pith. sign in

arxiv: 1906.10194 · v1 · pith:CD6ZLO2Xnew · submitted 2019-06-24 · 📡 eess.SP · cs.IT· math.IT

Deep Neural Network Based Resource Allocation for V2X Communications

Jin Gao , Muhammad R. A. Khandaker , Faisal Tariq , Kai-Kit Wong , Risala T. Khan This is my paper

Pith reviewed 2026-05-25 16:51 UTC · model grok-4.3

classification 📡 eess.SP cs.ITmath.IT
keywords V2XDNNWMMSEpower allocationresource allocationsupervised learningthroughput
0
0 comments X

The pith

A deep neural network trained on WMMSE outputs approximates the optimal transmit power allocation in V2X communications while greatly reducing computational overhead.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper focuses on solving the power allocation problem to maximize throughput in V2X systems. It presents the WMMSE algorithm using block coordinate descent to generate solutions. These solutions then serve as training targets for a DNN using supervised learning with mini-batch gradient descent. The central result is that the trained DNN approximates the WMMSE performance with significantly lower computation.

Core claim

The DNN algorithm can provide very good approximation of the iterative WMMSE algorithm reducing the computational overhead significantly.

What carries the argument

Deep neural network using supervised learning to mimic WMMSE-based power allocations.

If this is right

  • V2X systems can perform power allocation in real time.
  • Throughput is maintained close to the iterative optimum.
  • The training process is made efficient by mini-batch gradient descent.

Where Pith is reading between the lines

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

  • This approximation technique could be applied to other optimization tasks in communications.
  • It suggests potential for hardware implementation with lower power consumption.
  • Further work might explore the limits of generalization to unseen channel conditions.

Load-bearing premise

The solutions produced by the WMMSE algorithm are reliable near-optimal targets for training the DNN and the simulation scenarios are representative of real-world V2X channel conditions.

What would settle it

If measurements show that the DNN's throughput is substantially below that of WMMSE under realistic V2X conditions not covered in the simulations, the approximation claim would not hold.

Figures

Figures reproduced from arXiv: 1906.10194 by Faisal Tariq, Jin Gao, Kai-Kit Wong, Muhammad R. A. Khandaker, Risala T. Khan.

Figure 1
Figure 1. Figure 1: The proposed V2V and V2I communication system. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed deep neural network for approximating t [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of learning rate on training MSE. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of batch size on training epoch. The V2X network consists of one cellular base station, M = 8 CUs and K = 10 V2V transmitters. We construct a fully connected DNN for the system with one input layer, three hidden layers and one output layer. The input layer consists of N +K +N ×K neurons, the three hidden layers consist of 50, 22, 20 neurons, respectively, and the output layer has N + K neurons to pr… view at source ↗
Figure 7
Figure 7. Figure 7: The CDF of the achievable sum rates for the proposed DN [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The probability density function (PDF) of the propos [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
read the original abstract

This paper focuses on optimal transmit power allocation to maximize the overall system throughput in a vehicle-to-everything (V2X) communication system. We propose two methods for solving the power allocation problem namely the weighted minimum mean square error (WMMSE) algorithm and the deep learning-based method. In the WMMSE algorithm, we solve the problem using block coordinate descent (BCD) method. Then we adopt supervised learning technique for the deep neural network (DNN) based approach considering the power allocation from the WMMSE algorithm as the target output. We exploit an efficient implementation of the mini-batch gradient descent algorithm for training the DNN. Extensive simulation results demonstrate that the DNN algorithm can provide very good approximation of the iterative WMMSE algorithm reducing the computational overhead significantly.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes solving the non-convex sum-rate maximization problem for transmit power allocation in a V2X system via a block coordinate descent implementation of the weighted minimum mean square error (WMMSE) algorithm, then trains a DNN in a supervised manner to map channel realizations to the WMMSE power allocations. The central claim is that the resulting DNN supplies a close approximation to the iterative WMMSE solutions while substantially lowering computational cost, as supported by simulation results.

Significance. If the reported approximation quality holds with low error and the WMMSE targets are of high quality, the work demonstrates a practical route to real-time resource allocation in latency-sensitive V2X scenarios. The supervised-learning mimicry of an established iterative solver is a standard, reproducible technique that directly addresses the computational burden of BCD iterations.

major comments (3)
  1. [Abstract and §5] Abstract and §5: the claim that 'extensive simulation results demonstrate that the DNN algorithm can provide very good approximation' is unsupported by any quantitative error metrics (e.g., MSE on power vectors, average sum-rate gap, or CDF of throughput difference between DNN and WMMSE); only qualitative statements appear, leaving the central performance claim only moderately supported.
  2. [§3] §3: the WMMSE/BCD procedure is presented as the benchmark for optimal power allocation, yet the underlying problem is non-convex; no optimality-gap analysis, comparison against a global solver on small instances, or discussion of stationary-point quality is provided, so the practical value of the DNN approximation remains unclear even if empirical match to WMMSE is later shown.
  3. [§4] §4: the DNN architecture (number of hidden layers, neurons per layer, activation functions), input feature dimension, training-set size, mini-batch size, and convergence criteria are not specified, preventing assessment or reproduction of the supervised-learning results.
minor comments (2)
  1. [§5] The manuscript would benefit from an explicit statement of the system model parameters (number of vehicles, V2X channel model, noise variance, power constraints) in the simulation section to allow direct comparison with related V2X resource-allocation studies.
  2. Notation for the power vector p and the WMMSE weights should be introduced once and used consistently across the algorithm description and the DNN input/output definitions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and indicate the changes planned for the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract and §5] Abstract and §5: the claim that 'extensive simulation results demonstrate that the DNN algorithm can provide very good approximation' is unsupported by any quantitative error metrics (e.g., MSE on power vectors, average sum-rate gap, or CDF of throughput difference between DNN and WMMSE); only qualitative statements appear, leaving the central performance claim only moderately supported.

    Authors: We agree that quantitative metrics are needed to support the approximation claim. The revised manuscript will add tables and figures reporting MSE between DNN and WMMSE power vectors, average sum-rate gap, and CDFs of throughput differences. revision: yes

  2. Referee: [§3] §3: the WMMSE/BCD procedure is presented as the benchmark for optimal power allocation, yet the underlying problem is non-convex; no optimality-gap analysis, comparison against a global solver on small instances, or discussion of stationary-point quality is provided, so the practical value of the DNN approximation remains unclear even if empirical match to WMMSE is later shown.

    Authors: We will add explicit discussion in Section 3 noting that the problem is non-convex and that WMMSE yields stationary points (a standard benchmark in the literature). A full optimality-gap analysis or global-solver comparisons on large instances is beyond the current scope and would require prohibitive computation; we therefore treat this as a limitation rather than a core contribution. revision: partial

  3. Referee: [§4] §4: the DNN architecture (number of hidden layers, neurons per layer, activation functions), input feature dimension, training-set size, mini-batch size, and convergence criteria are not specified, preventing assessment or reproduction of the supervised-learning results.

    Authors: We will include the omitted implementation details (layer counts and widths, activations, input dimension, training-set size, mini-batch size, and convergence criteria) in the revised Section 4 to ensure reproducibility. revision: yes

Circularity Check

0 steps flagged

No circularity: DNN trained via standard supervised learning on independent WMMSE targets

full rationale

The paper generates power allocation targets using the established BCD-based WMMSE algorithm and trains the DNN via supervised learning to approximate those targets. This is explicit mimicry of an external iterative solver rather than any self-definitional loop, fitted parameter renamed as prediction, or self-citation chain. The reported approximation quality and complexity reduction are measured against the same independent WMMSE benchmark used for training, which is the intended and non-circular use of supervised learning. No load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper rests on the domain assumption that the WMMSE iterative solutions are suitable ground-truth targets and that simulated channels capture the relevant statistics; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption The WMMSE algorithm with block coordinate descent yields suitable target power allocations for supervised training.
    Explicitly used as the source of training labels for the DNN.

pith-pipeline@v0.9.0 · 5673 in / 1193 out tokens · 32193 ms · 2026-05-25T16:51:34.834537+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning the Wireless V2I Channels Using Deep Neural Networks

    eess.SP 2019-07 unverdicted novelty 3.0

    A deep neural network is trained on prior channel responses and pilots to predict future V2I channel states for improved system performance.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Connected roads of the future: Use cases, requirements, an d designcon- siderations for vehicle-to-everything communications,

    M. Boban, A. Kousaridas, K. Manolakis, J. Eichinger, and W. Xu, “Connected roads of the future: Use cases, requirements, an d designcon- siderations for vehicle-to-everything communications,” IEEE V eh.Tech. Magazine, vol. 13, pp. 110–123, Sep. 2018

    work page 2018

  2. [2]

    A speculative study on 6G,

    F. Tariq, M. R. A. Khandaker, K.-K. Wong, M. Imran, M. Bennis, and M. rouane Debbah, “A speculative study on 6G,” IEEE Commun. Magazine , June 2019 (submitted). Available: https://arxiv.org/pdf/1902.06700.pdf

    work page arXiv 2019

  3. [3]

    What will 5G be?

    J. G. A. et al., “What will 5G be?” IEEE J. Sel. Areas Commun. , vol. 32, pp. 1065–1082, June 2014

    work page 2014

  4. [4]

    Radio re- source management for D2D-based V2V communication,

    W. Sun, E. G. Str¨ om, F. Br¨ annstr¨ om, K. C. Sou, and Y . Sui, “Radio re- source management for D2D-based V2V communication,” IEEE Trans. V eh. Commun., vol. 65, pp. 6636–6650, Aug. 2016

    work page 2016

  5. [5]

    Resource allocation for D2D -enabled vehicular communications,

    L. Liang, G. Y . Li, and W. Xu, “Resource allocation for D2D -enabled vehicular communications,” IEEE Trans. Commun. , vol. 65, pp. 3186– 3197, July 2017

    work page 2017

  6. [6]

    Machine learning paradigms for next-generation wireless networks,

    C. Jiang, H. Zhang, Y . Ren, Z. Han, K.-C. Chen, , and L. Hanz o, “Machine learning paradigms for next-generation wireless networks,” IEEE Wireless Commun. , vol. 24, pp. 98–105, 2016

    work page 2016

  7. [7]

    Deep learning in mob ile and- wireless networking: A survey,

    C. Zhang, P . Patras, and H. Haddadi, “Deep learning in mob ile and- wireless networking: A survey,” IEEE Commun. Surveys & Tutorials , 2019

    work page 2019

  8. [8]

    Deep reinforcement learning for resou rce allocation in V2V communications,

    H. Y e and G. Y . Li, “Deep reinforcement learning for resou rce allocation in V2V communications,” in Proc. IEEE Int. Conf. Commun. (ICC) , Kansas City, MO, 2018

    work page 2018

  9. [9]

    Deep reinforcement learn ing based resource allocation for V2V communications,

    H. Y e, G. Y . Li, and B. F. Juang, “Deep reinforcement learn ing based resource allocation for V2V communications,” IEEE Trans. V eh. Technology, vol. 68, no. 3163-3173, Apr. 2019

    work page 2019

  10. [10]

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

    Q. Shi, M. Razaviyayn, Z. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,” IEEE Trans. Signal Process. , vol. 59, pp. 4331–4340, Sep. 2011

    work page 2011

  11. [11]

    D. P . Palomar and Y . Jiang, MIMO Transceiver Design via Majorization Theory. now Publishers, 2007

    work page 2007

  12. [12]

    S. M. Kay, Fundamentals of Statistical Signal Processing: Estimatio n Theory. Englewood Cilffs, NJ: Prentice Hall, 1993

    work page 1993

  13. [13]

    Joint transceiver optim ization for multiuser MIMO relay communication systems,

    M. R. A. Khandaker and Y . Rong, “Joint transceiver optim ization for multiuser MIMO relay communication systems,” IEEE Trans. Signal Process., vol. 60, pp. 5977–5986, Nov. 2012

    work page 2012

  14. [14]

    Boyd and L

    S. Boyd and L. V andenberghe, Convex Optimization. Cambridge, U. K.: Cambridge University Press, 2004

    work page 2004

  15. [15]

    Joint source and relay op timization for multiuser MIMO relay communication systems,

    M. R. A. Khandaker and Y . Rong, “Joint source and relay op timization for multiuser MIMO relay communication systems,” in Proc. 4th Int. Conf. Signal Process. Commun. Systems (ICSPCS’2010) , Gold Coast, Australia, Dec. 13-15, 2010

    work page 2010

  16. [16]

    Lecture 6a ov erview of mini-batch gradient descent,

    G. Hinton, N. Srivastava, and K. Swersky, “Lecture 6a ov erview of mini-batch gradient descent,” Coursera Lecture Slides, 20 12. [Online]. Available: https://class.coursera.org/neuralnets-2012-001/lecture

    work page 2012

  17. [17]

    Adaptive subgradien t methods for online learningand stochastic optimization,

    J. Duchi, E. Hazan, and Y . Singer, “Adaptive subgradien t methods for online learningand stochastic optimization,” J. Machine Learning Research, vol. 12, pp. 2121–2159, Jan. 2011

    work page 2011

  18. [18]

    An overview of gradient descent optimization algorithms

    S. Ruder, “An overview of gradient descent optimizatio n algorithms,” arXiv preprint. Available: arXiv:1609.04747 , June 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  19. [19]

    Learning to optimize: Training deep neural networks for wi reless resource management,

    H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropou los, “Learning to optimize: Training deep neural networks for wi reless resource management,” in Proc. IEEE 18th Int. W orkshop on Signal Process. Adv. Wireless Commun. (SPAWC) , Sapporo, 2017

    work page 2017