pith. sign in

arxiv: 1907.00036 · v1 · pith:STEX26LAnew · submitted 2019-06-26 · 📡 eess.SP

Novel Suboptimal approaches for Hyperparameter Tuning of Deep Neural Network [under the shelf of Optical Communication]

M. A. Amirabadi This is my paper

Pith reviewed 2026-05-25 15:46 UTC · model grok-4.3

classification 📡 eess.SP
keywords hyperparameter tuninggrid searchdeep neural networksoptical communicationfiber opticsfree space opticsmachine learning
0
0 comments X

The pith

Two new grid search variants allow practical hyperparameter tuning of deep neural networks for optical communication links.

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

The paper presents two suboptimal methods for searching hyperparameter grids in deep neural networks, one that scans dimensions marginally and another that alternates between them. These are applied to DNN models for both fiber-based and free-space optical communication systems. The methods are shown to cut computational cost substantially compared with exhaustive grid search while still delivering usable performance. The alternating approach is reported to outperform the marginal one on the tested tasks. The work also notes that hyperparameter tuning itself has received little attention in prior machine-learning studies of optical links.

Core claim

Two novel suboptimal grid search methods—marginal search and alternating search—are introduced for hyperparameter tuning of deep neural networks. When applied to DNN-based fiber optical communication and free-space optical communication systems, the methods reduce computation load while achieving favorable performance. The alternating search method is shown to deliver better performance than marginal grid search. The structures are presented as cost-effective and suitable for real-time applications, marking the first joint consideration of both optical link types in an ML-for-OC setting.

What carries the argument

Marginal and alternating suboptimal grid search methods that scan hyperparameter combinations one dimension at a time or by alternating dimensions instead of exhaustive enumeration.

If this is right

  • Hyperparameter tuning becomes feasible for DNNs in optical communication despite the high dimensionality of the search space.
  • The alternating search variant yields measurably better results than the marginal variant on the evaluated optical tasks.
  • Both methods are positioned as practical for real-time optical communication applications.
  • This constitutes the first reported joint examination of machine-learning models on both fiber and free-space optical links.

Where Pith is reading between the lines

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

  • The same search patterns could be tested on other high-dimensional tuning problems outside optical communication.
  • They might serve as lightweight baselines when comparing more sophisticated tuning algorithms such as Bayesian optimization.
  • Adoption could shorten the design cycle for new ML-based optical receivers by making repeated tuning runs cheaper.

Load-bearing premise

The reported performance gains and the advantage of alternating search over marginal search hold on the specific DNN tasks and quantitative metrics used for the fiber and free-space optical systems.

What would settle it

A side-by-side run on the same fiber and FSO DNN tasks showing that exhaustive grid search or standard alternatives produce clearly superior accuracy or BER at equal or lower total evaluations.

read the original abstract

Hyperparameter tuning is the main challenge of machine learning (ML) algorithms. Grid search is a popular method in hyperparameter tuning of simple ML algorithms; however, high computational complexity in complex ML algorithms such as Deep Neural Networks (DNN) is the main barrier towards its practical implementation. In this paper, two novel suboptimal grid search methods are presented, which search the grid marginally and alternating. In order to examine these methods, hyperparameter tuning is applied on two different DNN based Optical Communication (OC) systems (Fiber OC, and Free Space Optical (FSO) communication). The hyperparameter tuning of ML algorithms, despite its importance is ignored in ML for OC investigations. In addition, this is the first consideration of both FSO and Fiber OC systems in an ML for OC investigation. Results indicate that despite greatly reducing computation load, favorable performance could be achieved by the proposed methods. In addition, it is shown that the alternating search method has better performance than marginal grid search method. In sum, the proposed structures are cost-effective, and appropriate for real-time applications.

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

2 major / 2 minor

Summary. The manuscript proposes two novel suboptimal grid search methods for hyperparameter tuning of deep neural networks in optical communication applications: marginal grid search and alternating grid search. These are evaluated on DNN-based systems for Fiber OC and FSO communication, with the abstract asserting that they greatly reduce computational load while achieving favorable performance, and that the alternating method outperforms the marginal one. The work highlights the importance of hyperparameter tuning in ML for OC, which is often overlooked, and claims to be the first to consider both FSO and Fiber OC in such investigations.

Significance. If the performance claims are validated with quantitative evidence, the proposed methods could provide practical, computationally efficient alternatives to full grid search for tuning DNNs in real-time optical communication systems, potentially filling a gap in the ML-for-OC literature.

major comments (2)
  1. [Abstract] Abstract: The central claims that the proposed methods achieve 'favorable performance' despite greatly reducing computation load, and that alternating search has better performance than marginal grid search, are not accompanied by any quantitative metrics, error bars, baseline comparisons (e.g., to full grid search or random search), dataset details, or hyperparameter ranges. This absence makes it impossible to assess whether the methods deliver the asserted benefits on the Fiber OC and FSO tasks.
  2. [Results] Results section: No tables, figures, or numerical results (such as BER, accuracy, or computation time comparisons) are provided to support the assertions of reduced load with maintained performance or the superiority of alternating over marginal search, rendering the empirical validation of the core contribution unassessable.
minor comments (2)
  1. [Title] Title: The phrasing '[under the shelf of Optical Communication]' is unclear and appears to be a possible translation artifact or typo; revision for standard academic English (e.g., 'in the Context of Optical Communication') would improve readability.
  2. [Abstract] Abstract: Minor grammatical issues, such as the final sentence 'In sum, the proposed structures are cost-effective, and appropriate for real-time applications,' could be refined for precision and flow.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments highlighting the need for quantitative support. We agree that the current manuscript version does not provide sufficient numerical evidence, tables, or figures to substantiate the claims, and we will revise accordingly to include these elements for both the abstract and results section.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims that the proposed methods achieve 'favorable performance' despite greatly reducing computation load, and that alternating search has better performance than marginal grid search, are not accompanied by any quantitative metrics, error bars, baseline comparisons (e.g., to full grid search or random search), dataset details, or hyperparameter ranges. This absence makes it impossible to assess whether the methods deliver the asserted benefits on the Fiber OC and FSO tasks.

    Authors: We agree that the abstract lacks the required quantitative details. In the revised manuscript, the abstract will be updated to include specific metrics such as BER improvements, computation time reductions (e.g., percentage savings relative to full grid search), dataset sizes, hyperparameter ranges explored, and direct comparisons to full grid search and random search baselines, along with error bars where applicable. revision: yes

  2. Referee: [Results] Results section: No tables, figures, or numerical results (such as BER, accuracy, or computation time comparisons) are provided to support the assertions of reduced load with maintained performance or the superiority of alternating over marginal search, rendering the empirical validation of the core contribution unassessable.

    Authors: We acknowledge that the submitted manuscript's results section does not contain the supporting tables, figures, or numerical data. The revised version will incorporate detailed results for both Fiber OC and FSO systems, including BER values, accuracy metrics, computation load comparisons (with quantitative reductions), evidence of alternating search outperforming marginal search, baseline comparisons to full grid search, and appropriate error bars and dataset details. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method proposal with no derivation chain

full rationale

The paper proposes marginal and alternating grid search methods for DNN hyperparameter tuning and asserts favorable performance on Fiber OC and FSO tasks via internal experiments. No equations, first-principles derivations, or predictions appear in the provided text. None of the enumerated circularity patterns (self-definitional, fitted-input-as-prediction, self-citation load-bearing, etc.) are present because there is no mathematical chain that reduces to its own inputs. Performance claims rest on empirical results rather than any closed-form equivalence or self-referential construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no equations, fitted constants, background axioms, or new postulated entities are described.

pith-pipeline@v0.9.0 · 5719 in / 1080 out tokens · 23429 ms · 2026-05-25T15:46:59.952855+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

39 extracted references · 39 canonical work pages · 2 internal anchors

  1. [1]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). A novel hybrid FSO/RF communication system with receive diversity. Optik

    work page 2019

  2. [2]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). Performance analysis of hybrid FSO/RF communication systems with Alamouti Coding or Antenna Selection. The Journal of Engineering, 2019(5), 3433-3437

    work page 2019

  3. [3]

    Amirabadi, M. A. (2019). Performance comparison of two novel relay-assisted hybrid fso/rf communicati on systems. IET Communications

    work page 2019

  4. [4]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). Performance of a relay-assisted hybrid FSO/RF communication system. Physical Communication, 100729

    work page 2019

  5. [5]

    Bishop, C. M. (2006). Pattern recognition and machine learning. springer

    work page 2006

  6. [6]

    Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press

    work page 2016

  7. [7]

    T., Demuth, H

    Hagan, M. T., Demuth, H. B., Beale, M. H., & De Jesús, O. (1996). Neural network design (Vol. 20). Boston: Pws Pub

    work page 1996

  8. [8]

    T., Eriksson, T

    Jones, R. T., Eriksson, T. A., Yankov, M. P., & Zibar, D. (2018, September). Deep learning of geometric constellation shaping including fiber nonlinearities. In 2018 European Conference on Optical Communication (ECOC) (pp. 1-3). IEEE

    work page 2018

  9. [9]

    Wang, C., Fu, S., Xiao, Z., Tang, M., & Liu, D. (2019). Long short-term memory neural network (LSTM - NN) enabled accurate optical signal-to-noise ratio (OSNR) monitoring. Journal of Lightwave Technology

    work page 2019

  10. [10]

    N., Zhong, K., Al -Arashi, W

    Khan, F. N., Zhong, K., Al -Arashi, W. H., Yu, C., Lu, C., & Lau, A. P. T. (2016). Modulation format identification in coherent receivers using deep machine learning. IEEE Photonics Technology Letters , 28(17), 1886-1889

    work page 2016

  11. [11]

    N., Zhong, K., Zhou, X., Al -Arashi, W

    Khan, F. N., Zhong, K., Zhou, X., Al -Arashi, W. H., Yu, C., Lu, C., & Lau, A. P. T. (2017). Joint OSNR monitoring and modulation format identification in digital coherent receivers using deep neural networks. Optics express, 25(15), 17767-17776

    work page 2017

  12. [12]

    S., Bardenet, R., Bengio, Y., & Kégl, B

    Bergstra, J. S., Bardenet, R., Bengio, Y., & Kégl, B. (2011). Algorithms for hyper-parameter optimization. In Advances in neural information processing systems (pp. 2546-2554)

    work page 2011

  13. [13]

    (2015, September)

    Schilling, N., Wistuba, M., Drumond, L., & Schmidt -Thieme, L. (2015, September). Hyperparameter optimization with factorized multilayer perceptrons. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 87-103). Springer, Cham

    work page 2015

  14. [14]

    Hutter, F., Lücke, J., & Schmidt-Thieme, L. (2015). Beyond manual tuning of hyperparameters. KI- Künstliche Intelligenz, 29(4), 329-337

    work page 2015

  15. [15]

    Friedrichs, F., & Igel, C. (2005). Evolutionary tuning of multiple SVM parameters. Neurocomputing, 64, 11 107-117

    work page 2005

  16. [17]

    Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems (pp. 2951-2959)

    work page 2012

  17. [18]

    (2016, June)

    Luketina, J., Berglund, M., Greff, K., & Raiko, T. (2016, June). Scalable gradient-based tuning of continuous regularization hyperparameters. In International Conference on Machine Learning (pp. 2952-2960)

    work page 2016

  18. [19]

    Bergstra, J., & Bengio, Y. (2012). Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(Feb), 281-305

    work page 2012

  19. [20]

    K., Bansal, A., Garg, P., Tsiftsis, T., & Barrio s, R

    Sharma, P. K., Bansal, A., Garg, P., Tsiftsis, T., & Barrio s, R. (2017). Relayed FSO communication with aperture averaging receivers and misalignment errors. IET Communications, 11(1), 45-52

    work page 2017

  20. [21]

    V., Thang, T

    Trinh, P. V., Thang, T. C., & Pham, A. T. (2017). Mixed mmWave RF/FSO relaying systems over generalized fading channels with pointing errors. IEEE Photonics Journal, 9(1), 1-14

    work page 2017

  21. [22]

    J., Gomez, G., & Garrido -Balsells, J

    Lopez-Martinez, F. J., Gomez, G., & Garrido -Balsells, J. M. (2015). Physical-layer security in free -space optical communications. IEEE Photonics Journal, 7(2), 1-14

    work page 2015

  22. [23]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). On the performance of a multi-user multi-hop hybrid FSO/RF communication system. Optics Communications, 444, 172-183

    work page 2019

  23. [24]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2018). A new optimization problem in FSO communication system. IEEE Communications Letters, 22(7), 1442-1445

    work page 2018

  24. [25]

    Amirabadi, M. A. (2019). Performance evaluation of a novel relay assisted hybrid FSO/RF communication system with receive diversity. IET Optoelectronics

    work page 2019

  25. [26]

    Poggiolini, P. (2012). The GN model of non-linear propagation in uncompensated coherent optical systems. Journal of Lightwave Technology, 30(24), 3857-3879

    work page 2012

  26. [27]

    Dar, R., Feder, M., Mecozzi, A., & Shtaif, M. (2013). Properties of nonlinear noise in long, dispersion- uncompensated fiber links. Optics Express, 21(22), 25685-25699

    work page 2013

  27. [28]

    Carena, A., Bosco, G., Curri, V., Jiang, Y., Poggiolini, P., & Forghieri, F. (2014). EGN model of non-linear fiber propagation. Optics express, 22(13), 16335-16362

    work page 2014

  28. [29]

    Dar, R., Feder, M., Mecozzi, A., & Shtaif, M. (2014). Accumulation of nonlinear interference noise in fiber- optic systems. Optics express, 22(12), 14199-14211

    work page 2014

  29. [30]

    Source code,

    R. Jones, “Source code,” https://github.com/rassibassi/claude, 2018

    work page 2018

  30. [31]

    Bottou, L. (2010). Large-scale machine learning with stochastic gradient de scent. In Proceedings of COMPSTAT'2010 (pp. 177-186). Physica-Verlag HD

    work page 2010

  31. [32]

    Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:448-456, 2015

    work page 2015

  32. [33]

    C., Leung, C

    Leung, H. C., Leung, C. S., Wong, E. W., & Li, S. (2017). Extreme learning machine for estimating blocking probability of bufferless OBS/OPS networks. IEEE/OSA Journal of Optical Communications and Networking, 9(8), 682-692

    work page 2017

  33. [34]

    T., & Kumar, K

    Ahmad, S. T., & Kumar, K. P. (2016). Radial basis function neural network nonlinear equalizer for 16-QAM coherent optical OFDM. IEEE Photonics Technology Letters, 28(22), 2507-2510. 12

    work page 2016

  34. [35]

    F., & Fernandes, M

    de Sousa, T. F., & Fernandes, M. A. (2013, August). Multilayer perceptron equalizer for optical communication systems. In 2013 SBMO/IEEE MTT-S International Microwave & Optoelectronics Conference (IMOC) (pp. 1-5). IEEE

    work page 2013

  35. [36]

    Wang, D., Zhang, M., Li, J., Li, Z., Li, J., Song, C., & Chen, X. (2017). Intelligent constellation diagram analyzer using convolutional neural network-based deep learning. Optics express, 25(15), 17150-17166

    work page 2017

  36. [37]

    Mikolov, T., Karafiát, M., Burget, L., Černocký, J., & Khudanpur, S. (2010). Recurrent neural network based language model. In Eleventh annual conference of the international speech communication association

    work page 2010

  37. [38]

    Masters, T. (1993). Practical neural network recipes in C++. Morgan Kaufmann

    work page 1993

  38. [39]

    Zeiler, M. D. (2012). ADADELTA: an adaptive learning rate method. arXiv preprint arXiv:1212.5701

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  39. [40]

    Adam: A Method for Stochastic Optimization

    Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014