Neural Network Aided Low-Complexity Decoding for PAC Codes

Polarization-adjusted convolutional (PAC) codes are an advanced variant of polar codes, capable of approaching the dispersion bound in certain cases. However, to achieve the near-bound performance, the computational complexity is insufferable. In this letter, a neural network (NN)-aided low-complexity PAC decoding algorithm is proposed, which brings significant complexity reduction while maintaining the error-correction performance. Specifically, the NN is employed for the selection of the lowest complexity decoding algorithm from Fano sequential and list decoding algorithms. Moreover, we optimize the design of the NN to reduce the complexity it introduces. Simulation results show that the proposed algorithm reduces average complexity by up to 86.86% and 89.70% compared with the original Fano sequential and list decoding algorithms, respectively.