PhD defence of Toluwaleke Olutayo – Machine Learning Approaches to Symbol Detection in Massive MIMO Wireless Systems

Abstract

Massive Multiple-Input Multiple-Output (M-MIMO) systems are fundamental to achieving the high data rates and reliability required by future wireless networks. However, realizing the full potential of M-MIMO is hindered by significant challenges in symbol detection, particularly the high computational complexity and performance degradation under realistic, non-ideal operating conditions. Conventional detectors often struggle with complexity or rely on simplifying assumptions, such as stationary channels, white Gaussian noise, and perfect channel state information (CSI) knowledge, which do not hold in practice. This thesis addresses these critical limitations by developing and evaluating novel symbol detection algorithms tailored for practical M-MIMO deployments.

First, we introduce the Preconditioned Learned Conjugate Gradient Network (PrLcgNet), a learning-based detector that accelerates training convergence in stationary M-MIMO systems by incorporating a preconditioner during training. PrLcgNet achieves superior symbol error rate (SER) performance with reduced complexity compared to prior learning-based detectors. Building on this, we extend PrLcgNet to Dynamic Conjugate Gradient Network (DyCoGNet), tailored for time-varying channels. DyCoGNet leverages meta-learning and self-supervised learning guided by forward error correction (FEC), enabling rapid adaptation to unforeseen channel dynamics without labeled data, outperforming conventional and recent self-supervised benchmarks.

Second, we propose the Zero-Forcing based Latent Space Symbol Detector (ZF-LSSD) to effectively address the challenges associated with unknown or non-analytic noise distributions. ZF-LSSD combines zero-forcing initialization with score-based generative modeling using stochastic differential equations (SDEs) and a localized search strategy. This approach facilitates efficient approximate maximum likelihood detection within a latent space, circumventing the computational intractability posed by complex noise distributions in large-scale MIMO systems. Numerical simulations demonstrate that ZF-LSSD consistently outperforms existing benchmark methods across various massive MIMO configurations and diverse additive noise scenarios.

Finally, we introduce Attention-Based Successive Interference Cancellation (ASIC), a novel detection method tailored for massive MIMO systems experiencing imperfect CSI. ASIC integrates permutation-equivariant neural architectures with CSI-derived priors, dynamically adjusting the sequential decoding process to mitigate residual interference and channel uncertainty. Distinct from purely data-driven detectors, ASIC strikes a balance between model-based inference and learned attention mechanisms, significantly enhancing the robustness of SIC detectors under imperfect CSI conditions without incurring the substantial computational overhead typical of purely data-driven methods. Simulation results demonstrate that ASIC consistently maintains robustness across varying user counts and outperforms traditional SIC algorithms as well as learning-based baselines in scenarios with imperfect CSI.

Collectively, this thesis contributes novel M-MIMO symbol detection techniques that exhibit enhanced robustness and adaptability to practical channel impairments, thereby advancing the feasibility and performance of M-MIMO systems in real-world wireless communication scenarios.