Abstract
As the integration of unmanned aerial vehicles (UAVs) into visible light communications (VLC) can offer many benefits for massive-connectivity applications and services in 5G and beyond, this work considers a UAV-assisted VLC using non-orthogonal multiple-access (NOMA). We formulate a joint problem of power allocation and UAV placement to maximize the sum rate of all users. Since the problem is non-convex and NP-hard, we propose using the Harris Hawks Optimization (HHO) algorithm. Furthermore, we integrate HHO with artificial neural networks to create a real-time design that avoids local minima. Numerical results verify the effectiveness and superiority of the proposed HHO-based approach over existing metaheuristic algorithms.
Introduction: The Challenge of Massive Connectivity
As we transition toward 6G, the demand for "massive connectivity" requires innovative solutions. Visible Light Communication (VLC) is a promising candidate, utilizing the unregulated optical spectrum of LEDs for high-speed data transmission. However, VLC's reliance on line-of-sight (LoS) makes it vulnerable to blockages.
Deploying VLC transmitters on UAVs addresses this by providing mobile, high-altitude nodes that maintain clear LoS to ground users. To further enhance capacity, Non-Orthogonal Multiple Access (NOMA) is employed, allowing multiple users to share resources simultaneously.
The primary challenge lies in the joint optimization of UAV placement and power allocation. This problem is non-convex and NP-hard, particularly because the Lambertian emission model used in VLC channel modeling is mathematically complex, rendering conventional coordinate descent or convex approximation methods ineffective.
Methodology: Swarm Intelligence and Neural Networks
The researchers propose a two-stage solution to manage network resources efficiently:
1. Harris Hawks Optimization (HHOPAP)
The authors adapt Harris Hawks Optimization (HHO), a swarm intelligence metaheuristic inspired by the cooperative hunting strategies of hawks. HHO is well-suited for the "rugged" mathematical landscape of VLC because it does not require gradient information. It effectively navigates the search space to find near-optimal drone positions and power levels by mimicking the hawks' phases of exploration (searching for prey) and exploitation (pouncing).
2. HHO-trained Feedforward Neural Network (HHOFNN)
While direct optimization is effective, it is computationally slow for real-time applications. To solve this, the authors use HHO to train a Feedforward Neural Network (FNN). Unlike standard backpropagation, which often gets trapped in local minima, the HHO trainer explores the weight and bias space more globally to find superior network configurations.
Evaluation Datasets
The framework was evaluated using synthetic VLC data, simulating environments with 10–20 ground users in a 10m x 10m area. Additionally, the HHO trainer's performance was validated against the Iris and Cancer datasets from the UCI Machine Learning Repository to prove its generalizability.
Analysis and Key Results
The study yielded several critical insights into the behavior of UAV-VLC systems:
- The Altitude Paradox: There is an optimal "sweet spot" for UAV altitude. If the drone is too low, the angle of light incidence becomes too steep for ground sensors; if too high, the signal path loss becomes prohibitive.
- VLC-NOMA Synergy: NOMA thrives in high Signal-to-Noise Ratio (SNR) environments. The short-distance, LoS nature of VLC naturally provides the high SNR required for NOMA to maximize sum-rate gains.
- Performance Gains: The HHOPAP algorithm consistently outperformed standard OFDMA and other NOMA variants like Gain Ratio Power Allocation (GRPA).
- Real-time Feasibility: The HHOFNN achieved results within 10% of the direct optimization's performance but delivered them near-instantaneously, making it viable for dynamic real-world networks.
- Optimization Robustness: Compared to Particle Swarm Optimization (PSO) and Genetic Algorithms (GA), the HHO trainer achieved significantly lower Mean Squared Error (MSE) on benchmark tasks, demonstrating its ability to avoid local minima.
Conclusion and Future Directions
This work demonstrates that combining swarm intelligence with machine learning can effectively manage the complexities of next-generation wireless networks. However, several avenues for future research remain, including transitioning to fixed-wing UAV trajectory optimization, balancing data rates with strict battery constraints, and enhancing physical layer security against eavesdropping.
Summary updated