Machine learning for Reliable Communication in 5G and 6G wireless systems
Annapurna Pradhan22:112023
One lost packet in a thousand: why 5G reliability needs machine learning
Remote surgery and autonomous driving depend on wireless links that deliver data within a millisecond with almost no errors, yet the airwaves are noisy, crowded and open to eavesdroppers. PhD researcher Annapurna Pradhan of NIT Rourkela, a best paper award winner in wireless communication research, explains why traditional mathematical models struggle with this complexity. So where does machine learning genuinely help, and where should engineers resist using it?
Learn the 5G service classes, an ML pipeline for wireless data and the limits to respect
Annapurna Pradhan, a PhD research scholar at NIT Rourkela, explains how machine learning can help 5G and 6G networks deliver reliable, low-latency communication. She covers the three 5G service classes, why the noisy and fast-changing wireless medium suits learning models, a machine learning pipeline for cellular networks, and the challenges that remain. She ends by arguing that ML belongs only where complexity justifies it.
“It is the openness of the wireless medium because we are simply sending the data in the wireless medium and collecting it. And this wireless medium is not error free.”
“So we can take this challenge as an opportunity to implement this ML models.”
What you will learn
Distinguish the three 5G service classes (eMBB, mMTC and URLLC) and the reliability and latency each demands.
Explain why the noisy, dynamic wireless medium and exploding device counts push networks beyond traditional mathematical models.
Map a machine learning pipeline onto a cellular network, from problem framing and base station data to training and deployment.
Match supervised, unsupervised and reinforcement learning to wireless tasks such as demand prediction, clustering and resource scheduling.
Anticipate the main obstacles: missing standard datasets, unsuitable metrics, hyperparameter tuning and privacy, where federated learning helps.
Key takeaways from the speaker
Learn how to incorporate Machine Learning in wireless communication
More about 5G and 6G
Discuss about the use cases
Understanding market demand and technological trends for next generation wireless communication
Details
Level
Intermediate
Format
Talk
Technical depth
Introductory technical
For
Student, Early career, Mid-career
Length
22:11
Published
June 2023
Questions this talk answers
What are the three 5G service classes?
Enhanced mobile broadband for high data rates such as video streaming and VR, massive machine type communication for large numbers of IoT sensors with wide coverage, and ultra reliable low latency communication for near error-free delivery within one millisecond.
Why use machine learning instead of traditional mathematical models in wireless networks?
Connected devices, data traffic and network heterogeneity have grown so much that real-time optimization over massive datasets is very hard. Learning models handle complex problems and large datasets well.
How can models be trained when the wireless environment changes so fast?
Pradhan points to active or online learning, where models train as soon as data is generated, and to time series analysis for predicting network traffic over the next few days.
What are the biggest challenges of applying machine learning to 5G?
There are no standardized wireless datasets like those for image classification, evaluation metrics may not transfer, hyperparameter selection is difficult in a dynamic environment, and privacy and security must be protected, for example with federated learning.
Should machine learning be used everywhere in 5G and 6G networks?
No. Pradhan argues ML should be used where complexity arises and where it can actually be utilized.
About this talk
About the Presentation:
Machine Learning (ML) has shown a great potential in revolutionizing communication system worldwide. The ever-increasing complexity of the wireless network, and the emergence of novel use cases such as autonomous cars, industrial automation, virtual reality, e-health, and several intelligent applications, machine learning (ML) is expected to be essential to assist in making the 5G vision conceivable. The cognitive optimization of essential radio resources for efficient end-to-end communication, enabling improvement in conventional communication theories and new architectural development can be facilitated in 5G and 6G by the use of machine learning. Hence, a comprehensive overview of emerging studies on machine learning and deep learning (DL) based physical and MAC layer enhancements in 5G will be presented. To be specific a brief idea regarding the 5G service classes (i.e. eMBB, URLLC and mMTC) and the scope of ML and DL in facilitating the QoS requirements will be focused. In addition to this, the challenges and possible solutions in implementing ML and DL in 5G will be presented. The open research areas will be presented for motivating further developments in enabling 5G and beyond wireless communication.
Building an ML pipeline for wireless networks (89 s)
Don't use ML everywhere (71 s)
Speaker 1: Hello, everyone. Hello, everyone. I'm Anna Purna, PhD Research Scholar from India. So today I will be talking about 5G and 6G communication, basically the reliable communication in 5G and 6G, where machine learning can play a very important role in enabling those futuristic applications. Speaker 1: So let's get started. So first of all, the 5G application, just for the information that what are the 5G applications, why we are calling it as a 5G application, because whatever the applications we are looking around today, those we are approaching an era of futuristic applications like the factory of future autonomous driving intelligent transport system, smart home, smart home and smart city concept.
Speaker 1: So everything what is smart that needs to communicate with each other, the devices needs to communicate with each other, and it requires the wireless communication. So we can see that the application of 5G or 5G can enable these applications, and it can get the connectivity wherever and whenever it is required, with some very futuristic key performance indexes offered by 5G. So how 5G can facilitate these applications, like Speaker 1: Like remote surgery and autonomous driving, how it is connected to the base station, and everything which is connected, that should be reliable. There should not be any lag or glitch, and it should be perfect as 99.999%. So this can be only possible by some unique services, which is offered by 5G. Those are called, there are, it is categorized into three parts. What is called enhanced mobile broadband service, which is developed by 5G, which is developed by 5G.
Speaker 1: So it's called enhanced mobile broadband service, which is called enhanced mobile broadband service, which is this service is characterized by high broadband data like video streaming. Speaker 1: Is the video streaming live actions. So virtual reality augmented reality all these applications. And it enables the communication between devices and human and the smart devices. The next type of application and next type of service class is your massive machine type communication.
Speaker 1: Now IoT applications are the best example of these. So what are the massive machine type like a lot of smart devices that connect with each other through wirelessly. So those smart devices like sensors, they continuously transmit and receive signals in between each other. So for example, if you will take a smart phone system, there are sensors embedded into your home appliances. So they will be monitoring each and every.
Speaker 1: Operation of the appliances and inform you through the mobile app. So that means those sensors are collecting data. Speaker 1: Transmitting the data wirelessly and it is communicated through the mobile app. So there is a wireless communication. So when 5G comes into picture, it makes this more smooth. Speaker 1: By increasing the connectivity. So massive communication is characterized by large coverage.
Speaker 1: Now, the most important part of the 5G service classes offered are. Speaker 1: It is characterized by large coverage and the most important part of the 5G service classes. Speaker 1: Ultral reliable low latency communication. Speaker 1: Ultral reliable means it, as I have already told that it is provides very high reliability. Speaker 1: Reliability of service. Now how it provides the reliability of service.
Speaker 1: That means if I am sending. Speaker 1: A thousand data packets. Speaker 1: I am sending. Speaker 1: A thousand data packets. At most one data packet can be in error and all the 999 data packets. Speaker 1: That packets will be error free. So that that is just an example to make it a reliable low latency.
Speaker 1: That means that means that means that means that means that is a low latency. Speaker 1: That means that means that means that is a low latency. That means that means that is a low latency. Speaker 1: That means that means that is a low latency. That means that is a low latency. Speaker 1: That means whatever communication from the customer.
Speaker 1: This is a low latency. Speaker 1: That means that is a low latency. Speaker 1: That means that means that is a low latency. Speaker 1: That means that is a low latency. So that means that means that is a low latency. Speaker 1: It's very high reliability of service. Now how it provides the reliability of service.
Speaker 1: That means if I am sending a thousand data packets at most one data packet can be in error and all the 999 data packets will be error free. Speaker 1: So that that is just an example to make it a reliable low latency. That means whatever communication from the transmitter to receiver will happen. Speaker 1: That should be within one millisecond time period. So it provides a very faster and reliable support and it is offered by 5G and these applications will be these services will be enabling the next generation like 6G mobile services.
Speaker 1: Now comes to the fundamental of machine learning or AI. Now the artificial intelligence or machine learning Speaker 1: They are just revolutionized all the applications where it has been used, be it from the healthcare to the chemical engineering or any other any other field what we know. Speaker 1: Now, when it is applied to the wireless communication due to the inherent characteristic Speaker 1: of these machine learning models that are suitable for complex problem solving. So we are motivated to all the researchers are motivated to use it for the well enabling wireless communication.
Speaker 1: Now, how it is used and what are the areas that we will. Speaker 1: Now, as everyone knows, there are learning models like supervised and supervised and reinforcement learning model. So quickly I will go through this one. Speaker 1: So in supervised learning model, the output and input are mapped with some rules and the learning models.
Speaker 1: They extract automatic feature extraction from that and they decide the rule and they give the label or it giving you classification or regression type. Speaker 1: Similarly unsupervised where the labels are not there, but still models can learn the patterns and they can predict the new outputs from that. Speaker 1: And reinforcement learning is another which constantly learns from the environment and the agents will learn from the environment and take some action based and changes the states and it maximizes the reward.
Speaker 1: So those are the inherent quality of these learning models. So by understanding the models and understanding the challenges, we can apply it into wireless communication. Speaker 1: Now, as I have already told, this is just an example of showing how the agents learn and changes from one state to other and maximize the reward. So it maximizes reward depending on the correct action.
Speaker 1: Now, ML in 5G. So if we are considering ML, then why we are considering? 5G is the wireless communication. Now, in wireless communication, we have to understand what is the main thing. Speaker 1: Challenge in communication. It is the openness of the wireless medium because we are simply sending the data in the wireless medium and collecting it. And this wireless medium is not error free. It is a noisy wireless medium. And there are Speaker 1: Eva's droppers or there are interception in between interference can happen in between. So this characteristic of the wireless network is very different and difficult and it is a dynamic in nature and we cannot predict the region. So in those conditions, this machine learning models can, as we know,
Speaker 1: that they are very good in complex problem solving and they can automatically extract the features from the given data patterns. So in those conditions, we have seen the advantage or the use of ML in other other applications. Speaker 1: Like computer region, they are successfully utilized. Now, in case of wireless communication, we will define the challenge accordingly and utilize the expertise of learning models so that it can, it can give good results.
Speaker 1: Now, what, why we are going for machine learning models. Previously, there was, it was not yet. Speaker 1: Traditional mathematical models were used and it, it was giving good results. But why we are trying to do it because of complexity, complexity in everything like before 10 years, if you will think about the number of devices, which are connected in 4G and 3G and now in 5G, it's just increasing exponentially.
Speaker 1: So that means the traffic. So that means the traffic, which was generated before 10 years, the data traffic. Now it has increased multiple and the IoT devices, which were there before some years, some decades. Now it has been increased. Now that is the first difficulty. Second, the massive data set, which is generated by these devices, how to handle that, where to store, how to get the optimized value from that.
Speaker 1: So real-time optimization becomes very difficult and it becomes challenging. So on all those things previously, the network was simpler with small, with a very less number of cellular towers and users. Now it, it is heterogeneous with all the small scales and large scale networks. Now, due to this challenges, we are very much interested in using this machine learning models, which can handle these Speaker 1: large data sets and give the optimized result as per hour challenge. So we can take this challenge as an opportunity to implement this ML models. Now, as I have told that what is the big question? The big question is how to train these models with this dynamically changing environment, which is very fast. If a car is running very fast, then how we can transmit the data? Or with the mobility,
data or with the mobility how we can transmit a data reliability to these conditions so those are very challenging condition we have to understand each and every challenge in those things it can be done by some of the methods for example it is active learning that means learning on the go online learning you can think as soon as the data is generated the models will be trained on those datas and so that it can it can predict for the future for example uh time series analysis what we are doing for the network traffic prediction so for network traffic prediction if we are observing for some days then we can predict what can be for the next few days few couple of days similarly uh there are different kind of learning methods can be utilized to train those models in such environment
now i will uh i will quickly go through what is the machine learning machine learning pipeline as everyone knows that this machine learning pipeline uh depends on the problem statement like we cannot apply every we cannot apply machine learning to every problem that we are witnessing in wireless communication we have to understand the challenge whether machine learning models are utilized or whether it will give the good result or not or whether the problem is worth of utilizing those learning models so first understanding of the problem is the most important part then after understanding your problem then the related data collection to train the model is done but whatever data we are getting from this wireless environment all those are raw data so we need to curate those data according to our requirement and the data
data storage also we have to think about whether it is cached at the edge or at the central server and everything so all those uh whether it will be a centrally uh monitored data collection or decentralized data collection so we need to uh uh we need to uh we need to brainstorm into that and collect the data data we have to clean the data we have to clean the data and we have to transfer the data into a usable format then comes to the model training and after model training we at the due course of training we need to uh keep in mind some of the important factors like uh which model to be used which hyper parameters to be used then it is applied to the final stage like real-time data to be used in the
keep in mind some of the important factors like which model to be used, which hyperparameters to be used, then it is applied to the final stage like real-time applications. Now, this is just a basic, whatever I have told about how a machine learning pipeline is implemented to a wireless network, this is just an example. Now, here we can see that measurement data we can take directly from the cellular base stations or any drive test what we are doing or from any network equipment what we are using to monitor the spectrum, monitor the wireless channel. So, from there, we collect the data and we can store it at the data centers and use the data for the data leveling. Now, after the leveling, it is split and
given as the training and testing data. This is a supervised learning model. So, in this case, we are collecting from various different sources and then we use it for our problem statement. Then the model training, be it a classification model or a regression model, we can use it as a model training. Then we test and validate the model accordingly. Now, this is the pipeline of the wireless network. So, this is the pipeline in a wireless network, how we can use it. Similarly, what we can get when we are optimizing the things, for example, resource allocation, resource means frequency time allocation to the data what we are transmitting. So, frequency time power allocation is a very difficult task for the mobile network operator. Now, in those cases, if he knows about
what is the network condition, then they can collect the data. After collecting the data, they will classify or predict some of the values that after, for example, after certain time period, there will be search in the network usage. So, accordingly, they will pre-allocate some of the resources so that the reliable communication can happen. When there is a request connection request, the resources can be easily given so that there will be no loss and communication will happen without any drop. So, after that, it is used into the knowledge exploration and utilized in the real-time environment.
Now, the opportunities offered by ML in wireless communication, yes. It enables high data rate, which is very much important for the applications like HD video streaming and everything because it learns the statistics from the wireless channel. It predicts the wireless channel so that we can eliminate the noise or we can eliminate the effect of interference in advance. So, eliminating that, we can reallocate the resources or reserve the resources. For example, in case of network slicing, then there will be no packet drop and data rate will be increased or the throughput can be increased. Similarly, capacity improvement by connecting different devices or it can increase by using the planning like self-organizing networks, it can improve. Similarly, connection density will improve if the reliable communication will improve and it can also adapt to mobile environment. Like for example, if it is changing, if wireless
communication happening through an environment and these machine learning models have the capability to adapt to this changing environment. Like if some environment or if some signal to noise ratio or SNR condition is low, it automatically can increase the power from the cellular base station. So, that packet will not drop, data will not drop. So, power optimization can be done. So, it is called the adaptability is also possible through this one.
And of course, it gives you low latency communication because all the network prediction and everything are there, can be done proactive. Like before, the process will operate. It can proactively provide the network information depending on the data which we have collected in the past. For example, in supervised case, in machine learning models, how it is applicable to the wireless communication. For energy efficient solutions, for utilizing the energy harvesting, then bandwidth allocation like frequency allocation, depending on the number of users present at various time of the day, it can allocate the resources or bandwidth accordingly. If at the night time, it is high bandwidth communication, then definitely it will pre-allocate to these positions. For example, network demand prediction, as I have told, time series analysis can be done and classifier models can be used from this one.
Then, the unsupervised learning model, yes. Basically, clustering models can be used to have the data, extract the relationship between different data points that we are collecting and self-organizing network, of course, self-organizing map. So, these applications, these learning models are applied to various different applications. That means the problems, their challenges are there in the wireless communication, which cannot be solved by the conventional method. They can be easily solved. They can be easily modeled with this machine learning models. Now, again, the reinforcement learning, like it, it is very good in resource allocation, scheduling, scheduling of high priority data. For example, we are having various applications like autonomous driving. For autonomous driving, the data which is coming is very high priority data. So, it gives, it requires very much essential inputs from the reinforcement learning, like it adopts and
allocates according to the service requirement. Now, as I have told, it is always not beneficial. Also, there are some challenges when we are applying machine learning models to 5G. And definitely, there will be challenges, challenges like data collection. As there are some standard data for image classification and everything, but in case of wireless communication, we don't have those standardized data. If they are collected from some of the companies or some of the service providers, that is also not available. Similarly, for example, evaluation metric. In the output evaluation metric, we are evaluating the accuracy or the means to error or everything, all those values. So, that may not be possible, that may not be good for other applications also. So, in those conditions, those challenges are inherent. And we need to think about those challenges and
model and design the model or machinery model accordingly. Now, there are other challenges like hyperparameter tuning and privacy and security. So, whenever we are using this one, before, as I have told, as I have told, the wireless environment is very dynamic. So, this hyperparameter selection before training process also require how many epochs we have to take, what is the learning rate that we have to pre-define those things. So, according to the model requirement, we have to reduce that.
Now, the last thing that I want to share for reliable communication, the most important factor, most important factor is the privacy and security. So, for that, we need to go through this different kind of model, called federated learning model, which provides the security at different levels. Similarly, 6G applications, as I have told these, all the 6G applications require the machine learning model very much and it gives you the boost to this application in future. There are some of the key KPIs for these futuristic applications. So, if we will move forward, we will get these applications to be done and AI will play an important role in all of this. Now, last, we know that machine learning can provide a very good solution, but still some challenges are there. So, our aim is not to use ML everywhere. The aim should be,
we will use ML where the complexity arises and where it can be utilizable. So, thank you everyone and I would like to take some questions from the audience. I will share my LinkedIn. You can connect me on LinkedIn for any doubt or anything. I will share my LinkedIn. You can connect me on LinkedIn. You can connect me on LinkedIn.
Transcribed automatically; names checked against the speaker record. Select a time to jump there.
Article based on this talk
Where machine learning makes 5G and 6G networks more reliable
Researcher Annapurna Pradhan shows how learning models tame a noisy, crowded wireless medium, and why they should not be used everywhere.
In short
Machine learning can help 5G and 6G networks meet extreme demands for speed, reliability and low latency, argues Annapurna Pradhan, a PhD Research Scholar at the National Institute of Technology, Rourkela. She walks through the 5G service classes, explains why the wireless medium is such a hard problem, maps a machine learning pipeline onto a cellular network, and closes with a caution: use machine learning only where the complexity justifies it.
Key takeaways
Learn how to incorporate machine learning in wireless communication.
Learn more about 5G and 6G.
Discuss the use cases.
Understand market demand and technological trends for next-generation wireless communication.
Three service classes behind the 5G promise
Pradhan opens with the applications that make 5G necessary: factories of the future, autonomous driving, intelligent transport systems, smart homes and smart cities. All of these devices must talk to each other wirelessly, and for applications such as remote surgery the connection must be close to perfect, which she describes as 99.999% reliable, with no lag or glitch.
5G meets these needs through three service classes. Enhanced mobile broadband (eMBB) carries high data rates for video streaming, virtual reality and augmented reality. Massive machine type communication (mMTC) connects large numbers of IoT sensors, such as those embedded in home appliances that report to a mobile app, and is characterized by large coverage. Ultra reliable low latency communication (URLLC) is the most demanding of the three. She illustrates its reliability target in plain terms:
"That means if I am sending a thousand data packets at most one data packet can be in error and all the 999 data packets will be error free." (3:29)
Communication from transmitter to receiver should also happen within one millisecond, and she positions these services as the foundation for 6G mobile services.
Why the wireless medium needs machine learning
After a quick refresher on supervised, unsupervised and reinforcement learning, Pradhan turns to the core difficulty of wireless communication:
"It is the openness of the wireless medium because we are simply sending the data in the wireless medium and collecting it. And this wireless medium is not error free." (6:19)
Noise, interference and eavesdroppers make the channel dynamic and hard to predict. Traditional mathematical models served earlier generations well, she notes, but the number of connected devices has grown exponentially from 3G and 4G to 5G, data traffic has multiplied, and networks have become heterogeneous mixes of small and large cells. Real-time optimization over the resulting massive datasets becomes very hard. Her framing is optimistic:
"So we can take this challenge as an opportunity to implement this ML models." (8:44)
The big open question, she says, is how to train models in an environment that changes as fast as a moving car. She points to active or online learning, where models train as soon as data is generated, and to time series analysis that observes network traffic for some days to predict the next few.
A machine learning pipeline for a cellular network
Pradhan stresses that not every wireless problem is worth a learning model, so the pipeline begins with understanding the problem. Data then comes from cellular base stations, drive tests or equipment that monitors the spectrum and the wireless channel. That raw data has to be curated, cleaned and transformed into a usable format, and teams must decide whether it is cached at the edge or stored on a central server, and whether collection is centralized or decentralized. Training involves choosing the model and its hyperparameters, followed by testing, validation and use in real-time applications.
Her worked example is resource allocation: assigning frequency, time and power to transmitted data, which she calls a very difficult task for mobile network operators. If an operator can predict a surge in network usage, it can pre-allocate resources so that connection requests are served without loss or dropped communication.
What machine learning offers, and where it struggles
Because models learn the statistics of the wireless channel, they can anticipate noise and interference in advance. Pradhan links this to higher data rates, network slicing without packet drops, capacity gains through self-organizing networks and better connection density. Models can also adapt to changing conditions:
"Like if some environment or if some signal to noise ratio or SNR condition is low, it automatically can increase the power from the cellular base station." (15:43)
She maps learning types to tasks. Supervised models support energy-efficient solutions, bandwidth allocation by time of day and network demand prediction. Unsupervised clustering and self-organizing maps find relationships in collected network data. Reinforcement learning suits resource allocation and the scheduling of high-priority traffic, such as data for autonomous driving.
The challenges are real, and data comes first:
"As there are some standard data for image classification and everything, but in case of wireless communication, we don't have those standardized data." (18:35)
Common evaluation metrics may not suit every application, hyperparameters such as epochs and learning rate are hard to set in a dynamic environment, and privacy and security matter most for reliable communication. For that last challenge she points to federated learning, which provides security at different levels.
Use machine learning where complexity demands it
Pradhan closes by looking ahead to 6G, where she expects AI to play an important role in reaching the key performance indicators of futuristic applications. Her final message is a disciplined one:
"Now, last, we know that machine learning can provide a very good solution, but still some challenges are there. So, our aim is not to use ML everywhere." (20:19)
Instead, she argues, machine learning belongs where complexity arises and where it can genuinely be put to use.
Watch the talk
Machine learning for Reliable Communication in 5G and 6G wireless systems, by Annapurna Pradhan (22 minutes): https://www.womentech.net/video/machine-learning-reliable-communication-in-5g-and-6g-wireless-systems
About the speaker
Annapurna Pradhan is a PhD Research Scholar at the National Institute of Technology, Rourkela, India, whose research contributes to next-generation wireless communication, machine learning and quantum communication. She has won a best paper award and the Interscience young investigator award, consults for start-ups on machine learning-based microservices and is chief mentor of the machine learning club at NITR.
Questions this talk answers
What are the three 5G service classes?
Why use machine learning instead of traditional mathematical models in wireless networks?
How can models be trained when the wireless environment changes so fast?
What are the biggest challenges of applying machine learning to 5G?
Should machine learning be used everywhere in 5G and 6G networks?
Drafted with AI from the talk transcript. Editor review pending.
Frameworks and ideas mentioned
eMBB, mMTC and URLLC 5G service classes
Supervised learning
Unsupervised learning
Reinforcement learning
Federated learning
Online (active) learning
Time series analysis
Self-organizing networks
Self-organizing maps
Network slicing
Cite this talk
Annapurna Pradhan. "Machine learning for Reliable Communication in 5G and 6G wireless systems." WomenTech Network, 2023. https://www.womentech.net/video/machine-learning-reliable-communication-in-5g-and-6g-wireless-systems