Deep Reinforcement Learning-based LDPC Decoding Algorithms for Vehicular Communication Links

The rapid emergence of Cooperative Intelligent Transport Systems (C-ITS) and autonomous driving has necessitated ultra-reliable low-latency communication (URLLC) within highly dynamic vehicular environments. However, the prevalence of severe Doppler shifts and Non-Line-of-Sight (NLoS) conditions poses significant challenges for traditional Low-Density Parity-Check (LDPC) decoding algorithms. This paper proposes two novel Deep Reinforcement Learning (DRL)-based decoding frameworks that adaptively optimize error-correction performance over volatile Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) links. By leveraging the decision-making capabilities of DRL, the proposed algorithms—specifically the Bootstrap-Driven (B-D) DRL-Proposed Algorithm (2)—demonstrate a transformative improvement in both reliability and efficiency. Extensive simulations under Line-of-Sight (LoS) and NLoS scenarios reveal that the proposed frameworks significantly outperform the baseline DRL-WBF scheme. Most notably, B-D DRL-Proposed Algorithm (2) achieves a peak throughput exceeding 6 × 106 bps in LoS conditions and maintains a near-constant decoding latency of less than 0.005 ms, even at low Eb/No. Furthermore, it achieves a superior Bit Error Rate (BER) 10−5 in NLoS environments. These results underscore the potential of DRL to provide robust, real-time, and high-throughput decoding solutions for next-generation V2X communication standards. These results underscore the potential of DRL to provide robust, real-time, and high-throughput decoding solutions for next-generation V2X communication standards, thereby providing the deterministic low-latency foundation required to guarantee upper-layer statistical quality-of-service (QoS) in time-sensitive vehicular networks.