Comparative Evaluation of Model Based Deep Learning Receivers in Coded MIMO Systems

Deep learning (DL) methods have shown potential in tackling the performance-complexity trade-off in multiple-input multiple-output (MIMO) detection. Unlike most studies that evaluate state-of-the-art (SoA) DL receivers in uncoded scenarios, our paper focuses on realistic coded MIMO systems. After a comprehensive literature review, three representative SoA model-based DL receivers viz: DetNet, OAMPNet2, and DUIDD (MMSE-PIC and LoCo-PIC) were selected and comprehensively evaluated. Our findings indicate that DL receivers such as DetNet and OAMPNet2, which base their classical designs on the principle of symbol denoising, fail to sustain their superior performances from uncoded systems to coded systems due to inaccurate residual noise statistics. In contrast, DUIDD, specifically designed for coded systems, achieves effective interference cancellation, resulting in improved coded bit error rates across i.i.d. Gaussian channels, suggesting promising avenues for future research. However, LoCo-PIC, which simplifies MMSE-PIC with a linear solution, suffers performance degradation in correlated urban microcell channels, highlighting the importance of considering the non-linear correlation impacts during detection. Additionally, this simplification leads to further degradation in out-of-distribution channel scenarios, emphasizing the need to address these impacts in realistic wireless systems with varying channel conditions.