Optimizing Resource Allocation for QoS and Stability in Dynamic VLC-NOMA Networks via MARL

TL;DR

This paper develops multi-agent reinforcement learning frameworks to optimize resource allocation in dynamic VLC-NOMA networks, balancing QoS, stability, and interference management amid user mobility and dimming.

Contribution

It introduces tailored MARL frameworks for joint optimization of resource allocation and network stability in mobile VLC-NOMA systems, addressing complex dynamic conditions.

Findings

CTDE achieved 16% higher QoS satisfaction for high-priority users.

CTCE yielded 7 dB higher average SINR and 12% lower handover ratio.

Both frameworks effectively handle complex joint optimization in dynamic VLC-NOMA networks.

Abstract

Visible Light Communication (VLC) combined with Non-Orthogonal Multiple Access (NOMA) offers a promising solution for dense indoor wireless networks. Yet, managing resources effectively is challenged by VLC network dynamic conditions involving user mobility and light dimming. In addition to satisfying Quality of Service (QoS) and network stability requirements. Traditional resource allocation methods and simpler RL approaches struggle to jointly optimize QoS and stability under the dynamic conditions of mobile VLC-NOMA networks. This paper presents MARL frameworks tailored to perform complex joint optimization of resource allocation (NOMA power, user scheduling) and network stability (interference, handovers), considering heterogeneous QoS, user mobility, and dimming in VLC-NOMA systems. Our MARL frameworks capture dynamic channel conditions and diverse user QoS , enabling effective joint optimization. In these frameworks, VLC access points (APs) act as intelligent agents, learning to allocate power and schedule users to satisfy diverse requirements while maintaining network stability by managing interference and minimizing disruptive handovers. We conduct a comparative analysis of two key MARL paradigms: 1) Centralized Training with Decentralized Execution (CTDE) and 2) Centralized Training with Centralized Execution (CTCE). Comprehensive simulations validate the effectiveness of both tailored MARL frameworks and demonstrate an ability to handle complex optimization. The results show key trade-offs, as the CTDE approach achieved approximately 16\% higher for High priority (HP) user QoS satisfaction, while the CTCE approach yielded nearly 7 dB higher average SINR and 12\% lower ping-pong handover ratio, offering valuable insights into the performance differences between these paradigms in complex VLC-NOMA network scenarios.