A Novel Digital Twin Modeling for QoS-driven Heterogeneous Network Optimization

Abstract

Heterogeneous networks (HetNets) are one of the major technology trends for mobile communications. While they can exploit the complementary advantages of different access technologies, they come with complicated network topologies and cellular heterogeneity, causing significant challenges to resource allocation. In addition, the existing resource allocation methods mostly aim to maximize network capacity in the physical layer, which does not necessarily guarantee optimal quality of service (QoS) for users. Though there are some works studying QoSdriven resource allocation, they typically consider an ideal network environment to derive the mathematical expressions of QoS metrics. However, practical QoS metrics, which are usually acquired through a period of real-time measurement, are difficult to derive. The above two points collectively lead to great difficulty for unlocking the full potential of HetNets. To bridge this gap, we propose a novel learning-aided digital twin network (DTN) model to precisely and rapidly map the physical network deployment to QoS metrics. A GraphSAGEbased representation learning method is specially tailored to harness the complicated topology of HetNets, enabling inductive adaptation to varying numbers of access points and users. To handle the intricate relationships among multiple QoS metrics, a novel method is proposed which combines mixture-of-experts (MoE) with a hybrid residual learning mechanism. Results show that the proposed DTN offers both higher average prediction accuracy of multiple QoS metrics and a more lightweight design than state-of-the-art learning-based baselines. Index Terms-Heterogeneous network (HetNet), digital twin network (DTN), quality of service (QoS), GraphSAGE, mixtureof-experts (MoE)