion. The paper also leaves open questions about QoS feasibility of the learned policies and about robustness to imperfect CSI, both of which are relevant to the stated deployability claim.

major comments (3)
  1. [Section III-D and Section IV-C] The EDS benchmark does not optimize RIS phases. Section III-D states that during the exhaustive enumeration of (ω, φ) the transmit power is fixed to a uniform allocation and the RIS phase shifts are kept fixed; the SCA refinement in (56) optimizes only p. However, the original problem (20d) includes θ as a decision variable, and the DRL action (29) and the complexity expressions (58)-(59) explicitly include θ. Therefore the reported '96% of the benchmark reward' is a ratio against a θ-fixed baseline, not an upper bound on problem (20), and the 'near-optimal' wording in the abstract, Section IV-C, and the conclusion is unsupported. Please either augment the benchmark to optimize θ (e.g., alternating SCA over p and θ) or explicitly reframe all near-optimal claims as 'within 96% of EDS with fixed RIS phases'.
  2. [Section II-C, Eq. (20b) and Section III-A.4, Eq. (31)] The QoS constraints are hard constraints in the optimization problem (20b), but the DRL reward enforces them only through a penalty term λ_qos max(0, R^th - R). The EDS-SCA benchmark, by contrast, enforces the QoS constraints directly in (56b). The paper does not report QoS violation rates for the SAC or DDPG policies. If the learned policies violate the constraints, the penalized reward is not the same objective as (20a), so the comparison against the constrained benchmark is not apples-to-apples. Please report the fraction of users meeting R^th and, if necessary, restrict the comparison to feasible policies or use a constrained projection for the QoS requirements.
  3. [Section IV-B and Eq. (26)-(27)] The state space assumes perfect instantaneous CSI for all direct, BS-RIS, and RIS-user links at every time slot, and the simulator provides this information exactly. The paper's concluding claim is that the framework enables low-complexity real-time decision-making, but no experiment addresses CSI estimation error, delayed feedback, or mobility, even though cascaded RIS channel acquisition is known to be expensive and noisy. As it stands, the results are an idealized upper-bound study rather than evidence of deployability. Please either add a robustness experiment under imperfect CSI or explicitly state that the contribution is limited to the perfect-CSI idealized setting and soften the deployment claims accordingly.
minor comments (6)
  1. [Abstract] Typographical spacing issue: 'reaching up to96%' should read 'reaching up to 96%'.
  2. [Section III-D paragraph before Eq. (49)] The notation G_b,k^c depends on ω, φ, and θ but the dependence on the fixed θ is not written explicitly. This makes the SCA derivation harder to follow; please write G_b,k^c(θ) or state clearly that θ is fixed throughout the benchmark.
  3. [Section III-E, Eq. (58)] The action-dimension expression contains two identical terms V B_v K_v C, presumably representing scheduling and power, plus J M_j for RIS phases. However, the action vector in (29) also includes BS association φ and the text mentions RIS association d, which are not reflected in (58). Please clarify what is included in the action dimension and why φ is omitted.
  4. [Section II-B, Eq. (8)] The notation M_j is used both as the cardinality of the reflecting elements and as an index set through '∀m ∈ M_j'. Please use a distinct symbol, e.g., {1,...,M_j}, for the index set.
  5. [Section IV-D, Fig. 5] The text says the reusable subchannel case uses C_r=2, but the legend includes configurations such as C_d=0, C_r=3. The description and figure are inconsistent; please align the legend entries with the text.
  6. [Section II-C, Eq. (20c)] Constraint (20c) says 'ω, φ, ∀v,b,k,c,j', but j is not present in ω or φ. The 'j' is likely a typo; the RIS-association variable d, if used, should be explicitly included in the set of optimization variables.

Circularity Check

0 steps flagged

No circular reduction found; the 96% near-optimal claim is weakened by a benchmark that fixes RIS phases, but that is a benchmarking limitation, not circularity.

full rationale

The paper's chain is: formulate MINLP (20), recast as MDP with state (26)-(27), action (29), and reward (31), train SAC/DDPG, compare against the EDS+SCA benchmark (Section III-D), and report that SAC attains about 96% of the EDS reward for M1=16. I find no step in which an output is defined in terms of an input, no fitted parameter is relabeled as a prediction, and no load-bearing argument rests on a self-citation. The RL agent maximizes the same reward (31) on which it is later evaluated; this is standard RL practice and does not reduce the comparison to construction, because the EDS benchmark and the learned policies are computed independently on the same utility function and channel realizations. The closest concern is that the EDS benchmark keeps RIS phase shifts fixed during enumeration and its SCA refinement (56) optimizes only p, whereas the original problem (20d) treats the RIS phases as decision variables and the SAC action (29) includes them. Therefore the reported 96% does not rigorously establish near-optimality for the full problem (20); it establishes closeness to a RIS-phase-fixed reference. That is a benchmark-validity or overclaiming issue, not a circularity. Self-citations such as [8], [13]-[16], and [24] appear only as background and are not load-bearing, and no uniqueness theorem or ansatz is imported from the authors' prior work.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claims rest on a large set of hand-picked economic and channel parameters, plus perfect-CSI availability. The most consequential loose end is that the revenue coefficients beta_v and the scaling factors Phi_1, Phi_2 are never specified, so absolute reward comparisons are not fully pinned down. No new physical entities are introduced.

free parameters (6)
  • Profit coefficient beta_v and scaling factors Phi_1, Phi_2
    Appear in utility Eqs. (18)-(19) and entirely determine the reward, but their values are not given in Table II or the text. Absolute utility numbers and all comparisons depend on them.
  • Subchannel prices lambda_r, lambda_d = 0.2, 0.5
    Hand-set in Sec. IV-B. They control the economic tradeoff between dedicated and reusable subchannels, so the reported seven-fold utility gap is sensitive to these choices.
  • RIS leasing cost psi_j = 0.3
    Hand-set and paid only by VSP 1. This directly controls whether RIS phase optimization appears profitable in the reward.
  • Power cost coefficient alpha_power = 0.1
    Hand-set; changes the balance between achieved rate and transmit-power cost in the utility.
  • QoS penalty coefficient lambda_qos = 50
    Hand-set; large penalty makes QoS violations dominate the reward and shapes the learned policy.
  • Path-loss exponent beta = 2.5
    Assumed for all links with no sensitivity analysis. The size of the RIS benefit may be specific to this channel model.
axioms (6)
  • domain assumption Perfect instantaneous CSI of all direct, BS-RIS, and RIS-user links is available to the agent at every slot.
    State space in (26)-(27) includes H(t); no CSI acquisition or error model is considered. This is the weakest structural assumption for deployment.
  • domain assumption Links between users and non-associated RISs are negligible; each user benefits from at most one RIS.
    Remark after Eq. (13) assumes severe path loss, blockage, and cascaded double fading. This simplifies interference and may overstate RIS gains.
  • domain assumption The environment is quasi-static with i.i.d. Rayleigh fading; training and evaluation share the same distribution.
    Sec. IV-A uses independent CN(0,1) fading per subchannel and link. The learned policy is not tested under mobility or time-varying CSI.
  • domain assumption The feasibility projection Pi_F exactly enforces constraints (2)-(7) through thresholding, clipping, and normalization.
    Eqs. (34) and (42) describe the projection only abstractly; no proof is given that the projected discrete and continuous actions always satisfy scheduling and power constraints.
  • standard math Standard Shannon-rate SINR model and convex SCA lower bound are valid.
    Eqs. (14)-(15) and (51)-(54) use conventional rate expressions and first-order Taylor lower bounds; this is standard communication-theory machinery.
  • ad hoc to paper RIS phase shifts are frequency-independent and identical across all subchannels.
    Sec. IV-A states RIS phases are designed at the main carrier and identical across subchannels, a simplification that may not hold for wideband systems.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DRL-Based Spectrum Sharing for RIS-Aided Local High-Quality Wireless Networks." pith.science (2026). https://pith.science/paper/PAF5FKRM

@misc{pith2026260325332,
  author       = {Pith},
  title        = {Pith review of: DRL-Based Spectrum Sharing for RIS-Aided Local High-Quality Wireless Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PAF5FKRM}},
  note         = {Machine review of arXiv:2603.25332}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper investigates a smart spectrum-sharing framework for reconfigurable intelligent surface (RIS)-aided local high-quality wireless networks (LHQWNs) within a mobile network operator (MNO) ecosystem. Although RISs are often considered harmful due to interference, this work shows that properly controlled RISs can enhance quality of service (QoS). The proposed system enables temporary spectrum access for multiple vertical service providers (VSPs) by dynamically allocating radio resources. The spectrum is divided into dedicated subchannels assigned to individual VSPs and reusable subchannels shared among multiple VSPs, while RIS improves propagation conditions and zero-forcing (ZF) precoding is adopted at the multiantenna base station (BS) to cancel inter-user interference. We formulate a multi-VSP utility maximization problem that jointly optimizes subchannel assignment, transmit power, and RIS phase configuration while accounting for spectrum access costs, RIS leasing costs, and QoS constraints. The resulting mixed-integer non-linear program (MINLP) is modeled as a Markov decision process (MDP) and solved using deep reinforcement learning (DRL). Deep deterministic policy gradient (DDPG) and soft actor-critic (SAC) algorithms are developed and compared. Numerical results show that SAC generally outperforms DDPG in convergence, stability, and utility, particularly in larger-scale scenarios. In the reduced-scale ablation study, both joint DRL methods outperform the heuristic benchmark by at least 6.6% in final moving-average reward.

Figures

Figures reproduced from arXiv: 2603.25332 by Eduard A. Jorswieck, Hamid Reza Hashempour, Hien Quoc Ngo, Mina Khadem.

Figure 1
Figure 1. Figure 1: Functional use case for the integration of local high-quality wireless networks into an MNO ecosystem as service [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System model of an RIS-assisted multi-VSP wireless network within an MNO ecosystem with dedicated and reusable [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Convergence comparison of SAC and DDPG against [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impact of the learning rate µ on training performance. RIS assistance is enabled for VSP 1 with M1 = 8 reflecting elements. 0 5000 10000 15000 20000 25000 30000 Training step 10 15 20 25 30 35 40 45 Total reward B = 16 B = 32 B = 64 B = 128 B = 256 (a) DDPG 0 5000 10000 15000 20000 25000 30000 Training step 10 20 30 40 50 Total reward B = 16 B = 32 B = 64 B = 128 B = 256 (b) SAC [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 8
Figure 8. Figure 8: Impact of the mini-batch size B on training perfor￾mance. RIS assistance is enabled for VSP 1 with M1 = 8 reflecting elements. with M1 = 8 elements. Each curve is obtained using the same random seed and smoothed using a moving average window of 500 steps to highlight transient learning behavior [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Final average reward achieved by SAC with RIS ( [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 2 linked inside Pith

  1. [1]

    Cisco annual internet report 2018–2023,

    Cisco, “Cisco annual internet report 2018–2023,” 2020

    2018

  2. [2]

    Micro operators to boost local service delivery in 5G,

    M. Matinmikko, M. Latva-Aho, P. Ahokangas, S. Yrj ¨ol¨a, and T. Koivum¨aki, “Micro operators to boost local service delivery in 5G,” Wireless Personal Communications, vol. 95, no. 1, pp. 69–82, 2017

    2017

  3. [3]

    Reconfigurable radio systems (rrs); evolved licensed shared access (elsa); part 2: System architecture and high-level procedures,

    ETSI, “Reconfigurable radio systems (rrs); evolved licensed shared access (elsa); part 2: System architecture and high-level procedures,” Tech. Rep. ETSI TS 103 652-2 V1.1.1, European Telecommunications Standards Institute, Jan 2020

    2020

  4. [4]

    Cognitive radio: Making software radios more personal,

    J. Mitola and G. Q. Maguire, “Cognitive radio: Making software radios more personal,”IEEE Personal Commun. Mag., vol. 6, no. 4, pp. 13–18, 1999

    1999

  5. [5]

    Towards smart and reconfigurable environment: Intelligent reflecting surface-aided wireless networks,

    Q. Wu and R. Zhang, “Towards smart and reconfigurable environment: Intelligent reflecting surface-aided wireless networks,”IEEE Commun. Mag., vol. 58, no. 1, pp. 106–112, 2020

    2020

  6. [6]

    Reconfigurable intelligent surfaces for energy efficiency in wireless communication,

    C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wireless communication,”IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 4157–4170, 2019

    2019

  7. [7]

    An intelligent mechanism for dynamic spectrum sharing in 5G IoT networks,

    J. Xu, Z. Xu, W. Yao, W. Hu, A. Cabani, and X. Hu, “An intelligent mechanism for dynamic spectrum sharing in 5G IoT networks,”Expert Syst. Appl., vol. 252, p. 124122, 2024

    2024

  8. [8]

    Dynamic fairness-aware spectrum auction for enhanced licensed shared access in UA V-based networks,

    M. Khadem, M. Ansarifard, N. Mokari, M. R. Javan, H. Saeedi, and E. A. Jorswieck, “Dynamic fairness-aware spectrum auction for enhanced licensed shared access in UA V-based networks,”IEEE Trans. Commun., vol. 73, no. 5, pp. 3076–3092, 2025

    2025

  9. [9]

    Optimizing the spectrum and energy efficiency in dynamic licensed shared access systems,

    S. O. Onidare, O. A. Tiamiyu, Q. R. Adebowale, O. T. Ajayi, K. B. Adewole, and A. A. Ayeni, “Optimizing the spectrum and energy efficiency in dynamic licensed shared access systems,”Int. J. Electr. Eng. Inform., vol. 15, no. 3, pp. 368–386, 2023

    2023

  10. [10]

    Spectral efficiency of dynamic licensed shared access,

    S. O. Onidare, K. Navaie, and Q. Ni, “Spectral efficiency of dynamic licensed shared access,”IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 15149–15161, 2020

    2020

  11. [11]

    Multi- block ascending auctions for effective 5G licensed shared access,

    A. Chouayakh, A. Bechler, I. Amigo, L. Nuaymi, and P. Maill ´e, “Multi- block ascending auctions for effective 5G licensed shared access,”IEEE Trans. Mobile Comput., vol. 21, no. 11, pp. 4051–4063, 2021

    2021

  12. [12]

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

    M. Di Renzo, A. Zappone, M. Debbah, M.-S. Alouini, C. Yuen, J. De Rosny, and S. Tretyakov, “Smart radio environments empowered by reconfigurable intelligent surfaces: How it works, state of research, and the road ahead,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, 2020

    2020

  13. [13]

    Qos-driven spectrum sharing for reconfigurable intelligent surfaces (RISs) aided vehicular networks,

    Y . Chen, Y . Wang, J. Zhang, and M. Di Renzo, “Qos-driven spectrum sharing for reconfigurable intelligent surfaces (RISs) aided vehicular networks,”IEEE Wireless Commun., vol. 20, no. 9, pp. 5969–5985, 2021

    2021

  14. [14]

    Secure SWIPT in the multiuser STAR-RIS aided MISO rate splitting downlink,

    H. R. Hashempour, H. Bastami, M. Moradikia, S. A. Zekavat, H. Behroozi, G. Berardinelli, and A. L. Swindlehurst, “Secure SWIPT in the multiuser STAR-RIS aided MISO rate splitting downlink,”IEEE Trans. Veh. Technol., vol. 73, no. 9, pp. 13466–13481, 2024

    2024

  15. [15]

    Power efficient cooperative communication within IIoT subnetworks: Relay or RIS?,

    H. R. Hashempour, G. Berardinelli, R. Adeogun, and E. A. Jorswieck, “Power efficient cooperative communication within IIoT subnetworks: Relay or RIS?,”IEEE Internet Things J., 2024

    2024

  16. [16]

    Secure rate splitting in STAR- RIS assisted downlink MISO systems,

    H. R. Hashempour and G. Berardinelli, “Secure rate splitting in STAR- RIS assisted downlink MISO systems,” inIEEE MeditCom 2024, pp. 529–534, IEEE, 2024

    2024

  17. [17]

    QoS-aware resource allocation of RIS-aided multi-user MISO wireless communications,

    Y . Gao, C. Lu, Y . Lian, X. Li, G. Chen, D. B. da Costa, and A. Nal- lanathan, “QoS-aware resource allocation of RIS-aided multi-user MISO wireless communications,”IEEE Trans. Veh. Technol., vol. 73, no. 2, pp. 2872–2877, 2023

    2023

  18. [18]

    Weighted sum- rate maximization for reconfigurable intelligent surface aided wireless networks,

    H. Guo, Y .-C. Liang, J. Chen, and E. G. Larsson, “Weighted sum- rate maximization for reconfigurable intelligent surface aided wireless networks,”IEEE Trans. Wireless Commun., vol. 19, no. 5, pp. 3064– 3076, 2020

    2020

  19. [19]

    Joint power control and passive beamforming in IRS-assisted spectrum sharing,

    X. Guan, Q. Wu, and R. Zhang, “Joint power control and passive beamforming in IRS-assisted spectrum sharing,”IEEE Commun. Letters, vol. 24, no. 7, pp. 1553–1557, 2020

    2020

  20. [20]

    Intelligent reflecting surface- aided spectrum sensing for cognitive radio,

    S. Lin, B. Zheng, F. Chen, and R. Zhang, “Intelligent reflecting surface- aided spectrum sensing for cognitive radio,”IEEE Wireless Commun. Letters, vol. 11, no. 5, pp. 928–932, 2022

    2022

  21. [21]

    Private 5G networks: Concepts, archi- tectures, and research landscape,

    M. Wen, Q. Li, K. J. Kim, D. L ´opez-P´erez, O. A. Dobre, H. V . Poor, P. Popovski, and T. A. Tsiftsis, “Private 5G networks: Concepts, archi- tectures, and research landscape,”IEEE J. Sel. Topics Signal Process., vol. 16, no. 1, pp. 7–25, 2021

    2021

  22. [22]

    5G non-public networks: Standardization, architectures and challenges,

    J. Prados-Garzon, P. Ameigeiras, J. Ordonez-Lucena, P. Mu ˜noz, O. Adamuz-Hinojosa, and D. Camps-Mur, “5G non-public networks: Standardization, architectures and challenges,”IEEE Access, vol. 9, pp. 153893–153908, 2021

    2021

  23. [23]

    Spectrum sharing in multi-tenant 5G cellular networks: Modeling and planning,

    O. Al-Khatib, W. Hardjawana, and B. Vucetic, “Spectrum sharing in multi-tenant 5G cellular networks: Modeling and planning,”IEEE Access, vol. 7, pp. 1602–1616, 2018

    2018

  24. [24]

    AI-enabled pri- ority and auction-based spectrum management for 6G,

    M. Khadem, F. Zeinali, N. Mokari, and H. Saeedi, “AI-enabled pri- ority and auction-based spectrum management for 6G,” inProc. IEEE Wireless Commun. Networking Conference (WCNC), 2024

    2024

  25. [25]

    Utility-based interference coordination for local spec- trum licensing in 6G,

    A. Basaure, A. S. De Sena, M. Matinmikko-Blue, S. Yrj ¨ol¨a, and P. Ahokangas, “Utility-based interference coordination for local spec- trum licensing in 6G,” inIEEE DySPAN 2025, pp. 1–8, IEEE, 2025

    2025

  26. [26]

    Compete or coordinate? analysis of spectrum sharing strategies for local wireless services,

    K. Mu, Z. Xie, C. E. C. Bastidas, I. Kadota, W. Lehr, and R. Berry, “Compete or coordinate? analysis of spectrum sharing strategies for local wireless services,” inIEEE DySPAN 2025, pp. 1–10, IEEE, 2025

    2025

  27. [27]

    The frontiers of deep reinforcement learning for resource management in future wireless HetNets: Techniques, challenges, and research direc- tions,

    A. Alwarafy, M. Abdallah, B. S. Ciftler, A. Al-Fuqaha, and M. Hamdi, “The frontiers of deep reinforcement learning for resource management in future wireless HetNets: Techniques, challenges, and research direc- tions,”IEEE Open J. Commun. Soc., vol. 3, pp. 322–365, 2022

    2022

  28. [28]

    A heterogeneous-agent deep rein- forcement learning approach for dynamic spectrum access in cognitive wireless networks,

    Q. Wang, W. Xu, and H.-H. Chen, “A heterogeneous-agent deep rein- forcement learning approach for dynamic spectrum access in cognitive wireless networks,”IEEE Trans. Cogn. Commun. Netw., 2025

    2025

  29. [29]

    Resource man- agement in dynamic shared spectrum networks,

    E. Atimati, T. Nyasulu, D. Crawford, and R. Stewart, “Resource man- agement in dynamic shared spectrum networks,” inIEEE DySPAN 2025, pp. 13–19, IEEE, 2025

    2025

  30. [30]

    An offline multi-agent reinforcement learning framework for radio resource management,

    E. Eldeeb and H. Alves, “An offline multi-agent reinforcement learning framework for radio resource management,”arXiv preprint arXiv:2501.12991, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Deep reinforcement learning for resource allocation in RIS-assisted NOMA- MEC vehicular networks,

    S. Wang, W. Yu, C. H. Foh, Q. Ni, Q. Cheng, and L. Wen, “Deep reinforcement learning for resource allocation in RIS-assisted NOMA- MEC vehicular networks,” inProc. 52th Annual Int. Veh. Technol. Conf., pp. 1–7, IEEE, 2025

    2025

  32. [32]

    Soft actor-critic algorithms and applications,

    T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V . Kumar, H. Zhu, A. Gupta, P. Abbeel,et al., “Soft actor-critic algorithms and applications,”arXiv preprint arXiv:1812.05905, 2018

    Pith/arXiv arXiv 2018

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.