Research On Decoding Algorithm Of Polar Codes Based On Neural Network

Polar codes have attracted much attention in the communication field because of their unique recursive structure characteristics and low encoding and decoding complexity.In terms of its decoding,the improved algorithm of Successive Cancellation(SC)decoding has excellent performance.But for long codewords,its latency and throughput rate are not sufficient for high-speed communication because of its inherent serial structure.Therefore,it is only suitable for short code scenarios in 5G communication.Due to the powerful learning ability and offline operation of neural network,it has been proposed in the current research to combine various types of neural networks with polar codes to improve decoding performance and latency.However,due to the considerably training complexity,the performance of neural network decoder for long codes need to be further studied.Firstly,the thesis uses Multilayer Perceptron(MLP),Convolutional Neural Network(CNN),Long Short Term Memory(LSTM)and Gated Recurrent Unit(GRU)as decoders to decode polar codes.The following parameters are mainly discussed: activation function,loss function,optimization function,the number of layer nodes,signal-to-noise ratio and the size of the convolution kernel unique to CNN.The results are simulated in the case of code length N=16 and information length K=8.Simulation results show that:(1)In terms of activation function,Re LU performs well among the MLP,LSTM and GRU,and the Swish activation function is more suitable for CNN;(2)In terms of loss function,binary＿crossentropy is suitable as the loss function of MLP and CNN,and log＿cosh is suitable as the loss function of LSTM,and MSE better matches GRU;(3)In terms of optimization function,Adam and Nadam are suitable for MLP and RMSprop for CNN and GRU,while Nadam achieves excellent performance in LSTM;(4)In terms of the number of layer nodes,combining of the decoeding performance,parameter amount,calculation,the number of nodes in each layer [256,128,62] used in three layers of MLP and CNN shows the best performance;(5)In terms of signal-to-noise ratio,training set signal-to-noise ratio 0d B is suitable for MLP network,2d B is suitable for CNN network,1d B is suitable for LSTM network and 3d B is suitable for GRU network;(6)In terms of the size of the convolution kernel,when the number of convolution kernel [256,128,64],it shows the best performance that the size of convolution kernels is 1×4.The simulation results of this experiment show that the performance of the four network decoders based on the best parameters are better than SC decoder.Secondly,this thesis deeply studies the neural network decoders based on the long polar codes by combines the MLP neural network with SC decoder.A SC decoder for long polar codes can be divided into several SC decoders for short codes,which can be repalced by the MLP neural network with the best parameters selected in the previous section.Simulation results indicate that the decoding performances of MLP-SC on N=32,K=16 and N=64,K=32outperform traditional SC decoder.Finally,this thesis improve the learning ability of the entire network by increasing the number of network layers,and proposes a decoder Res Net-MLP-SC that combines residual network and MLP-SC.Quick connection of the residual network is used to make the neural network to avoid losing information in the process of transmission,and to solve the problem of network performance degradation,so that the network achieves better decoding performance.The simulation results show that the decoding performance of Res Net-MLPSC based on N=32,K=16 and N=64,K=32 achieve slightly better than that of MLP-SC in the case of corresponding code length.After comprehensive analysis,it is concluded that the decoding performance of Res Net-MLP-SC is approximately 0.25 dB to 0.35 dB gain compared to SC decoding performance.