Performance Improvement in Rayleigh Faded Channel using Deep Learning

A Neural Network representing an end to end communication system is an exciting concept. This study has been made only for systems affected by Additive White Gaussian Noise where the neural network's performance is comparable to that of a communication system that uses conventional signal processing techniques at the transmitter and at the receiver. We in this paper represent a very fast time varying Rayleigh faded wireless system by a feed forward deep neural network. A comparison of the Block Error Rate vs Bit-Energy to Noise Ratio plot of the neural network with that of a conventional wireless system shows significant improvement in Block Error Rate. We use the t-distributed Stochastic Neighbourhood Embedding algorithm which reveals that Deep Learning suggests using a non-regular and non-symmetric two dimensional constellation for a Rayleigh faded channel. The effect of hyper parameters of the neural network on the performance of the end-to-end communication system has been studied and analysed in details here as well.