Optimizing Wireless Resource Management and Synchronization in Digital Twin Networks

Haijian Sun; Hanzhi Yu; Mingzhe Chen; Yuchen Liu; Zhaohui Yang

REVIEW 5 major objections 5 minor 33 references

Optimizing Wireless Resource Management and Synchronization in Digital Twin Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a distributed reinforcement learning scheme combining GRU-based digital twin prediction with value decomposition can jointly optimize user data rates and twin synchronization, outperforming independent Q-learning by…

desk verdict A legitimate joint formulation of DNT synchronization and resource allocation, but the convergence proof assumes the very decomposition it needs to establish; referee it with that caveat in mind. read the letter →

arxiv 2502.05116 v1 pith:J757AB4V submitted 2025-02-07 cs.NI cs.LGcs.SYeess.SY

Hanzhi Yu , Yuchen Liu , Zhaohui Yang , Haijian Sun , Mingzhe Chen This is my paper · ORCID

classification cs.NIcs.LGcs.SYeess.SY

keywords digitalnetworktwinresourceallocationgatedrecurrentunitsvaluedecompositionmulti-agentreinforcementlearningsynchronizationOFDMAHungarianalgorithm

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

A digital twin of a wireless network is only useful if it stays close to the physical network, but feeding it status updates costs spectrum that could otherwise serve users. This paper argues that base stations can learn when to skip those updates, relying on a gated recurrent unit (GRU) at the cloud server to predict missing status, while a value decomposition network (VDN) lets each base station choose its user associations and transmission decisions from its own partial view. The two pieces are joined by a Hungarian-algorithm step that turns each base station's resource-block allocation into a bipartite matching problem. The paper reports that in a simulated three-base-station network, the GRU+VDN method improves the weighted sum of user data rates and physical-to-twin similarity by up to 28.96% over a GRU+independent-Q-learning baseline. If the result holds, distributed spectrum management can treat twin synchronization as a tunable objective rather than a fixed overhead.

What carries the argument

The load-bearing machinery is the additive value decomposition of Eq. (19), $Q_{tot}(s_t,a_t)=\sum_{m=1}^{M} Q_{\theta_m}(s_t^m,a_t^m)$, which makes the global team-reward optimization tractable for base stations that observe only local user positions. Around it sit the GRU predictor at the cloud, which estimates missing physical-network status from the last $K$ twin states, and the Hungarian-algorithm subroutine that converts each base station's resource-block allocation into a bipartite matching problem once user associations are fixed. The team reward in Eq. (16) combines the negative twin-synchronization error with the sum of user data rates, with a penalty for a user being served by multiple base stations. This GRU+VDN combination is what lets the paper claim partial observability can be overcome without a central controller.

What would settle it

Run the same GRU+VDN training in a network with strong inter-cell interference, for example users concentrated at cell edges with scarce resource blocks, and measure the gap between the team reward actually received and the sum of the local Q estimates; if the gap is large, the additive-decomposition assumption fails and the learned policy would be suboptimal. Alternatively, replace the random-walk mobility with correlated group mobility and check whether the GRU prediction error rises enough to erase the reported 28.96% improvement.

Watch

Extended reading notes

Core claim

The central claim is that the resource management problem in a digital-twin-enabled network, where each base station must split limited resource blocks between serving users and reporting physical-network status to the cloud, can be solved by a combination of recurrent prediction and cooperative multi-agent reinforcement learning. The GRU at the cloud lets the twin extrapolate user positions from recent history, so a base station that withholds its report does not immediately desynchronize the twin. The VDN makes the global Q-function the sum of per-base-station Q-functions, allowing each base station to update its own deep Q-network using only its local observation while still optimizing the team reward. Given the chosen user associations, the remaining resource-block assignment is solved optimally per base station with the Hungarian algorithm. On the paper's simulated network, this design improves the weighted sum of user data rates and twin similarity by up to 28.96% relative to the independent-Q-learning baseline.

Load-bearing premise

The scheme's load-bearing premise is that the team's overall Q-function is exactly the sum of the base stations' individual Q-functions, even though base stations interfere and share one twin; the convergence proof only recovers this by assuming the decomposition error is zero.

Editorial extensions

If this is right

If the DNT can predict user positions accurately, base stations can skip frequent status uploads and use the freed resource blocks for user data without immediately losing twin accuracy.
Distributed base stations that observe only their own coverage area can still learn a joint policy through additive Q decomposition, avoiding the need for a centralized controller.
The Hungarian-based resource-block allocation keeps each base station's learned action space at $O(U)$ instead of $O(NU)$, reducing the complexity of the reinforcement learning problem in larger networks.
In the simulated settings, the proposed method outperforms independent Q-learning by up to 28.96% in the weighted objective, and by up to 31.79% in average user data rate when the weight $ε$ is 0.8.
The reported tradeoff between data rate and DNT error as $ε$ varies suggests that operators can tune the balance between user service and twin accuracy through a single weight parameter.

Reading between the lines

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

Beyond the paper, the additive decomposition assumption in Eq. (19) deserves direct testing: if the true team-reward function is not close to additive under strong inter-cell interference, the learned policies could become arbitrarily suboptimal in networks denser than the simulated three-base-station setup.
Beyond the paper, the reported 28.96% gain is tied to random-walk mobility; replacing it with correlated or group mobility would likely change GRU prediction error and could shrink or erase the advantage, making this a natural stress test for the method.
Beyond the paper, the weighted objective does not bound the worst-case DNT error, so a system designer who needs a hard synchronization guarantee would have to add an explicit constraint rather than rely on the average-similarity term.
Beyond the paper, the same framework could be extended to nonlinear value mixing functions to see whether the improvement persists when the global Q-function is no longer forced to be a plain sum of local values.

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

The real contribution here is the problem formulation: having each BS decide when to send physical network information to the cloud, trading DNT accuracy against user data rates, with the DNT using a GRU to predict missing states. That joint treatment is genuinely absent from the cited synchronization papers [12]-[15], and the VDN-plus-Hungarian structure is a sensible way to keep the action space tractable. The paper is honest about what it does: it is an incremental extension of known building blocks, and it says so.

What it does well: the system model is detailed, the optimization problem is well specified, and the choice to let the RL agents pick user association while a Hungarian algorithm handles RB allocation is a clean decomposition. The simulations consistently show an improvement over the IQL baseline, both for the headline 28.96% and in the trend plots as epsilon and user count vary. The GRU prediction errors are reported per user and look plausible.

The soft spots are real but not fatal to the empirical claim. Lemma 1's convergence proof is circular: it defines the gap between the true Q and the VDN's additive Qtot, then requires that gap to vanish to get a contraction. That is exactly the assumption in question, and the gamma-to-1 escape hatch does not apply with gamma=0.2 in Table II. The action space size is also miscounted (|A| = 2U+1 does not match an action containing a transmit indicator and a U-bit user association vector), and the penalty rho appears in the reward but is never given a value. Constraint (8g) is present in the formulation and then ignored in the solution. None of these destroy the comparison against IQL, but they do mean the theoretical support is weaker than presented, and the 28.96% gain should be treated as a simulation result without a proven convergence guarantee.

The stress-test note is accurate on the circularity. I disagree that this makes the central claim unreliable; it makes the theory unreliable, and the empirical comparison plausible. The reader's conditional verdict is fair.

Who is this for? Researchers working on digital twin synchronization or MARL for radio resource allocation in a niche subfield. It deserves a serious referee: the problem formulation is worth publishing, but the paper needs correction of the convergence lemma (or its removal), clarification of the action encoding and missing parameters, and ideally a second baseline and error bars. Send it to review with those demands.

Referee Report

5 major / 5 minor

Summary. The paper studies a digital twin network (DNT) in which a set of base stations (BSs) must allocate limited spectrum resources both to serve mobile users and to transmit partially observed physical-network state to a cloud server that maintains the DNT. The goal is to jointly maximize user data rates and the similarity between the DNT and the physical network. The authors formulate this as a constrained optimization problem, propose a hybrid solution that combines GRU-based state prediction at the cloud with a value-decomposition network (VDN) at the BSs, and use a Hungarian algorithm to assign resource blocks after each BS selects its user association and synchronization action. Simulation results are reported for a 3-BS, 12-user scenario, showing an improvement over an independent Q-learning (IQL) baseline.

Significance. The problem is timely and the core idea—using GRU prediction to reduce the frequency of physical-to-virtual synchronization transmissions, and using VDN to coordinate BSs despite partial observations—is interesting and practically motivated. The main empirical claim, a 28.96% improvement in the weighted objective over IQL in the simulated configuration, is not circular: it comes from a direct simulation comparison, not from fitting a parameter. The decomposition of RB allocation via the Hungarian algorithm is a pragmatic way to reduce the RL action space. However, the theoretical convergence analysis (Lemma 1) is not sound as stated, the action-space cardinality used in the complexity analysis is inconsistent with the definition of the action, and constraint (8g) is not enforced in the RL reward. These issues undermine the paper's convergence and complexity claims, though they do not by themselves refute the empirical improvement.

major comments (5)

[IV-B, Lemma 1 and Eq. (36)] The convergence proof of Lemma 1 requires either ε→0 or γ→1. The ε→0 condition is effectively the assumption that the VDN decomposition error vanishes, i.e., that the global Q function equals the sum of local Q functions, which is exactly the additivity property that needs to be established. The γ→1 escape does not provide a Banach contraction (a contraction coefficient of 1 does not yield a fixed point by the Banach theorem) and is inconsistent with the simulation setting, where Table II sets γ=0.2. Thus Lemma 1 does not establish convergence in the regime used for the simulations. The paper should either remove the convergence claim, prove a suboptimality bound that holds for nonzero ε, or clearly label the lemma as a heuristic justification.
[III-C3 and IV-A, Eqs. (19) and (25)–(26)] The action is defined as a_m^t = [|y_mt|, z_mt] with z_mt ∈ {0,1}^U, so the number of possible actions is at least 2^{U+1} (or (N+1)·2^U if the choice of RB for synchronization matters), not |A| = 2U+1 as claimed in Section IV-A. The complexity comparison between the proposed method and the standard VDN, and the statement that the proposed method 'can significantly reduce the complexity,' rely on this cardinality. The definition of A and the stated |A| need to be reconciled; if the Q-network outputs a factored policy (e.g., separate heads for association and transmission), this should be stated explicitly.
[II-D and III-C5, constraint (8g) and reward (16)] The delay constraint T_mt ≤ α in (8g) is part of the formulated optimization problem, but it does not appear in the team reward (16) or in the RL training procedure. The learned policy may therefore violate (8g). The paper does not specify the value of α in Table II or report whether the constraint is satisfied in the simulations. To make the RL solution faithful to problem (8), the delay constraint should be incorporated into the reward or a justification should be given that (8g) is inactive in the simulated regime.
[III-C5, Eq. (16) and Table II] The penalty ρ for a user being served by multiple BSs is introduced in the reward function and is used to enforce constraint (8b), but its value is never specified in Table II or anywhere in the text. Since the penalty directly affects the tradeoff between synchronization and data rate, the empirical results are not reproducible without this value. Please report ρ and, if the penalty is meant to be a soft constraint, discuss how close the learned policies are to satisfying (8b).
[V-B, Figs. 3–6] The reported improvements (28.96% and 31.79%) appear to come from a single simulation run without error bars or multiple random seeds. Given the stochastic nature of user mobility and neural network training, the significance of the improvement is unclear. Please report mean and standard deviation over several independent runs, and state the number of runs used.

minor comments (5)

[I, paragraph after contributions] There is a typo 'the the similarity' in the sentence beginning 'Simulation results show that our proposed method improves...'.
[III-A3, after Eq. (13)] 'By Given the hidden state ht' should read 'Given the hidden state ht'.
[III-D, Algorithm 2] 'reply memory' should be 'replay memory'.
[V-A, baseline description] The baseline IQL uses a local reward r_m(s_m^t, a_m^t) in Eq. (37), but it is not specified how this local reward is computed from the global reward (16). Please clarify whether the baseline uses the same team reward or a locally decomposed version, since this affects the fairness of the comparison.
[IV-A, complexity of Hungarian algorithm] The complexity O(Σ_m |U_m|^2 N) for the Hungarian algorithm assumes |U_m| ≤ N; if |U_m| > N, the complexity would depend on the rectangular assignment problem. This condition is not stated.

Circularity Check

1 steps flagged · score 4.0 of 10

The 28.96% empirical comparison is non-circular, but Lemma 1's convergence proof is circular: it requires ε→0, which is exactly the additive-decomposition assumption of Eq. (19) restated.

self definitional [Section IV-B, Lemma 1, Eq. (36)]
"From (36), we see that when ε → 0 or γ → 1, the VDN Bellman operator HV satisfies (30). Based on the Banach fixed-point theorem, the VDN in our proposed method will converge to Q∗ tot."

ε is defined in Eq. (27) as Qπ(s,a) − Qπ_tot(s,a), i.e., exactly the error of the additive decomposition introduced in Eq. (19). Requiring ε→0 is thus equivalent to assuming that the global Q-function is exactly the sum of local Q-functions, which is the very property VDN is supposed to justify. The alternative γ→1 is not available in the simulated regime (Table II sets γ=0.2) and would make the contraction coefficient in (30) equal to 1, so the Banach fixed-point theorem would not apply. Hence the only operative condition is the additivity assumption restated, and no independent convergence guarantee or suboptimality bound is provided for nonzero ε.

full rationale

The paper's headline quantitative result—up to 28.96% improvement over GRU+IQL—is an empirical simulation comparison against a baseline, not a quantity derived from the model's assumptions or from a fitted parameter; it is therefore not circular. The GRU training procedure (Eqs. (14)–(15)) and the Hungarian-based RB allocation are also self-contained algorithmic components. The only genuine circular step is Lemma 1 in Section IV-B. The proof defines the decomposition error ε in Eq. (27) as the difference between the true global Q-function and the additive Q_tot used by VDN, then concludes convergence when ε→0. But ε→0 is precisely the additive-decomposition assumption of Eq. (19) restated; the theorem therefore does not provide an independent convergence guarantee for the simulated setting, where γ=0.2 and no bound on ε is supplied. This circularity does not undermine the empirical comparison vs. IQL, but it leaves the theoretical support for VDN's optimality conditional on the very property at issue. Several citations are to the authors' own prior work ([22], [28], [30], [32], [33]), but they are used for standard complexity, Bellman-operator, and parameter conventions, not as load-bearing evidence for the main claim.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central claim rests on the VDN additivity assumption, the GRU's ability to predict user mobility from K past states, a simplified synchronization model where transmission implies perfect DNT accuracy, and an unspecified soft penalty for multi-BS association. The numeric parameters epsilon, rho, alpha, and RL hyperparameters are chosen by hand. No new physical entities are introduced.

free parameters (8)

Tradeoff weight epsilon = 0.25, 0.3, 0.8, 0.99 in simulations
Weights data rate against DNT synchronization error in objective (8) and reward (16); chosen by hand, not derived.
Multi-BS penalty rho = Not specified, must be negative
Penalty in reward (16) for a user being associated with multiple BSs; acts as a soft version of constraint (8b).
Delay threshold alpha = Not specified
Constraint (8g) requires transmission delay T_mt <= alpha, but the algorithm does not explain how alpha is enforced; value never given.
Discount factor gamma = 0.2
Standard RL hyperparameter in Table II; low value, chosen without sensitivity analysis.
GRU input length K = 5
Number of past DNT states fed to the GRU (Table II); chosen by hand.
Learning rates lambda_G, lambda_Q = 1e-3, 1e-4
Learning rates for GRU and Q networks (Table II).
Hidden sizes N_h, theta_h = 128
Hidden layer sizes for GRU and Q networks (Table II).
Training epochs G = 75
Number of training epochs for VDN (Table II and Algorithm 2).

assumptions (7)

domain assumption The global Q function is additively decomposable into per-BS local Q functions (Eq. 19)
Section III.D assumes Qtot is the sum of Qtheta_m; the convergence lemma further requires the decomposition error to vanish. This is not validated for the coupled interference and synchronization reward.
ad hoc to paper Constraint (8b) (one BS per user) is adequately enforced by penalty rho in reward (16)
The decomposed subproblem (17) omits the global constraint (8b); the algorithm relies on a negative reward penalty to discourage violations, which can be violated during training.
domain assumption The GRU predictor trained on random-walk trajectories generalizes to the online DNT prediction task
Section III claims no reliance on user mobility models, yet Section V.A trains the GRU on 2000 trajectories generated from the random walk model (1); long-horizon prediction error when units are not transmitted is not analyzed.
domain assumption The DNT is exactly synchronized when a BS transmits, with no other source of error (Eq. 7)
Section II.C defines DNT status as equal to physical status if sum y_mt,n=1; transmission errors, quantization, and cloud processing delay are ignored.
domain assumption Rayleigh fading with path loss exponent 2 and known interference model
Transmission model in Section II.B uses channel gain h=o_u d^-2 and interference equations (3),(5).
domain assumption Random walk mobility model with fixed transition probabilities p_u
Section II.A equation (1); probabilities p_u are not specified in Table II.
ad hoc to paper Convergence proof may assume the VDN decomposition error vanishes (epsilon -> 0)
Lemma 1 states the method converges if epsilon -> 0 or gamma -> 1; letting the approximation error go to zero assumes the additivity is exact, which is the property to be proven.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimizing Wireless Resource Management and Synchronization in Digital Twin Networks." pith.science (2026). https://pith.science/paper/J757AB4V

@misc{pith2026250205116,
  author       = {Pith},
  title        = {Pith review of: Optimizing Wireless Resource Management and Synchronization in Digital Twin Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J757AB4V}},
  note         = {Machine review of arXiv:2502.05116}
}

read the original abstract

In this paper, we investigate an accurate synchronization between a physical network and its digital network twin (DNT), which serves as a virtual representation of the physical network. The considered network includes a set of base stations (BSs) that must allocate its limited spectrum resources to serve a set of users while also transmitting its partially observed physical network information to a cloud server to generate the DNT. Since the DNT can predict the physical network status based on its historical status, the BSs may not need to send their physical network information at each time slot, allowing them to conserve spectrum resources to serve the users. However, if the DNT does not receive the physical network information of the BSs over a large time period, the DNT's accuracy in representing the physical network may degrade. To this end, each BS must decide when to send the physical network information to the cloud server to update the DNT, while also determining the spectrum resource allocation policy for both DNT synchronization and serving the users. We formulate this resource allocation task as an optimization problem, aiming to maximize the total data rate of all users while minimizing the asynchronization between the physical network and the DNT. To address this problem, we propose a method based on the GRUs and the value decomposition network (VDN). Simulation results show that our GRU and VDN based algorithm improves the weighted sum of data rates and the similarity between the status of the DNT and the physical network by up to 28.96%, compared to a baseline method combining GRU with the independent Q learning.

Figures

Figures reproduced from arXiv: 2502.05116 by the authors.

**Figure 1.** The considered DNT enabled network. server, and RB allocation, optimizing the data rate of all users in the physical network and ensuring an accurate synchronization between the physical network and the DNT. Compared to other RL methods [19], the VD-RL method allows each BS to use its partial observation, specifically, the locations of the users in its coverage, to collaboratively find a globally optimal solution fo… view at source ↗

**Figure 2.** The prediction of the user movement trajectories via the GRU model. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Convergence of VDN and IQL (ϵ = 0.3, N = 12). B. Simulation Results and Analysis [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: The average DNT error and the data rate as the number of users varies [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: The number of BSs updating the DNT and the DNT errors at each [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

[12]

Edge Continual Learning for Dynamic Digital Twins over Wireless Networks

O. Hashash, C. Chaccour, and W. Saad, “Edge continual learning for dynamic digital twins over wireless networks,” arXiv preprint arXiv:2204.04795, April 2022

work page Pith review arXiv 2022
[15]

A dynamic hierarchical framework for IoT-assisted digital twin synchronization in the metaverse,

Y . Han, D. Niyato, C. Leung, D. I. Kim, K. Zhu, S. Feng, X. Shen, and C. Miao, “A dynamic hierarchical framework for IoT-assisted digital twin synchronization in the metaverse,” IEEE Internet of Things Journal , vol. 10, no. 1, pp. 268–284, January 2023

work page 2023
[1]

Digital twin networks: A survey,

Y . Wu, K. Zhang, and Y . Zhang, “Digital twin networks: A survey,”IEEE Internet of Things Journal , vol. 8, no. 18, pp. 13789–13804, September 2021

work page 2021
[2]

Digital twin for networking: A data-driven performance modeling perspective,

L. Hui, M. Wang, L. Zhang, L. Lu, and Y . Cui, “Digital twin for networking: A data-driven performance modeling perspective,” IEEE Network, vol. 37, no. 3, pp. 202–209, May/June 2023

work page 2023
[3]

Digital network twins for next-generation wireless: Creation, optimization, and challenges,

Y . Liu, Z. Peng, Z. Zhang, H. Yu, and M. Chen, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” arXiv preprint arXiv:2410.18002 , October 2024

arXiv 2024
[4]

Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,

L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong, “Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,” IEEE Communications Surveys & Tutorials , vol. 24, no. 4, pp. 2230–2254, August 2022

work page 2022
[5]

Five disruptive technologies in 6G to support digital twin networks,

Q. Guo, F. Tang, T. K. Rodrigues, and N. Kato, “Five disruptive technologies in 6G to support digital twin networks,” IEEE Wireless Communications, vol. 31, no. 1, pp. 149–155, February 2024

work page 2024
[6]

Mapping Wireless Networks into Digital Reality through Joint Vertical and Horizontal Learning

Z. Zhang, M. Chen, Z. Yang, and Y . Liu, “Mapping wireless networks into digital reality through joint vertical and horizontal learning,” arXiv preprint arXiv:2404.14497, April 2024

work page Pith review arXiv 2024

Show all 33 references

[7]

Digital twins: State of the art theory and practice, challenges, and open research questions,

A. Sharma, E. Kosasih, J. Zhang, A. Brintrup, and A. Calinescu, “Digital twins: State of the art theory and practice, challenges, and open research questions,” Journal of Industrial Information Integration , vol. 30, pp. 100383, September 2022

2022
[8]

6G digital twin networks: From theory to practice,

X. Lin, L. Kundu, C. Dick, E. Obiodu, T. Mostak, and M. Flaxman, “6G digital twin networks: From theory to practice,” IEEE Communications Magazine, vol. 61, no. 11, pp. 72–78, June 2023

2023
[9]

Deep learning for hybrid 5G services in mobile edge computing systems: Learn from a digital twin,

R. Dong, C. She, W. Hardjawana, Y . Li, and B. Vucetic, “Deep learning for hybrid 5G services in mobile edge computing systems: Learn from a digital twin,” IEEE Transactions on Wireless Communications , vol. 18, no. 10, pp. 4692–4707, July 2019

2019
[10]

A Bayesian framework for digital twin-based control, monitoring, and data collection in wireless systems,

C. Ruah, O. Simeone, and B. M. Al-Hashimi, “A Bayesian framework for digital twin-based control, monitoring, and data collection in wireless systems,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 10, pp. 3146–3160, August 2023

2023
[11]

Digital twin- empowered network planning for multi-tier computing,

C. Zhou, J. Gao, M. Li, X. S. Shen, and W. Zhuang, “Digital twin- empowered network planning for multi-tier computing,” Journal of Communications and Information Networks , vol. 7, no. 3, pp. 221–238, September 2022

2022
[13]

Resource allocation for aerial assisted digital twin edge mobile network,

Q. Guo, F. Tang, and N. Kato, “Resource allocation for aerial assisted digital twin edge mobile network,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 10, pp. 3070–3079, October 2023

2023
[14]

Data synchronization in vehicular digital twin network: A game theoretic approach,

J. Zheng, T. H. Luan, Y . Zhang, R. Li, Y . Hui, L. Gao, and M. Dong, “Data synchronization in vehicular digital twin network: A game theoretic approach,” IEEE Transactions on Wireless Communications , vol. 22, no. 11, pp. 7635–7647, November 2023

2023
[16]

Learning phrase representations using RNN encoder-decoder for statistical machine translation,

K. Cho, “Learning phrase representations using RNN encoder-decoder for statistical machine translation,” arXiv preprint arXiv:1406.1078 , June 2014

2014 arXiv
[17]

Empirical evaluation of gated recurrent neural networks on sequence modeling,

J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, December 2014

2014 arXiv
[18]

Value-decomposition networks for cooperative multi-agent learning,

P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V . Zambaldi, M. Jaderberg, M. Lanctot, N. Sonnerat, J. Z. Leibo, K. Tuyls, et al., “Value-decomposition networks for cooperative multi-agent learning,” arXiv preprint arXiv:1706.05296 , June 2017

2017 arXiv
[19]

Multi-agent reinforcement learning: Independent vs. cooperative agents,

M. Tan, “Multi-agent reinforcement learning: Independent vs. cooperative agents,” in Proc. International Conference on Machine Learning , Amherst, MA, USA, June 1993, pp. 330–337

1993
[20]

Fundamentals of mobility- aware performance characterization of cellular networks: A tutorial,

H. Tabassum, M. Salehi, and E. Hossain, “Fundamentals of mobility- aware performance characterization of cellular networks: A tutorial,” IEEE Communications Surveys & Tutorials , vol. 21, no. 3, pp. 2288– 2308, March 2019

2019
[21]

A survey of algorithmic methods for partially observed Markov decision processes,

William S Lovejoy, “A survey of algorithmic methods for partially observed Markov decision processes,” Annals of Operations Research , vol. 28, no. 1, pp. 47–65, 1991

1991
[22]

A joint learning and communications framework for federated learning over wireless networks,

M. Chen, Z. Yang, W. Saad, C. Yin, H. V . Poor, and S. Cui, “A joint learning and communications framework for federated learning over wireless networks,” IEEE Transactions on Wireless Communications , vol. 20, no. 1, pp. 269–283, January 2021

2021
[23]

Distributed and distribution-robust meta reinforcement learning (D 2-RMRL) for data pre-storage and routing in cube satellite networks,

Y . Hu, X. Wang, and W. Saad, “Distributed and distribution-robust meta reinforcement learning (D 2-RMRL) for data pre-storage and routing in cube satellite networks,” IEEE Journal of Selected Topics in Signal Processing, vol. 17, no. 1, pp. 128–141, January 2023

2023
[24]

Improving the Hungarian assignment algorithm,

R. Jonker and T. V olgenant, “Improving the Hungarian assignment algorithm,” Operations Research Letters , vol. 5, no. 4, pp. 171–175, October 1986

1986
[25]

Human-level control through deep reinforcement learning,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, February 2015

2015
[26]

Lin, Reinforcement learning for robots using neural networks, Carnegie Mellon University, 1992

L. Lin, Reinforcement learning for robots using neural networks, Carnegie Mellon University, 1992

1992
[27]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. Conference on Neural Information Processing Systems , Long Beach, CA, USA, December 2017, vol. 30

2017
[28]

Deep learning for optimal deployment of UA Vs with visible light communications,

Y . Wang, M. Chen, Z. Yang, T. Luo, and W. Saad, “Deep learning for optimal deployment of UA Vs with visible light communications,” IEEE Transactions on Wireless Communications , vol. 19, no. 11, pp. 7049–7063, November 2020

2020
[29]

A theoretical analysis of deep Q-learning,

J. Fan, Z. Wang, Y . Xie, and Z. Yang, “A theoretical analysis of deep Q-learning,” in Proc. Learning for Dynamics and Control , Feburary 2020, pp. 486–489

2020
[30]

Distributed reinforcement learning for age of information minimization in real-time IoT systems,

S. Wang, M. Chen, Z. Yang, C. Yin, W. Saad, S. Cui, and H. V . Poor, “Distributed reinforcement learning for age of information minimization in real-time IoT systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 3, pp. 501–515, April 2022

2022
[31]

Stochastic approximation for nonexpansive maps: Application to Q-learning algorithms,

J. Abounadi, D. P. Bertsekas, and V . Borkar, “Stochastic approximation for nonexpansive maps: Application to Q-learning algorithms,” SIAM Journal on Control and Optimization , vol. 41, no. 1, pp. 1–22, March 2002

2002
[32]

Optimization of image transmission in cooperative semantic communication networks,

W. Zhang, Y . Wang, M. Chen, T. Luo, and D. Niyato, “Optimization of image transmission in cooperative semantic communication networks,” IEEE Transactions on Wireless Communications , vol. 23, no. 2, pp. 861–873, 2023

2023
[33]

Rethinking the implementation tricks and monotonicity constraint in cooperative multi- agent reinforcement learning,

J. Hu, S. Jiang, S. A. Harding, H. Wu, and S. Liao, “Rethinking the implementation tricks and monotonicity constraint in cooperative multi- agent reinforcement learning,” arXiv preprint arXiv:2102.03479, Feburary 2021

2021 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [12]

Edge Continual Learning for Dynamic Digital Twins over Wireless Networks

O. Hashash, C. Chaccour, and W. Saad, “Edge continual learning for dynamic digital twins over wireless networks,” arXiv preprint arXiv:2204.04795, April 2022

work page Pith review arXiv 2022

[2] [15]

A dynamic hierarchical framework for IoT-assisted digital twin synchronization in the metaverse,

Y . Han, D. Niyato, C. Leung, D. I. Kim, K. Zhu, S. Feng, X. Shen, and C. Miao, “A dynamic hierarchical framework for IoT-assisted digital twin synchronization in the metaverse,” IEEE Internet of Things Journal , vol. 10, no. 1, pp. 268–284, January 2023

work page 2023

[3] [1]

Digital twin networks: A survey,

Y . Wu, K. Zhang, and Y . Zhang, “Digital twin networks: A survey,”IEEE Internet of Things Journal , vol. 8, no. 18, pp. 13789–13804, September 2021

work page 2021

[4] [2]

Digital twin for networking: A data-driven performance modeling perspective,

L. Hui, M. Wang, L. Zhang, L. Lu, and Y . Cui, “Digital twin for networking: A data-driven performance modeling perspective,” IEEE Network, vol. 37, no. 3, pp. 202–209, May/June 2023

work page 2023

[5] [3]

Digital network twins for next-generation wireless: Creation, optimization, and challenges,

Y . Liu, Z. Peng, Z. Zhang, H. Yu, and M. Chen, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” arXiv preprint arXiv:2410.18002 , October 2024

arXiv 2024

[6] [4]

Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,

L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong, “Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,” IEEE Communications Surveys & Tutorials , vol. 24, no. 4, pp. 2230–2254, August 2022

work page 2022

[7] [5]

Five disruptive technologies in 6G to support digital twin networks,

Q. Guo, F. Tang, T. K. Rodrigues, and N. Kato, “Five disruptive technologies in 6G to support digital twin networks,” IEEE Wireless Communications, vol. 31, no. 1, pp. 149–155, February 2024

work page 2024

[8] [6]

Mapping Wireless Networks into Digital Reality through Joint Vertical and Horizontal Learning

Z. Zhang, M. Chen, Z. Yang, and Y . Liu, “Mapping wireless networks into digital reality through joint vertical and horizontal learning,” arXiv preprint arXiv:2404.14497, April 2024

work page Pith review arXiv 2024