부산대학교 도서관

The min-sum (MS) algorithm can decode Low-densityparity-check (LDPC) codes with low computational complexity atthe cost of slight performance loss. It is an effective way to realizehardware implementation of the min-sum decoder by quantizingthe floating belief messages (i.e., check-to-variable messages andvariable-to-check messages) into low-resolution (i.e., 2–4 bits) ver sions. However, such a way can lead to severe performance degra dation due to the finite precision effect. In this paper, we proposea neural-network optimized low-resolution decoding (NOLD) al gorithm for LDPC codes to deal with the problem. Specifically,the optimization of decoding parameters (i.e., scaling factors andquantization step) is achieved in a hybrid way, in which we con catenate a NOLD decoder with a customized neural network. Alllearnable parameters associated with the decoding parameters areassigned to each neuron in the proposed method. What’s more, wedesign a new activation function whose outputs are close to the em ployed quantizer ones when network parameters are finally opti mized off-line. Finally, the performance of the proposed method isverified by numerous experiments. For the case of 2-bit decoding,the proposed approach significantly outperforms several conven tional decoders at the expense of slightly increased off-line trainingtime. Besides, the proposed method with 4-bit quantization incursonly 0.1 dB performance loss compared with the floating min-sumdecoder at the coded bit-error-rate of 10−5. Moreover, we showthat the proposed NOLD decoder works over a wide range of chan nel conditions for regular and irregular LDPC codes. Simulationcode for reproductive results is publicly available1.