Deep Learning-Based Resource Allocation for Visible Light Communication Networks

Abstract

To fulfill the ultra-low latency and high-reliability requirements of sixth-generation (6G) hyperconnectivity, this paper proposes VLC-Net, a supervised deep learning framework for real-time resource allocation in multicell visible light communication (VLC) networks employing non-orthogonal multiple access (NOMA). A feedforward deep neural network (DNN) with a Bayesian-optimized architecture was trained using twenty thousand near-optimal power allocation labels generated by a fairness-centric constrained genetic algorithm (CGA). Extensive Monte Carlo simulations for networks with four, six, eight, and ten users show that VLC-Net achieves an inference time as low as 2.65 milliseconds for ten users and an average inference time of 13.71 milliseconds across all evaluated user densities. This performance represents an approximately 388 times speedup over iterative evolutionary solvers. Furthermore, the proposed model improves Jain’s fairness index by 66 percent compared with the greedy Water Filling approach while maintaining an energy efficiency of 10.87 megabits per second per watt under high-interference conditions with ten users. Paired t tests yield probability values below 0.05 for all evaluated user densities, confirming statistically significant performance differences. These results demonstrate that VLC-Net offers a robust, scalable, and low-latency solution for practical intelligent power management in future dense indoor optical wireless network deployments.