Predictive Data Replication for XR Applications in Multi-Connectivity Enabled mmWave Networks

Abstract

Emerging applications such as Extended Reality (XR) require a fundamental change in the way network architecture and functions are designed and optimized due to strict Quality of Service (QoS) requirements, especially with respect to hard deadlines. Fortunately, multi-connectivity enabled mmWave networks provide us with the capability of catering to such stringent constraints. In a multi-connectivity enabled network where users (UEs) connect to multiple base stations (gNBs) that can simultaneously and rapidly switch data connection between them for optimal data delivery, it is vital to carefully select which gNBs the data should be placed at pre-emptively. Selectively placing data at multiple base stations (gNBs) can lead to a network which is more resilient to blockages and which minimizes data plane interruptions. In this paper, we use a Deep Learning agent that encodes a complex system state and then uses a Deep Q-Network (DQN) to find the optimal selection of gNBs where the UEs’ data should be placed. Our results show that using our Deep Learning agent, which essentially uses a vast amount of state information to pre-emptively predict the best selection of gNBs for future transmissions, delivers markedly better performance than other heuristic selection algorithms.