Demodulation Network Training with Non-Gaussian LLR Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning-based iterative demodulation algorithms using MSE or BCE loss functions do not necessarily minimize the bit error rate (BER) of the decoder output, as they are optimized for demodulator outputs rather than decoder outputs, and assume a Gaussian noise distribution, which may not align with the actual non-Gaussian noise components in radio communication systems.
Innovation Solution
A method involving training a network by unfolding an iterative algorithm for demodulation or demodulation using a machine learning technique with a loss function that accounts for the non-Gaussianity of the log likelihood ratio (LLR) distribution, generating learned parameters to improve the BER of the decoder output by making the LLR distribution more Gaussian, suitable for error correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning is used to optimize demodulator parameters using MSE or BCE loss functions, then demodulation accuracy is improved, but decoder output BER is not necessarily minimized
Solution Approach 1:
The patent changes the optimization parameter from MSE/BCE loss (which optimizes demodulator output) to a loss function that directly optimizes decoder output BER. This involves modifying the objective function to account for the entire detection chain including the decoder, not just the demodulator stage.
Solution Approach 2:
The patent introduces an intermediary step of analyzing the LLR distribution characteristics and its impact on decoder performance. By examining how the demodulator output distribution affects the decoder's error correction capability, the system bridges the gap between demodulation accuracy and final decoder BER.
2Ease of manufacture
If traditional loss functions (MSE/BCE) are used for training, then training simplicity is maintained, but error correction capability is not optimized
Solution Approach 1:
The patent modifies the loss function parameters to include terms that evaluate the Gaussianity of the LLR distribution and its impact on decoder performance. This involves adding new parameters related to distribution characteristics while maintaining the fundamental gradient descent optimization framework.
3Device complexity
If Gaussian noise assumption is used for optimization, then mathematical tractability is improved, but performance in non-Gaussian noise conditions deteriorates
Solution Approach 1:
The patent makes the system adaptive by dynamically adjusting the loss function based on the actual noise conditions observed in the channel. Rather than assuming a fixed Gaussian model, the system evaluates the actual LLR distribution characteristics and adapts the optimization criterion accordingly.
Solution Approach 2:
The patent introduces feedback mechanisms where the decoder performance is evaluated and used to adjust the demodulator parameters. The loss function incorporates feedback about how well the current parameters are actually performing in terms of decoder BER, creating a closed-loop optimization system.
Data Source
AI summary
A method implemented in a computer system includes training a network, which is obtained by unfolding an iterative algorithm for demodulation or demodulation and decoding, using a machine learning technique with a loss function that takes into account non-Gaussianity of a log-likelihood ratio (LLR) distribution calculated from an output of the network. The method further includes producing a first set of learned parameters of that iterative algorithm.