Support Vector Machine-Based Soft Decision for Consecutive-Symbol-Expanded 4-Dimensional Constellation in Underwater Visible Light Communication System

Round 1

Reviewer 1 Report

The paper proposes the Support Vector Machine for underwater VLC. Though interesting ideas are presented in the article, the overall critical analysis of the work is relatively shallow and not very convincing from a practical point of view. Hence, I recommend the authors improve the paper significantly.

The abstract emphasized the use of the SVM for QAM. However, the authors have used CAP in the practical work. One cannot use QAM in the IMDD system as IMDD requires a real and positive signal. Ensure that the abstract accurately reflects the actual experimental work.

The authors used a 4-D constellation and justified the use of SVM as ‘the correlation between the two consecutive symbols will introduce crosstalk’. However, the paper needs to clarify

-          Why is there crosstalk/ISI if the system already has ‘hardware forward equalization’ and ‘post-equalization’? An appropriately designed equalizer must compensate the ISI and remove any correlation between symbols.

-          Since the overall system is nonlinear, how was LMS post-equalization designed and optimized?

-          Why two consecutive symbols only? Why not a larger number of consecutive symbols.

The motivation for the use of the Sobel operator is not clear nor fully justified. Why do you need a 2-D spatial gradient measurement? Furthermore, the performance system depends on the selection of the threshold value for the Sobel edge detector. Is this practical to optimize the threshold in a real-time system with varying channel conditions if performance is sensitive to the threshold?

What is box constraint? Why are there different trends for different data rates? This clearly indicates that the system needs to be optimized for each operating condition (e.g. data rate, bias condition, received power etc.) which is not practical.

Since the overall system is nonlinear and non-Gaussian (as stated by the authors), the use of the Q-factor to compare the performance is not appropriate. Also, Q-factor cannot be translated to BER in a nonlinear system. Furthermore, the authors have not compared the performance of the proposed system with other nonlinear equalizers/solutions that are cited by the authors (some of these works are from the co-authors). There must be a fair comparison, and hence authors should compare this work with other nonlinear equalizers (Volterra and ML)

The paper needs to improve the overall description of the technology (i.e., provide a better technical explanation of the system e.g., SVM, why the experimental system used ‘hardware forward equalization’ and what are their characteristics, how LMS equalizer was optimized (e.g., number of taps, training algorithms, etc.), the justification of the use of Sobel operation, key contribution and difference with other work (e.g. work in [12] which proposes a similar technique).

The paper needs to be proof-read as there are many sentences which are not technically sound (for example,

-          ‘To reduce the complexity of we apply the bit-based binary SVM classifier as the multi-class strategy.’ (How will ‘multi-class strategy’ reduces the complexity? It is a statement without justification.

-          ‘With the carrierless amplitude and phase (CAP) modulation, the signal is loaded into an arbitrary waveform generator (Tektronix AWG710B, 4.2GSa/s)’. (The sentence does not make any sense. Rewrite)

-          Then the signal passes through a hardware forward equalization circuit to compensate for the high-frequency component to resist the fading of the VLC channel. (‘Resist the fading of the VLC channel ?’, ‘compensate for the high-frequency component?’, these terms are not technically correct.)

-          ‘In the offline DSP, LMS post-equalization is used to compensate for distortion caused by Inter-symbol interference (ISI) after downsampling process’. (?)

-          To make a trade-off between BER performance and complexity, the training data size is 26214, 10% of the whole receiving data.(? is it practical to have 10% of data for training, and how the training data is related to complexity and performance?)

-          How was the practical system optimized?

-          What is the performance gain for other modulation (e.g. 8, 32, 64-QAM)?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

In this paper, the authors proposed a novel SVM-based soft decision method for consecutive-symbol-expanded 4-D constellation in underwater visible light communication system. The idea is innovative and the experimental results have shown great improvement in Q-value performance. There are some minor problems that should be addressed:

1.     The inset (i) in figure 3 is not clear, please replace with a clearer picture

2.     Please provide the parameters of lens used in the experiment.

3.     The roll-off factor of the CAP filter is not given.

4.     As the transmission function of light in water is largely determined by the quality of the water. Please specify the water used in the experiment. Is this tap water? Do you consider the influence of turbulence?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

It can be accepted in present form.