Research On A Deep Learning Based Method For Min-sum LDPC Decoding Under Quantization

The next-generation wireless communication system is conceived to be ultra-fast,low-latency,and ultra-reliable.Since the low-density parity-check(LDPC)codes have low decoding complexity,low error floor,and are capable of parallel operating,it is a decoding technology of great potential and broad prospects in the future communication system.It has been theoretically proved that the LDPC codes can achieve the performance near the Shannon limit under belief propagation(BP)decoding algorithms.The most common BP algorithms are the sum-product(SP)algorithm and the min-sum(MS)algorithm.The MS algorithm is more suitable for hardware implementation and its performance loss can be compensated by approximating the node message to the SP algorithm.In the hardware implementation,a quantizer is utilized to reduce the storage cost and power consumption but causes severe quantization noise which degrades the performance.To overcome this problem,we introduce an adaptive method to choose the quantization step size that fits the node message.The simulations show that the adaptive quantization outperforms the fixed quantization.We can further improve the quantized decoding performance when considering a joint optimization of the scaling factors and quantization step size instead of a signle optimization.Since the adaptive method is not so effective for joint optimization,we introduce the density evolution(DE)method.The DE method computes the theoretical error rate and searches for the parameters that minimize the error rate which means that the performance of the DE method depends on the robustness and the efficiency of the search algorithm.However,it's difficult to design a efficient search algorithm as the decoding parameters increase.Moreover,the prerequisites of the DE,that is,the independent and symmetric conditions are not satisfied in real scenarios.To this end,we propose a deep learning(DL)based approach with low network complexity.The simulations show that the learning-based method has better decoding performance compared with the adaptive method and the DE method under additive white Gaussian noise(AWGN).The performance can get close to the unquantized MS algorithm under 4-bit quantization.Numerical trials shows that the learning-based method also possesses good preformance under Rayleigh fading noise and correlated Gaussian noise.