| In 2008,polar codes were first proposed by professor Arikan of Bilkent University.Due to its low complexity of coding and decoding,it has been proved theoretically that polar codes can reach Shannon limit when the code length goes to infinity,which has the potential to change the modern coding scheme.Successive Cancellation(SC)decoding algorithm and Belief Propagation(BP)decoding algorithm are polarization yards of two kinds of mainstream decoding algorithm.However,the two algorithms are limited by decoding delay and decoding performance,which are contrary to the requirement of the next generation communication system for high speed and low delay,so this needs to explore new decoding is proposed.In recent years,Deep Neural Network(DNN)has made breakthrough achievements in many fields,especially in image classification,speech recognition and other aspects,which has attracted extensive attention of the academic community.The general channel decoding problem can be regarded as a classification problem,so the neural network theory can be applied to the channel coding and decoding problem.It is of great theoretical and practical significance to use neural network technology to solve the decoding problem of polar codes and explore its decoding characteristics and potential performance gains.In this thesis,neural network technology is applied to the study of decoding polar codes.The main work contents are as follows:Firstly,the decoders of polar codes based on Multilayer Perceptron(MLP),Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN)are designed,and the model framework and parameter setting are given.Secondly,for the activation function of neural network for the further research,respectively,Hard_sigmoid function,Softsign function and Swish function are applied to the three decoding of neural network study,and make comparison with the current widely used Re LU function,exploring the Swish function has more excellent performance than Re LU function.The parameter values of Swish function are also discussed.Thirdly,based on MLP,CNN and RNN,the decoding performance and characteristics of three types of decoders are given,and further exploration is made on the basis of previous studies.Furthermore,the decoding performance of MLP under different activation functions,different parameter sizes,different codeword structures and different channels are compared,the decoding framework of CNN is improved,the decoding performance and decoding complexity of five RNN are compared,and the decoding performance of Gated Recurrent Unit(GRU)is explored.Fourthly,on the basis of decoding performance and decoding delay,the Neural Network Decoders(NND)are compared with the traditional BP and SC decoders to verify that the Neural network decoders have considerable decoding performance and low decoding delay,and can become a competitive decoding tool in the next generation of communication systems. |
