Partial learning for MIMO detection

Abstract

Reliable and efficient multiple-input multiple-output (MIMO) detection remains a central challenge in modern wireless receivers. Optimal maximum-likelihood (Max-L) detection delivers the best performance. However, its exponential complexity is prohibitive, while linear schemes such as zero-forcing (ZF) and minimum mean square error (MMSE) are computationally attractive yet they suffer from poor performance. Fully learned detectors improve robustness but introduce substantial parameter counts and computational complexity. Building on prior work on partial learning (PL), this thesis contributes a unified detection framework based on PL that addresses these trade-offs by applying learning only where it yields the most benefits: a subset of the weakest symbol streams, with the remaining streams detected using low-complexity linear detection.

The first part of the thesis designs a soft-output PL demapper implemented with a small fully connected neural network (FCNN) for quasi-static channels and embeds it into an iterative detection. The inner MIMO detector produces log-likelihood ratios (LLRs) that are exchanged with an outer convolutional decoder. EXIT charts and decoding trajectories are used to analyze convergence. Across representative 2×2 and 4×4 quadrature phase-shift keying (QPSK) systems, the iterative PL (Iter-PL) technique closes most of the gap to iterative Max-L and full-learning detectors while operating at a fraction of their complexity. Operation counts are reported and related to the number of learning-assisted streams d, demonstrating explicit performance versus complexity trade-off.

The second part extends Iter-PL to time-varying channels, while also considering channel state information (CSI) error. The same FCNN-based soft demapper is trained using CSI errors. Results show that Iter-PL retains its iterative gains under 5% CSI error and remains markedly superior to purely linear detection. An adaptive PL strategy is further introduced to select d based on the average received signal-to-noise ratio (SNR), thereby achieving a near-constant target bit error rate (BER) with reduced average complexity.

The final part addresses scalability in dynamic multi-user uplinks. A graph neural network (GNN)–based PL detector is proposed, where an approximate message passing (AMP) frontend supplies soft symbols and variance estimates to the GNN. The GNN then detects only the d weakest users, while ZF detects the remaining users. By operating on user graphs, the model generalizes across changing activity masks without requiring retraining and maintains a low parameter count. Simulations over multiple activity patterns consistently confirm low BER and favorable performance–complexity trade-offs.

Overall, the thesis demonstrates that PL enables near-optimal soft detection, accompanied by clear and quantifiable reductions in complexity, and that GNN-based partial learning offers the same benefits in multi-user scenarios. The proposed technique offers a practical approach to scalable, low-latency MIMO detection, making it suitable for evolving wireless systems.

More information

Identifiers

ORCID for Abdulaziz Mohammed A Babulghum: orcid.org/0009-0000-2837-8652

ORCID for Mohammed El-Hajjar: orcid.org/0000-0002-7987-1401

ORCID for Michael Ng: orcid.org/0000-0002-0930-7194

ORCID for Chao Xu: orcid.org/0000-0002-8423-0342

Catalogue record

Export record