Research On Decoding Algorithms Of Polar Codes Based On Machine Learning

Polar codes are the first capacity-achieving code family,which have been adopted in the 5th generation wireless system(5G),representing the advanced trend in the channel coding field.Nevertheless,conventional polar decoding algorithms still need to be improved in terms of error performance and latency.Despite mitigating performance defects,decoding with machine learning(ML)techniques lacks the corresponding theoretical basis and incurs extra storage and training burdens.The current research on the theoretical framework of ML-based polar decoding algorithms and the neural decoders with low complexity and high reliability is far from sufficient.On this account,based on the channel polarization theory,the thesis inves-tigates the analytical theory and critical techniques of ML in polar decoding.Guided by the theoretical framework,ML-based polar code high-performance float-point decoding,quantization decoding,reinforcement learning based de-coding and blind decoding(blind detection)are comprehensively studied from two perspectives:model-driven and data-driven respectively,which ensures an efficient and reliable transmission in polar coding.The innovation work of this thesis consists of the following four aspects:1)The thesis investigates and designs a high-performance ResNet-based BP decoding algorithm for polar codes.Since the conventional model-driven decoders have numerous trainable parameters,they usually lead to additional storage and training complexities.For this reason,mutual information is used as the analysis tool of decoding performance.From the message-passing point of view,we build a residual BP neural decoder by introducing ResNet structure to the polar BP factor graph.During the process of message update,part of the previous information is inherited through skip connections,thus increasing the information amount delivered in BP decoding.Trainable damping factors are introduced to adjust the ratio of current messages to historical messages.Simu-lation and numerical results indicate that the residual BP decoder can effectively improve the error performance and accelerate the convergence of standard po-lar BP decoding.Furthermore,we demonstrate that the error performance is insensitive to the individual optimization of the damping factors and just one shared constant damping factor can achieve comparable performance to the ful-ly parameterized structure.2)Based on weighted decoding trellis,the thesis designs high-performance model-driven decoders to improve the quantization decoding performance of polar codes.Quantization is essential for the hardware implementation of the decoder,but it causes BLER loss.To tackle this issue,we carefully design the ML-based quantization polar decoders to compensate for the performance loss caused by quantization.Following the model-driven idea,we build a weighted SC neural decoder by introducing trainable weights to the SC trellis.Besides,we analyze the redundant operations in SC decoding and simplify the struc-ture of the weighted SC decoder.By sharing weights,the extra storage and training burdens caused by trainable parameters are alleviated.Moreover,we extend the weight-sharing scheme to the weighted BP decoder to optimize the error performance and the convergence speed under one-bit quantization.Fi-nally,we analyze the structural features of the neural decoders and propose a symmetry-based analysis method that accounts for why the model-driven neu-ral decoders can be trained with a single codeword.Simulation results verify the effectiveness of the proposed weighted SC and BP decoders.3)The thesis investigates the RL-based decoding algorithms for polar codes and the bit decision strategy is explored.First of all,we establish the theoretical framework to model polar decoding as a Markov decision process.Under this framework,we design a state-behavior space and a reward strategy based on the path metric(PM)value for polar decoding.On this basis,we design two RL polar decoding schemes based on Q-learning and deep-Q network(DQN),respectively.In the Q-learning scheme,the storage overhead is reduced by the low-complexity Q-table construction method.In the DQN scheme,we con-struct an efficient decision network.By fitting the Q-function,the proposed RL decoder achieves a comparable performance to the existing polar decoders.4)The thesis utilizes neural network to design a high performance blind decoding algorithm for polar codes.First,we take a performance analysis of the conventional polar code blind decoding algorithm in 5G.By calculating the false alarm rate(FAR)error,the essential cause of rate loss in the conventional method is unveiled.On this account,we utilize a data-driven fully-connected neural network(FCNN)to improve the accuracy of the decision in blind decod-ing.The FCNN autonomously extracts the features of the ratios of the squared Euclidean distance between the received signal and the valid transmitted signal to that between the received signal and the null signal and determines whether the transmitted signal contains target downlink control information through the classification threshold.Further,we propose a quantitative analysis theory by tracking the probability distribution of the above features.This theory can quickly calculate the classification threshold when the block error rate(BLER)increment or the FAR decrement is given.Simulation results show that the pro-posed blind decoding algorithm can effectively reduce the CRC overhead and reduce the rate loss when reaching the target BLER and FAR.In summary,on the basis of channel polarization theory and machine learn-ing,the thesis designs a model-driven high-performance decoding algorithm for polar codes,and extends the model driven idea to the quantization decod-ing.This thesis further gets rid of the limitation of the decoding trellis,and designs a data-driven RL-based decoding algorithm and a blind decoding al-gorithm.The innovative work ensures an efficient and reliable transmission in polar coding.