Cell- and Area-based ML Models: Unlocking High Precision Models for Radio Access Networks

Abstract

Cellular networks evolve towards future generations, facing unprecedented levels of device and network programmability. At the same time, the new vision of the cyber-physical continuum will rely on diverse network architectures in extremely dense deployments. The emergence of new types of cells, like mobile and drone ones, would rely on instantly available AI/ML algorithms to provide a service within a few seconds after being powered on, avoiding long periods of data collection and training.In this work, we discuss how cell-specific characteristics, like the radio environment, can impact area-based ML models. Even though area-based models simplify the management of ML workflows considerably, there is also a need for cell-based models as these tend to provide better performance. Moreover, we show that area-based models can be part of ML workflow as they can complement cell-based ones. We finalize our work by discussing the possibility of reusing available ML models from other cells as a way of reducing the time needed for applying ML algorithms in newly deployed cells. We provide initial insights on the model re-usability and performance assessment and highlight the need for more research in this direction.In our work, we utilize the data from a test network, allowing us to explore the dynamics of real networks and provide results with increased confidence.