RL-Based Parallel LDPC Decoding with Clustered Scheduling.

Area of Science:

• Coding Theory
• Machine Learning
• Digital Communications

Background:

• Parallel decoding of Low-Density Parity-Check (LDPC) codes is crucial for high-throughput communication systems.
• Balancing error-correction performance, decoding latency, and memory conflicts is a key challenge.
• Existing reinforcement learning (RL)-based scheduling methods face significant storage complexity.

Purpose of the Study:

• To develop an RL-based decoding framework for high-throughput parallel LDPC code decoding.
• To address the trade-off between performance and latency in parallel LDPC decoders.
• To reduce the storage complexity of RL-based scheduling methods.

Main Methods:

• Constructing clusters of check nodes with a two-edge independence property for conflict-free belief propagation.
• Training an RL agent offline to assign Q-values and prioritize cluster updates.
• Introducing the Q-Sum method to approximate cluster Q-values, reducing storage complexity.
• Proposing an On-the-Fly clustering strategy for dynamic two-edge independence enforcement.

Main Results:

• The proposed methods improve the latency-versus-performance trade-off for parallel LDPC decoders.
• Achieved lower decoding latency and higher throughput compared to existing methods.
• Maintained error rates comparable to state-of-the-art decoding techniques.

Conclusions:

• The RL-based framework with Q-Sum and On-the-Fly clustering offers an efficient solution for parallel LDPC decoding.
• These advancements enable faster and more efficient communication systems.
• The proposed techniques provide flexibility and improved performance in various decoding scenarios.