Model-Aware End-to-End Learning for SISO Optical Wireless Communication Over Poisson Channel

Abstract

Faced with the challenge of transceiver design over the Poisson channel, we leverage the deep-learning technique and devise two novel end-to-end learning schemes to fulfill the design task in this paper. One of the schemes accords with the basic principle of the currently available autoencoder (AE) but is specially designed for the Poisson channel with the aid of the square root (SR) transform. The other scheme, following a different design philosophy from AE, is developed based on a double neural network (DNN) model, which regards the receiver and the transmitter as two separate networks. By these designs, the end-to-end learning task can be conducted over Poisson channel. Extensive computer simulations reveal that 1) the transceiver learned by the DNN scheme always performs better than or comparably to the currently available artificially designed transceivers, and 2) compared with the transceiver learned by DNN, the transceiver learned by SR-AE suffers performance loss in some cases, but the SR-AE scheme has a lower complexity to compute the loss function and fewer network parameters. This study takes the first step toward applying end-to-end learning techniques in the field of the Poisson channel and lays a foundation for further works on this topic.