Machine Learning-Based Solutions for Handover Decisions in Non-Terrestrial Networks
Round 1
Reviewer 1 Report
In this work, the authors propose solutions to make handover decisions using ML in NTN. Are these decisions or solutions made on board the payload? What difference is there with the processing techniques, or do they coincide with the methods proposed in the following work?
Ortiz, F.; Monzon Baeza, V.; Garces-Socarras, L.M.; Vásquez-Peralvo, J.A.; Gonzalez, J.L.; Fontanesi, G.; Lagunas, E.; Querol, J.; Chatzinotas, S. Onboard Processing in Satellite Communications Using AI Accelerators. Aerospace 2023, 10, 101. https://doi.org/10.3390/aerospace10020101
They should carry out a more complete state of the art that allows understanding of the need and advantages of this work. How is the training data obtained to carry out the networks with ML? Which of the ML techniques is more suitable needs to be clarified at the end of the paper.
Author Response
Response to Reviewer 1 Comments
Point 1: In this work, the authors propose solutions to make handover decisions using ML in NTN. Are these decisions or solutions made on board the payload? What difference is there with the processing techniques, or do they coincide with the methods proposed in the following work?
Ortiz, F.; Monzon Baeza, V.; Garces-Socarras, L.M.; Vásquez-Peralvo, J.A.; Gonzalez, J.L.; Fontanesi, G.; Lagunas, E.; Querol, J.; Chatzinotas, S. Onboard Processing in Satellite Communications Using AI Accelerators. Aerospace 2023, 10, 101. https://doi.org/10.3390/aerospace10020101
They should carry out a more complete state of the art that allows understanding of the need and advantages of this work. How is the training data obtained to carry out the networks with ML? Which of the ML techniques is more suitable needs to be clarified at the end of the paper.
Response 1:
Thank you very much for taking the time to read and comment on our manuscript and for providing an interesting recent article about Onboard Processing in Satellite communication.
The architecture NTN presented in Figure 1 is based on a regenerative payload, since the regenerative payload is possible with on-board processing (OBP), then the proposed handover decisions in a non-terrestrial network will be made onboard. In addition, since on-device machine learning can be used to perform inference with models directly on, we think that once the model is trained then it can be deployed on a device.
The main goal of the recommended paper “Onboard Processing in Satellite Communications Using AI Accelerators” is about identifying the key requirements for onboard AI/ML processing, and defining a reference architecture, especially what type of chipset could be suitable for onboard-based satellite payload. However, our paper focuses on handover decision-making in NTN by simulating the user and satellite communication through a service link.
We have updated the manuscript in response to the reviewers' suggestions and criticisms as follows:
We have added content to the section “4.2 Performance result of the classification Algorithm” by explaining machine learning model inference.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
A very disappointing article, because it promised to show a new use of Machine Learning and a new concept of "hand-over in reverse", which at no time manages to convey the validity of what is presented.
The easiest part of evaluating the format of the text: it is practically impossible to understand an explanation of something complex when said explanation is supported by graphics that are three pages before or after the text to which they refer!!!
Only with the numbering of figures and tables, the first recommendation is that the authors review in detail... There are figures and tables with bad references, with errors in the number or even without a number (???) In addition, there are figures displaced up to three pages with respect to the reference in the text... Not to mention that some of the figures either have a questionable quality (they look blurry) or do not contribute anything (the figure numbered 9 looks like a blur, and nothing is explained what does it mean, or what does it represent...UE positions? What were the UEs everywhere inside the NTN beam?)
About the study... poor explanation of dataset!!! The most important thing in ML!!! UE's are decrypted in a very general way... Are all of them static??? Only the NTN moves??? Handover is always consider from the point of view of NTN placement??? How have the positions of the EUs been generated? What mobility model has been used? Very poorly explained!
Why do you use a KNN and then skip to typical ML models??? What are the parameters used in the different ML models and why? How many layers are assumed in the case of the neural model and why?
Honestly, once I reach the conclusions, I still have no idea what was sought in the article, because I don't understand the point of making a classification of the static UE positions, no matter how much the NTN moves, which By the way, it only considers a cluster of 3 nodes... The reasoning supposedly given by the authors is not well explained...
Author Response
Response to Reviewer 2 Comments
Point 1: A very disappointing article, because it promised to show a new use of Machine Learning and a new concept of "hand-over in reverse", which at no time manages to convey the validity of what is presented.
The easiest part of evaluating the format of the text: it is practically impossible to understand an explanation of something complex when said explanation is supported by graphics that are three pages before or after the text to which they refer!!!
Only with the numbering of figures and tables, the first recommendation is that the authors review in detail... There are figures and tables with bad references, with errors in the number or even without a number (???) In addition, there are figures displaced up to three pages with respect to the reference in the text... Not to mention that some of the figures either have a questionable quality (they look blurry) or do not contribute anything (the figure numbered 9 looks like a blur, and nothing is explained what does it mean, or what does it represent...UE positions? What were the UEs everywhere inside the NTN beam?)
About the study... poor explanation of dataset!!! The most important thing in ML!!! UE's are decrypted in a very general way... Are all of them static??? Only the NTN moves??? Handover is always consider from the point of view of NTN placement??? How have the positions of the EUs been generated? What mobility model has been used? Very poorly explained!
Why do you use a KNN and then skip to typical ML models??? What are the parameters used in the different ML models and why? How many layers are assumed in the case of the neural model and why?
Honestly, once I reach the conclusions, I still have no idea what was sought in the article, because I don't understand the point of making a classification of the static UE positions, no matter how much the NTN moves, which By the way, it only considers a cluster of 3 nodes... The reasoning supposedly given by the authors is not well explained...
Response:
Thank you very much for taking the time to read and comment on our manuscript.
We have updated the manuscript in response to the reviewers' suggestions and criticisms as follows:
Point 1. A: Only with the numbering of figures and tables, the first recommendation is that the authors review in detail... There are figures and tables with bad references, with errors in the number or even without a number (???)
Response 1. A:
We have updated the Figures and Tables’ numbering.
Point 1. B: In addition, there are figures displaced up to three pages with respect to the reference in the text... Not to mention that some of the figures either have a questionable quality (they look blurry) or do not contribute anything (the figure numbered 9 looks like a blur, and nothing is explained what does it mean, or what does it represent...UE positions? What were the UEs everywhere inside the NTN beam?)
Response 1. B:
We organized the order of the figure according to their explanation in the text, we have already explained figure 9 and what it means in the section “4.1 Performance of the proposed clustering Algorithm”, and we also replaced figure 9 with a clear picture and updated its numbering. UEs (user equipment) position is the position of the users at a given time. The users are everywhere in the NTN beam, and in our proposed architecture, the cells are overlapped. For convenience, we replaced UEs with UE. In the NTN Beam, UE is the user equipment that is connected to the Satellite through a service link.
Point 1.C: About the study... poor explanation of dataset!!! The most important thing in ML!!! UE's are decrypted in a very general way... Are all of them static??? Only the NTN moves??? Handover is always consider from the point of view of NTN placement??? How have the positions of the EUs been generated? What mobility model has been used? Very poorly explained!
Response 1. C:
We added more details about the generated dataset and added a Figure showing the UE Dataset. We consider the simulation in the earth-moving beams scenario; in NTN we have earth-fixed cells/beams and earth-moving cells; however, in our study, we did not consider earth-fixed beams. Therefore, we consider earth-moving cells where cells are moving on the ground, so the NTN platforms (LEO-based satellite), and the beams are all moving but the UE is static.
Point 1.D: Why do you use a KNN and then skip to typical ML models??? What are the parameters used in the different ML models and why? How many layers are assumed in the case of the neural model and why?
Response 1. D:
We did not use KNN but we have typical ML classifier models such as logistic regression, naive Bayes, support vector machine, and random forest. In addition, we built a classifier using a neural network. We have only described the parameter of the random forest algorithm, in Section “4.2. Performance of the classification Algorithm”, because it outperformed all the models in terms of accuracy. We have also described the hyperparameters for the neural network (21 neurons in the input layer, 2 hidden layers with 4 neurons each, and the output layer). We used a simple neural network because the number of features in the dataset was 21, so it can be easily computed by that simple neural network.
Point 1. E: Honestly, once I reach the conclusions, I still have no idea what was sought in the article, because I don't understand the point of making a classification of the static UE positions, no matter how much the NTN moves, By the way, it only considers a cluster of 3 nodes... The reasoning supposedly given by the authors is not well explained...
Response 1. E:
In the NTN, there are a lot of challenges due to the movement of the Satellite and the beams, among those challenges, the signaling storm created by handing over all UEs in a cell to a new cell is a big problem. Therefore, all users in a cell are expected to experience a change of cell due to handover every few seconds. In this case, UE location is a crucial parameter in a serving. In our study, we propose a machine learning-based Handover decision in non-terrestrial networks for intra-satellite handover in order to reduce signaling storm handover by first, grouping users with certain similarities; then using the feature distance, we compute the distance between UE and its cell center. The UE is static but the NTN platforms and the beams are all moving. The use of a classification algorithm is that UE far from the cell center is ready to hand over and UE close to its cell center is not ready to hand over, this situation gives us a binary classification problem. Moreover, we don’t have 3 nodes but we designed 3 types of cells in our simulation (Ref. Figure 3). Serving cell, Target cell, and Candidate cell.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have misdirected my comments from the previous round. Requested clarifications are not covered.
Furthermore, the methods proposed in this manuscript are not justified from the point of view of the training data. In their response, the authors say that decisions are made by simulation. This is not reliable for a real handover scenario, a method cannot be trained by simulation, so this proposal requires more work to be clarified and published.
Unfortunately, the work does not require sufficient quality to be published due to a lack of explanation, justification, analysis of results, and comparison with the background to stand out from existing works. For this reason, I cannot recommend its publication.
Author Response
Response to Reviewer 1 Comments
Point 1: Furthermore, the methods proposed in this manuscript are not justified from the point of view of the training data. In their response, the authors say that decisions are made by simulation. This is not reliable for a real handover scenario, a method cannot be trained by simulation, so this proposal requires more work to be clarified and published.
Unfortunately, the work does not require sufficient quality to be published due to a lack of explanation, justification, analysis of results, and comparison with the background to stand out from existing works. For this reason, I cannot recommend its publication
Response 1:
Thank you very much for taking the time to read and comment on our manuscript again.
In our comments, we did not say that the decision is made by the simulation but we said that we collected the data from the simulation. We consider the simulation in the earth-moving beams scenario; in NTN we have earth-fixed cells/beams and earth-moving cells; however, in our study, we did not consider earth-fixed beams. Therefore, we consider earth-moving cells where cells are moving on the ground, so the NTN platforms (LEO-based satellite), and the beams are all moving but the UE is static. In the NTN, there are a lot of challenges due to the movement of the Satellite and the beams, among those challenges, the signaling storm created by handing over all UEs in a cell to a new cell is a big problem. Therefore, all users in a cell are expected to experience a change of cell due to handover every few seconds. In this case, UE location is a crucial parameter in a serving. In our study, we propose a machine learning-based Handover decision in non-terrestrial networks for intra-satellite handover in order to reduce signaling storm handover by first, grouping users with certain similarities; then using the feature distance, we compute the distance between UE and its cell center. The UE is static but the NTN platforms and the beams are all moving. The use of a classification algorithm is that UE far from the cell center is ready to hand over and UE close to its cell center is not ready to hand over, this situation gives us a binary classification problem. Moreover, we don’t have 3 nodes but we designed 3 types of cells in our simulation (Ref. Figure 3). Serving cell, Target cell, and Candidate cell.
We have updated the manuscript in response to the reviewers' suggestions and criticisms as follows:
Massive amounts of data are being produced while 5G networks are growing more sophisticated. To effectively manage the 5G networks, data-driven management, and AI solutions are essential. Research on AI-enabled NG-RAN was carried out by 3GPP for Version 17. The study examined the high-level principles, functional framework, potential use cases, and related solutions for AI-enabled RAN intelligence. For a selection of AI-based use cases, including as network energy conservation, load balancing, and mobility optimization, 3GPP Release 18 will specify data collection upgrades and signaling support.
In addition to enhancing current data collection features, 3GPP Release 18 will look at how AI approaches might increase air interface capabilities. Moreover, 3GPP Release 18 will improve NR data collecting in the context of the self-organizing network (SON)/minimization of drive testing (MDT). By automating RAN planning, configuration, management, optimization, and healing, SON reduces the need for human involvement. MDT gives operators the ability to set up typical UEs to gather and submit measurement data, reducing the need for traditional drive testing. The work on Release-18 will take care of SON features left over from Release 17 and data gathering for random access channel (RACH) optimization.
Author Response File: Author Response.pdf
Reviewer 2 Report
This new version presents some changes that improve the result a little.
The ordering and numbering of the graphics now seems more correct, and the quality of some of the modified figures is appreciated.
Regarding the corrections in the text, the most there are changes in the capital letters of the acronyms, and the structure of the original article has not been touched.
The explanations that have been given in some points, regarding the comments of the review, are correct, although they still leave some gaps in some parts of the paragraphs... I agree with some of the explanations of the authors, although I do not I am clear that what has been written now continues to make a new reader obtain the same conclusions as those I had in my first reading.
I am still not clear that the "mobility model" used is sufficient for the proposed simulation, but it is a matter of opinion rather than correction.
Author Response
Response to Reviewer 2 Comments
Point 1: The ordering and numbering of the graphics now seems more correct, and the quality of some of the modified figures is appreciated.
Regarding the corrections in the text, the most there are changes in the capital letters of the acronyms, and the structure of the original article has not been touched.
The explanations that have been given in some points, regarding the comments of the review, are correct, although they still leave some gaps in some parts of the paragraphs... I agree with some of the explanations of the authors, although I do not I am clear that what has been written now continues to make a new reader obtain the same conclusions as those I had in my first reading.
I am still not clear that the "mobility model" used is sufficient for the proposed simulation, but it is a matter of opinion rather than correction
Response:
Thank you very much for taking the time to read and comment on our manuscript again.
We have updated the manuscript in response to the reviewers' suggestions and criticisms as follows:
The mobility management in the non-terrestrial network is very different from the traditional network or terrestrial; In NTN, based on the architecture, we have many types of handover such as inter-satellite handover, intra-satellite handover, and inter-access network handover. However, in our study, we focus on an intra-satellite handover; in our proposed architecture, both cells/beams are served by the same satellite, and no other satellite is involved in the handover process. In LEO satellite systems, intra-satellite handovers are the most typical kind of handovers encountered because of the small area covered by beams and the rapid satellite speed. Thus, we can consider the user mobility negligible compared to high satellite speed. Mobility management of LEO satellites is therefore much more challenging than GEO or MEO systems. With a few exceptions, terrestrial network systems and LEO satellite systems have mobility that is somewhat comparable. In both systems, the relative position between the cells and the UE changes continuously, requiring the handover of the UE between adjacent cells. In terrestrial network systems, the UE moves through the cells, while in LEO systems the cells move through the UE. The cell size of LEO satellite systems is larger compared to terrestrial network systems. Moreover, the speed of the UE can be ignored in LEO satellite systems, since that speed is negligible compared to the rotational speed of the LEO satellite.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
Unfortunately, the quality of the article has not improved in any round. The authors have not directed any of my comments correctly in any of the rounds, even refusing to include proposals for improvement.
The background of the paper is not addressed adequately, not justifying this work compared to other existing ones.
In addition, the title they propose is misleading with the content. It is a very generic title in which several solutions can be expected, while the authors only propose one and it is based on k-means, which is a well-known method and not new to propose.
This work requires a lot of content to clarify. After two rejections, and neither including my recommendations, I am not going to change my opinion about this version. The paper needs to be rewritten and do more simulations and work. I don't see this as the right solution for all the challenges that arise in NTN handover. It is not enough to add a couple of paragraphs as they have done in the previous rounds. Most of the formulations are trigonometric formulas that do not add novelty. The ML formulas are classic, they don't add anything new either. All this makes the mathematics about the article, challenging importance and subtracts content.
As previously stated, the contribution of the paper is not sufficient to be published
Author Response
Response to Reviewer 1 Comments
Point 1: Unfortunately, the quality of the article has not improved in any round. The authors have not directed any of my comments correctly in any of the rounds, even refusing to include proposals for improvement.
The background of the paper is not addressed adequately, not justifying this work compared to other existing ones.
In addition, the title they propose is misleading with the content. It is a very generic title in which several solutions can be expected, while the authors only propose one and it is based on k-means, which is a well-known method and not new to propose.
This work requires a lot of content to clarify. After two rejections, and neither including my recommendations, I am not going to change my opinion about this version. The paper needs to be rewritten and do more simulations and work. I don't see this as the right solution for all the challenges that arise in NTN handover. It is not enough to add a couple of paragraphs as they have done in the previous rounds. Most of the formulations are trigonometric formulas that do not add novelty. The ML formulas are classic, they don't add anything new either. All this makes the mathematics about the article, challenging importance and subtracts content.
As previously stated, the contribution of the paper is not sufficient to be published
Response:
Thank you very much for taking the time to read and comment on our manuscript again.
We have updated the manuscript in response to the reviewers' suggestions and criticisms as follows:
We think that the background of this work is justified because in section 1 “the Introduction”, we presented a detailed description of the NTN and mobility management. Moreover in the section “Related Work”, we presented some works which use machine learning and deep learning for handover solutions in terrestrial networks. However, the mobility management in the non-terrestrial network is very different from the traditional network or terrestrial; In NTN, based on the architecture, we have many types of handover such as inter-satellite handover, intra-satellite handover, and inter-access network handover. Therefore, in our study, we focus on an intra-satellite handover; and in our proposed architecture, both cells/beams are served by the same satellite, and no other satellite is involved in the handover process. In LEO satellite systems, intra-satellite handovers are the most typical kind of handovers encountered because of the small area covered by beams and the rapid satellite speed. Thus, we can consider the user mobility negligible compared to high satellite speed. Mobility management of LEO satellites is therefore much more challenging than GEO or MEO systems. With a few exceptions, terrestrial network systems and LEO satellite systems have the mobility that is somewhat comparable. In both systems, the relative position between the cells and the UE changes continuously, requiring the handover of the UE between adjacent cells. In terrestrial network systems, the UE moves through the cells, while in LEO systems the cells move through the UE. The cell size of LEO satellite systems is larger compared to terrestrial network systems. Moreover, the speed of the UE can be ignored in LEO satellite systems, since that speed is negligible compared to the rotational speed of the LEO satellite.
The 3GPP work is being done to create 5G wireless technology to support non-terrestrial satellite networks and they did not provide a dataset for UE information. In addition to enhancing current data collection features, 3GPP Release 18 will look at how AI approaches might increase air interface capabilities. Moreover, 3GPP Release 18 will improve NR data collecting in the context of the self-organizing network (SON) by automating RAN planning, configuration, management, optimization, and healing, SON reduces the need for human involvement. Minimization of drive testing (MDT) gives operators the ability to set up typical UEs to gather and submit measurement data, reducing the need for traditional drive testing.
In our study, we generated a dataset by simulating communication between users and LEO-satellite, and proposed an intra-satellite handover decision in NTN and not in TN. In the NTN, there are a lot of challenges due to the movement of the Satellite and the beams, among those challenges, the signaling storm created by handing over all UEs in a cell to a new cell is a big problem. Therefore, all users in a cell are expected to experience a change of cell due to handover every few seconds. In this case, UE location is a crucial parameter in a serving. In our study, we propose a machine learning-based Handover decision in non-terrestrial networks for intra-satellite handover in order to reduce signaling storm handover by first, grouping users with certain similarities; then using the feature distance, we compute the distance between UE and its cell center.
We did not only propose a K-means algorithm but also a classification algorithm; the use of a classification algorithm is that UE far from the cell center is ready to hand over and UE close to its cell center is not ready to hand over, this situation gives us a binary classification problem. Moreover, we designed 3 types of cells in our simulation (Ref. Figure 3). Serving cell, Target cell, and Candidate cell.
Author Response File: Author Response.pdf
