Adaptive Learned Belief Propagation for Decoding Error-Correcting Codes

Area of Science:

• Coding Theory
• Digital Communications
• Machine Learning

Background:

• Weighted belief propagation (WBP) is a decoding technique for linear block codes.
• Traditional WBP uses static, offline-optimized weights.
• Adaptive methods are needed to improve decoding performance dynamically.

Purpose of the Study:

• To introduce and investigate adaptive weighted belief propagation (WBP) decoders.
• To evaluate the performance of adaptive WBP variants against static WBP.
• To demonstrate the effectiveness of adaptive WBP in practical communication scenarios.

Main Methods:

• Unrolling the Tanner graph of a code for belief propagation iterations.
• Assigning and optimizing edge weights for a recurrent network.
• Developing parallel WBP decoders with discrete weights.
• Implementing a two-stage decoder using a neural network for dynamic weight determination.

Main Results:

• Adaptive WBP decoders achieve up to an order of magnitude lower bit error rates (BERs) than static WBP.
• Performance improvements are observed across various codes (BCH, polar, QC-LDPC) and channel conditions (AWGN).
• In a concatenated code application for optical fiber channels, adaptive WBP provided a 0.8 dB coding gain.

Conclusions:

• Adaptive WBP offers significant performance enhancements over static WBP for linear block code decoding.
• The proposed adaptive decoders maintain comparable computational complexity and decoding latency.
• Adaptive WBP is a promising technique for improving the reliability of digital communication systems.