International Conference on Communications
Deep Reinforcement Learning Aided Scheduling Design for Pipelined Layered QC-LDPC Decoder
Xia An, Chao Zhang, Jian Song, Jintao Wang
Abstract
Due to the widespread use of Low-Density Parity-Check (LDPC) codes in modern wireless communication and broadcasting standards, the architectures of their decoders have been extensively discussed in the literature. In this paper, we propose a unified decoding framework for pipelined layered quasi-cyclic LDPC decoders, which encompasses various decoder architectures as special cases, each corresponding to a specific parameter setting within a three-dimensional parallel architecture. Based on this framework, we derive an optimal grid re-scheduling method to mitigate memory conflicts between layers under a fixed decoding order. We also explore the impact of both adjacent check nodes and previously processed check nodes on the number of inserted NOPs, which leads to a reformulation of the scheduling problem as an episodic Markov decision process. To solve this problem, we apply classical heuristic algorithms, such as genetic programming and Beam Search, and propose a Deep Reinforcement Learning aided approach to further enhance the performance. Experimental results demonstrate that our proposed method outperforms traditional approaches in terms of memory conflict resolution.