Learning Stabilization in Deep Unfolding of Generalized Approximate Message Passing

Generalized approximate message passing (GAMP) achieves near-optimal performance in detecting spatially multiplexed massive multiple-input multiple-output (MIMO) signals with significantly reduced computational complexity under independent and identically distributed (i.i.d.) measurements. However, in spatially correlated MIMO channels where the ideal assumption of large-scale uncorrelated observation does not hold, the detection capability severely deteriorates. This performance degradation can be compensated for by optimizing GAMP with embedded learnable parameters via deep unfolding (DU) techniques, i.e., data-driven tuning; however, the learning process becomes quite unstable. To address this issue, we propose a novel method that achieves stable learning by incorporating a monotonic increase constraint on the reliability of propagated messages by learning the differential (incremental) values of the learnable parameters between two consecutive iterations. The efficacy of the proposed method is confirmed through numerical results in terms of loss trajectory in the learning process and bit error rate (BER) of massive MIMO detection.