While each generalization domain moves us closer to an AI-native RAN, the untapped potential of AI lies in a design that incorporates all generalization dimensions, leading the path toward the vision of a single AI algorithm regulating and replacing numerous RAN processes at the same time.
Generalization over the RAN environment
Training AI models to generalize across the RAN environment, rather than specializing to meet specific conditions (such as a cell environment), provides the necessary robustness to deploy a single AI model for a given task over the whole network. This boosts scalability and minimizes the complexity of AI functionalities and operations. For example, model life-cycle management (LCM) becomes less complex when there is just one AI model per radio feature or RAN protocol layer (which is ideal).
AI generalization enables a single AI model to learn from the collective experience of thousands of network elements. Distributing training-data generation across space and time (with thousands of network entities each contributing a few training samples) greatly reduces the impact of the learning process on RAN performance, such as for reinforcement learning (RL) algorithms performing suboptimal actions during training in live networks.
The RAN environment is defined by a combination of static and semi-static network deployment and configuration information, as well as more dynamic information about the radio environment, traffic, load, unique user equipment (UE) characteristics, and so on. At Ericsson, we studied two key design methods for generalizing AI models throughout the RAN environment. The first approach is based on traditional feature engineering, in which domain knowledge is used to select the most relevant information needed to recreate the RAN environment for specific AI applications. Techniques like feature sensitivity analysis allow you to reduce model complexity by filtering out redundant information (based on correlation metrics, for example).
A more systematic approach would be to learn an embedding of static and semi-static RAN environment features independently and then mix it with dynamic RAN environment information tailored to individual use cases. Graph neural networks (GNNs) [3] are a promising candidate for encoding the RAN environment's static and semi-static elements. These include network configurations and relations between different RAN entities (such as relations between one cell and a group of direct or indirect nearby cells), which may represent broader network entity relations that are outside the scope of typical feature-engineering approaches. Such embedding could be learned and reused across multiple use cases (for example, from L1 to L3). This approach has been researched at Ericsson in the context of L3 features such as secondary carrier.
Generalization over RAN intents
A single key performance indicator (KPI) such as throughput, spectral efficiency, latency, packet loss, dependability, or resource utilization cannot characterize RAN performance. Instead, it is defined by a number of, sometimes conflicting, KPIs. As a result, many RAN control problems, such as RRM, do not provide a single optimum-for-RAN KPI, but rather a Pareto frontier of RAN KPIs, each of which trades off against another.