<p>Visible light communication (VLC) networks modeled through a scheduling-level IEEE 802.15.7-oriented PHY/MAC abstraction are sensitive to line-of-sight blockage, receiver orientation, ambient-light noise, heterogeneous traffic loads, and inter-luminaire optical interference, which limits the effectiveness of fixed medium access control (MAC) policies in dense deployments. This study presents a federated deep reinforcement learning framework for distributed MAC scheduling in multi-luminaire IEEE 802.15.7-oriented VLC networks. Each luminaire is modeled as a local scheduling agent that selects the served receiver, transmission slot, optical power level, and physical-layer (PHY) mode from local queue, channel, blockage, interference, and illumination states. Instead of sharing raw observations, luminaires periodically exchange model parameters with a federated aggregation server to coordinate policy updates while preserving data locality. The proposed method is evaluated in a custom discrete-time simulator for a <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(5\times 5\times 3~\textrm{m}^3\)</EquationSource></InlineEquation> indoor VLC scenario with four ceiling luminaires, 8–32 receivers, stochastic traffic arrivals, receiver mobility, ambient-light noise, and line-of-sight blockage. Results averaged over 30 independent runs show that, at <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(K=32\)</EquationSource></InlineEquation> receivers, the proposed scheduler reduces average packet latency from 45 ms to 31 ms and 95th-percentile latency from 92 ms to 66 ms relative to the implemented resource-constrained centralized deep Q-network (DQN) baseline. Under a blockage probability of 0.3, the packet delivery ratio increases from 0.858 to 0.902, while under high ambient-light noise the packet error rate decreases from 0.064 to 0.049. The method also achieves a Jain fairness index of 0.96, reduces average synchronization overhead from <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(18.5\%\)</EquationSource></InlineEquation> to <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(4.7\%\)</EquationSource></InlineEquation> at a synchronization interval of 10 episodes, and shortens convergence time from 940 to 670 episodes at <InlineEquation ID="IEq5"><EquationSource Format="TEX">\(N=9\)</EquationSource></InlineEquation> luminaires. Illumination and flicker diagnostics show that executed actions satisfy the normalized feasibility mask after filtering. These results indicate that, within the adopted IEEE 802.15.7-oriented simulation abstraction, periodic federated parameter sharing improves the scheduling trade-off by reducing delay and coordination cost while preserving mask-enforced lighting feasibility, improving empirical reliability and fairness, and showing favorable multi-luminaire scalability trends.</p>

Distributed MAC scheduling in IEEE 802.15.7-oriented VLC networks via federated deep reinforcement learning

期刊论文
出版模式：

开放获取
发表日期：

2026年7月8日

查看全文

Iván Sánchez Salazar,
Pablo Palacios Játiva,
María Camila Reyes,
Amjad Ali,
Carlos Saavedra Arancibia,
Ismael Soto

摘要

Visible light communication (VLC) networks modeled through a scheduling-level IEEE 802.15.7-oriented PHY/MAC abstraction are sensitive to line-of-sight blockage, receiver orientation, ambient-light noise, heterogeneous traffic loads, and inter-luminaire optical interference, which limits the effectiveness of fixed medium access control (MAC) policies in dense deployments. This study presents a federated deep reinforcement learning framework for distributed MAC scheduling in multi-luminaire IEEE 802.15.7-oriented VLC networks. Each luminaire is modeled as a local scheduling agent that selects the served receiver, transmission slot, optical power level, and physical-layer (PHY) mode from local queue, channel, blockage, interference, and illumination states. Instead of sharing raw observations, luminaires periodically exchange model parameters with a federated aggregation server to coordinate policy updates while preserving data locality. The proposed method is evaluated in a custom discrete-time simulator for a \(5\times 5\times 3~\textrm{m}^3\) indoor VLC scenario with four ceiling luminaires, 8–32 receivers, stochastic traffic arrivals, receiver mobility, ambient-light noise, and line-of-sight blockage. Results averaged over 30 independent runs show that, at \(K=32\) receivers, the proposed scheduler reduces average packet latency from 45 ms to 31 ms and 95th-percentile latency from 92 ms to 66 ms relative to the implemented resource-constrained centralized deep Q-network (DQN) baseline. Under a blockage probability of 0.3, the packet delivery ratio increases from 0.858 to 0.902, while under high ambient-light noise the packet error rate decreases from 0.064 to 0.049. The method also achieves a Jain fairness index of 0.96, reduces average synchronization overhead from \(18.5\%\) to \(4.7\%\) at a synchronization interval of 10 episodes, and shortens convergence time from 940 to 670 episodes at \(N=9\) luminaires. Illumination and flicker diagnostics show that executed actions satisfy the normalized feasibility mask after filtering. These results indicate that, within the adopted IEEE 802.15.7-oriented simulation abstraction, periodic federated parameter sharing improves the scheduling trade-off by reducing delay and coordination cost while preserving mask-enforced lighting feasibility, improving empirical reliability and fairness, and showing favorable multi-luminaire scalability trends.