PaperlayerView exact revision on arXiv

Source-linked AI summary

Model-Driven Deep Learning for Physical Layer Communications

Hengtao He, Shi Jin, Chao-Kai Wen, Feifei Gao, Geoffrey Ye Li, Zongben Xu

arXiv:1809.06059v2cs.ITcs.LG

TL;DR

Physical-layer DL often treats communication systems as black boxes and demands large datasets and substantial training resources. This article surveys model-driven DL, which combines physical-layer knowledge with learnable networks across transmission, receiver design, and CSI recovery. The surveyed work reports competitive performance with fewer trainable variables than black-box architectures and identifies settings where data-driven DL remains necessary.

  • Problem

    Existing data-driven DL approaches treat communication systems as black boxes and require large datasets, while their structures lack a unified theoretical foundation.

  • Method

    The article provides a comprehensive overview of model-driven DL applications in physical-layer transmission schemes, receiver design, and CSI estimation and feedback.

  • Results

    Model-driven DL shows competitive performance with fewer trainable variables than black-box architectures.

  • Takeaways & Limitations

    The survey identifies model-driven DL as promising for intelligent communications while highlighting open issues for further research.

Abstract

from arXiv · show

Intelligent communication is gradually considered as the mainstream direction in future wireless communications. As a major branch of machine learning, deep learning (DL) has been applied in physical layer communications and has demonstrated an impressive performance improvement in recent years. However, most of the existing works related to DL focus on data-driven approaches, which consider the communication system as a black box and train it by using a huge volume of data. Training a network requires sufficient computing resources and extensive time, both of which are rarely found in communication devices. By contrast, model-driven DL approaches combine communication domain knowledge with DL to reduce the demand for computing resources and training time. This article reviews the recent advancements in the application of model-driven DL approaches in physical layer communications, including transmission scheme, receiver design, and channel information recovery. Several open issues for further research are also highlighted after presenting the comprehensive survey.

I. INTRODUCTION

Future wireless networks must handle large data volumes, adapt to complex environments, and meet demanding speed and accuracy requirements. The article therefore examines model-driven DL as a domain-informed alternative to data-driven DL for physical-layer communications.

  • 5G scenarios require wireless systems to process large data volumes, adapt dynamically to complex environments, and provide high speed and accurate processing.
  • Data-driven DL treats communication systems as black boxes and requires large datasets and long training times, resources often scarce in wireless communications.
  • Model-driven DL constructs network topologies from physical mechanisms and domain knowledge, requiring less training data and shorter training time.
  • The article reviews model-driven DL applications in physical-layer transmission schemes and receiver design while highlighting open research issues.

II. WHY MODEL-DRIVEN DL ?

Model-driven DL addresses data-driven DL’s dependence on large labeled datasets and opaque network structures by embedding physical-layer knowledge into explainable, learnable architectures. Its benefits include lower training-data demand, reduced overfitting risk, and rapid implementation, while model accuracy still affects performance.

  • Data-driven DL can require large labeled datasets, while limited theoretical understanding makes neural-network structures unexplainable and unpredictable.
  • Model-driven DL constructs networks from domain knowledge rather than relying heavily on large labeled datasets, making them more explainable and predictable.
  • Model-driven DL combines a rough physical model, an algorithm, and a task-specific network with learnable parameters trained by back-propagation.
  • Model-driven DL can exploit imperfect physical models while DL compensates for inaccuracies in models and predetermined parameters.
  • Model accuracy still influences model-driven DL performance, and inaccurate models can limit its effectiveness.
  • Model-driven DL offers lower training-data demand, reduced overfitting risk, and rapid implementation.

III. TRANSMISSION SCHEMES

Model-driven DL integrates communication structure and expert knowledge into transmission schemes, extending autoencoder-based designs and addressing practical OFDM impairments. Reported applications include RTN, OFDM-autoencoder, and PRNet designs.

  • Autoencoder-based transceiver design jointly optimizes transmitter and receiver DNNs over an AWGN channel but treats the communication system as a black box.
  • RTN: The RTN incorporates expert knowledge through parameter estimation, parametric transformation, and learned discrimination, supporting adaptation to hardware impairments.
  • RTN: The autoencoder with RTN consistently outperforms and converges faster than the original autoencoder.
  • OFDM-autoencoder: The OFDM-autoencoder retains conventional OFDM structure and inherits robustness to sampling-synchronization errors while simplifying equalization over multipath channels.
  • OFDM-autoencoder: The OFDM-autoencoder combined with RTN can handle carrier-frequency offset directly in the time domain with slight performance degradation.
  • PRNet: PRNet adaptively learns constellation mapping and demapping while retaining traditional OFDM blocks such as FFT, and it outperforms conventional schemes in PAPR.

IV. RECEIVER DESIGN

Model-driven DL receiver designs embed established communication algorithms into learnable architectures for OFDM channel estimation and detection. Reported results include improved accuracy, faster convergence, fewer parameters, and robustness to SNR mismatch.

  • Traditional OFDM receivers use pilots and received symbols for channel estimation and signal detection in a block-by-block process.
  • A fully connected DNN receiver replaces traditional OFDM blocks including FFT, channel estimation, signal detection, and QAM demodulation, but treats the receiver as a black box.
  • OFDM receiver: Model-driven DL integrates into OFDM receiver design to address the large-parameter and high-complexity problems of FC-DNN receivers.
  • OFDM receiver: ComNet combines expert knowledge with two subnets for channel estimation and signal detection in a block-by-block architecture.
  • OFDM receiver: The initialization of each subnet is an accessible communication algorithm, linking the learned receiver to conventional signal-processing procedures.
  • OFDM receiver: The model-driven approach provides more accurate CE than LMMSE under linear and nonlinear cases and outperforms traditional MMSE and FC-DNN methods.
  • OFDM receiver: The approach converges relatively faster, requires fewer parameters than the FC-DNN OFDM receiver, and is robust to SNR mismatch.

B. Model-Driven DL for MIMO Detection

Model-driven deep learning applies iterative detection principles to MIMO detection, addressing distributional assumptions and fixed architectures. DetNet and OAMP-Net reduce reliance on conventional iterative methods while maintaining or improving detection performance.

  • Limitations of iterative detection: Conventional iterative detectors assume particular channel distributions and may perform suboptimally in complicated environments.Examples include Kronecker-correlated and Saleh-Valenzuela channels; their architectures and parameters are also fixed.
  • DetNet: DetNet unfolds projected gradient descent for maximum-likelihood detection and recovers transmitted signals from received signals and perfect CSI.The network incorporates the detection procedure into a neural architecture.
  • DetNet: DetNet outperforms iterative algorithms and has SER performance comparable with the K-best sphere decoder.Its robustness is demonstrated for deterministic ill-conditioned fixed channels and varying channels with known distributions.
  • DetNet: DetNet accuracy is equal to AMP and greater than semidefinite relaxation accuracy.
  • OAMP-Net: OAMP-Net unfolds the iterative OAMP detector using few trainable parameters whose number depends on layers rather than antenna count.This parameterization is intended to support large-scale problems such as massive MIMO detection.
  • OAMP-Net: OAMP-Net improves training stability and convergence speed, handles time-varying channels with one training, and improves OAMP performance in Rayleigh and correlated MIMO channels.

V. CSI ESTIMATION AND FEEDBACK

Model-driven deep learning supports CSI estimation in massive MIMO, where accurate channel models are difficult to obtain and limited RF chains make recovery challenging. LDAMP combines AMP-based recovery with learned denoising and can be analytically characterized through state evolution.

  • Motivation: Accurate CSI is important for realizing massive MIMO advantages, but CSI estimation and feedback are challenging when specific channel models are unavailable.A data-driven network may otherwise require large amounts of data to learn massive MIMO channel features.
  • Beamspace mmWave CSI estimation: Beamspace mmWave massive MIMO reduces RF-chain requirements, but channel estimation remains extremely challenging with large antenna arrays and few RF chains.Dedicated RF chains for every antenna can require expensive hardware or high power consumption.
  • LDAMP: LDAMP treats the channel matrix as a 2D natural image and incorporates a denoising CNN into the AMP algorithm.The network is based on compressed signal recovery and iterative signal recovery, making it model-driven.
  • LDAMP: LDAMP learns channel structure from training data and demonstrates excellent performance even with a small number of receiver RF chains.
  • LDAMP: LDAMP performance can be accurately predicted quickly using an analytical framework derived from AMP state evolution.This avoids time-consuming Monte Carlo simulation for performance prediction.

B. Data-Aided Model-Driven DL for Channel Feedback

Model-driven and data-aided deep learning reduce CSI feedback overhead in massive MIMO by combining compressed-sensing structure, convolutional reconstruction, and temporal correlation. CsiNet-LSTM recovers CSI at low compression ratios with only a slight resolution decrease.

  • Motivation: Massive MIMO produces excessive CSI feedback overhead, while conventional model-driven schemes face channel-model inaccuracies and antenna-dependent overhead.
  • CsiNet: CsiNet mimics compressed-sensing architecture with a CNN autoencoder containing an encoder for compressive sensing and a decoder for reconstruction.RefineNet units use residual connections to reduce gradient-vanishing problems.
  • CsiNet-LSTM: CsiNet-LSTM exploits temporal correlation in time-varying massive MIMO channels to reduce feedback overhead over time.CSI matrices within the coherence time are grouped into sequences for recurrent processing.
  • CsiNet-LSTM: CsiNet-LSTM compresses the first CSI matrix at a higher compression ratio and the remaining similar matrices at a low compression ratio.The first matrix is reconstructed with high resolution, while later matrices use their limited additional information to reduce overhead.
  • CsiNet-LSTM: CsiNet-LSTM is robust to compression-ratio reduction and recovers CSI under a low compression ratio with a slight decrease in resolution.

of training time. Although previous studies have obtained promising results, the model-driven

The survey identifies theoretical analysis, online training, model accuracy, and specialized architectures as open issues for model-driven deep learning. Analytical prediction and online adaptation are presented as promising responses to these challenges.

  • Theoretical analysis: DL communication algorithms show competitive performance but lack a unified theoretical foundation and framework, limiting widespread usage.
  • Theoretical analysis: Model-driven networks enable theoretical analysis because their underlying models can produce rigorous results within performance limits.Removing the DL elements reduces such networks to model-driven approaches whose performance can be analyzed.
  • Theoretical analysis: State evolution equations can predict channel-estimation network performance across layers, replacing time-consuming Monte Carlo simulation and supporting system optimization.
  • Online training: Offline-trained networks may deteriorate when deployed under different channel conditions, making adaptation to environmental alterations necessary.An indoor-trained network can perform worse than traditional LMMSE channel estimation when used outdoors.
  • Online training: Online training updates network weights during operation, and labeled data can be recovered from error-correcting codes.The described design uses sub-networks for indoor and outdoor channels plus trainable parameters updated online.
  • Online training: Rapid online training produces a large gain for the outdoor channel by adapting the contributions of indoor- and outdoor-trained networks to channel alterations.

C. Effect of Model Accuracy

Model-driven DL performance depends on model accuracy: accurate models can reach optimal solutions, whereas particularly inaccurate models can cause serious deterioration. Specialized architectures remain an open need for physical layer communication modules.

  • Effect of Model Accuracy: Model accuracy influences model-driven DL performance, while DL can compensate for inaccuracies using side information from data.
  • Effect of Model Accuracy: Accurate models can enable the optimal solution, whereas particularly inaccurate or totally wrong models seriously deteriorate performance.
  • Effect of Model Accuracy: The effects of model accuracy and DL's ability to compensate for inaccuracies should be investigated in future work.
  • Specialized Architectures: CNN and RNN architectures were designed for image and speech signals, respectively, rather than specifically for physical layer communications.
  • Specialized Architectures: LDAMP can be interpreted as a specialized model-driven DL architecture for channel estimation.
  • Specialized Architectures: Specialized model-driven DL architectures for physical layer communication modules remain desirable because existing networks mainly use plain DL architectures or iterative algorithms.

VII. CONCLUSION

The article surveys model-driven DL for physical layer communication challenges and highlights achievements across transmission, receiver design, and CSI estimation and feedback. It reports competitive performance with fewer trainable variables than black-box architectures, indicating potential for intelligent communications.

  • VII. CONCLUSION: The article provides a comprehensive overview of model-driven DL for challenges in physical layer communications.
  • VII. CONCLUSION: The survey highlights achievements in transmission schemes, receiver design, and CSI estimation and feedback.
  • VII. CONCLUSION: Model-driven DL delivers competitive performance with fewer trainable variables than black-box architectures and shows potential in intelligent communications.

Fink Overview Paper Award, and Distinguished ECE Faculty Achievement Award from Georgia

The surveyed model-driven DL designs combine communication models, expert knowledge, and neural networks across autoencoder transmission, receiver design, and channel-information recovery. Examples include ComNet, CsiNet, and OFDM receivers with offline or online training.

  • Autoencoder-based communication represents the transmitter and receiver with fully connected networks for transmission over an AWGN channel.
  • ComNet uses two subnets for channel estimation and signal detection, incorporating traditional communication solutions as initializations refined by DL networks.
  • The reviewed approaches report relatively faster convergence, more accurate channel estimation than linear minimum methods, and a robust candidate for binary-symbol recovery.
  • CsiNet uses an encoder-decoder architecture with convolutional layers and RefineNet modules to reconstruct compressed channel information.
  • OFDM receiver designs compare offline training with online updates, where additional trainable parameters adapt the receiver and online training positively affects BER performance.
Loading 1809.06059v2…