TL;DR
This paper introduces two novel deep neural network (DNN) based detectors for massive MIMO systems. These detectors leverage the Alternating Direction Method of Multipliers (ADMM) framework to achieve improved performance and significantly reduced computational complexity, particularly for higher-order QAM signals.
What Problem Does This Paper Address?
Massive MIMO (Multiple-Input Multiple-Output) systems, which employ a large number of antennas at the base station, offer significant improvements in wireless communication capacity and reliability. However, accurately detecting the transmitted signals in these systems is computationally challenging, especially with higher-order modulation schemes like QAM. Traditional methods, including those based on ADMM, struggle with the optimal selection of penalty parameters, limiting their performance. Existing deep learning approaches often face limitations with specific modulation types or still incur high computational costs. This paper aims to bridge this gap by proposing DNN-based ADMM detectors that are both efficient and effective for high-order QAM signals.
Methods
The authors propose two novel DNN-based detectors by "unfolding" the iterations of the PS-ADMM algorithm.
- ADMM-PSNet: This detector uses a DNN architecture to learn the optimal penalty parameters of the PS-ADMM algorithm directly from data. This approach bypasses the need for manual tuning of these parameters, which is a known challenge in traditional ADMM methods.
- ADMM-HNet: This detector focuses on reducing computational complexity. It introduces a hidden layer MLP within each ADMM iteration to approximate the computationally expensive matrix inversion and multiplication operations. Crucially, it proposes a novel feature extraction method that avoids using raw channel and received signal data directly in each layer, preventing the network from getting stuck in local minima. Instead, it uses residual variables and derived features that evolve with each iteration.
Both methods are trained using stochastic gradient descent (SGD).
Datasets
The paper does not explicitly mention specific named datasets. Instead, it describes training the proposed detectors on synthetically generated data. The ADMM-PSNet is trained on 10,000 samples, and the ADMM-HNet is trained on 90,000 samples. The evaluation is performed using Monte Carlo simulations.
Key Insights and Ideas
- Learning Penalty Parameters: The ADMM-PSNet demonstrates that DNNs can effectively learn the crucial penalty parameters for ADMM, leading to improved performance without manual tuning.
- Complexity Reduction via Feature Engineering: The ADMM-HNet's innovative feature extraction strategy is key to its efficiency. By avoiding direct use of raw channel and received signal data in each layer and instead using evolving residual and derived features, it enables a more effective and less computationally intensive learning process.
- Layer-wise Training for Efficiency: The ADMM-HNet's layer-wise training approach, combined with its efficient feature extraction, allows it to achieve competitive performance with fewer layers compared to other methods.
- Handling Higher-Order QAM: Both proposed methods are designed to handle higher-order QAM signals, a significant advantage over some prior ADMM-based deep learning detectors limited to BPSK/QPSK.
Key Results
- Performance Improvement: Both ADMM-PSNet and ADMM-HNet show improved symbol-error-rate (SER) performance compared to existing methods, especially under higher SNR regimes.
- ADMM-HNet Outperforms: The ADMM-HNet, in particular, outperforms other detectors, including ADMM-PSNet, in SER performance, even when using fewer layers (e.g., 20 layers for ADMM-HNet vs. 30 for others).
- Significant Complexity Reduction: The ADMM-HNet achieves substantially lower computational complexity. Table 1 shows that the average run time for ADMM-HNet is significantly less than PS-ADMM, DetNet, and ADMIN. For example, for 16-QAM, 128x64 MIMO, ADMM-HNet takes 0.0029 seconds, while PS-ADMM takes 0.0093 seconds. This is attributed to avoiding matrix inversion and multiplication in each layer.
Limitations and Future Work
- ADMM-HNet Parameter Learning: The paper notes that explicitly learning the penalty parameters within the ADMM-HNet's hidden layers is not straightforward. Therefore, they pre-train these parameters using the ADMM-PSNet model. Future work could explore methods for end-to-end learning of all parameters within the ADMM-HNet.
- Generalization to More Complex Channels: While the paper focuses on flat fading Rayleigh channels, future work could investigate the performance of these detectors in more complex channel environments.
- Hardware Implementation: Exploring the practical hardware implementation of these low-complexity DNN detectors would be a valuable next step.
Summary updated