Robust CDMA multiuser detection using a neural-network approach

Area of Science:

• Signal Processing
• Machine Learning
• Neural Networks

Background:

• A robust linear decorrelating detector (LDD) using Huber's M-estimation was recently proposed.
• Implementing LDD typically requires matrix inversion, which can be computationally intensive.

Purpose of the Study:

• To implement LDD and M-decorrelating detectors (MDD) using a three-layer recurrent neural network (RNN) without matrix inversion.
• To demonstrate computational savings and performance gains, particularly in impulsive noise environments.

Main Methods:

• Utilized a three-layer RNN to iteratively minimize a computational energy function for LDD implementation.
• Incorporated sigmoidal neurons in the RNN for MDD implementation.
• Modeled impulsive noise using non-Gaussian alpha-stable distributions and employed the geometric signal-to-noise ratio (G-SNR).

Main Results:

• Successfully implemented LDD and MDD using RNNs, eliminating the need for matrix inversion.
• Demonstrated significant computational savings in realistic network scenarios.
• Achieved further performance enhancements for subspace-based blind MDD by using robust initial estimates.

Conclusions:

• Recurrent neural networks offer an efficient alternative for implementing robust decorrelating detectors.
• The proposed RNN-based detectors show promise for signal processing in non-Gaussian noise environments.
• The method provides a computationally advantageous approach to robust signal detection.