Deep Unfolding-Aided Gaussian Belief Propagation for Correlated Large MIMO Detection

Abstract

This paper proposes a deep unfolding-aided belief propagation (BP) for large multi-user multi-input multi-output (MU-MIMO) detection under correlated fading channels. A BP-based detector is a well-known strategy for realizing large-scale MU detection (MUD) with low-complexity and high-accuracy. However, its convergence property is severely degraded under insufficient large-system conditions and spatial fading correlation among RX antenna elements. To compensate for this drawback, we design a trainable Gaussian BP (T-GaBP) having well-organized trainable internal parameters based on the BP structure. These parameters are optimized by the deep learning techniques in the signal-flow graph of unfolded GaBP; this approach is referred to as data-driven tuning. By training the parameters according to the system model, T-GaBP can maintain the high detection capability even in practical system configurations that differ from the ideal uncorrelated massive MIMO assumption. Numerical results show that the proposed detector improves the convergence property and achieves a comparable detection performance to the cutting-edge expectation propagation (EP) detector in correlated MUD, with a lower computational cost.