Towards 6G: Machine Learning Empowered Resource Management in Hybrid LiFi and WiFi Networks

The future sixth-generation (6G) network envisions unprecedented capabilities, including Tbps-level data rates, Kbps/Hz spectral efficiency, and µs-level latency. Achieving this holistic TKµ vision requires groundbreaking technologies, such as higher frequency spectrum utilisation, ultra-dense small-cell deployments, low-complexity energy-efficient algorithms, and the native integration of artificial intelligence (AI) into network operations. Over the past decade, heterogeneous networks (HetNets) have garnered significant attention for combining diverse wireless technologies across various frequency bands, offering complementary advantages over homogeneous networks. Among HetNets, hybrid light fidelity (LiFi) and wireless fidelity (WiFi) networks (HLWNets) stand out as a promising paradigm, particularly for indoor wireless local area networks. By delivering ubiquitous WiFi coverage and high-capacity LiFi transmission, HLWNets emerge as a strong candidate for realizing the TKµ vision of 6G networks. Despite their potential, resource management in HLWNets remains a significant challenge due to heterogeneous APs and dynamic mobile environments. This involves three interdependent problems: (i) access point selection (APS), (ii) resource allocation (RA), and (iii) mobility management. The interconnected nature of these problems complicates solutions, as APS and RA influence mobility management, which in turn impacts APS and RA. Moreover, resource management in HLWNets is typically formulated as a mixed-integer nonlinear programming (MINLP) problem, which becomes computationally infeasible at scale. Therefore, this thesis explores ML-empowered resource management for HLWNets at three main stages: APS, RA, and mobility management.
First, an adaptive target-condition neural network (A-TCNN) is introduced for load balancing (LB) in transmission control protocol (TCP)-based HLWNets, where each user equipment (UE) connects to a single AP. Unlike conventional deep neural networks (DNNs) that handle joint AP selection for all UEs, A-TCNN operates per target UE, accounting for the state of other UEs. Additionally, an adaptive mechanism accommodates varying UE numbers by splitting data rate requirements without altering AP selection outcomes. Simulations show that A-TCNN achieves near-optimal performance with sub-millisecond inference time and adapts well to dynamic UE scenarios. Second, the study extends TCP-based HLWNets to multipath TCP (MPTCP) scenarios, where UEs can connect to multiple APs simultaneously, enhancing throughput, reliability, and flexibility. While APS in this context is straightforward using signal strength-based methods, RA remains challenging due to its inherent nonlinearity. To address this, a graph attention network (GAT)-based approach is proposed, modeling APs and UEs as nodes and their subflows as edges. A multi-head attention mechanism aggregates information from neighboring nodes, capturing complex topologies in MPTCP-based HLWNets. Results demonstrate that the GAT-based method outperforms conventional optimization, DNNs, and non-attentive graph neural networks. Finally, a mobility-supporting neural network (MSNN) is developed to dynamically adjust update intervals for each UE based on four key inputs: associated AP, channel quality, movement direction, and speed. This model is integrated with the A-TCNN framework, forming a mobility-supporting A-TCNN (MS-ATCNN) that jointly addresses resource management and mobility. Furthermore, a throughput degradation criterion is defined to generate sufficient training labels, ensuring controlled throughput reduction while optimizing update intervals. Last, extensive simulation results indicate that MS-ATCNN enhances network throughput by up to 215\% compared to baseline methods while maintaining sub-millisecond inference time comparable to A-TCNN.