AI-RAN Training Data Secured Through a ‘Wireless Digital Twin’
IT DAILY ·
✦ AI Summary
Securing training data was presented as a key challenge in AI-RAN research, and it was noted that collecting enough data in actual communication networks is difficult.
As an alternative, the use of a wireless digital twin was proposed, introducing a method that recreates the movement of cities, buildings, and devices in a virtual environment for generating training data and validating algorithms.
The content was presented on the 13th by Prof. Choi Gye-won of Sungkyunkwan University's Department of Electrical and Computer Engineering at a joint meeting of the AINA AI-RAN Computing WG and the Industrial Convergence Committee.
One of the key challenges in AI-RAN research was identified as securing training data. The target of AI applications is wireless communication networks, and large-scale data is needed to learn diverse radio propagation environments. However, it was noted that collecting enough data in actual networks is difficult.
As an alternative, a way to use a Wireless Digital Twin was proposed. A wireless digital twin is a method that reproduces cities, buildings, and device mobility in a virtual environment. The approach was introduced as a way to generate training data and to use it for algorithm validation as well.
The details were presented on the 13th at a joint meeting of the AINA AI-RAN Computing WG and the Industrial Convergence Committee, held as an event. The presenter was Prof. Choi Gye-won of the Department of Electrical and Computer Engineering at Sungkyunkwan University. The presentation focused on how to secure data for AI-RAN training and validation through a digital twin.
AI-RAN is a technology that applies AI to the Radio Access Network (RAN) to improve the efficiency of wireless resource management and enhance communications performance. During the presentation, Prof. Choi Gye-won of the Department of Electrical and Computer Engineering at Sungkyunkwan University explained why a digital twin is needed for AI-RAN.
Professor Choi first pointed to the limitations of statistical channel models such as CDL and TDL. CDL and TDL are wireless channel models that represent multipath variations in radio waves and are used in communications algorithm and equipment performance simulation.
He said those channel models are overly simplified and insufficient for reflecting the diversity and spatial characteristics of real environments. He explained that these limitations lead to a problem in AI-RAN implementation, where real-world conditions cannot be sufficiently reproduced.
Professor Choi also addressed the challenge of collecting data in real environments. Because the data available in actual environments is limited, there are constraints in reflecting reality during the validation and implementation process required for AI-RAN.
He also mentioned the limitations of existing network simulators. He explained that existing network simulators rely on abstract protocol models that differ from reality, making it difficult to sufficiently reproduce real environments. The photo is by Reporter Seong Won-young.
A wireless digital twin aims to compensate for these limitations by virtually reproducing real radio environments and communication operations. To do this, terrain, building layouts, and device mobility are modeled as components. Then, Ray Tracing is used to calculate radio propagation and reflection paths, and wireless channel data are generated based on that.
The environment created in this way can be linked with actual communication protocol stacks. After linkage, communication behavior can be tested according to changes in the radio environment, and data for AI training and validation can also be secured.
A representative digital twin platform is NVIDIA's Aerial Omniverse Digital Twin (AODT). AODT has functions for configuring wireless environments in virtual space, calculating radio paths based on Ray Tracing, running communication link simulations, and running network simulations. It is also used for developing base station algorithms, training AI models, and evaluating AI model performance based on data that reflects the radio environment of specific areas.
A wireless digital twin can be used not only for training individual AI-RAN algorithms but also for research on a Wireless Foundation Model. The approach to a Wireless Foundation Model is similar to that of an LLM.
The model follows a process of Pre-training on massive data before performing task-specific fine-tuning. In the Pre-training stage, it learns common characteristics and patterns of channels based on diverse wireless channel data.
After that, the model can be fine-tuned for specific purposes or further trained for each task. Examples of application tasks include beam prediction and channel restoration. This training method is characterized by reusing diverse features obtained through Pre-training instead of learning from scratch for each specific task.
For such uses, scale, diversity, and consistency with the real environment were presented as conditions for securing digital twin data. Professor Choi said that simply increasing the amount of data in a specific city environment has limits when training models to respond to diverse radio environments. He said different building layouts, device mobility conditions, and characteristics of the actual application environment must therefore be reflected.
As a validation task is needed to confirm whether a virtual-environment learning model delivers the same performance in actual communication networks, model performance can decline if there are differences between the virtual environment and the actual network in the radio characteristics of building materials or in how transmitting and receiving equipment operates in practice. To reduce this, Professor Choi proposed reflecting training under a variety of conditions and conducting additional learning based on measured data.
Source: IT DAILY · Seong Won-young
Original: https://www.itdaily.kr/news/articleView.html?idxno=242173
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.