Vehicle Position Nowcasting with Gossip Learning

Nowcasting, i.e., short-term forecasting, of end user location is becoming increasingly important for anticipatory resource management in radio access networks (RAN). In this paper, we look at the case of vehicles moving in dense urban environments, and we tackle the location nowcasting problem with a particular class of machine learning (ML) algorithms that goes under the name Gossip Learning (GL). GL is a peer-to-peer machine learning approach based on direct, opportunistic exchange of models among nodes via wireless device-to-device (D2D) communications, and on collaborative model training. It has recently proven to scale efficiently to large numbers of static nodes, and to offer better privacy guarantees than traditional centralized learning architectures. We present new decentralized algorithms for GL, suitable for setups with dynamic nodes. In our approach, nodes improve their personalized model instance by sharing it with neighbors, and by weighting neighbors' contributions according to an estimate of their marginal utility. Our results show that the proposed GL algorithms are capable of providing accurate vehicle position predictions for time horizons of a few seconds, which are sufficient to implement effective anticipatory radio resource management.