Skip navigation
Like what you’re reading?

Pioneering within graph neural networks for increased optimization of networks

Artificial Intelligence is opening new possibilities in optimizing cellular networks, with Ericsson’s Graph Neural Networks (GNNs) being a pioneering example. This breakthrough technology swiftly solves uplink interference and offers several other benefits such as precise predictions and efficient optimization, seamlessly integrating top-tier research into real-world telecom solutions. 

Data Scientist working in Graph Neural Networks

Pioneering within graph neural networks for increased optimization of networks

Data Scientist working in Graph Neural Networks

Data Scientist working in Graph Neural Networks

Cellular network technology is globally used by mobile phones for transmitting and receiving information. They are activated by cells, that are just geographical areas with one or more antennas with common configuration parameters. We can have several problems in these cells, for example, the uplink interference that needs to be resolved for the entire network to be running smoothly. One of the challenges, however, is that optimizing the network with an expert means always having a domain specialist available. On the other hand, using traditional machine learning algorithms such as Reinforcement Learning can take many weeks and degrade the customer network in the process. Moreover, the optimization process is complicated since each cell has its own parameters that modify its own behavior, but when changed, they can affect their neighbours’ behavior, even degrading their performance. A Graph Neural Network can resolve these issues by simulating the behavior of a real network.  

What are Graph Neural Networks and why are we using them? 

Graph Neural Networks is a new AI technology developed over the last few years that has a major impact in certain domains such as social networks, telecommunications networks, and drug discovery. Examples of this can be the prediction of a connection between users, forecasting the future value of a KPI, or predicting if a drug will have a certain property. All these examples use neural networks that receive a graph as an input. Internally, they predict a certain value for each node of the graph by considering its own information and that of its neighbors. In this case, our focus is on a new optimization framework that- optimizes the network offline and in a single shot. The objective is to construct a model that is accurate enough with KPIs and input parameters that can be changed and adapted as per our requirement, and a performance indicator we can improve as the output. By modifying these parameters to improve the model output, we can optimize the model, which represents the real world. Two considerations that we need to keep in mind for optimizing the network include: 

  • First, we need an accurate model for the described process which puts a need on having an algorithm that is capable of modeling the behavior of the network. The main challenge here is that the cells we are trying to optimize have a close relationship with their neighbouring cells. If we change the parameters of one cell, we can affect the others. And that is precisely what GNNs solve, being capable of supporting as many neighbours as a cell has and merging all the information for making predictions. Therefore, we need to construct a model that predicts the performance indicator we want to improve in the network with KPIs and parameters as inputs. 
  • Secondly, once we have identified the issued cells/or/cells having an issue or default or cells that have a certain problem and need to be optimized, we can apply any algorithm to change the parameters with the objective of improving the performance indicator of those cells. 
PyTorch blog_Optimization Framework

Figure 1: Optimization framework with GNN

Our use-case: The uplink interference

In this case, what we want to solve is the uplink interference problem or in other words, the interference that a user experiences when he tries to send information to the network such as uploading a file or starting a new phone call. To achieve this goal, we will simulate the behavior of the cellular network with a GNN, using the fixed KPIs of the network and the parameters we change as inputs for each cell. As the output of the model, we can use a performance indicator correlated with the problem we are trying to solve, which in this case, is the SINR (Signal to Interference and Noise Ratio) that needs to be predicted. Then, we will use this model as a simulator and optimize the configurations with any algorithm type (which can be iterative, genetic, and so on) to improve the SINR and therefore solve the uplink interference problem detected in some problematic cells (issue cells) as illustrated in figure 1.

Introducing our solution to the world: PyTorch conference and Stanford 

The issue case was presented at two global Artificial Intelligence conferences, showcasing Ericsson’s cutting-edge technology, a reference point in the industry for the adoption of this type of technology. The use case also successfully bridges the academic and the real-world scenario and represents the competence of our research team who used the Ericsson Cognitive Software to blend Artificial Intelligence with radio expertise. 

PyTorch Conference 

The conference has united AI experts from all over the world, presenting the latest topics on Artificial Intelligence both in academia and industry, and was attended by the top tech companies in this domain, who discussed the challenges they were facing in their business. It was a fantastic opportunity for Ericsson to showcase our leadership in AI as applied to telecommunications since we were the only telecom company advanced enough to present a use case at the event. We also had a speaker’s slot called Lightening Talk, dedicated to explaining the use case to the audience, followed by a  Q&A session. It was amazing to see the interest that this use case awakened among researchers globally since we are one of the very first companies bringing this latest research in AI and graph learning to create a solution around it. This event has provided us the opportunity to learn from world-class researchers and share with them the latest knowledge about AI’s usability in the telecom domain.  

Lightning Talk: Uplink Interference Optimizer, How to Optimize a Cellular Network...- Oscar Gonzalez - YouTube 

Stanford Conference 

In this second conference, the topic was more specific, focused on Geometric Deep Learning, a specific use case in the Artificial Intelligence field. It is a family of algorithms that can learn from examples when the input is a graph. Stanford is one of the universities that is leading the research in this field and the event is sponsored by PyTorch Geometric, the programming library that researchers and industry practitioners are using for constructing this algorithm type, having hosted 7000 attendees last year. This conference is attended by a small number of speakers and presenters, with only a few companies besides the Stanford Research Group having the honor of presenting a poster there.  

At Stanford, we presented our latest upgraded version of the optimization framework, which used a Bayesian GNN to offer our customers confidence intervals in the predictions instead of a single value. Our research team has been working on this topic over the last year, merging two concepts in literature, Bayesian Neural Networks, and GNNs to bring together the benefits of both worlds. Now, instead of providing customers with only a single predicted value for the gain that they will be experiencing with our framework, we can compute different intervals with different confidence levels. These intervals are different for each cell and have distinct confidence so the customers can choose between them. This development opens possibilities of including all the benefits of a probabilistic model along with the accurate prediction that a GNN provides. Moreover, the topic awakens the greatest interest of researchers across all domains, since Ericsson is not just presenting a new use case but showcasing breakthrough developments that can have an impact across all industries. Going forward, we have plans for publishing a library to make this technology easily accessible to probable future users. 

Solutions and challenges to optimize a cellular network with Graph Neural Networks - YouTube

Stanford Workshop photo:

Stanford Workshop photo
The Ericsson Blog

Like what you’re reading? Please sign up for email updates on your favorite topics.

Subscribe now

At the Ericsson Blog, we provide insight to make complex ideas on technology, innovation and business simple.