[Paper Review] A Model-Driven Deep Learning Network for MIMO Detection

Hengtao He, Chao-Kai Wen+2 more|arXiv (Cornell University)|Sep 25, 2018

Wireless Signal Modulation Classification16 references4 citations

TL;DR

This paper proposes OAMP-Net, a model-driven deep learning network for MIMO detection that unfolds the orthogonal approximate message passing (OAMP) algorithm into a neural network with trainable parameters per layer. By optimizing these parameters via back-propagation, the network achieves significant performance gains over conventional OAMP, especially in Rayleigh and correlated MIMO channels, with gains of up to 2.97 dB in BER performance at SNR levels for MIMO systems with 8x8 antennas.

ABSTRACT

In this paper, we propose a model-driven deep learning network for multiple-input multiple-output (MIMO) detection. The structure of the network is specially designed by unfolding the iterative algorithm. Some trainable parameters are optimized through deep learning techniques to improve the detection performance. Since the number of trainable variables of the network is equal to that of the layers, the network can be easily trained within a very short time. Furthermore, the network can handle time-varying channel with only a single training. Numerical results show that the proposed approach can improve the performance of the iterative algorithm significantly under Rayleigh and correlated MIMO channels.

Motivation & Objective

• To address the performance-complexity trade-off in MIMO detection by integrating deep learning with iterative detection algorithms.
• To develop a trainable, model-driven neural network that inherits the structure of the OAMP algorithm for improved detection accuracy.
• To enable fast and stable training with minimal trainable parameters, suitable for time-varying channels.
• To achieve soft-output detection for compatibility with modern wireless systems.
• To demonstrate superior performance over conventional OAMP and existing deep learning-based detectors in both independent and correlated MIMO fading channels.

Proposed method

• The OAMP-Net architecture is derived by unfolding the iterative OAMP algorithm into T layers, with each layer incorporating trainable parameters (γt, θt) to refine estimation.
• The network uses a real-valued representation of the complex MIMO system to enable standard deep learning training in the real domain.
• Trainable parameters are optimized via back-propagation using an L2 loss function and Adam optimizer on synthetic training data with QPSK modulation.
• The network is trained once and can generalize to time-varying channels without retraining, due to the minimal number of trainable variables.
• The method enables soft decision outputs by leveraging the posterior mean estimates from each layer.
• The network structure is designed to maintain the Bayesian optimality of OAMP while enhancing performance through end-to-end learning.

Experimental results

Research questions

• RQ1 Can deep learning improve the performance of the OAMP algorithm in MIMO detection while maintaining low complexity?
• RQ2 How does the performance of the OAMP-Net compare to the original OAMP and other deep learning-based detectors in Rayleigh fading MIMO channels?
• RQ3 Can the OAMP-Net generalize to time-varying channels with only a single training?
• RQ4 What performance gain does the OAMP-Net achieve over OAMP in correlated MIMO channels?
• RQ5 How does the number of trainable parameters affect training speed and convergence stability?

Key findings

• The OAMP-Net achieves a BER improvement of approximately 1.37 dB over the OAMP algorithm at BER = 10⁻³ for a 4×4 MIMO system under Rayleigh fading.
• For an 8×8 MIMO system, the OAMP-Net provides a performance gain of about 2.97 dB over OAMP at the same BER level.
• In a 64×64 MIMO system, the OAMP-Net still achieves a 0.82 dB gain over OAMP, demonstrating scalability.
• Under correlated MIMO channels with ρ = 0.5, the OAMP-Net achieves a performance gain of over 2.15 dB, exceeding the 1.86 dB gain in independent Rayleigh channels.
• The OAMP-Net maintains high performance even when the channel is correlated, with only 6.05 dB SNR loss at BER = 10⁻², compared to 6.34 dB for the OAMP algorithm.
• The network is trained in only 10,000 epochs with a small number of trainable parameters, enabling fast and stable convergence.