AI Modeling Will Bolster Future 5G and 6G Systems
Sep 02, 2024
AI Modeling Will Bolster Future 5G and 6G Systems
Artificial intelligence and machine learning are proving to be an invaluable tool in many engineering design tasks, enabling engineers to reduce iteration time and consider more variables. The design of wireless 5G and 6G systems, which involve complex calculations of many variables, is no exception.
In a recent e-mail interview, Dr. Houman Zarrinkoub, Principal Product Manager at MathWorks, discusses how AI-native systems can facilitate the complex tasks involved with designing wireless systems, whether they be cellular, WiFi, Bluetooth, or satellite communication systems.
Dr. Houman Zarrinkoub, Principal Product Manager at MathWorks
What benefits do AI-native systems offer over traditional models in wireless engineering?
Zarrinkoub: AI native systems benefits are as follows:
-
More adaptable: Learn faster and more responsive to the dynamic nature of operational situations
-
More efficient implementation: Less computational complexity to achieve the same performance level as traditional models
-
More robust: If designed well, AI models can handle multi-variate/non-linear design challenges better
Why is the adoption of AI-native technologies necessary in the development of next-generation wireless standards, like 5G Advanced and 6G?
Zarrinkoub: Optimal operations of 5G-Advanced and 6G requires supporting:
Related:Antenna Design Needs Optimization for 6G Systems
-
Substantially larger sets of users
-
Different types of use cases such as high-throughput cellular communications, autonomous driving and satellite communications
-
Substantially more dynamic and customized service based on instantaneous channel conditions
What tools and knowledge are needed to develop AI-native wireless systems?
Zarrinkoub:
-
AI systems are statistical and control-based systems. So, to understand, design and implement them engineers need a good background in mathematical and statistical interference. Engineers also need knowledge of systems theory, which is how systems connect input sensory data into actions.
-
Engineers also need tools that allow them to easily develop and train AI networks. AI systems have a large set of parameters. Engineers provide input to the system in training phases to set parameters and train the network.
-
These tools are usually computer programming tools written in languages like MATLAB and Python.
Why can’t traditional tools be used to develop these systems?
Zarrinkoub:
-
Traditional rule-based systems have always been designed on “best effort” philosophy and never promised adaptability to dynamic changes of the environment.
-
User expectations on adaptability and emergence of new use cases has increased enormously, making use of traditional approaches and tools impossible.
-
For example, as wireless users travel throughout the world, they expect at any moment either to be connected to cellular, Wi-Fi, Bluetooth, or satellite communications systems, depending on which system has more signal strength at that moment. These types of expert systems can be designed using AI techniques.
Related:6G Development Efforts Off to A Good Start, Say Researchers
What are the key steps involved in designing and integrating an AI-native wireless system?
Zarrinkoub: To apply AI-native models to the design process engineers need to:
-
Have lots of data. Not only do engineers need tools to acquire live data over the air with RF devices, but also to generate synthetic data that matches what the systems usually produces.
-
Knowledge about machine learning systems in general, albeit deep learning systems or reinforcement learning simulation models. Software tools that allow engineers to look deep inside these learning networks and models and provide “explainability” of what they are doing is necessary.
-
Easily plug-in the designed-trained AI network into a larger system that uses this AI-based sub-component. Engineers need to test with a set of data “not used in the training process” to make sure AI-subcomponent reaches the performance they want. So, they need multi-disciplinary knowledge of “AI plus the area of applications of AI”.
-
When the AI system is integrated with a larger model and works well, engineers need to implement it on hardware, cloud, edge devices etc.
What are the common hurdles engineers face when integrating AI into their wireless system design?
Zarrinkoub: There are few hurdles and challenges in developing AI systems:
-
Robustness: Almost all AI systems work very well within the normal operating ranges of the systems, where system parameters are within averages. That’s because usually lots of representative data is available for the normal cases. However, they usually fail at outlier or edge cases where there isn’t much representative input training data available. To overcome this problem, engineers should have tools that help them have many representative data sets for an average as well as outlier cases both as synthetic data and live acquired data.
-
Explainability: One major concern of engineers using AI-native systems is the lack of explanation on why the design is optimal outside verification using data. Lack of a model describing why the AI systems work in all cases comes from one source: AI systems combine many systems’ subcomponents to make one single optimal set. This impedes the ability to identify which part of the system was responsible when a system fails.