Neural-Enhanced Quantized Min-Sum Successive Cancellation Decoding for Polar Codes

Neural-Enhanced Quantized Min-Sum Successive Cancellation Decoding for Polar Codes

Abstract—Polar codes have been adopted as the channel coding scheme for the 5G NR control channel, yet achieving hardware-efficient decoding without sacrificing error-correction performance remains a challenge. The Min-Sum(MS) approximation simplifies the successive cancellation (SC) f-function at a cost of systematic magnitude overestimation, conventional correction schemes—Normalized Min-Sum(NMS) and Offset Min-Sum(OMS)—rely on fixed parameters optimized under full-precision conditions, rendering them structurally unable to compensate for the nonlinear distortions introduced by ultra-low bit-width quantization. This paper proposes the Neural-network-enhanced Quantized Min-Sum SC (NQ-MS-SC) decoder, in which a lightweight, bit-width-adaptive neural network predicts input-dependent normalization and offset parameters directly from quantized LLRs, jointly trained with a quantization-aware strategy using the Straight Through Estimator to simultaneously address both Min-Sum approximation error and quantization induced distortion. Simulation results demonstrate that NQ-MS-SC outperforms conventional quantized Min-Sum decoders, with gains that intensify at higher SNR, coarser quantization, and longer code lengths.

Keywords-Successive; Cancellation; Decoding; Fixed-Point Quantization; Neural Network