Neural-Model-Augmented Hybrid NMS-OSD Decoders for Near-ML in Short Block Codes
To address the fundamental trade-off among throughput, latency, and computational complexity in near-maximum-likelihood (ML) decoding of short codes (e.g., LDPC, BCH, RS), this paper proposes an NMS-OSD hybrid decoding architecture. The method integrates four key innovations: (1) a CNN-based soft-information aggregation model to enhance initial parity-check reliability; (2) an adaptive OSD path selection mechanism for dynamic pruning of low-probability candidates; (3) a sliding-window-assisted early termination strategy to reduce average iteration depth; and (4) an undetected-error identifier tailored for high-rate codes to improve reliability decision-making. Experimental results demonstrate that the proposed scheme achieves frame error rates approaching the ML bound while maintaining low complexity and high throughput. It significantly outperforms state-of-the-art decoders across multiple short codes, reducing average latency by up to 40% and computational load by 35%.