Leveraging direct learning and reinforcement learning architectures for pre-equalization in high-speed VLC systems

Recently, neural network (NN)-based approaches have garnered significant attention for mitigating nonlinear impairments in visible light communication (VLC) systems. Deploying a pre-equalization NN at the transmitter side can reduce receiver complexity. However, existing pre-equalizers, which primarily utilize an indirect learning architecture (ILA), face performance challenges when combating severe noise and nonlinear distortions in high-speed scenarios. In this paper, two pre-equalization schemes employing direct learning architecture (DLA) and reinforcement learning architecture (RLA), respectively, are proposed to enable a high-speed VLC with bandwidth-limited light-emitting diodes (LEDs). DLA employs an auxiliary network modeling system, characterized by jointly optimized end-to-end for direct pre-equalizer training. In contrast, RLA is a model-free approach that trains the pre-equalizer through continuous system interaction based on rewards. Simulations show both RLA and DLA outperform ILA in noise and nonlinearity tolerance, with RLA achieving the best performance. In a 1.5-m VLC prototype at 2.1 Gbps, RLA maintains excellent robustness under high-order modulation, demonstrating significant potential for future high-speed communications.