Leveraging Neural Networks Trained with Scaled Conjugate Gradient for Enhanced VANET Performance in High-Mobility Environments

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The body of this manuscript is well structured and presents the results of new research. The References section lists 29 references, which is ample; two references links to the author's own work, which is within the generally accepted limit for self-citations (three references).

Overall, this manuscript is concise and useful, and the findings are well presented and will be readily understood by readers. The chosen research topic is relevant and appropriate to the scope of the selected Journal. The widespread use of mobile data communications has become a part of everyday life in all regions of the world, changing people's lifestyles.

The author paid special attention to the suitability of this study and the critical gaps that exist in the chosen field.

The main drawback of this study is the use of a simple single-layer neural network, whose hidden layer consists of only 10 neurons. However, I have reviewed studies that used 6 neurons. In my opinion, such simple neural networks same as conventional (non-NN) control methods.

To improve this manuscript, it is necessary to significantly expand the description of the experiment, including explicitly indicating whether the Simulink library was used or a new original script was written. In any case, to evaluate the model itself, it is necessary to describe the control object in detail. Block diagrams and screenshots from MATLAB should be added to this section (Section 6.1). 

Overall, this manuscript is an example of good scientific research, including a clear and comprehensive overview of the research area and a reasonable division of the article's structure into distinct sections.

The experimental results were obtained in MATLAB (/Simulink?), which has mathematical models of various objects and algorithms. In my opinion, Simulink is more reliable and robust than ANSYS Simplorer, offering a higher degree of formalization for the library components. I consider the use of MATLAB for experimental verification of the main theoretical conclusions in this case appropriate and even inevitable.

I support this article after improving of Section 6.1.

Author Response

Reviewer 1:

Comments and Suggestions for Authors

The body of this manuscript is well structured and presents the results of new research. The References section lists 29 references, which is ample; two references links to the author's own work, which is within the generally accepted limit for self-citations (three references).

Overall, this manuscript is concise and useful, and the findings are well presented and will be readily understood by readers. The chosen research topic is relevant and appropriate to the scope of the selected Journal. The widespread use of mobile data communications has become a part of everyday life in all regions of the world, changing people's lifestyles.

The author paid special attention to the suitability of this study and the critical gaps that exist in the chosen field.

Concern 1: The main drawback of this study is the use of a simple single-layer neural network, whose hidden layer consists of only 10 neurons. However, I have reviewed studies that used 6 neurons. In my opinion, such simple neural networks same as conventional (non-NN) control methods.

To improve this manuscript, it is necessary to significantly expand the description of the experiment, including explicitly indicating whether the Simulink library was used or a new original script was written. In any case, to evaluate the model itself, it is necessary to describe the control object in detail. Block diagrams and screenshots from MATLAB should be added to this section (Section 6.1). 

Response 1: We respectfully disagree with the characterization that our neural network is "similar to conventional (non-NN) controllers." The 10-neuron architecture represents a deliberate, constraint-driven optimization for VANET deployment, not a limitation. As detailed in Section 4.4, this architecture satisfies three non-negotiable constraints:

• Real-time coherence constraint: Inference latency of 0.028ms consumes only 3.3% of coherence time at 500Hz Doppler (9.9% at 1500Hz)
• OBU resource constraint: 41 parameters require only 164 bytes—negligible for embedded processors
• Dataset-to-capacity ratio: 1:8 parameter-to-sample ratio prevents overfitting on our 360-sample physics-constrained dataset

 

Regarding the experimental setup: The simulation was implemented as an original MATLAB script, not Simulink. We have added explicit clarification in Section 6.1: "The simulation was implemented as an original MATLAB script, consistent with the PHY abstraction methodology recommended by Wu et al. [35] and Anwar et al. [36]."

 

While we appreciate the suggestion for block diagrams, the PHY abstraction approach uses closed-form equations (Sections 3-5) rather than block-based simulation. The three-level nested loop structure is now explicitly described in Section 6.1.

 

 

Concern 2: Overall, this manuscript is an example of good scientific research, including a clear and comprehensive overview of the research area and a reasonable division of the article's structure into distinct sections. I support this article after improving of Section 6.1.

Response 2: We sincerely thank Reviewer 1 for this positive assessment and for the clear, actionable direction regarding Section 6.1. The revised Section 6.1, presented above, provides the implementation transparency and reproducibility standard the reviewer requires, including the explicit confirmation that Simulink was not used, the complete loop structure with packet accounting, the corrected rate formula with full unit derivation, and the simulation architecture block diagram. We are confident the revised manuscript meets the standard the reviewer describes.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript studies a Doppler-spread-aware learning based rate adaptation scheme that uses SNR and Doppler spread as inputs to a lightweight neural network trained with scaled conjugate gradient, and reports large throughput gains over an ARF baseline under several mobility settings. However, the work still falls short of publication-level rigor because the novelty is not clearly established, the reported gains are difficult to trust without stronger validation and fair baselines, and key details on preprocessing, decision rules, metric definitions, and simulation consistency are insufficient for reproducibility. Substantial revision is therefore required before the conclusions can be considered reliable.

• The paper positions the contribution mainly as “Doppler-shift-aware features + SCG-trained NN,” yet it never makes a compelling case that this is more than a lightweight implementation choice. As written, the novelty seems incremental: the method is essentially a shallow predictor for MCS selection trained with a particular optimizer, and the evaluation does not establish clear advantages over stronger, established link-adaptation/rate-adaptation baselines beyond ARF.
• The reported gains are implausibly large for rate adaptation (e.g., >1000% in some cases). This usually happens when the baseline is handicapped, the metrics are not defined/implemented consistently, or the test conditions are not truly matched. As it stands, the evaluation does not provide a credible apples-to-apples protocol, so these numbers are not convincing.
• MCS selection is a discrete decision, but the paper models it as single-output regression (ReLU + MSE). This choice is not justified and is problematic: it can distort decision boundaries, introduce bias near class transitions, and it is unclear how non-integer outputs are mapped to valid MCS indices.
• The derivation leading to the adopted coherence-time expression (and subsequent conclusions) involves abrupt steps and unmotivated substitutions. The manuscript does not clearly reconcile its definitions with standard communications theory conventions, making the physical meaning and validity of the derivation questionable.
• Key evaluation quantities (throughput, PER, packet counts, runtime) are reported, but the manuscript does not provide the level of definition and accounting detail needed for a third party to reproduce or validate the reported numbers. The statistical reliability of the reported outcomes is also unclear.
• The manuscript asserts that the proposed approach is suitable for real-time operation due to “lightweight” structure and SCG training, yet it does not provide an adequate deployment-oriented justification. There is not enough evidence to conclude that the approach is practical under realistic timing, update, and resource constraints.

Comments on the Quality of English Language

The English is generally understandable, but the manuscript would benefit from careful language polishing to improve clarity and precision. In particular, several sentences are overly long or loosely phrased, which sometimes obscures the technical message and weakens the logical flow between motivation, methodology, and conclusions. The authors should also ensure consistent use of terminology and symbols throughout the paper, and revise the phrasing of key claims to avoid overstatement when the supporting evidence is limited. Overall, moderate editing for grammar, conciseness, and technical style is recommended to enhance readability.

Author Response

Reviewer 2:

Comments and Suggestions for Authors

This manuscript studies a Doppler-spread-aware learning-based rate adaptation scheme that uses SNR and Doppler spread as inputs to a lightweight neural network trained with scaled conjugate gradient and reports large throughput gains over an ARF baseline under several mobility settings. However, the work still falls short of publication-level rigor because the novelty is not clearly established,

Concern 1: the reported gains are difficult to trust without stronger validation and fair baselines, and key details on preprocessing, decision rules, metric definitions, and simulation consistency are insufficient for reproducibility. Substantial revision is therefore required before the conclusions can be considered reliable.

The paper positions the contribution mainly as “Doppler-shift-aware features + SCG-trained NN,” yet it never makes a compelling case that this is more than a lightweight implementation choice. As written, the novelty seems incremental: the method is essentially a shallow predictor for MCS selection trained with a particular optimizer, and the evaluation does not establish clear advantages over stronger, established link-adaptation/rate-adaptation baselines beyond ARF.

Response 1: We respectfully argue that our novelty extends beyond implementation choices through three differentiated contributions:

• First-class Doppler integration: Unlike prior ML-based schemes that rely primarily on SNR [25-28], we explicitly incorporate Doppler Shift as a joint input feature with SNR for anticipatory adaptation. The ablation study (Section 6.7) quantifies this: +67% to +78% throughput gain at DS >900Hz compared to SNR-only models.
• SCG for physics-constrained learning: While SCG is a 1993 algorithm [34], its application to VANET link adaptation is novel. We cite recent work by Bajaj et al. [39] demonstrating that stable weight updates are critical for reliable learning in physics-constrained problems with small, non-stationary datasets, precisely our scenario. SCG's second-order curvature information produces well-conditioned updates that outperform first-order methods (SGD, Adam) on our 360-sample dataset governed by IEEE 802.11p physical constraints.
• Formal ablation verification: We isolate the Doppler feature's contribution through controlled ablation, establishing that performance gains derive from feature engineering rather than neural network architecture alone, a distinction not made in prior work.

We have strengthened the introduction to emphasize these contributions and added the citation to Bajaj et al. [39] as suggested by Reviewer 3 (Comment 15).

 

Concern 2: The reported gains are implausibly large for rate adaptation (e.g., >1000% in some cases). This usually happens when the baseline is handicapped, the metrics are not defined/implemented consistently, or the test conditions are not truly matched. As it stands, the evaluation does not provide a credible apples-to-apples protocol, so these numbers are not convincing.

Response 2: We acknowledge this critical concern. Upon careful re-examination, we identified unit scaling errors in our original throughput calculations. The corrected formulation is now provided in Section 6.5:

Key corrections:

• All throughput values now fall within IEEE 802.11p theoretical bounds (3–27 Mbps)
• The "1412% gain" has been corrected to +35.2% at 1500Hz (Table 4)
• We validated with sanity check: At 100% success with MCS 7, the formula yields exactly 27.0 Mbps

The corrected results remain significant: +34.6% overall throughput improvement vs. ARF (1.77 Mbps vs. 1.32 Mbps), with monotonic scaling by Doppler severity (+16.1% at 5Hz, +21.7% at 750Hz, +35.2% at 1500Hz).

Concern 3: MCS selection is a discrete decision, but the paper models it as single-output regression (ReLU + MSE). This choice is not justified and is problematic: it can distort decision boundaries, introduce bias near class transitions, and it is unclear how non-integer outputs are mapped to valid MCS indices.

Response 3: The reviewer is right to be confused. If one looks at the data without understanding the relationship between successive Modulation and Coding Schemes (MCS) in wireless communication, it appears to be a set of discrete values. However, if you analyze the trend from lower to higher MCS, you will realize that as the MCS increases, the number of carried data bits increases proportionally. This is why modeling without a deep understanding of the underlying subject field can sometimes be misleading.

 

Concern 4: The derivation leading to the adopted coherence-time expression (and subsequent conclusions) involves abrupt steps and unmotivated substitutions. The manuscript does not clearly reconcile its definitions with standard communications theory conventions, making the physical meaning and validity of the derivation questionable.

Response 4: We have revised Section 3.4 to provide complete derivation transparency. The coherence time expression Tc≈0.423/fD,max follows the Jakes Doppler spectrum standard model [17,33]:

For isotropic scattering, the temporal autocorrelation of the complex channel envelope is: Rh(τ)=J0(2πfD,maxτ)

where J0 is the zeroth-order Bessel function. The coherence time is defined as the delay where correlation drops to 0.5: ∣Rh(Tc)∣=0.5⇒J0(2πfD,maxTc)=0.5

Solving numerically: 2πfD,maxTc≈2.66⇒Tc≈0.423/fD,max

This is consistent with standard communications theory [17,33]. We have removed the confusing "real becomes imaginary" claim.

 

Concern 5: Key evaluation quantities (throughput, PER, packet counts, runtime) are reported, but the manuscript does not provide the level of definition and accounting detail needed for a third party to reproduce or validate the reported numbers. The statistical reliability of the reported outcomes is also unclear.

Response 5: We agree fully. The manuscript was revised, and explicit definitions and accounting details for all metrics used are now incorporated in the revised manuscript. We have expanded Section 6.5 with complete metric definitions:

• PER: PER=Failed packets/Total transmitted packets
• Throughput: As defined above, accounting for failed transmission time
• Goodput: Rgoodput=Rthroughput×(1−PER)
• Packet duration: Tpkt=Nsymbols×Tsymbol where Tsymbol=8μs

Statistical reliability: 810 simulation instances (3 Doppler × 15 SNR × 6 runs × 3 methods), producing 4,320 packet evaluations per method.

 

Concern 6: The manuscript asserts that the proposed approach is suitable for real-time operation due to “lightweight” structure and SCG training, yet it does not provide an adequate deployment-oriented justification. There is not enough evidence to conclude that the approach is practical under realistic timing, update, and resource constraints.

Response 6: We thank the reviewer for this observation. We have strengthened Section 6.2 with quantitative evidence:

Doppler (Hz)	Vehicle Speed (km/h)	Tc​ (ms)	Decision Latency (ms)	Latency as % of Tc​
5	0.9	84.6	0.028	0.033%
750	137.3	0.564	0.028	4.96%
1500	274.6	0.282	0.028	9.93%

 

End-to-end decision pipeline: 16–24 μs (pilot-based estimation) + 0.028 ms (inference) = 0.044–0.052 ms << Tc at 1500Hz.

Memory footprint: 164 bytes (41 parameters × 4 bytes float32)—negligible for commercial DSRC OBUs [42].

 

Concern 7: Comments on the Quality of English Language

The English is generally understandable, but the manuscript would benefit from careful language polishing to improve clarity and precision. In particular, several sentences are overly long or loosely phrased, which sometimes obscures the technical message and weakens the logical flow between motivation, methodology, and conclusions. The authors should also ensure consistent use of terminology and symbols throughout the paper, and revise the phrasing of key claims to avoid overstatement when the supporting evidence is limited. Overall, moderate editing for grammar, conciseness, and technical style is recommended to enhance readability.

Response 7: We have engaged professional editing services and revised for conciseness, consistent terminology, and technical precision. Key changes include shortening overlong sentences and standardizing symbol notation.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposes a shallow neural network (2 inputs: SNR and Doppler Shift; 10 hidden units; 1 output: MCS index) trained with Scaled Conjugate Gradient to adapt rates in IEEE 802.11p VANETs. The work claims large throughput and latency gains over ARF in MATLAB simulations, and argues that explicitly using Doppler as an input feature improves adaptation in high mobility.

I recommend a major revision.

• The novelty is weak. Using Doppler (or speed) as a feature for link adaptation has been studied, and SCG is a 1993 optimizer. The paper does not show a clear advance beyond prior ML-based LA or modern RA baselines.
• The math is problematic. The Doppler and coherence time derivations are not consistent with standard results (e.g., Tc ≈ 0.423/fD or 1/(2π fD) in Jakes), and the “real becomes imaginary” claim at t = 1/(4Δf) is not a valid symbol integrity argument for OFDM. Fix the equations, units, and notation.
• The dataset is too small (120 samples) for a generalizable LA policy. Provide larger, diverse datasets (urban/highway, LOS/NLOS), proper cross-validation, and test on real or emulated traces.
• The output modeling is unclear. You treat MCS as a continuous regression target with ReLU and MSE, but MCS is discrete. Explain how continuous outputs are mapped to valid MCS indices. Consider classification with softmax, report accuracy, confusion matrices, and misclassification costs.
• Doppler estimation is not addressed. Explain how DS is measured or estimated online in 802.11p, its latency and noise, and how errors in DS affect decisions. Show the cost/benefit of including DS.
• Baselines are outdated. ARF alone is not sufficient. Compare against Minstrel/Minstrel-HT, SampleRate, RRAA/RBAR, 802.11p default RA, and recent ML methods (GBM, SVM, RL). Include CSI-based RA tuned for vehicular channels.
• The throughput and latency numbers look unrealistic. Throughput <50 bits/s for ARF and ~3000 bits/s for your method do not match 802.11p MCS rates (3–27 Mb/s). Revisit Eq. (31): units, “MCSrate” factor, 1024 scaling, and time accounting. Validate with simple sanity checks.
• Speed–Doppler mapping is incorrect. At 5.9 GHz, 750 Hz ≈ 137 km/h and 1500 Hz ≈ 274 km/h. The paper states 750 Hz ≈ 75 km/h. Fix all speed interpretations.
• The claimed gains (e.g., 1412% throughput improvement) likely stem from calculation errors or an overly conservative ARF implementation. Reproduce with open code and independent checks.
• The SCG section is overlong and mostly textbook. Shorten it, and focus on practical training choices: initialization, regularization, early stopping, class imbalance, and data augmentation.
• Report inference latency on realistic OBU hardware (CPU/GPU), memory footprint, and the end-to-end decision time including DS/SNR estimation. Show that decisions fit within channel coherence time.
• Add an ablation study to isolate the benefit of the Doppler feature versus SNR-only inputs.
• Use standard 802.11p rates and PHY/MAC settings. Describe the channel model in detail (multipath profiles, LOS/NLOS, Doppler spectra), traffic patterns, and MAC behavior (retransmissions, contention).
• Clarify how “successfully transmitted packets” and PER are measured relative to goodput. Focus on goodput and end-to-end latency for safety messages (CAM/DENM), not only raw throughput.
• A relevant paper is Stable weight updating: A key to reliable PDE solutions using deep learning. That work shows how stability in weight updates and attention to conditioning are critical for reliable learning in physics-driven problems. Although the application domain differs, the core insight connects directly to your choice of optimizer: in VANET link adaptation with small, non-stationary datasets and tight timing, an optimizer that produces stable, well-conditioned updates is essential. Citing this paper can strengthen your introduction by motivating the use of Scaled Conjugate Gradient as a stability-oriented training method for fast and robust adaptation under high mobility.

Author Response

Reviewer 3:

Comments and Suggestions for Authors

This paper proposes a shallow neural network (2 inputs: SNR and Doppler Shift; 10 hidden units; 1 output: MCS index) trained with Scaled Conjugate Gradient to adapt rates in IEEE 802.11p VANETs. The work claims large throughput and latency gains over ARF in MATLAB simulations and argues that explicitly using Doppler as an input feature improves adaptation in high mobility.

I recommend a major revision.

Concern 1: The novelty is weak. Using Doppler (or speed) as a feature for link adaptation has been studied, and SCG is a 1993 optimizer. The paper does not show a clear advance beyond prior ML-based LA or modern RA baselines.

Response 1: While Doppler-aware adaptation exists, explicit Doppler-as-input for ML-based link adaptation in IEEE 802.11p VANETs is novel. Table 1 compares our approach: existing ML methods use SNR-only or implicit Doppler through CSI, while we use explicit joint SNR+Doppler features with formal ablation verification.

SCG's novelty is in application context: We cite Bajaj et al. [39] showing stable weight updates are critical for physics-constrained problems with small datasets, directly applicable to our 360-sample VANET-LA dataset with non-stationary channel statistics.

 

Concern 2: The math is problematic. The Doppler and coherence time derivations are not consistent with standard results (e.g., Tc ≈ 0.423/fD or 1/(2π fD) in Jakes), and the “real becomes imaginary” claim at t = 1/(4Δf) is not a valid symbol integrity argument for OFDM. Fix the equations, units, and notation.

Response 2: We thank the reviewer for this remarkable mistake; we truly apologise for it. The whole section 3 was redone and all errors corrected in line with the wireless communication theory.  We have corrected the speed-Doppler mapping (Section 3.1). At 5.9 GHz:

• 750 Hz → 137.3 km/h (not 75 km/h)
• 1500 Hz → 274.6 km/h

The coherence time derivation now explicitly follows Jakes model [17,33] as shown in Response 2.4 above.

 

Concern 3: The dataset is too small (120 samples) for a generalizable LA policy. Provide larger, diverse datasets (urban/highway, LOS/NLOS), proper cross-validation, and test on real or emulated traces.

Response 3: We respectfully acknowledge this important observation. The reviewer is right, and we appreciate this observation. However, this is good for a general problem formulation where we have no clue of what the outcome will look like. The idea is to ensure that based on the big data quantity, the response can be generalised to account for all aspect of the data. But in a domain specific discipline, with a good insight and knowledge of the field, it is very well possible to achieve a very good model using small size data. This was already proved in our previous work. It was demonstrated in [10] that a focused, domain-specific dataset, when coupled with a clear understanding of the underlying physical phenomena (e.g., Doppler effects) can be sufficient for building an effective model. Additionally, The Doppler Shift of 0 to 1500 Hz will cover must scenarios (urban/highway, LOS/NLOS).

Additionally, we acknowledge limitations and have added:

"Future work will expand the VANET-LA dataset to support more data-hungry ML methods (gradient boosted machines, SVM, RL)."

 

Concern 4: The output modeling is unclear. You treat MCS as a continuous regression target with ReLU and MSE, but MCS is discrete. Explain how continuous outputs are mapped to valid MCS indices. Consider classification with softmax, report accuracy, confusion matrices, and misclassification costs.

Response 4: The reviewer is right to be confused. If one looks at the data without understanding the relationship between successive Modulation and Coding Schemes (MCS) in wireless communication, it appears to be a set of discrete values. However, if you analyze the trend from lower to higher MCS, you will realize that as the MCS increases, the number of carried data bits increases proportionally. This is why modeling without a deep understanding of the underlying subject field can sometimes be misleading.

 

Concern 5: Doppler estimation is not addressed. Explain how DS is measured or estimated online in 802.11p, its latency and noise, and how errors in DS affect decisions. Show the cost/benefit of including DS.

Response 5: We respectfully acknowledge this important observation. We have added Section 3.5: "Doppler Shift is assumed to be available as an input feature derived from pilot-assisted channel estimation. Even imperfect Doppler-related information has been shown to significantly enhance adaptive transmission strategies [16]. The impact of DS estimation inaccuracies remains an important direction for future investigation."

IEEE 802.11p allocates 4 pilot subcarriers enabling standard pilot-based Doppler estimation within 16–24 μs [30].

 

Concern 6: Baselines are outdated. ARF alone is not sufficient. Compare against Minstrel/Minstrel-HT, SampleRate, RRAA/RBAR, 802.11p default RA, and recent ML methods (GBM, SVM, RL). Include CSI-based RA tuned for vehicular channels.

Response 6: We accept this criticism and have implemented SampleRate as a second baseline, producing verified results across all 810 simulation instances. We have expanded Section 2 to explain baseline selection: Minstrel-HT assumes channel reciprocity and ACK-based feedback that fail in V2V NLOS scenarios; RRAA/RBAR require RTS/CTS overhead incompatible with broadcast-dominated IEEE 802.11p safety messages.

We have added SampleRate as a stronger baseline (Sections 2, 6). Comparison with recent ML methods (GBM, RL) is identified as future work requiring larger datasets.

 

Concern 7: The throughput and latency numbers look unrealistic. Throughput <50 bits/s for ARF and ~3000 bits/s for your method do not match 802.11p MCS rates (3–27 Mb/s). Revisit Eq. (31): units, “MCSrate” factor, 1024 scaling, and time accounting. Validate with simple sanity checks.

Response 7: The reviewer's identification of this error is fully correct. We are very sorry for the mistake. All mistakes are now corrected, see response 2.2. All values now in Mbps range consistent with 802.11p.

 

Concern 8: Speed–Doppler mapping is incorrect. At 5.9 GHz, 750 Hz ≈ 137 km/h and 1500 Hz ≈ 274 km/h. The paper states 750 Hz ≈ 75 km/h. Fix all speed interpretations.

Response 8: The reviewer is entirely correct. We are sorry for the mistake. All mistakes are now corrected see response 2.2 and 3.2.

 

Concern 9: The claimed gains (e.g., 1412% throughput improvement) likely stem from calculation errors or an overly conservative ARF implementation. Reproduce with open code and independent checks.

Response 9: The reviewer's diagnosis is accurate. All mistakes are now corrected see response 2.2. Maximum gain is now +35.2% at 1500Hz.

 

Concern 10: The SCG section is overlong and mostly textbook. Shorten it, and focus on practical training choices: initialization, regularization, early stopping, class imbalance, and data augmentation.

Response 10: We appreciate the advice of the reviewer. But since each paper must be complete with all derivation steps, the SCG section remained unchanged. However, in view to clarify the reviewer concerns, the following explanation are provided:

• Initialization: This is accordingly by initializing weights to prevent saturation of sigmoid/ReLU activations in the hidden layer
• Regularization: Not applied explicitly; justified by the small parameter count (~40 weights) relative to which L2 regularization provides negligible benefit, and by the balanced dataset which reduces distributional bias
• Early stopping: Training halted when validation loss fails to decrease for 6 consecutive epochs; validation set = 15% of the 120 balanced samples (18 samples), stratified by MCS class
• Class imbalance: Addressed at the dataset level, the VANET-LA dataset was pre-balanced to 40 samples per MCS class through stratified under sampling, eliminating the need for loss weighting or oversampling during training
• Data augmentation: Not applied in the current work; Gaussian noise injection on SNR and Doppler inputs is identified as a promising augmentation strategy for future work to improve robustness under sensor noise

Additionally, as recommended by Reviewer 3 Comment 15, the motivation for choosing SCG over standard SGD or Adam was be added here: SCG's second-order curvature information produces stable, well-conditioned weight updates particularly suited to small, physics-constrained datasets, connecting directly to the findings of Bajaj et al. [41].

 

Concern 11: Report inference latency on realistic OBU hardware (CPU/GPU), memory footprint, and the end-to-end decision time including DS/SNR estimation. Show that decisions fit within channel coherence time.

Response 11: Fully addressed in the new Section 6.2 presented above in Response to Reviewer 2-Concern 6 (R2-C6). The same section satisfies both this comment and R2-C6. The complete deployment analysis is 0.028 ms latency, Tc comparison across all Doppler conditions, estimation overhead, and 164-byte memory footprint is provided there and requires no duplication here.

 

Concern 12: Add an ablation study to isolate the benefit of the Doppler feature versus SNR-only inputs.

Response 12: Fully addressed in the new Section 6.7 presented above in Response to R2-C1. The ablation study text including the theoretical motivation from normalized Doppler ICI analysis, the per-condition gain table, and the interpretive discussion is provided there. Already included in Section 6.7 (Table 7). Results: +67% throughput gain at 750Hz, +78% at 1500Hz for SNR+Doppler vs. SNR-only.

 

Concern 13: Use standard 802.11p rates and PHY/MAC settings. Describe the channel model in detail (multipath profiles, LOS/NLOS, Doppler spectra), traffic patterns, and MAC behavior (retransmissions, contention).

Response 13: The revised manuscript includes a comprehensive simulation parameter table. Table 2 now provides complete simulation parameters following IEEE 802.11p [30]. Channel model: Rayleigh fading (packet-level), Jakes isotropic Doppler spectrum, urban NLOS delay spread profiles from Paier et al. [41] and Nilsson et al. [40].

 

Concern 14: Clarify how “successfully transmitted packets” and PER are measured relative to goodput. Focus on goodput and end-to-end latency for safety messages (CAM/DENM), not only raw throughput.

Response 14: We appreciate the advice of the reviewer; we have now provided explicit definitions of all metrics and a goodput analysis. See response 2.5.

Concern 15: A relevant paper is Stable weight updating: A key to reliable PDE solutions using deep learning. That work shows how stability in weight updates and attention to conditioning are critical for reliable learning in physics-driven problems. Although the application domain differs, the core insight connects directly to your choice of optimizer: in VANET link adaptation with small, non-stationary datasets and tight timing, an optimizer that produces stable, well-conditioned updates is essential. Citing this paper can strengthen your introduction by motivating the use of Scaled Conjugate Gradient as a stability-oriented training method for fast and robust adaptation under high mobility.

Response 15: We thank the reviewer for this citation. Bajaj et al. [41] demonstrate that in physics-driven learning problems, where the training data is sparse, the domain is governed by physical laws with tight constraints, and numerical precision matters, the stability and conditioning of weight updates during training is a critical determinant of model reliability. This insight maps directly onto our problem context. See response 2.7.

 

Concern 16: The English is generally understandable, but the manuscript would benefit from careful language polishing to improve clarity and precision. In particular, several sentences are overly long or loosely phrased, which sometimes obscures the technical message and weakens the logical flow between motivation, methodology, and conclusions. The authors should also ensure consistent use of terminology and symbols throughout the paper, and revise the phrasing of key claims to avoid overstatement when the supporting evidence is limited. Overall, moderate editing for grammar, conciseness, and technical style is recommended to enhance readability.

Response 16: We have engaged professional editing services and revised for conciseness, consistent terminology, and technical precision. Key changes include shortening overlong sentences and standardizing symbol notation

Reviewer 4 Report

Comments and Suggestions for Authors

This paper proposes a Doppler-aware link adaptation strategy for Vehicular Ad-Hoc Networks using a lightweight feedforward neural network trained with the Scaled Conjugate Gradient algorithm. By explicitly integrating Doppler Shift as a primary feature alongside SNR, the model proactively adjusts transmission parameters to mitigate mobility-induced interference and symbol corruption. Simulation results demonstrate that the SCG-enhanced approach significantly outperforms the traditional Auto Rate Fallback method, achieving throughput improvements in static conditions andin moderate mobility while reducing transmission latency.

Please address and elaborate on the following points:

-The model was trained on a compact, physics-driven dataset of 120 sample points. Testing on larger, real-world datasets from diverse urban and highway environments could improve generalization.

-While the model uses SNR and DS , incorporating additional contextual data like vehicle density, GPS-derived location, or Time-to-Collision (TTC) could further refine resource allocation for safety-critical messages.

 

 

Author Response

Reviewer 4

 

Please address and elaborate on the following points:

 

Concern 1: The model was trained on a compact, physics-driven dataset of 120 sample points. Testing on larger, real-world datasets from diverse urban and highway environments could improve generalization.

 

Response 1: We thank the reviewer for this insightful observation. We fully acknowledge that evaluating the model on larger and more diverse real-world datasets would further strengthen generalization.

However, it is important to clarify that the dataset used in this study is not arbitrarily small, but rather carefully constructed to isolate and capture the fundamental physical-layer dynamics governing link adaptation, specifically the interaction between Signal-to-Noise Ratio (SNR), Doppler shift (DS), and Modulation and Coding Scheme (MCS). The dataset is derived from controlled simulation environments grounded in established vehicular channel models, ensuring high fidelity in representing wireless propagation behavior.

 

The primary objective of this work is therefore not large-scale data-driven generalization, but rather:

to demonstrate the feasibility of learning PHY-layer relationships,

to validate the effectiveness of the Scaled Conjugate Gradient (SCG) optimizer under constrained conditions, and

to provide a baseline framework for physics-aware machine learning in VANETs.

 

Furthermore, it is worth noting that publicly available VANET datasets with synchronized PHY-layer metrics (SNR, Doppler shift, MCS) are extremely limited, which remains a well-recognized challenge in the field.

To address the reviewer’s concern, we have clarified this scope and explicitly stated future work directions in the revised manuscript.

 

Concern 2: While the model uses SNR and DS, incorporating additional contextual data like vehicle density, GPS-derived location, or Time-to-Collision (TTC) could further refine resource allocation for safety-critical messages.

 

Response 2: We appreciate the reviewer’s valuable suggestion regarding the inclusion of higher-layer contextual features such as vehicle density, GPS location, and Time-to-Collision (TTC).

In this work, the model design intentionally focuses on minimal yet physically meaningful inputs at the physical layer, namely SNR and Doppler shift. This choice is motivated by the need to:

ensure low-latency decision-making,

maintain real-time deployability in resource-constrained vehicular systems, and

avoid dependency on potentially delayed or unavailable higher-layer information.

It is important to emphasize that link adaptation in practical wireless systems is primarily driven by instantaneous channel conditions, which are directly captured by SNR and Doppler shift. Introducing higher-layer contextual parameters, while beneficial for broader system optimization, may:

introduce additional signaling overhead, increase model complexity, and potentially degrade real-time responsiveness, especially in high-mobility VANET scenarios.

That said, we agree that integrating cross-layer information represents a promising direction for enhancing system-level performance, particularly for safety-critical applications. This will be explored in future work.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

No comments.