Model-Driven Deep Learning ADMM Decoder for Irregular Binary LDPC Codes

Abstract

In this letter, a model-driven deep learning (DL) decoder for irregular binary low-density parity-check (LDPC) codes is proposed via the alternating direction method of multipliers (ADMM) technique. Our technical contributions three twofold: 1) we formulate the maximum likelihood decoding problem as a non-convex quadratic program with a new penalty term and present an ADMM based decoder; 2) to alleviate hyper-parameter tuning, we employ the deep unfolding strategy to derive a model-driven ADMM-DL decoder; and 3) to save the number of learning parameters, we propose an initialization strategy which can apply to different codeword structures. Specifically, the designing procedure for the DL network structure and DL network parameters training algorithm are presented in detail for efficient implementation. Numerical results demonstrate that the proposed model-driven ADMM-DL decoder for irregular binary LDPC codes is competitive in comparison with the state-of-the-arts.