New Min-Sum Decoders Based on Deep Learning for Polar Codes

• Published in: 2018 IEEE International Workshop on Signal Processing Systems (SiPS) pp. 252 - 257
• Main Authors: Dai, Bin, Liu, Rongke, Yan, Zhiyuan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2018
• Subjects: Approximation algorithms; belief propagation; Complexity theory; Decoding; Deep learning; Iterative decoding; Neural networks; Polar codes; Signal processing algorithms
• ISSN: 2374-7390
• DOI: 10.1109/SiPS.2018.8598384

In this paper, we propose two novel min-sum (MS) decoding algorithms based on deep learning for polar codes, an offset min-sum (OMS) and a scaling offset min-sum (SOMS) algorithm. The parameters of both algorithms are different from iteration to iteration, and are obtained by training over a deep neural network. Our simulation results show that the OMS algorithm has roughly the same error performance as a previously proposed multiple scaling min-sum (MSMS) algorithm, and that the SOMS algorithm performs better than all existing BP-based algorithms. Since the OMS algorithm requires only an addition as opposed to a multiplication in the MSMS algorithm, the OMS algorithm is more suitable for hardware implementation. The two proposed decoding algorithms provide a tradeoff between complexity and error performance.