Foundation Model-Aided Deep Reinforcement Learning for RIS-Assisted Wireless Communication

Akram Bin Sediq; Ali Afana; Han Zhang; Melike Erol-Kantarci; Mohammad Ghassemi; Sara Farrag Mobarak

REVIEW 4 major objections 5 minor 19 references

Foundation Model-Aided Deep Reinforcement Learning for RIS-Assisted Wireless Communication

T0 review · 4 major / 5 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 By replacing raw channel matrices with embeddings from a fine-tuned wireless foundation model, the paper's FMDRL agent jointly learns BS beamforming and RIS phase shifts and raises spectral efficiency by 9.89% over a raw-CSI DRL agent and…

desk verdict The FMDRL integration is plausible and clearly written, but the evaluation is circular: the SE reward is computed on learned embeddings that are never validated against the physical channel, so the reported gains are not anchored to reality. read the letter →

arxiv 2506.09855 v1 pith:JBZLYQ7E submitted 2025-06-11 eess.SP

Mohammad Ghassemi , Sara Farrag Mobarak , Han Zhang , Ali Afana , Akram Bin Sediq , Melike Erol-Kantarci This is my paper

classification eess.SP

keywords reconfigurableintelligentsurfacefoundationmodellargewirelessdeepreinforcementlearningchannelembeddingbeamformingspectralefficiencyRISphase-shiftoptimization

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 establish that a pre-trained foundation model for wireless channels can move from passive inference into active network control: the fine-tuned Large Wireless Model (LWM) produces channel embeddings that a deep reinforcement learning agent consumes to choose base-station beamforming vectors and reconfigurable-intelligent-surface (RIS) phase shifts, with the goal of maximizing spectral efficiency. In simulations on a realistic outdoor urban scenario, this FMDRL agent reaches 48.28 bps/Hz at 30 dBm with ten users, 9.89% above a deep RL agent that uses raw channel state information and 43.66% above exhaustive beam sweeping. The advantage holds as transmit power rises from 30 to 45 dBm and as the user count rises from 5 to 20, with the FMDRL edge over raw-CSI DRL staying above 15% for user-count scaling. The paper reads this as evidence that modality-specific foundation models can be fine-tuned and embedded in decision-making loops, not just used for prediction or regression. This matters because RIS-assisted systems are limited by channel-estimation overhead, and learned embeddings are a candidate way to turn coarse channel representations into directly actionable configurations.

What carries the argument

The load-bearing machinery is the LWM embedding pipeline feeding a DDPG agent. Raw channel matrices are split into real and imaginary parts, flattened, and cut into patches; a masked channel-modeling loss (mean squared error over 15% of randomly masked patches, with 80% mask tokens, 10% random noise, 10% unchanged) fine-tunes the last layer on an outdoor urban scenario so the model's output embeddings capture RIS-specific patterns. The CLS token aggregates patch information into a global channel representation, and this representation replaces raw CSI in the DDPG state; the actor-critic then outputs continuous actions for F and θ under power and unit-modulus constraints.

What would settle it

Evaluate the learned beamforming and phase-shift policy on the original raw channel matrices instead of the embedded H_e matrices: if the spectral efficiency computed from the true channels does not reproduce the reported FMDRL advantage over raw-CSI DRL, the gain is an artifact of the embedding substitution rather than a better policy.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that feeding LWM channel embeddings into a deep deterministic policy gradient (DDPG) agent yields better joint decisions than the same agent consuming raw CSI or than exhaustively sweeping beam pairs. The environment is a RIS-aided downlink with a 32-antenna BS, 32-element RIS, and K single-antenna users; the state is the embedded channel triple {H_e,k, G_e,k, H_e,BR}, the action is the beamforming matrix F and phase-shift vector θ, and the reward is the achieved spectral efficiency. The reported result at 30 dBm with 10 users is 48.28 bps/Hz for FMDRL versus 43.93 bps/Hz for DRL with raw CSI and 33.63 bps/Hz for beam sweeping; at 45 dBm the corresponding values are 58.92, 54.98, and 39.27 bps/Hz, and across 5 to 20 users the FMDRL advantage over raw-CSI DRL stays between 15.48% and 15.66%. The paper concludes that a pre-trained FM for radio signal understanding can be fine-tuned and integrated with DRL for effective decision-making in wireless networks.

Load-bearing premise

The load-bearing premise is that LWM's compressed channel embeddings can stand in for the physical channels in the received-signal and spectral-efficiency equations; the paper does not define the raw-to-embedded mapping or verify that the embedding preserves the information needed to compute true spectral efficiency.

Editorial extensions

If this is right

At 30 dBm with 10 users, FMDRL achieves 48.28 bps/Hz, a 9.89% gain over DRL with raw CSI (43.93 bps/Hz) and a 43.66% gain over beam sweeping (33.63 bps/Hz).
At 45 dBm, FMDRL reaches 58.92 bps/Hz, staying 7.16% ahead of raw-CSI DRL and 50.02% ahead of beam sweeping, so the gain does not vanish at higher transmit power.
With 5 to 20 users at 35 dBm, FMDRL keeps a 15.48% to 15.66% advantage over raw-CSI DRL and a 45.52% to 62.71% advantage over beam sweeping, indicating the method scales with user count.
Fine-tuning only the last LWM layer on a realistic DeepMIMO urban scenario is sufficient for the resulting embeddings to improve both the DDPG agent's cumulative reward and its convergence behavior relative to raw-CSI DRL.
The consistent margin over two very different baselines supports the paper's broader claim that wireless foundation models can be integrated into optimization loops rather than confined to prediction tasks.

Reading between the lines

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

Because the received-signal and spectral-efficiency expressions in Eqs. (1) and (3) are written directly in terms of the embedded channels H_e rather than the physical channels, an independent check should re-evaluate the learned F_opt and θ_opt on raw channel matrices; if the margin shrinks or reverses, the reported SE gains are partly an artifact of the embedding substitution.
The same fine-tuned-embedding-plus-DDPG pattern should transfer to other continuous wireless control problems with high-dimensional CSI, such as power allocation, user scheduling, or integrated sensing and communication; this is a natural testbed, not something the paper simulates.
The convergence improvement in Fig. 3 hints that embeddings shrink the effective state space, so a testable prediction is that FMDRL reaches a target spectral efficiency in fewer training episodes than raw-CSI DRL; the paper reports final performance but does not quantify episode-level efficiency.
Since LWM and DeepMIMO are both openly available, the 9.89% and 43.66% margins are directly reproducible, and the stability of those margins across random seeds and channel realizations would clarify whether the advantage is structural or tuned to one configuration.

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

Quick take: this paper is a clean integration of a channel foundation model with DDPG for RIS control, and the writing is readable. But be careful with the evaluation. The received-signal and spectral-efficiency equations (Eqs. 1 and 3) are written using embedded channel matrices H_e, and the DRL reward is computed from that same embedded model. The paper never defines how the LWM output is reshaped into an N_r×N_t or M×N_t matrix, nor validates that H_e approximates the physical channel. So the reported 9.89% and 43.66% gains are measured against an artificial model, not the true channel. The stress-test concern holds up on reading.

Credit where due: the specific combination of fine-tuned LWM embeddings with DDPG for joint beamforming and RIS phase selection is not in the cited prior art. The idea is plausible, and the simulation sweeps over power and user counts are broad. The related work is relevant, and the paper is generally clear about its architecture.

The soft spots, in order of severity. First, the load-bearing circularity above. Second, the abstract promises reduced channel estimation overhead, but the experiments use perfect CSI from DeepMIMO and simply pass it through the LWM. No channel estimation is simulated. Third, the fine-tuning objective is masked patch reconstruction (Eq. 6), not channel estimation per se, so calling the output 'accurate CSI' is unsubstantiated. Fourth, there are no error bars, no code, and no comparison against a simpler feature extractor or an upper bound. Those omissions make it hard to assess whether the gains come from the foundation model or just from the DRL framework.

Minor issues: the reward definition r_t = sum over T SE_t is ambiguous, and the paper leans a bit on self-citations, but the cited prior work is relevant.

The central modeling premise is the problem, not the authors' execution. A corrected paper would define the embedding-to-channel map, train or fine-tune with a channel-estimation loss, and then evaluate the DRL actions on the physical channel rather than on the embedding. That would make the results meaningful and testable.

For peer review: I'd send this to a serious referee, mainly because the integration is novel and the flaw is fixable, but the current version should be rejected unless the authors can re-anchor the evaluation. If you want a paper that illustrates the difference between optimizing a surrogate model and optimizing the true objective, this is a good reading-group example.

Referee Report

4 major / 5 minor

Summary. The paper proposes FMDRL, a framework that combines a fine-tuned Large Wireless Model (LWM) with a deep deterministic policy gradient (DDPG) agent to jointly optimize BS beamforming and RIS phase shifts in a multi-user RIS-assisted system. The LWM is fine-tuned on DeepMIMO channels using masked channel modeling, and its embeddings are used as the state for DDPG, whose reward is the spectral efficiency (SE) computed from those embedded channels. Simulations report that FMDRL outperforms DRL with raw CSI and beam sweeping, e.g., 9.89% and 43.66% higher SE at 30 dBm with K=10, with consistent gains across power levels and user counts.

Significance. The direction of integrating a large wireless foundation model with reinforcement learning for network control is timely and potentially valuable. The paper is clearly written, the problem formulation is standard, and the authors provide a specific DeepMIMO scenario and hyperparameter details. However, the central technical premise is not validated: Eqs. (1) and (3) substitute embedded channel matrices H_e,k, G_e,k, and H_e,BR, whose mapping from the physical channel is never defined, into the received-signal and SE expressions. The evaluation is therefore self-referential, and the reported gains are not anchored to a physical channel. No error bars, code, or independent validation of the embedding are provided. If the embedding were properly defined and validated, the work could be a meaningful step toward FM-aided wireless optimization, but as presented the performance claims are unsupported.

major comments (4)

[Section III.B, Eqs. (1)-(3)] The received signal model and spectral-efficiency expression use embedded channel matrices H_e,k, G_e,k, and H_e,BR as if they were the physical propagation matrices. Section IV.A only describes patching, CLS tokens, and masked reconstruction; it never defines the transformation from the raw complex channel H to the matrices inserted into Eq. (1), nor states how the LWM output is reshaped to dimensions N_r x N_t or M x N_t. As a result, Eq. (3) is not the physical spectral efficiency, the DDPG reward is computed on an artificial model, and the learned policy may maximize embedded SE while degrading true SE. To support the claims in Section V.C, the paper must specify the embedding mapping, validate that H_e preserves the information needed for SINR computation (e.g., by comparing SE computed with H_e against SE computed with H on hold-out channel realizations), and report all results using the physical-channel SE.
[Section IV.A, Eq. (6)] The fine-tuning objective is a patch-level MSE reconstruction of masked patches. This loss does not guarantee that the resulting embedding preserves the per-element phase and amplitude information required for the interference and beamforming terms in Eq. (3). At minimum, the paper should provide a quantitative fidelity check, such as the relative error between SE(H_e,F,Theta) and SE(H,F,Theta) over a test set, before any performance comparison is meaningful.
[Section V.C, Figs. 3-5] All performance numbers are reported as single deterministic curves with no error bars, confidence intervals, or multiple random seeds. DDPG is known to be seed-sensitive, so the claimed 9.89% and 43.66% improvements cannot be assessed for statistical significance. Additionally, if the DRL-with-raw-CSI baseline is trained and evaluated on physical SE while FMDRL is trained and evaluated on embedded SE, the comparison in Figs. 4-5 is not on equal footing; both policies should be evaluated on the same physical SE metric.
[Section V.A and Eq. (5)] The patch-size parameters are inconsistent with the stated channel dimensions. With N_t=32, N_r=1, and P=32, Eq. (5) gives L = 2XY/P = 2, while Section V.A states a patch size of L=32. The paper should clarify how the complex channel is padded or segmented and how the final embedding is converted back into a channel matrix of the required dimensions for use in Eqs. (1) and (3).

minor comments (5)

[Eq. (3)] The Euclidean norm in the SINR expression should use a Hermitian transpose when N_r > 1; the current notation is only correct for the scalar case N_r=1, which should be stated explicitly.
[Section IV.B] The reward is written as r_t = sum_{t=1}^T SE_t, which reuses the time index t and is dimensionally inconsistent; it should be the sum over users of SE_k at time step t.
[Algorithm 1] The line 'a(t) <- scale actions(a(t))' is confusing; the action-scaling mechanism, including how the power constraint in Eq. (4) is enforced, should be described explicitly.
[References] The DDPG algorithm is not cited; reference [4] is cited for the replay buffer but it is a prior work on a different ISAC problem. A standard DDPG reference should be added.
[Section V.A] The fine-tuning description mentions a learning rate of 1e-5 and then later AdamW with learning rate 5e-5; clarify which learning rate applies to the LWM fine-tuning and which to the DDPG.

Circularity Check

1 steps flagged · score 6.0 of 10

The SE metric in Eq. (3) is defined on the LWM's own embeddings, so the reported gains are measured against the method's self-supplied channel model rather than physical channels.

self definitional [Section III.B, Eqs. (1) and (3); used again in Section IV.B (state and reward) and Section V.C (reported SE gains).]
"yk = (He,k +G e,kΘHe,BR) Σ u=1..K f_u x_u + n_k, (1) ... where H e,k ∈C Nr×Nt is the embedded direct channel between the BS and the kth user. The subscript e denotes an embedding transformation applied to the original channel representation. ... SEk = log2( 1 + ||(He,k +Ge,k ΘHe,BR )f_k||^2 / ( Σ u≠k ||(He,k +Ge,k ΘHe,BR )f_u||^2 + σ^2_k ) ). (3)"

The spectral efficiency that the DDPG agent maximizes and that the paper reports as its headline result is, by Eq. (3), a function of the learned matrices H_e,k, G_e,k, and H_e,BR. The paper only says that the subscript e denotes an embedding transformation; it never defines how the LWM's patch or CLS output is reshaped into an N_r×N_t or M×N_t matrix, nor validates that H_e preserves the SINR of the physical DeepMIMO channel. The DRL state s_t = {H_e,k, G_e,k, H_e,BR} and reward r_t = Σ SE_t in Section IV.B are built from the same H_e objects, so the reported 9.89% and 43.66% improvements are, by construction, improvements of an objective evaluated on the proposed component's own output, not on an independent physical channel.

full rationale

The paper's central derivation chain is: LWM takes raw channel H and produces an embedding H_e; DDPG observes H_e and optimizes F and Θ; the reward is SE_t computed from H_e via Eq. (3); and the reported gains are those reward/SE values. Because Eq. (3) defines SE on H_e, and H_e is produced by the very model whose benefit the paper claims, the evaluation is self-referential: the metric is co-defined with the proposed component. The paper neither specifies the raw-to-embedded mapping nor benchmarks the final action against physical raw CSI, so the numerical claims are not anchored to an independent physical truth model. This is not a full 8-10 because the gain is not purely tautological—it depends on the learned H_e and the DDPG optimization—and there is no load-bearing self-citation or fitted-parameter chain; the circularity is that the target metric itself is defined through the method's output.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central result is an empirical ML comparison. It rests on a large set of learned weights and hand-selected hyperparameters, and on the unstated assumption that the LWM embedding can stand in for the physical channel in the received signal and SE equations. There are no external benchmarks, no code, and no error bars, so the ledger is dominated by ad hoc modeling choices.

free parameters (5)

Fine-tuned LWM weights (including projection W_dec) = not released
Fine-tuned on 70% of the DeepMIMO O1 dataset; the reported gains depend on these fitted weights and they are not shipped.
DDPG actor and critic network weights = not released
Trained over 150,000 episodes in the same simulator used for evaluation; no weights or code provided.
LWM patch size and patch count = P=32, L=32
Chosen in Section V.A without ablation or sensitivity analysis.
Masking ratio and optimization hyperparameters = mask 15%; LR 1e-5 and 5e-5; AdamW; batch 64
Selected by hand; no sensitivity study, so the reported margin may depend on them.
DDPG hyperparameters = actor LR 1e-4, critic LR 1e-3, gamma 0.99, tau 0.005, hidden 64x64
Set in Section V.A without tuning analysis; not shown to be optimal or comparable across baselines.

assumptions (4)

ad hoc to paper The LWM embedding can replace the physical channel in the received signal and SE equations (Eqs (1) and (3)).
No definition or justification is given for the subscript-e channels; the physical validity of this substitution is the core unstated assumption.
domain assumption Perfect CSI is available at the input of the LWM.
The simulation uses DeepMIMO channel matrices directly; the channel estimation overhead discussed in the introduction is never modeled.
domain assumption The DeepMIMO O1 dataset provides valid channels for a RIS-assisted system.
The paper does not describe how BS-RIS and RIS-user channels are generated; DeepMIMO is a ray-tracing MIMO dataset, not a RIS dataset.
domain assumption DDPG converges to a near-optimal policy for the non-convex problem in Eq (4).
The paper relies on standard DRL convergence behavior without guarantees or convergence certificates.

invented entities (1)

Embedded channel matrices H_e,k, G_e,k, H_e,BR
purpose: Used as both the DRL state and as the channel matrices inside the received signal and SE equations.
The paper introduces the subscript-e notation but never specifies the transformation from raw channels to embeddings, and provides no evidence that the embedding preserves the physical information needed for spectral efficiency.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Foundation Model-Aided Deep Reinforcement Learning for RIS-Assisted Wireless Communication." pith.science (2026). https://pith.science/paper/JBZLYQ7E

@misc{pith2026250609855,
  author       = {Pith},
  title        = {Pith review of: Foundation Model-Aided Deep Reinforcement Learning for RIS-Assisted Wireless Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBZLYQ7E}},
  note         = {Machine review of arXiv:2506.09855}
}

read the original abstract

Reconfigurable intelligent surfaces (RIS) have emerged as a promising technology for enhancing wireless communication by dynamically controlling signal propagation in the environment. However, their efficient deployment relies on accurate channel state information (CSI), which leads to high channel estimation overhead due to their passive nature and the large number of reflective elements. In this work, we solve this challenge by proposing a novel framework that leverages a pre-trained open-source foundation model (FM) named large wireless model (LWM) to process wireless channels and generate versatile and contextualized channel embeddings. These embeddings are then used for the joint optimization of the BS beamforming and RIS configurations. To be more specific, for joint optimization, we design a deep reinforcement learning (DRL) model to automatically select the BS beamforming vector and RIS phase-shift matrix, aiming to maximize the spectral efficiency (SE). This work shows that a pre-trained FM for radio signal understanding can be fine-tuned and integrated with DRL for effective decision-making in wireless networks. It highlights the potential of modality-specific FMs in real-world network optimization. According to the simulation results, the proposed method outperforms the DRL-based approach and beam sweeping-based approach, achieving 9.89% and 43.66% higher SE, respectively.

Figures

Figures reproduced from arXiv: 2506.09855 by the authors.

**Figure 1.** System architecture for FMDRL. system consists of a BS equipped with Nt antennas, a RIS comprising M reflecting elements, and K User Equipment (UE) devices, each with Nr receive antennas. The BS serves UEs via line-of-sight (LoS) transmissions, while the RIS creates enhanced propagation paths by reconstituting non-line-of-sight (NLoS) components into quasi-LOS links. With this setup, we aim to enhance the SE with t… view at source ↗

**Figure 2.** System architecture for optimizing the BS beamforming matrix and the RIS phase-shift matrix by leveraging FMDRL [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison of average cumulative reward between the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: SE performance comparison of the FMDRL, DRL with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: SE performance comparison of the FMDRL, DRL with [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 13 canonical work pages

[1]

Cooperative Hierarchical Deep Reinforcement Learning Based Joint Sleep and Power Control in RIS-Aided Energy-Efficient RAN,

H. Zhou, M. Elsayed, M. Bavand, R. Gaigalas, S. Furr, and M. Erol- Kantarci, “Cooperative Hierarchical Deep Reinforcement Learning Based Joint Sleep and Power Control in RIS-Aided Energy-Efficient RAN,”IEEE Transactions on Cognitive Communications and Net- working, 2024

work page 2024
[2]

Channel estimation for RIS-aided multiuser millimeter-wave sys- tems,

G. Zhou, C. Pan, H. Ren, P. Popovski, and A. L. Swindlehurst, “Channel estimation for RIS-aided multiuser millimeter-wave sys- tems,”IEEE Transactions on Signal Processing, vol. 70, pp. 1478– 1492, 2022

work page 2022
[3]

Digital Twin Aided RIS Communi- cation: Robust Beamforming and Interference Management,

S. Alikhani and A. Alkhateeb, “Digital Twin Aided RIS Communi- cation: Robust Beamforming and Interference Management,”2024 IEEE 100th V ehicular Technology Conference (VTC2024-Fall), pp. 1–6, 2024

work page 2024
[4]

Sum Rate Enhancement using Machine Learning for Semi-Self Sensing Hybrid RIS-Enabled ISAC in THz Bands

S. F. Mobarak, T. Bao, and M. Erol-Kantarci, “Sum Rate Enhance- ment using Machine Learning for Semi-Self Sensing Hybrid RIS- Enabled ISAC in THz Bands,”arXiv preprint arXiv:2501.12353, 2025

work page Pith review arXiv 2025
[5]

Hierar- chical Codebook-Based Beam Training for RIS-Assisted mmWave Communication Systems,

J. Wang, W. Tang, S. Jin, C.-K. Wen, X. Li, and X. Hou, “Hierar- chical Codebook-Based Beam Training for RIS-Assisted mmWave Communication Systems,”IEEE Transactions on Communications, vol. 71, no. 6, pp. 3650–3662, 2023

work page 2023
[6]

DRL-based Joint Beamforming and BS-RIS-UE Association Design for RIS- Assisted mmWave Networks,

Y . Zhu, M. Li, Y . Liu, Q. Liu, Z. Chang, and Y . Hu, “DRL-based Joint Beamforming and BS-RIS-UE Association Design for RIS- Assisted mmWave Networks,”2022 IEEE Wireless Communications and Networking Conference (WCNC), pp. 345–350, 2022

work page 2022
[7]

Multi- Agent Deep Reinforcement Learning for Beam Codebook Design in RIS-Aided Systems,

A. Abdallah, A. Celik, M. M. Mansour, and A. M. Eltawil, “Multi- Agent Deep Reinforcement Learning for Beam Codebook Design in RIS-Aided Systems,”IEEE Transactions on Wireless Communi- cations, vol. 23, no. 7, pp. 7983–7999, 2024

work page 2024
[8]

Llama 2: Open Foundation and Fine-Tuned Chat Models,

H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open Foundation and Fine-Tuned Chat Models,”arXiv preprint arXiv:2307.09288, 2023

arXiv 2023

Show all 19 references

[9]

Large Wireless Model (LWM): A Foundation Model for Wireless Channels,

S. Alikhani, G. Charan, and A. Alkhateeb, “Large Wireless Model (LWM): A Foundation Model for Wireless Channels,”arXiv preprint arXiv:2411.08872, 2024

2024 arXiv
[10]

Deep Learning for mmWave Beam and Blockage Prediction Using Sub-6 GHz Channels,

M. Alrabeiah and A. Alkhateeb, “Deep Learning for mmWave Beam and Blockage Prediction Using Sub-6 GHz Channels,”IEEE Transactions on Communications, vol. 68, no. 9, pp. 5504–5518, 2020

2020
[11]

Generative AI-enabled Blockage Prediction for Robust Dual-Band mmWave Communication,

M. Ghassemi, H. Zhang, A. Afana, A. B. Sediq, and M. Erol-Kantarci, “Generative AI-enabled Blockage Prediction for Robust Dual-Band mmWave Communication,”arXiv preprint arXiv:2501.11763, 2025

2025 arXiv
[12]

RIS-Aided Cell-Free Massive MIMO System: Joint Design of Transmit Beamforming and Phase Shifts,

S.-N. Jin, D.-W. Yue, and H. H. Nguyen, “RIS-Aided Cell-Free Massive MIMO System: Joint Design of Transmit Beamforming and Phase Shifts,”IEEE Systems Journal, vol. 17, no. 2, pp. 3093–3104, 2022

2022
[13]

Enabling Efficient Blockage-Aware Handover in RIS-Assisted mmWave Cel- lular Networks,

L. Jiao, P. Wang, A. Alipour-Fanid, H. Zeng, and K. Zeng, “Enabling Efficient Blockage-Aware Handover in RIS-Assisted mmWave Cel- lular Networks,”IEEE Transactions on Wireless Communications, vol. 21, no. 4, pp. 2243–2257, 2021

2021
[14]

Joint Cross Layer Radio Downlink Beamforming and RIS Configuration via Deep Reinforcement Learning in RIS-Aided MISO Video Commu- nications,

S.-M. Tseng, R.-Y . Chen, Y .-F. Chen, and C. Fang, “Joint Cross Layer Radio Downlink Beamforming and RIS Configuration via Deep Reinforcement Learning in RIS-Aided MISO Video Commu- nications,”IEEE Transactions on Cognitive Communications and Networking, 2025

2025
[15]

Deep reinforcement learning for energy-efficient networking with recon- figurable intelligent surfaces,

G. Lee, M. Jung, A. T. Z. Kasgari, W. Saad, and M. Bennis, “Deep reinforcement learning for energy-efficient networking with recon- figurable intelligent surfaces,”ICC 2020-2020 IEEE international conference on communications (ICC), pp. 1–6, 2020

2020
[16]

Wifo: Wire- less foundation model for channel prediction,

B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “Wifo: Wire- less foundation model for channel prediction,”arXiv preprint arXiv:2412.08908, 2024

2024 arXiv
[17]

Multi-Modal Transformer and Reinforcement Learning- Based Beam Management,

M. Ghassemi, H. Zhang, A. Afana, A. B. Sediq, and M. Erol- Kantarci, “Multi-Modal Transformer and Reinforcement Learning- Based Beam Management,”IEEE Networking Letters, 2024

2024
[18]

Beam Selection in ISAC using Contextual Bandit with Multi-modal Transformer and Transfer Learning,

M. Farzanullah, H. Zhang, A. B. Sediq, A. Afana, and M. Erol- Kantarci, “Beam Selection in ISAC using Contextual Bandit with Multi-modal Transformer and Transfer Learning,”arXiv preprint arXiv:2503.08937, 2025

2025 arXiv
[19]

Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,

A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,”arXiv preprint arXiv:1902.06435, 2019

1902 arXiv

Pith tools

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

[1] [1]

Cooperative Hierarchical Deep Reinforcement Learning Based Joint Sleep and Power Control in RIS-Aided Energy-Efficient RAN,

H. Zhou, M. Elsayed, M. Bavand, R. Gaigalas, S. Furr, and M. Erol- Kantarci, “Cooperative Hierarchical Deep Reinforcement Learning Based Joint Sleep and Power Control in RIS-Aided Energy-Efficient RAN,”IEEE Transactions on Cognitive Communications and Net- working, 2024

work page 2024

[2] [2]

Channel estimation for RIS-aided multiuser millimeter-wave sys- tems,

G. Zhou, C. Pan, H. Ren, P. Popovski, and A. L. Swindlehurst, “Channel estimation for RIS-aided multiuser millimeter-wave sys- tems,”IEEE Transactions on Signal Processing, vol. 70, pp. 1478– 1492, 2022

work page 2022

[3] [3]

Digital Twin Aided RIS Communi- cation: Robust Beamforming and Interference Management,

S. Alikhani and A. Alkhateeb, “Digital Twin Aided RIS Communi- cation: Robust Beamforming and Interference Management,”2024 IEEE 100th V ehicular Technology Conference (VTC2024-Fall), pp. 1–6, 2024

work page 2024

[4] [4]

Sum Rate Enhancement using Machine Learning for Semi-Self Sensing Hybrid RIS-Enabled ISAC in THz Bands

S. F. Mobarak, T. Bao, and M. Erol-Kantarci, “Sum Rate Enhance- ment using Machine Learning for Semi-Self Sensing Hybrid RIS- Enabled ISAC in THz Bands,”arXiv preprint arXiv:2501.12353, 2025

work page Pith review arXiv 2025

[5] [5]

Hierar- chical Codebook-Based Beam Training for RIS-Assisted mmWave Communication Systems,

J. Wang, W. Tang, S. Jin, C.-K. Wen, X. Li, and X. Hou, “Hierar- chical Codebook-Based Beam Training for RIS-Assisted mmWave Communication Systems,”IEEE Transactions on Communications, vol. 71, no. 6, pp. 3650–3662, 2023

work page 2023

[6] [6]

DRL-based Joint Beamforming and BS-RIS-UE Association Design for RIS- Assisted mmWave Networks,

Y . Zhu, M. Li, Y . Liu, Q. Liu, Z. Chang, and Y . Hu, “DRL-based Joint Beamforming and BS-RIS-UE Association Design for RIS- Assisted mmWave Networks,”2022 IEEE Wireless Communications and Networking Conference (WCNC), pp. 345–350, 2022

work page 2022

[7] [7]

Multi- Agent Deep Reinforcement Learning for Beam Codebook Design in RIS-Aided Systems,

A. Abdallah, A. Celik, M. M. Mansour, and A. M. Eltawil, “Multi- Agent Deep Reinforcement Learning for Beam Codebook Design in RIS-Aided Systems,”IEEE Transactions on Wireless Communi- cations, vol. 23, no. 7, pp. 7983–7999, 2024

work page 2024

[8] [8]

Llama 2: Open Foundation and Fine-Tuned Chat Models,

H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open Foundation and Fine-Tuned Chat Models,”arXiv preprint arXiv:2307.09288, 2023

arXiv 2023