Study On Decoding Algorithms Of Polar Codes Based On Deep Learning

In the channel coding schemes,polar codes have low complexity and have been strictly proved to reach the Shannon limit.Therefore,polar codes have very high research value.In the common decoding algorithms of polar codes,successive cancelation(SC)decoding algorithm has the lowest complexity,so SC decoding algorithm has always attracted the attention.However,since the calculation process of the SC decoding algorithm is recursive,it has a large decoding delay.Deep learning has a strong computational advantage.The neural network is static after training and the data only needs to pass through the network once.Therefore,deep learning is applied to decoding process of polar codes,which can effectively reduce decoding latency and improve decoding efficiency.This paper mainly studies decoding algorithms of polar codes based on deep learning,and improves the SC decoding algorithm in terms of decoding latency.The specific work is summarized as follows:(1)The basic principles of encoding and decoding of polar codes and deep learning are studied,and the SC decoding algorithm is simulated.The simulation results show that when the length of polar codes is longer,the bit error rate performance of SC decoding algorithm is better.(2)Neural network decoders(NND)based on multi-layer perceptron(MLP),convolutional neural network(CNN)and recurrent neural network(RNN)for polar codes are designed and trained respectively.The network design and parameter setting in the neural network decoders are discussed,and the decoding performance of the neural network decoders based on different network models is compared.By comparison,the CNN-based decoder has better performance with fewer parameters.(3)Partitioned neural network decoding algorithm of polar codes is designed.According to the coding characteristics of polar codes,the long code words are partitioned for decoding.All sub-blocks are decoded by neural network decoders,and SC decoding algorithm is used to couple each sub-block to complete the decoding.The PNND of polar codes with code length of 32,64,128 and code rate of 0.5 is implemented,and the decoding efficiency is improved and decoding latency is reduced when the bit error rate performance is very close to that of SC decoding algorithm.(4)The PNND with fixed-length partitions is designed,and it can decode polar codes of arbitrary code length and arbitrary code rate.And it can effectively reduce decoding latency.In this paper,the PNND with the fixed-length partitions that each sub-block has 16 bits is implemented,and the simulation result of decoding for polar codes that length is 2048 and rate is 0.5 is given.The simulation result shows that the decoding efficiency is improved when the bit error rate performance is very close to that of SC decoding algorithm.