Pre-equalization scheme for visible light communications with trial-and-error learning.

In this Letter, we propose a novel, to the best of our knowledge, neural network pre-equalizer based on the trial-and-error (TE) mechanism for visible light communication. This approach, unlike indirect learning (IL) architecture, does not require an additional auxiliary post-equalizer. Instead, it allows the pre-equalizer to be trained directly from the transmitter side through continuous interaction with the actual system. In a 1.95-Gbps 64-QAM carrier-less amplitude phase (CAP) free space optical transmission platform, the proposed scheme demonstrates superior nonlinear approximation capabilities and noise resilience. Specifically, the TE-recurrent neural network (RNN)-based pre-equalizer exhibits signal-to-noise ratio (SNR) gains of 0.8 dB and 1.8 dB over the IL-RNN-based and IL-Volterra-based pre-equalizers, respectively. We believe this is the first application of trial-and-error learning for training pre-equalizer in visible light communications.