Deep Ensemble of Weighted Viterbi Decoders for Tail-Biting Convolutional Codes

Area of Science:

• Coding Theory
• Machine Learning
• Digital Communications

Background:

• Tail-biting convolutional codes offer flexible termination compared to zero-termination codes.
• Efficient decoding is crucial, especially in short code length regimes like LTE.
• Existing methods like the Circular Viterbi Algorithm (CVA) provide a baseline for decoding.

Purpose of the Study:

• To develop a machine learning-based decoder for tail-biting convolutional codes.
• To improve decoding performance in the short code length regime, relevant to standards like LTE.
• To enhance the state-of-the-art in decoding tail-biting codes, particularly in the waterfall region.

Main Methods:

• Parameterization of the Circular Viterbi Algorithm (CVA) into weighted decoders.
• Ensemble creation combining specialized decoders covering the entire state space.
• Implementation of a non-learnable gating mechanism and CRC-based expert selection.

Main Results:

• Achieved up to 0.75 dB Frame Error Rate (FER) improvement over CVA in the waterfall region.
• Demonstrated effectiveness across multiple code lengths.
• Introduced negligible computational complexity increase compared to CVA at high SNRs.

Conclusions:

• The proposed machine learning approach significantly enhances tail-biting convolutional code decoding.
• The ensemble method with gating and CRC filtering offers a practical improvement over standard CVA.
• This technique is particularly beneficial for short code lengths and improves performance in critical regions.