Dynamic Conjugate Gradient Unfolding for Symbol Detection in
Time-Varying Massive MIMO
Abstract
This work addresses symbol detection in timevarying Massive
Multiple-Input Multiple-Output (M-MIMO) systems. While conventional
symbol detection techniques often exhibit subpar performance or impose
significant computational burdens in such systems, learning-based
methods have shown potential in stationary scenarios but often struggle
to adapt to nonstationary conditions. To address these challenges, we
introduce a hierarchy of extensions to the Learned Conjugate Gradient
Network (LcgNet) M-MIMO detector. Firstly, we present Preconditioned
LcgNet (PrLcgNet), which incorporates a preconditioner during training
to enhance the uplink M-MIMO detectorâ\euro™s filter matrix. This
enhancement enables the detector to achieve faster convergence with
fewer layers compared to the original, nonpreconditioned approach.
Secondly, we introduce an extension of PrLcgNet, known as the Dynamic
Conjugate Gradient Network (DyCoGNet), specifically designed for
time-varying environments. DyCoGNet leverages self-supervised learning
with forward error correction, enabling autonomous adaptation without
the need for explicit labeled data during training. It also employs
metalearning, facilitating rapid adaptation to unforeseen channel
conditions. Our simulation results demonstrate that PrLcgNet achieves
faster convergence, lower residual error, and comparable symbol error
rate (SER) performance to LcgNet in stationary scenarios. Furthermore,
in the time-varying context, DyCoGNet exhibits swift and efficient
adaptation, achieving significant SER performance gains compared to
baseline cases without metalearning and online self-supervised learning.