Machine Learning at the Edge: A Data-Driven Architecture with Applications to 5G Cellular Networks

The fifth generation of cellular networks (5G) will rely on edge cloud deployments to satisfy the ultra-low latency demand of future applications. In this paper, we argue that an edge-based deployment can also be used as an enabler of advanced Machine Learning (ML) applications in cellular networks, thanks to the balance it strikes between a completely distributed and a centralized approach. First, we will present an edge-controller-based architecture for cellular networks. Second, by using real data from hundreds of base stations of a major U.S. national operator, we will provide insights on how to dynamically cluster the base stations under the domain of each controller. Third, we will describe how these controllers can be used to run ML algorithms to predict the number of users, and a use case in which these predictions are used by a higher-layer application to route vehicular traffic according to network Key Performance Indicators (KPIs). We show that prediction accuracy improves when based on machine learning algorithms that exploit the controllers' view with respect to when it is based only on the local data of each single base station. The next generation of cellular networks (5G) is being designed to satisfy the massive growth in capacity demand, number of connections and the evolving use cases of a connected society for 2020 and beyond [1]. Michele Polese and Michele Zorzi are with the Department of Information Engineering (DEI), University of Padova, Italy. In order to meet these requirements, a new approach in the design of the network is required, and new paradigms have recently emerged [3]. First, the densification of the network will increase the spatial reuse and, combined with the usage of mmWave frequencies, the available throughput. On the other hand, this will introduce new challenges related to mobility management [4].