Decoding Optimization for 5G LDPC Codes by Machine Learning

TL;DR: The proposed generalized minimum-sum decoding algorithm using a linear approximation (LAMS) for protograph-based low-density parity-check (PB-LDPC) codes with quasi-cyclic (QC) structures shows noticeable improvement over the normalized and the offsetMinimum-sum algorithms and even better performance than the belief propagation algorithm in some high signal-to-noise ratio regions.

Abstract: In this paper, we propose a generalized minimum-sum decoding algorithm using a linear approximation (LAMS) for protograph-based low-density parity-check (PB-LDPC) codes with quasi-cyclic (QC) structures The linear approximation introduces some factors in each decoding iteration, which linearly adjust the check node updating and channel output These factors are optimized iteratively using machine learning, where the optimization can be efficiently solved by a small and shallow neural network with training data produced by the LAMS decoder The neural network is built according to the parity check matrix of a PB-LDPC code with a QC structure which can greatly reduce the size of the neural network Since, we optimize the factors once per decoding iteration, the optimization is not limited by the number of the iterations Then, we give the optimized results of the factors in the LAMS decoder and perform decoding simulations for PB-LDPC codes in fifth generation mobile networks (5G) In the simulations, the LAMS algorithm shows noticeable improvement over the normalized and the offset minimum-sum algorithms and even better performance than the belief propagation algorithm in some high signal-to-noise ratio regions