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

Area of Science:

• Optical Communications
• Machine Learning
• Signal Processing

Background:

• Visible Light Communication (VLC) systems face nonlinear impairments, especially at high speeds.
• Existing pre-equalization methods using indirect learning architecture (ILA) struggle with severe noise and distortion.
• Reducing receiver complexity by deploying pre-equalization at the transmitter is a key challenge.

Purpose of the Study:

• To propose novel neural network-based pre-equalization schemes for high-speed VLC.
• To address the limitations of indirect learning architectures in noisy and nonlinear environments.
• To enhance the performance and robustness of VLC systems using advanced learning techniques.

Main Methods:

• Developed two pre-equalization schemes: Direct Learning Architecture (DLA) and Reinforcement Learning Architecture (RLA).
• DLA utilizes an auxiliary network for end-to-end joint optimization.
• RLA employs a model-free approach with continuous system interaction and rewards for training.

Main Results:

• Both DLA and RLA significantly outperform the traditional ILA in tolerance to noise and nonlinearity.
• RLA demonstrates the best overall performance among the proposed methods.
• A 1.5-m VLC prototype operating at 2.1 Gbps showed RLA's robustness with high-order modulation.

Conclusions:

• DLA and RLA are effective strategies for mitigating nonlinear impairments in high-speed VLC.
• RLA offers superior performance and robustness, making it highly promising for future high-speed optical wireless communications.
• The proposed methods enable bandwidth-limited light-emitting diodes (LEDs) to support higher data rates.