Characterizing throughput and convergence time in dynamic multi-connectivity 5G deployments

Fifth-generation (5G) mobile communications are expected to integrate multiple radio access technologies (RATs) within a unified access network by allowing the user equipment (UE) to utilize them concurrently. As a consequence, mobile users face even more heterogeneous connectivity options, which creates challenges for efficient decision-making when selecting a network dynamically. In this work, with the tools of queuing theory, integral geometry, and optimization theory, we develop a novel mobility-centric analytical methodology for multi-RAT deployments. Particularly, we first contribute a framework for optimal data rate allocation in the network-assisted regime. Then, we characterize the convergence time of the distributed optimization algorithms based on reinforcement learning to reduce the signaling overheads. Our findings suggest that network-assisted strategies may improve the UE throughput by up to 60% depending on the considered deployment, where the gains increase with a higher density of millimeter-wave New Radio (NR) base stations. A user-centric solution based on reinforcement learning mechanisms is capable of approaching the performance of the network-assisted scheme. However, the associated convergence time may be prohibitive, on the order of several minutes. To improve the latter, we further propose and evaluate a transfer learning-based algorithm that allows to decrease the convergence time by up to 10 times, thus becoming a simple solution for rate-optimized operation in future 5G NR deployments.