Research On LDPC Quantized Decoding Algorithm Based On Neural Network

In recent years,deep learning(DL)has been widely applied and achieved remarkable performance in many fields,such as computer vision,speech processing,and so on.The application of DL methods has been gradually extended to the network layer and physical layer of communication framework.Since the unfolded structure of the iterative decoding algorithms for low-density parity-check(LDPC)codes is similar to a neural network,it seems to be natural to attempt the use of neural network on the LDPC decoding.However,in existing work messages are all in floating-point precision,without any consideration on hardware implementations in real systems.As a necessary consideration for hardware implementations,the introduction of quantization directly affects the decoding performance.This dissertation investigates quantized LDPC decoding algorithms based on neural network to achieve the performance optimization.For 5G NR LDPC codes,this dissertation investigates the neural floating-point decoding algorithms,neural quantized decoding algorithm and the optimization of quantized decoding performance.Firstly,the dissertation implements the neural floating-point decoding algorithm,and the weights attached to the edges in the Tanner graph are trained,which efficiently reduces the correlation in message passing caused by the short cycles.Consequently,a significant performance gain can be achieved for short codes.Based on this,the dissertation proposes the neural quantized decoding algorithm which solves two problems:one is that weight quantization is not considered in the neural decoder,and the second is that the gradient vanishing makes the network untrainable.The proposed decoder employs differentiable soft quantization(DSQ)function during the training phase to overcome gradient vanishing problem.In addition,it uses the parameter sharing and an iteration-by-iteration training method to realize weight quantization.Finally,to achieve the optimization of neural quantized decoding,this dissertation evaluates the decoding performance of LDPC codes from the perspectives such as training method,signal-tonoise ratio and so on.And we look into the sensitivity of the neural decoder to these factors.On the other hand,the T-EXIT is used as a theoretical tool to optimize the quantized decoding process,so that the best tradeoff can be obtained between decoding performance and quantized bit width.Simulation results show that the proposed decoder clearly outperforms the traditional BP decoding algorithm for short codes.And the quantized neural decoder with as small quantization bit width as possible outperforms the MS decoding algorithm.The above results are all validated by trajectory-based extrinsic information transfer(T-EXIT)charts from a theoretical point of view.