PABAFT: Channel Prediction Approach Based on Autoregression and Flexible TDD for 5G Systems
Round 1
Reviewer 1 Report
All the abbreviations must be explained at their first appearance (for, example URRLC etc.). Some of the abbreviations are introduced several times (for example, UMa 30, lines 165, 185, 186).
Care is needed in reference section formatting: no page numbers in conference papers. [22], [25] have improper name formatting. [20] – “,” instead of “.” is present. Revise the reference formatting carefully.
In their literature review authors omitted one of the mainstream approaches to channel prediction – ML-based methods. See, for instance:
- E. Bogale, X. Wang and L. B. Le, "Adaptive Channel Prediction, Beamforming and Scheduling Design for 5G V2I Network: Analytical and Machine Learning Approaches," in IEEE Transactions on Vehicular Technology, vol. 69, no. 5, pp. 5055-5067, May 2020, doi: 10.1109/TVT.2020.2975818.
- Luo, J. Ji, Q. Wang, X. Chen and P. Li, "Channel State Information Prediction for 5G Wireless Communications: A Deep Learning Approach," in IEEE Transactions on Network Science and Engineering, vol. 7, no. 1, pp. 227-236, 1 Jan.-March 2020, doi: 10.1109/TNSE.2018.2848960.
- Jiang and H. D. Schotten, "Neural Network-Based Fading Channel Prediction: A Comprehensive Overview," in IEEE Access, vol. 7, pp. 118112-118124, 2019, doi: 10.1109/ACCESS.2019.2937588.
and many other papers on that topic.
Authors should include this class of methods and comment on their relevance (or irrelevance, then why?) to the assumed problem.
Moreover, the Reference section can be sufficiently expanded by including (and discussing) a wide range of modern technical sources on that topic.
In the literature review, section authors state that the “Proni method … can obtain accurate channel parameters until the environment change from user mobility makes the initial model obsolete” citing
[9] H. Yin, H. Wang, Y. Liu and D. Gesbert, "Addressing the Curse of Mobility in Massive MIMO With Prony-Based Angular-Delay Domain Channel Predictions," in IEEE Journal on Selected Areas in Communications, vol. 38, no. 12, pp. 2903-2917, Dec. 2020, doi: 10.1109/JSAC.2020.3005473.
But in [9] a method of combating this “curse” is proposed. Moreover, in [9] scenarios with moving UE with speeds: 30 km/h, 60 km/h, and even 90 km/h were assumed; while in the current submission authors limit themselves to only 30 km/h. Moreover, in [9] it was demonstrated that the proposed algorithms yielded good results (for fast-moving UE). Authors of the current submission disregard those achievements. This should be commented on and cleared out.
It is advised to add a table with simulation parameters, rather than scatter them throughout the paper. This would make the simulation part more solid.
The AR model order parameter (very important for AR performance) is not discussed. Some abuse of notation is present: at first, it is called p, but afterwards, it is renamed K. Moreover, it is unclear how it relates to # of samples? It is critical in view of [9], where authors had claimed that their “method is able to achieve asymptotically error-free prediction, provided that only two accurate channel samples are available”. And in the current submission, one sees from 200 to 500 samples. This should be explained.
In Section 4.2 authors compare Last SRS with PABAFT without including information on what a number of samples were used.
The description of the results is very brief, especially concerning Fig. 9-11 and can be expanded.
The conclusion section should be expanded by adding numerical results obtained by the authors.
Author Response
Comment#1: All the abbreviations must be explained at their first appearance (for, example URRLC etc.). Some of the abbreviations are introduced several times (for example, UMa 30, lines 165, 185, 186).
Response: Thank you for your comment! We explained all abbreviations at their first appearance and removed duplicates. Also, we extended the table with abbreviations.
Comment#2: Care is needed in reference section formatting: no page numbers in conference papers. [22], [25] have improper name formatting. [20] – “,” instead of “.” is present. Revise the reference formatting carefully.
Response: We revised the reference section and added page numbers for conference papers.
Comment#3: In their literature review authors omitted one of the mainstream approaches to channel prediction – ML-based methods. See, for instance:
- Bogale, X. Wang and L. B. Le, "Adaptive Channel Prediction, Beamforming and Scheduling Design for 5G V2I Network: Analytical and Machine Learning Approaches," in IEEE Transactions on Vehicular Technology, vol. 69, no. 5, pp. 5055-5067, May 2020, doi: 10.1109/TVT.2020.2975818.
Luo, J. Ji, Q. Wang, X. Chen and P. Li, "Channel State Information Prediction for 5G Wireless Communications: A Deep Learning Approach," in IEEE Transactions on Network Science and Engineering, vol. 7, no. 1, pp. 227-236, 1 Jan.-March 2020, doi: 10.1109/TNSE.2018.2848960.
Jiang and H. D. Schotten, "Neural Network-Based Fading Channel Prediction: A Comprehensive Overview," in IEEE Access, vol. 7, pp. 118112-118124, 2019, doi: 10.1109/ACCESS.2019.2937588.
and many other papers on that topic.
Authors should include this class of methods and comment on their relevance (or irrelevance, then why?) to the assumed problem.
Moreover, the Reference section can be sufficiently expanded by including (and discussing) a wide range of modern technical sources on that topic.
Response: We agree that the ML methods (in particular, the Deep Learning methods) were not represented in our initial literature review.
In the revised version, we have included state-of-the-art ML methods (including references provided above). We outlined the advantages and disadvantages of ML methods. In particular, the key disadvantage of ML methods is their high computational complexity and need for a very high amount of training data, which significantly complicates their usage at gNBs with limited computational resources in an online manner.
Additionally, for the better comparison of various existing methods/approaches, we added Table 1, which classifies various approaches and summarizes their key features.
Comment#4: In the literature review, section authors state that the “Proni method … can obtain accurate channel parameters until the environment change from user mobility makes the initial model obsolete” citing
[9] H. Yin, H. Wang, Y. Liu and D. Gesbert, "Addressing the Curse of Mobility in Massive MIMO With Prony-Based Angular-Delay Domain Channel Predictions," in IEEE Journal on Selected Areas in Communications, vol. 38, no. 12, pp. 2903-2917, Dec. 2020, doi: 10.1109/JSAC.2020.3005473.
But in [9] a method of combating this “curse” is proposed. Moreover, in [9] scenarios with moving UE with speeds: 30 km/h, 60 km/h, and even 90 km/h were assumed; while in the current submission authors limit themselves to only 30 km/h. Moreover, in [9] it was demonstrated that the proposed algorithms yielded good results (for fast-moving UE). Authors of the current submission disregard those achievements. This should be commented on and cleared out.
Response:
Thank you for this valuable comment! We changed the description of paper [9] in Section 2.
We acknowledge that the method proposed in [9] can provide high accuracy of channel prediction at high speeds and a high number of antennas. However, the Prony method has high complexity. With the limited computational capabilities at gNB, the model retraining may require sufficient time, which may be not applicable to scenarios with URLLC traffic having very tight latency requirements.
PABAFT uses the Least Squares method to train the AR model, which has significantly lower complexity and, thus, can be executed in real-time at the gNB.
Following reference [9], we extend our performance evaluation results by considering 30 km/h, 60 km/h, and 90 km/h for both link-level and system-level simulations.
Comment#5: It is advised to add a table with simulation parameters, rather than scatter them throughout the paper. This would make the simulation part more solid.
Response: We added Table 2 with the main simulation parameters.
Comment#6: The AR model order parameter (very important for AR performance) is not discussed. Some abuse of notation is present: at first, it is called p, but afterwards, it is renamed K. Moreover, it is unclear how it relates to # of samples? It is critical in view of [9], where authors had claimed that their “method is able to achieve asymptotically error-free prediction, provided that only two accurate channel samples are available”. And in the current submission, one sees from 200 to 500 samples. This should be explained.
Response: Thank you for noting this notation error! We have corrected it. Now, the AR model order is denoted as K. In our experiments, following the results from numerous papers on AR, we set K = 30.
As for the number of training samples S, in Section 3, we explain that we can build a single AR model for multiple antenna pairs and multiple neighboring frequencies. The performance evaluation results show that for various UE speeds S~1000 is enough to train the model (see Fig. 4). Further, in Fig. 5, we show that it is possible to aggregate samples from all antenna pairs (64 in our experiment) and from up to 8 neighboring precoder groups. Thus, S=1000 samples can be obtained during a very short period of time, i.e., two TDD frames or 10 ms. So, the time needed to collect the samples is similar to that in [9].
Comment#7: In Section 4.2 authors compare Last SRS with PABAFT without including information on what a number of samples were used.
Response: Based on the link-level results presented in Fig. 4, we had selected S=1024 samples to use in further experiments. We also added the information about used number of samples in Fig.6.
Comment#8: The description of the results is very brief, especially concerning Fig. 9-11 and can be expanded.
Response: We have significantly revised the Performance evaluation section. First, we added more results for different UE speeds. Second, we extended the analysis of the obtained results.
Comment#9: The conclusion section should be expanded by adding numerical results obtained by the authors.
Response: We revised and extended the Conclusion section: we added the summary of the key obtained numerical results and the discussion on the future research directions.
Reviewer 2 Report
This paper proposes a linear regression model for predicting the channel in massive MIMO based 5G networks. The paper has many drawbacks that prevent me from recommending its publication as follows:
1- The novelty and technical content of this paper are too low, the is no analysis of the proposed scheme and a prove of its effectiveness.
2- The comparisons are done against baseline schemes, and no advanced schemes proposed in literature are involved in comparisons .
3- The paper seems like a technical report more than a scientific paper.
Author Response
Comment#1: The novelty and technical content of this paper are too low, the is no analysis of the proposed scheme and a prove of its effectiveness.
Response: We significantly revised and extended the description of the proposed approach in Section 3 and provided the detailed analysis of related works in Section 2. The key novelties of the proposed approach are as follows. First, we model the channel in each slot between two consequent SRS transmissions with a vector autoregressive model, while other existing approaches model only slots in which SRS is transmitted. Second, to fit the vector autoregressive model, we use the novel feature of the 5G system called flexible TDD. With flexible TDD, it is possible to tune frame structure (i.e., the number of UL/DL slots) to obtain fine-grained measurements. When the model is fitted, the base station can switch back to the legacy frame structure with fewer SRS transmissions and use the model to predict the channel state.
We significantly extended the performance evaluation (Section 4). First, we compared the proposed PABAFT approach with the other advanced channel prediction approaches (see the response to Comment#2). Second, we evaluated the performance of various approaches under various UE speeds. The results show that for high speeds, the performance of all approaches degrades. However, the proposed approach still provides significant gain with respect to other considered approaches. Third, we studied the influence of aggregation of training samples over neighboring frequencies on the performance of the proposed approach. The results show that it is possible to aggregate measurements from several neighboring subbands without loss of model accuracy. Thus, the duration of the training phase can be significantly shortened.
Comment#2: The comparisons are done against baseline schemes, and no advanced schemes proposed in literature are involved in comparisons.
Response: Thank you for your comment! We included in the analysis several advanced channel prediction approaches described in the literature (they are described in Section 2): (i) the “Geodesic” approach, (ii) the “AR-interpolation” approach based on a single-valued autoregressive model and interpolation. We significantly extended our results in Section 4. Both link-level and system-level results show the advantage of the proposed PABAFT approach with respect to other approaches. Besides that, we compare the proposed approach with the upper bound given by the ideal prediction. The results show that for the moderate speed of 30 kmph, PABAFT provides results very close to the upper bound. The gap to the upper bound increases with a higher UE speed, but the proposed approach still provides the results significantly better than other considered approaches.
Comment#3: The paper seems like a technical report more than a scientific paper.
Response: We have significantly revised and extended the paper.
- We revised the literature review section. We classified various existing approaches and analyzed their pros and cons. The summary is provided in Table 1.
- We revised and extended the description of the proposed PABAFT approach in Section 3, highlighting its peculiarities and key novelties.
- We extended the performance evaluation (Section 4). We added more approaches to compare and analyze them under various conditions.
- We enhanced the conclusion section by adding the description of the key obtained numerical results and the discussion on the directions for future research.
Reviewer 3 Report
The paper presents a novel method for channel prediction based on the Autoregression methodology. The simulation results show that the proposed algorithm outperforms the last SRS, while reducing the channel estimation overhead. The below are some notes that aim to improve the paper quality and presentation.
1. The authors sometime uses an initial capital for the acronymous or initial small letters (e.g. user equipment in line 26 and Channel State Information (CSI) in line 25. I suggest to use initial small letters for all ACRONYMS definitions and to be consistent and NOT to define the acronym more than one time (e.g. SNR). The words Ground truth appeared once initial capital and one initial small.
2. The paper needs some typos and format checking, for example, in line 26, (..With the first approach).
3. Can the authors explain more what they mean by rare measurements (line 65)?
4. Can you expand the literature review section to include more recent works on channel status prediction algorithms? Can you also prepare a comparison table to compare between your approach with what have been proposed in the literature?
5. Can you provide more justifications on why the A
6. in addition to the line colouring of the figures' results, can you please add symbols (* or square or dotted line, or dashed lines), so this way, if someone (like me) print the paper on a white/color printer, he will be able to identify the different results and lines?
6. In reality. do we need a training data for each environment? can you please explain more if we want to practically implement this system, what should be done? should we create a new training data for different environments? does the used training data cover wide range of environments/scenarios? will we be able to use the developed model for any environment (i.e. indoor, outdoor with many blocking objects, outdoor with less blocking objects, etc.),
7. What are the cons of the proposed algorithm?
Author Response
Comment#1: The authors sometime uses an initial capital for the acronymous or initial small letters (e.g. user equipment in line 26 and Channel State Information (CSI) in line 25. I suggest to use initial small letters for all ACRONYMS definitions and to be consistent and NOT to define the acronym more than one time (e.g. SNR). The words Ground truth appeared once initial capital and one initial small.
Response: Thank you for your comment! We have fixed all acronyms and deleted all duplicate definitions.
Comment#2: The paper needs some typos and format checking, for example, in line 26, (..With the first approach).
Response: We have significantly revised the text and fixed the format.
Comment#3: Can the authors explain more what they mean by rare measurements (line 65)?
Response: Thank you for the question.
With the PABAFT approach, the TDD frame structure is changed with time. During the so-called training phase, we use the TDD structure shown in Fig.1(c) and measure the channel in each slot to obtain the necessary dataset to fit the AR model. During the inference phase, the downlink slots are prevalent (see the example in Fig.1(a)), and therefore we obtain the measurements less frequently (e.g., each 10th slot). Hence, during the training phase, we have frequent measurements while during the inference phase we have “rare measurements” with a higher period.
We have added the necessary explanation in the text.
Comment#4: Can you expand the literature review section to include more recent works on channel status prediction algorithms? Can you also prepare a comparison table to compare between your approach with what have been proposed in the literature?
Response: We have significantly improved the Literature review section and emphasized the novelty of our approach with respect to the prior art. Specifically, we significantly extended the list of related papers (i.e., by including recent machine learning based approaches). Additionally, we added Table 1 which classifies various approaches and summarizes their key features.
Comment#5: Can you provide more justifications on why the A.
Response: Unfortunately, we cannot understand your comment.
Comment#6: In addition to the line coloring of the figures' results, can you please add symbols (* or square or dotted line, or dashed lines), so this way, if someone (like me) print the paper on a white/color printer, he will be able to identify the different results and lines?.
Response: We added the markers to all plots for better readability.
Comment#7: In reality. do we need a training data for each environment? can you please explain more if we want to practically implement this system, what should be done? should we create a new training data for different environments? does the used training data cover wide range of environments/scenarios? will we be able to use the developed model for any environment (i.e., indoor, outdoor with many blocking objects, outdoor with less blocking objects, etc.).
Response: Thank you for this question. We propose periodically switching between the training and the inference phases. During the training phase, we collect the necessary number of samples to (re)train the AR model. With this approach, even if the environment is changing with time, after each training phase, the AR model is retrained to take into account the new state of the environment. The open question which is the direction of our future research is how to dynamically decide whether we need to run the training phase rather than doing it strictly periodically. For example, for static environments, we do not need to (re)train the model.
Our simulation results presented in Section 4 show that with proper aggregation of measurements from neighboring frequencies, we can provide the duration of the training phase below two frames. If the period for running training is 100 frames, the overhead is only 2%.
Comment#8: What are the cons of the proposed algorithm?
Response: In the current version of the proposed model, we use fixed values of: (i) the number of aggregated PGs (with simulations, we have shown that 8 PGs can be aggregated without loss of accuracy), (ii) the period of running the training phase. In our future works, we will develop methods to adaptively select this parameter depending on the scenario (environment, UE speed, etc.).
We added this direction for future research in Conclusion.
Round 2
Reviewer 2 Report
The authors did all required modifications
