User-Centric Machine Learning for Resource Allocation in MPTCP-Enabled Hybrid LiFi and WiFi Networks

TL;DR

This paper introduces a user-centric machine learning approach for load balancing in MPTCP-enabled hybrid LiFi and WiFi networks, significantly improving throughput and flexibility over traditional network-centric methods.

Contribution

It proposes a novel user-centric learning model tailored for MPTCP scenarios, addressing challenges of load balancing in heterogeneous networks with enhanced performance.

Findings

Up to 40% throughput increase compared to TCP-based methods

Low complexity and high flexibility of the user-centric approach

Outperforms network-centric learning methods in simulations

Abstract

As an emerging paradigm of heterogeneous networks (HetNets) towards 6G, the hybrid light fidelity (LiFi) and wireless fidelity (WiFi) networks (HLWNets) have potential to explore the complementary advantages of the optical and radio spectra. Like other cooperation-native HetNets, HLWNets face a crucial load balancing (LB) problem due to the heterogeneity of access points (APs). The existing literature mostly formulates this problem as joint AP selection and resource allocation (RA), presuming that each user equipment (UE) is served by one AP at a time, under the constraint of the traditional transmission control protocol (TCP). In contrast, multipath TCP (MPTCP), which allows for the simultaneous use of multiple APs, can significantly boost the UE's throughput as well as enhancing its network resilience. However, the existing TCP-based LB methods, particularly those aided by machine learning, are not suitable for the MPTCP scenario. In this paper, we discuss the challenges when developing learning-aided LB in MPTCP-enabled HLWNets, and propose a novel user-centric learning model to tackle this tricky problem. Unlike the conventional network-centric learning methods, the proposed method determines the LB solution for a single target UE, rendering low complexity and high flexibility in practical implementations. Results show that the proposed user-centric approach can greatly outperform the network-centric learning method. Against the TCP-based LB method such as game theory, the proposed method can increase the throughput of HLWNets by up to 40\%.