ADMM Neural Channel Decoding for Low-SNR Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classical linear programming (LP) decoders for channel decoding in wireless communications have high computational complexity and lower error correction performance, especially in low signal-to-noise ratio (SNR) regions compared to belief propagation (BP) decoders, and existing deep learning methods lack efficient training processes.
Innovation Solution
A deep learning-based channel decoding method using the Alternating Direction Method of Multipliers (ADMM) is proposed, where ADMM iterations are unfolded to construct a network, and penalty parameters are converted into network parameters for training, enabling improved decoding performance with reduced computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a linear programming (LP) decoder is used for channel decoding, then theoretical decoding performance is guaranteed, but computational complexity increases and error correction performance decreases in low SNR regions
Solution Approach 1:
The patent replaces the traditional linear programming decoder with a deep neural network that learns optimal decoding strategies. The neural network substitutes the mechanical LP optimization process with learned patterns, achieving comparable reliability with reduced computational complexity during actual decoding operations.
Solution Approach 2:
The patent performs preliminary training of the neural network offline using training data to learn optimal decoding parameters. This preliminary action allows the network to store learned decoding strategies, eliminating the need for complex real-time LP optimization during actual channel decoding while maintaining theoretical performance guarantees.
2Device complexity
If a classical belief propagation (BP) decoder is used, then computational complexity is reduced, but error correction performance decreases in low signal-to-noise ratio regions
Solution Approach 1:
The patent transforms the fixed-parameter BP decoder into a neural network with learnable parameters. By changing from static BP parameters to dynamic learned parameters, the system achieves both low computational complexity (like BP) and high error correction performance (surpassing traditional BP in low SNR regions).
Solution Approach 2:
The neural network applies different learned parameters to different input conditions and error patterns, providing locally optimized decoding strategies. This allows the system to adapt to specific low SNR conditions rather than using uniform BP parameters, improving error correction where it matters most while maintaining overall computational efficiency.
3Reliability
If deep learning techniques are applied to channel decoding, then decoding performance can be improved, but training complexity and computational resources increase
Solution Approach 1:
The patent performs all complex training operations as a preliminary offline action. The neural network is trained once using training data to learn optimal decoding parameters, and these learned parameters are then deployed for actual channel decoding. This separates the complex training phase from the simple deployment phase, making deep learning practical for real-time applications.
Solution Approach 2:
The patent creates a simplified copy of the complex training process by using the trained neural network parameters directly in deployment. Instead of repeating complex training during actual decoding, the system copies the learned parameters into the decoder, achieving high performance with minimal real-time computational resources.
Data Source
AI summary
A channel decoding method includes constructing a maximum likelihood decoding problem including an objective function and a parity check constraint; converting the parity check constraint in the maximum likelihood decoding problem into a cascaded form, converting a discrete constraint into a continuous constraint, and adding a penalty term to the objective function to obtain a decoding optimization problem with the penalty term; obtaining ADMM iterations according to a specific form of the penalty term, and obtaining a channel decoder based on the ADMM with the penalty term; constructing a deep learning network according to the ADMM iterations, and converting a penalty coefficient and a coefficient contained in the penalty term into network parameters; training the deep learning network with training data offline and learning the network parameters; and loading the learned network parameters in the channel decoder based on the ADMM with the penalty term, and performing real-time channel decoding.