pith. sign in

arxiv: 1906.09981 · v1 · pith:Q6LLRB6Rnew · submitted 2019-06-21 · 📡 eess.SP · stat.ML

Optimal WDM Power Allocation via Deep Learning for Radio on Free Space Optics Systems

Zhan Gao , Mark Eisen , Alejandro Ribeiro This is my paper

Pith reviewed 2026-05-25 18:43 UTC · model grok-4.3

classification 📡 eess.SP stat.ML
keywords power allocationWDMRoFSOdeep learningprimal-dualfree space opticsradio over opticscapacity maximization
0
0 comments X

The pith

A primal-dual deep learning method learns power allocations for WDM radio-over-free-space-optics that work without any system model.

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

The paper studies power allocation across wavelengths in radio-on-free-space-optics links to maximize a weighted sum of transmission capacities while respecting a total power budget and eye-safety limits. It first supplies an exact model-based solver that uses stochastic dual gradients and exploits a null duality gap. It then supplies a model-free alternative that represents the allocation rule as a deep neural network and trains the network with a primal-dual procedure that needs only observed data. Simulations show both methods improve on uniform power splitting, and the deep-learning version requires no prior knowledge of the channel or noise statistics.

Core claim

The paper shows that the weighted capacity maximization problem for WDM RoFSO can be solved exactly by a stochastic dual gradient algorithm when the system model is known, and that the same objective can be optimized without any model by parametrizing the allocation map as a deep neural network and applying a primal-dual learning procedure that updates the network weights from observed performance.

What carries the argument

The Primal-Dual Deep Learning algorithm, which parametrizes the power allocation policy with a deep neural network and trains it directly from data under the stated constraints.

If this is right

  • Power allocations can be computed without an explicit model of the optical channel or noise.
  • The learned allocations respect both the total power limit and the eye-safety constraint while exceeding uniform allocation.
  • The model-based stochastic dual gradient method recovers the exact optimum whenever the system equations are available.
  • The same primal-dual training loop can be rerun whenever the link statistics change.

Where Pith is reading between the lines

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

  • The same training procedure could be used in any wireless system whose performance can be measured but whose analytic model is unavailable or too complex.
  • If the neural network generalizes across different weather or turbulence conditions, the method would support online adaptation without repeated model re-derivation.
  • The exact model-based solver supplies a practical benchmark for measuring how close any learned policy comes to optimality in simulation.

Load-bearing premise

The numerical simulations used for testing match the statistics of actual RoFSO channels and the training procedure reaches a policy whose performance is close to the true optimum.

What would settle it

Apply the trained neural network to a fully known analytical RoFSO model, compute the achieved weighted capacity, and compare it to the capacity obtained by the model-based stochastic dual gradient solver on the same model; a large gap would falsify the claim that the learned policy is near-optimal.

Figures

Figures reproduced from arXiv: 1906.09981 by Alejandro Ribeiro, Mark Eisen, Zhan Gao.

Figure 1
Figure 1. Figure 1: The objective function value (left) and the constraint [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The objective function value over learning iterations [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
read the original abstract

Radio on Free Space Optics (RoFSO), as a universal platform for heterogeneous wireless services, is able to transmit multiple radio frequency signals at high rates in free space optical networks. This paper investigates the optimal design of power allocation for Wavelength Division Multiplexing (WDM) transmission in RoFSO systems. The proposed problem is a weighted total capacity maximization problem with two constraints of total power limitation and eye safety concern. The model-based Stochastic Dual Gradient algorithm is presented first, which solves the problem exactly by exploiting the null duality gap. The model-free Primal-Dual Deep Learning algorithm is then developed to learn and optimize the power allocation policy with Deep Neural Network (DNN) parametrization, which can be utilized without any knowledge of system models. Numerical simulations are performed to exhibit significant performance of our algorithms compared to the average equal power allocation.

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

1 major / 0 minor

Summary. The paper claims to solve the weighted total capacity maximization problem for WDM power allocation in RoFSO systems subject to total power and eye-safety constraints. It first presents a model-based Stochastic Dual Gradient algorithm that exploits the null duality gap for exact solution, then introduces a model-free Primal-Dual Deep Learning algorithm that parametrizes the power allocation policy with a DNN and can be used without knowledge of the system model. Numerical simulations are said to demonstrate significant performance gains over equal power allocation.

Significance. If the model-free claim holds with training performed via black-box interactions only, the Primal-Dual Deep Learning approach would offer a practical advantage for RoFSO systems where channel models are unavailable or inaccurate. The explicit model-based benchmark using null duality gap is a strength that allows direct comparison of the learning method's performance.

major comments (1)
  1. [Abstract / Primal-Dual Deep Learning algorithm description] Abstract and the section describing the Primal-Dual Deep Learning algorithm: the central claim that the algorithm 'can be utilized without any knowledge of system models' is load-bearing for the paper's contribution, yet the training procedure is not shown to rely solely on black-box interactions; if the weighted capacity objective or gradients during DNN training are computed from the analytical RoFSO fading/attenuation expressions (as is typical for numerical results), the method is model-free only at inference time, weakening the stated distinction from the Stochastic Dual Gradient method.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful review and for highlighting an important point of clarification regarding the model-free claim in our work. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract / Primal-Dual Deep Learning algorithm description] Abstract and the section describing the Primal-Dual Deep Learning algorithm: the central claim that the algorithm 'can be utilized without any knowledge of system models' is load-bearing for the paper's contribution, yet the training procedure is not shown to rely solely on black-box interactions; if the weighted capacity objective or gradients during DNN training are computed from the analytical RoFSO fading/attenuation expressions (as is typical for numerical results), the method is model-free only at inference time, weakening the stated distinction from the Stochastic Dual Gradient method.

    Authors: We agree that the presentation requires clarification to avoid ambiguity. In the numerical results, the weighted capacity objective and associated gradients for DNN training are indeed computed using the analytical RoFSO channel expressions. Consequently, the algorithm is model-free at inference/deployment time, where the trained policy operates without explicit model knowledge, in contrast to the Stochastic Dual Gradient method that requires the model at every iteration. We will revise the abstract and the algorithm description section to explicitly distinguish between training (which can be performed via black-box system interactions in practice) and inference, and to qualify the model-free claim accordingly. This revision will strengthen rather than weaken the contribution by making the distinction precise. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation chain is self-contained

full rationale

The paper first presents a model-based Stochastic Dual Gradient method that solves the weighted capacity maximization exactly by exploiting the null duality gap property. It then introduces a separate Primal-Dual Deep Learning algorithm that parametrizes the policy with a DNN and is explicitly described as usable without knowledge of system models. No equations or steps in the provided text reduce a claimed prediction or result to a fitted parameter or self-citation by construction; the two algorithms are presented as distinct, with the learning method positioned as model-free at the algorithmic level. This matches the default expectation that most papers are not circular, yielding a score of 0 with no load-bearing reductions identified.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated beyond standard optimization duality and DNN training assumptions.

axioms (1)
  • domain assumption Null duality gap holds for the weighted capacity maximization problem
    Invoked to claim the model-based algorithm solves the problem exactly.

pith-pipeline@v0.9.0 · 5674 in / 1098 out tokens · 19921 ms · 2026-05-25T18:43:07.529288+00:00 · methodology

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

22 extracted references · 22 canonical work pages

  1. [1]

    Radio over fiber link design for next generation wireless systems,

    D. Wake, A. Nkansah, and N. J. Gomes, “Radio over fiber link design for next generation wireless systems,” Journal of Lightwave Technology, vol. 28, no. 16, pp. 2456–2464, 2010

    work page 2010

  2. [2]

    Free-space optical communications,

    W. S. C Chang, “Free-space optical communications,” Journal of Lightwave Technology, vol. 24, no. 12, pp. 4750–4762, 2006

    work page 2006

  3. [3]

    L. C. Andrews and R. L. Phillips, Laser beam propagation through random media 2nd ed. , Bellingham : SPIE Press, 2005

    work page 2005

  4. [4]

    Rofso: A universal platform for convergence of fiber and free-space optical communication networks,

    K. Kazaura, K. Wakamori, M. Matsumoto, T. Higashino, K. Tsukamoto, and S. Komaki, “Rofso: A universal platform for convergence of fiber and free-space optical communication networks,” IEEE Communica- tions Magazine, vol. 48, no. 2, pp. 130–137, 2010

    work page 2010

  5. [5]

    Performance analysis of free-space optical communication systems over atmospheric turbulence channels,

    H. E. Nistazakis, T. A. Tsiftsis, and G. S. Tombras, “Performance analysis of free-space optical communication systems over atmospheric turbulence channels,” IET Communications , vol. 3, no. 8, pp. 1402– 1409, 2009

    work page 2009

  6. [6]

    Beam spread and wander of gaussian beam through anisotropic non-kolmogorov atmospheric turbulence for optical wireless communication,

    Z. Gao, J. Zhang, and A. Dang, “Beam spread and wander of gaussian beam through anisotropic non-kolmogorov atmospheric turbulence for optical wireless communication,” in IEEE International Conference on Communications (ICC) Workshops, 2017

    work page 2017

  7. [7]

    Beam wander effects on scintillation the- ory of gaussian beam through anisotropic non-kolmogorov atmospheric turbulence for optical wireless communication,

    Z. Gao, Z. Li, and A. Dang, “Beam wander effects on scintillation the- ory of gaussian beam through anisotropic non-kolmogorov atmospheric turbulence for optical wireless communication,” in IEEE International Conference on Communications (ICC) Workshops , 2018

    work page 2018

  8. [8]

    Performance evaluation of an advanced dwdm rofso system for transmitting multiple rf signals,

    A. Bekkali, P. T. Dat, K. Kazaura, K. Wakamori, M. Matsumoto, T. Higashino, K. Tsukamoto, and S. Komaki, “Performance evaluation of an advanced dwdm rofso system for transmitting multiple rf signals,” IEICE Transactions on Fundamentals of Electronics, vol. E92.A, no. 11, pp. 2697–2705, 2009

    work page 2009

  9. [9]

    1.28 terabit/s (32x40 gbit/s) wdm transmission system for free space optical communications,

    E. Ciaramella, Y . Arimoto, G. Contestabile, M. Presi, A. D’Errico, V . Guarino, and M. Matsumoto, “1.28 terabit/s (32x40 gbit/s) wdm transmission system for free space optical communications,” IEEE Journal on Selected Areas in Communications, vol. 27, no. 8, pp. 1639– 1645, 2009

    work page 2009

  10. [10]

    Novel adaptive transmission algorithms for free-space optical links,

    M. Karimi and M. Uysal, “Novel adaptive transmission algorithms for free-space optical links,” IEEE Transactions on Communications , vol. 27, no. 12, pp. 3808–3815, 2012

    work page 2012

  11. [11]

    Free-space optics optimization models for building sway and atmospheric interference using variable wavelength,

    X. Liu, “Free-space optics optimization models for building sway and atmospheric interference using variable wavelength,”IEEE Transactions on Communications, vol. 57, no. 2, pp. 492–498, 2009

    work page 2009

  12. [12]

    Wdm optical power allocation method for adaptive radio on free space optics system design,

    K. Kim, T. Higashino, K. Tsukamoto, and S. Komaki, “Wdm optical power allocation method for adaptive radio on free space optics system design,” in International Topical Meeting on Microwave Photonics jointly held with the 2011 Asia-Pacific Microwave Photonics Confer- ence, 2011

    work page 2011

  13. [13]

    Optical power allocation for adaptive transmissions in wavelength-division multiplexing free space optical networks,

    H. Zhou, S. Mao, and P. Agrawal, “Optical power allocation for adaptive transmissions in wavelength-division multiplexing free space optical networks,” Digital Communications and Networks , vol. 1, no. 3, pp. 171–180, 2015

    work page 2015

  14. [14]

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

    H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos, “Learning to optimize: Training deep neural networks for wireless resource management,” IEEE Transactions on Signal Processing , vol. 66, no. 20, pp. 5438–5453, 2018

    work page 2018

  15. [15]

    A deep reinforcement learning based framework for power-efficient resource allocation in cloud RANs,

    Z. Xu, Y . Wang, J. Tang, J. Wang, and M. C. Gursoy, “A deep reinforcement learning based framework for power-efficient resource allocation in cloud RANs,” in IEEE International Conference on Communications (ICC), 2017

    work page 2017

  16. [16]

    Learning optimal resource allocations in wireless systems,

    Mark Eisen, Clark Zhang, Luiz FO Chamon, Daniel D Lee, and Alejandro Ribeiro, “Learning optimal resource allocations in wireless systems,” IEEE Transactions on Signal Processing, vol. 67, no. 10, pp. 2775–2790, 2019

    work page 2019

  17. [17]

    Optimal resource allocation in wireless communication and networking,

    A. Ribeiro, “Optimal resource allocation in wireless communication and networking,” EURASIP Journal on Wireless Communications and Networking, vol. 1, pp. 1–19, 2012

    work page 2012

  18. [18]

    On relay selection and power allo- cation in cooperative free-space optical networks,

    H. Zhou, S. Mao, and P. Agrawal, “On relay selection and power allo- cation in cooperative free-space optical networks,” Photonic Network Communications, vol. 29, no. 1, pp. 1–11, 2015

    work page 2015

  19. [19]

    Error rate performance analysis of coded free-space optical links over gamma-gamma atmospheric turbulence channels,

    M. Uysal, J. Li, and M. Yu, “Error rate performance analysis of coded free-space optical links over gamma-gamma atmospheric turbulence channels,” IEEE Transactions on Wireless Communications, vol. 5, no. 6, pp. 1229–1233, 2006

    work page 2006

  20. [20]

    Studies on characterizing the transmission of rf signals over a turbulent fso link,

    P. T. Dat, A. Bekkali, K. Kazaura, K. Wakamori, T. Suzuki, M. Mat- sumoto, T. Higashino, K. Tsukamoto, and Komaki S., “Studies on characterizing the transmission of rf signals over a turbulent fso link,” Optics Express, vol. 17, no. 10, pp. 7731–7743, 2009

    work page 2009

  21. [21]

    Multilayer feedforward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural networks, vol. 2, no. 5, pp. 359–366, 1989

    work page 1989

  22. [22]

    Policy gradient methods for reinforcement learning with function approxima- tion,

    R. S. Sutton, D. A. McAllester, S. P. Singh, and Y . Mansour, “Policy gradient methods for reinforcement learning with function approxima- tion,” in Advances in Neural Information Processing Systems (NIPS) , 2000

    work page 2000