Artificial Intelligence (often ML) is a promising technology whereby many challenges can be solved that before were hard or impossible.
Artificial Intelligence brings new capabilities as a groundbreaking new way to design systems.
So, where should you apply AI in RAN in order to achieve the best payback, and where not? The industry has concluded that 6G will be built heavily with AI/ML technologies including cutting-edge GenAI. However, it is far less obvious how to apply AI/ML and what the realistic expectations should really be. We could start by outlining different areas to which AI/ML could add value.
RAN is an area that is full of algorithms. Much of the efficiency of the RAN comes from how accurate those algorithms are functioning and their performance. O-RAN has defined new network elements (Radio Intelligent Controllers) that create placeholders for algorithms in the form of Apps, overseeing the operation of the RAN. These Apps could be based on AI/ML and are managed in parallel with the life cycle of the RAN, thus making it possible to collect data and train AI functions specifically for the operational context in which they are supposed to operate. Such AI functions depend heavily on the ability to transfer data to and from the RAN.
The next area to explore would be the algorithms that reside inside the RAN, i.e., those that implement functionality specified by 3GPP. As there’s no defined App concept here, the life cycle management of such AI models would be trickier and subject to vendor management systems and MNO data for training and MLOps.
Some of these algorithms are close to being mathematically optimal, whilst others are obvious candidates for AI/ML. It would typically be algorithms with a lot of unknown parameters, i.e., where it is hard to mathematically describe the problem at hand. Such problems are often dealing with multiple “cells”, such as CoMP or D-MIMO, or higher in the stack, such as MAC Scheduling.
RAN as well as most other domains are increasingly SW-defined. It means that the SW components, responsible for the RAN implementation, need to find their (optimal) place in a real network and require a proper configuration.
This introduces challenges in Day-0, Day-1, and Day-2 operations that are suitable candidates for AI/ML (including GenAI) implementations. These functions deal with the placement and life cycle management of SW components (including AI model artifacts). They also deal with performance optimization through tuning of configuration parameters depending on traffic conditions and policies for a particular MNO and are related to the provisioning of services as well as their assurance. The latter may include the ability for humans to express their intent in a more natural way and have GenAI-based functionality translating that into machine-readable instructions. We could also see more advanced what-if-scenario simulations or prediction capabilities that emerge by utilizing GenAI.
The challenge of properly understanding the real situation in an operative network is huge. For operational staff to mitigate situations occurring during operation, they first need to be provided with an accurate view of the network and any abnormalities that may exist. This is a task that is well suited for AI/ML but requires proper training which then becomes a challenge.
Not only is the RAN of interest. Any application that depends on communication based on 5G (or later 6G) may be subject to smart control through AI/ML.
This is a huge field that promises savings and changes in the R&D process. IT professionals can make their daily lives much easier with the help of GenAI when designing high-quality program codes or configurations (as mentioned above). Such improvements may ultimately change the value chain of R&D, opening it up for more actors as the task of training models becomes distinctively separated from that of developing the system in which the model is executed.
Another related area is testing. AI/ML can assist in defining and selecting proper test cases as well as the interpretation of test results. As network products get increasingly complex this translates into the corresponding challenge to determine whether a specific test result is correct or not. It becomes specifically challenging if some functionality is implemented through the use of AI/ML. Such capabilities can improve to pinpoint, which parts of a complex system are faulty as part of a continuous testing process and, thereby, reduce cycle times for testing.
Tietoevry already participates in research programs and PoCs together with its partners, applying its solid in-house capabilities in RAN as well as wide experience in applying AI/ML.
Reach out to us to learn more and how we can assist in Your journey to a more intelligent RAN.
Mats Eriksson leads business development and sales in the telecom and radio access sector in Tietoevry Product Development Services. He has previously co-founded technology companies and held managerial positions in various companies. He has a background in academia where he was in charge of a research cooperation institute and founded an EU innovation initiative.