Researchers have moved away from using convolutional neural networks (CNNs) and toward a new AI architecture based on transformers to more efficiently manage traffic, which will lead to significantly higher speeds in future 5G and 6G networks. (Image credit: AerialPerspective Images/Getty Images)
Scientists are creating artificial intelligence (AI) models that could help the next generation of wireless networks, such as 6G, provide faster and more reliable connections.
In a study published in the December 2024 issue of IEEE Transactions on Wireless Communications, the authors detailed an AI system that reduces the amount of information transmitted between a device and a wireless base station (such as a cell tower) by focusing on key data such as angles, delays, and signal strength.
By optimizing signal data in wireless networks using high-frequency millimeter waves (mmWave bands of the electromagnetic spectrum), the researchers noted that the number of connection errors was significantly reduced, and the AI system improved data reliability and communication in various conditions, such as urban areas with moving vehicles and pedestrians.
“To meet the growing demand for data in next-generation wireless networks, it is necessary to utilize the abundant frequency resource in millimeter-wave bands,” said lead author Byungju Lee, a professor in the Department of Telecommunications at Incheon National University in South Korea.
“Our approach enables precise beamforming, allowing signals to connect seamlessly to devices even when users are on the move,” Lee added.
More efficient methods of wave formation
The problem faced by networks using high-frequency radio spectrum such as mmWave is that they rely on a large group of antennas working together via massive multiple-input multiple-output (MIMO). For the process to be successful, accurate information – “channel state information” (CSI) – is needed to ensure communication between base stations and mobile devices with compatible antennas.
This situation is compounded by changes in the network environment, such as antennas moving with people and vehicles, or obstacles that block line-of-sight between devices and cell towers. This leads to “link aging” — a mismatch between the expected state of the link and its actual state, which in turn degrades performance, such as throughput and signal quality.
To overcome these challenges, the authors of the study applied a new type of AI model known as a transformer. Although convolutional neural networks (CNNs) can be used to predict and optimize wireless network traffic by recognizing signal patterns and classifying them.
But the researchers took a different approach: by using a transformer model instead of a CNN in their network analysis method, they were able to track both short-term and long-term patterns in signal changes. The resulting AI system, called “parametric CSI feedback with transformer,” could adjust the wireless network in real time to improve the connection between the base station and the user, even if the user was moving quickly.
The improvement is explained by the difference between CNNs and transformers. Both are neural network models that analyze visual patterns like images — in this case, patterns in the electromagnetic spectrum — but CNNs are typically trained on smaller data sets and focus on “local” features, while transformer models use larger data sets and have a self-perception mechanism that allows them to determine the importance of different input elements and their relationships at both the global and local levels.
In simple terms, the transformer model looks at the image as a whole, while CNNs focus on details like edges and textures. Transformers see the big picture, so to speak.
However, transformer models are more computationally intensive than CNNs. But if they can enable reliable next-generation wireless networks, they could be the key to high-speed
Sourse: www.livescience.com
