Source-linked AI summary
A Model-Driven Deep Learning Network for MIMO Detection
Hengtao He, Chao-Kai Wen, Shi Jin, Geoffrey Ye Li
TL;DR
MIMO detection requires methods that avoid the intractable computation of Bayesian MMSE estimation while retaining strong detection performance. The paper unfolds OAMP into OAMP-Net, adding trainable parameters to improve iterative detection. Simulations show significant gains under independent Rayleigh and correlated MIMO channels, with few trainable variables and support for time-varying channels.
Problem
Bayesian MMSE detection is computationally intractable because marginal posterior probabilities require high-dimensional integration.
Method
OAMP-Net unfolds the OAMP iterative detector into a model-driven deep-learning network with learnable scalar parameters in each layer.
Results
OAMP-Net significantly outperforms OAMP under independent Rayleigh and correlated MIMO channels, including more than 2.15 dB improvement under correlated channels.
Takeaways & Limitations
With only 2T trainable variables, OAMP-Net can be trained quickly, independently of antenna dimensions, and handle time-varying channels.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
MIMO detection must balance near-optimal performance against practical complexity, motivating iterative and model-driven deep-learning approaches. The paper introduces OAMP-Net to improve OAMP detection while supporting time-varying channels and short training.
- ML detection achieves optimal performance but requires exhaustive search whose complexity grows exponentially with the number of decision variables.
- AMP-based detection is simple and practical, but performs well under specific i.i.d. sub-Gaussian channel conditions.
- EP-based detection can achieve Bayes-optimal performance for unitarily invariant channels, but has higher complexity than AMP-based detection.
- Model-driven deep learning unfolds iterative algorithms into trainable networks, as demonstrated by DetNet for projected gradient descent.
- OAMP-Net adds adjustable parameters to iterative detection, supports time-varying channels, enables soft decisions, and requires few parameters for short training.
II. PROBLEM DESCRIPTION AND ALGORITHM REVIEW
The paper formulates MIMO detection through Bayesian inference, but direct Bayesian MMSE estimation is computationally intractable because it requires a high-dimensional marginalization.
- The MIMO detector seeks to recover transmitted signals from received observations using Bayesian inference.
- The transmitted symbol vector passes through a Rayleigh fading channel and is corrupted by additive white Gaussian noise.
- Deep learning uses an equivalent real-valued representation by separating the real and imaginary parts of complex signals and matrices.
- The Bayesian MMSE estimate is obtained from the posterior probability P(x|y, H).
- The Bayesian MMSE estimator is not computationally tractable because its marginal posterior involves a high-dimensional integral.
B. OAMP-based detector
The OAMP-based detector iteratively combines linear estimation and posterior-mean denoising to recover signals from received data, using decorrelation properties to support the algorithm.
- OAMP iteratively decouples the joint posterior into scalar posteriors for individual signal components.
- Algorithm 1 takes the received signal, channel matrix, and noise level as input, then outputs recovered signal estimates.
- The linear estimator W_t may use the transpose, pseudoinverse, or linear MMSE matrix of the channel.
- When W_t is decorrelated, residual errors are uncorrelated with the signal and mutually uncorrelated, while B_t = I − W_tH.
- OAMP combines a linear estimator with a posterior-mean estimator to decouple the model into parallel AWGN channels and obtain Bayesian MMSE estimates.
- A deep-learning approach provides appropriate step sizes for updating the iterative estimates and variance.
III. OAMP-NET
The paper presents OAMP-Net as a model-driven network derived from the OAMP algorithm and analyzes its modules and computation complexity.
- OAMP-Net is introduced as a network for MIMO detection whose modules and trainable variables are analyzed alongside its computation complexity.
A. OAMP-Net architecture
OAMP-Net unfolds the OAMP detector into cascade layers and adds learnable scalar variables to adjust its iterative updates. It processes received signals and channel information through linear estimation and MMSE denoising to produce a final signal estimate.
- OAMP-Net consists of T cascade layers, each containing an MMSE denoiser and error-variance estimation components.
- The network takes the received signal y and an initial estimate, then outputs the final estimate of signal x.
- At layer t, the network combines the previous estimated signal with y to perform data detection.
- The linear estimator uses an MMSE matrix, while the denoiser is selected according to the prior distribution of the transmitted signal.
- The learnable variables γ_t and θ_t provide adjustable update step sizes for the MMSE denoiser's mean and variance.
- Error-variance estimators supply the variance estimates required by the MMSE denoiser under the stated residual-error assumptions.
B. Computation complexity
OAMP-Net has per-iteration complexity O(N 3), dominated by a matrix inverse, while its trainable-variable count depends only on the layer count rather than antenna dimensions.
- B. Computation complexity: OAMP-Net requires O(N 3) computation complexity per iteration, similar to OAMP.The matrix inverse in (10) dominates computation in each iteration.
- B. Computation complexity: Each OAMP-Net layer contains two adjustable variables, giving 2T trainable variables overall.
- B. Computation complexity: The trainable-variable count is independent of antenna numbers N and M and is determined only by the number of layers T.The passage identifies this as advantageous for large-scale, high-dimensional MIMO detection.
IV. SIMULATION RESULTS
The simulation-results section presents simulated OAMP-Net results for MIMO detection and uses signal-to-noise ratio to measure noise level.
- IV. SIMULATION RESULTS: This section provides simulated results of OAMP-Net for MIMO detection.
- IV. SIMULATION RESULTS: The system signal-to-noise ratio is defined for the simulations.
- IV. SIMULATION RESULTS: Signal-to-noise ratio is used to measure the noise level.
A. Implementation details
The implementation uses TensorFlow with a 10-layer OAMP-Net trained on randomly generated QPSK-based data for 10,000 epochs.
- A. Implementation details: OAMP-Net is implemented in TensorFlow with the number of layers T set to 10.
- A. Implementation details: Training data consists of randomly generated pairs (x, y), with x generated from QPSK modulation symbols.
- A. Implementation details: The network is trained for 10,000 epochs using training and validation sets of 5,000 and 1,000 samples per epoch.Test data is generated until bit errors exceed 1,000.
B. Rayleigh MIMO channel
Under independent and correlated MIMO channels, OAMP-Net improves BER over OAMP, while channel correlation degrades both methods. The reported gains remain substantial across the tested channel settings.
- Rayleigh MIMO channel: The independent Rayleigh experiment assumes a time-varying channel whose entries are zero-mean complex Gaussian with variance 1/M.The tested system has equal numbers of transmitting and receiving antennas, M=N.
- Rayleigh MIMO channel: OAMP-Net outperforms OAMP across the tested Rayleigh-channel settings by learning parameters that are fixed in the original algorithm.For BER=10^-3, the reported gains are 1.37 dB, 2.97 dB, and 0.82 dB for M=N=4, 8, and 64, respectively.
- Correlated MIMO channel: The correlated channel follows a Kronecker model with receiver and transmitter spatial correlation matrices generated by an exponential model.The correlation coefficient is ρ=0.5, and the Rayleigh fading matrix is A.
- Correlated MIMO channel: Channel correlation degrades the BER performance of both OAMP-Net and OAMP relative to the independent Rayleigh channel.At BER=10^-2, the reported losses are approximately 6.34 dB for OAMP-Net and 6.05 dB for OAMP.
- Correlated MIMO channel: OAMP-Net retains a larger performance gain under correlation, exceeding 2.15 dB compared with the 1.86 dB gain in the independent Rayleigh channel.The correlated-channel experiment uses M=N=4 with correlation coefficient ρ=0.5.
V. CONCLUSION
The paper develops a model-driven deep learning network for MIMO detection that combines an iterative signal-recovery algorithm with trainable parameters. It reports significant improvements over OAMP under independent Rayleigh and correlated MIMO channels, with few parameters and support for time-varying channels.
- V. CONCLUSION: The proposed model-driven network combines an iterative signal-recovery algorithm with deep learning techniques for MIMO detection.Its parameters are optimized to improve detection performance.
- V. CONCLUSION: The network is easy and fast to train because it requires optimizing only a few adjustable parameters.
- V. CONCLUSION: The network can handle time-varying channels.
- V. CONCLUSION: Simulation results show that the network significantly outperforms OAMP under independent Rayleigh and correlated MIMO channels.