Deep Learning in Encoding and Decoding of Polar Codes

In this paper, we apply neural networks to the encoding and decoding of polar codes. Polar codes in 5G communication which meets the simple and ultra-low latency can also be improved. This paper aims to improve the polar codes to achieve simplicity and hardware friendliness. After trained by the neural network, we get a simple math formula to estimate the reliability of the encoding in the polar codes, which can achieve the same effect as GA (Gaussian Algorithm). In terms of decoding, by using NND (Neural Network Decoding) to decode polar codes, we achieve almost the same effect as the conventional SC (Successive Cancellation) algorithm. In order to achieve further improvements, we use the neural network to optimize BP (Belief Propagation) algorithm. Thanks to it, a small number of iterations can be enough to decode the polar codes, even with an iteration of one. The simulation shows that though the early stage of training is full of difficulties, the result can be used to achieve one-step and fool-type decoding.