PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial intelligence enabled radio propagation for communications-Part I: Channel characterization and antenna-channel optimization

Chen Huang, Ruisi He, Bo Ai, Andreas F. Molisch, Buon Kiong Lau, Katsuyuki Haneda, Bo Liu, Cheng-Xiang Wang, Mi Yang, Claude Oestges, Zhangdui Zhong

arXiv:2111.12227v1eess.SP

TL;DR

Propagation-channel study faces data and complexity challenges, while antenna-channel optimization must account for complex channel conditions. This paper surveys ML approaches to channel characterization and antenna-channel optimization, and reports application-specific comparisons including near-exhaustive-search performance with reduced feedback and beam-selection accuracy approaching an optimum allocation.

Abstract

from arXiv · show

To provide higher data rates, as well as better coverage, cost efficiency, security, adaptability, and scalability, the 5G and beyond 5G networks are developed with various artificial intelligence techniques. In this two-part paper, we investigate the application of artificial intelligence (AI) and in particular machine learning (ML) to the study of wireless propagation channels. It firstly provides a comprehensive overview of ML for channel characterization and ML-based antenna-channel optimization in this first part, and then it gives a state-of-the-art literature review of channel scenario identification and channel modeling in Part II. Fundamental results and key concepts of ML for communication networks are presented, and widely used ML methods for channel data processing, propagation channel estimation, and characterization are analyzed and compared. A discussion of challenges and future research directions for ML-enabled next generation networks of the topics covered in this part rounds off the paper.

I. INTRODUCTION

The paper introduces AI/ML as tools for addressing the scale and complexity of propagation-channel analysis and frames this part around channel characterization and antenna-channel optimization. It outlines the ML foundations and channel-processing tasks covered, with scenario identification and channel modeling reserved for Part II.

I. INTRODUCTION

  • Massive device, antenna, and bandwidth scenarios challenge measurement campaigns and data analysis, while classical techniques may incur overhead or miss relationships in the resulting data.The paper identifies channel study as fundamental to wireless-system design, optimization, and evaluation.Source: §I · p. 1

A. Development of AI and ML

  • ML includes supervised, unsupervised, and reinforcement learning; deep learning is a subset that relies on neural networks and requires more data and computing resources than traditional ML.The paper frames regression, classification, and clustering as mathematical problem types to which AI/ML techniques can be applied.Source: A. Development of AI and ML · p. 2

B. Application of ML to Propagation Channels

  • Channel feature extraction supports analysis and modeling, including parameter estimation, MPC clustering, and tracking for time-varying channels.The paper also covers antenna-channel optimization, while scenario identification and channel modeling are assigned to Part II.Source: B. Application of ML to Propagation… · pp. 2–4
  • ML can model complex antenna and channel parameters, and offline training can support low-complexity online antenna optimization.The review focuses on far-field antenna properties and their interaction with the propagation channel, rather than speeding up classical antenna-design simulations.Source: B. Application of ML to Propagation… · p. 3

A. Channel Parameter Estimation

Channel parameter estimation ranges from conventional high-resolution and Fourier-based methods to ML regression models. The paper highlights verification in fast-changing, non-stationary channels as a challenge for ML-based estimators.

A. Channel Parameter Estimation

  • High-resolution estimators overcome classical Fourier or beamforming resolution limits for multidimensional MPC parameters, but iterative computation makes them complex.Fourier-transform filters are typically less complex but provide a spectrum; SAGE decomposes joint estimation into one-dimensional problems, while RiMAX also models dense multipath components.Source: A. Channel Parameter Estimation · p. 4
  • ML estimation learns a regression model from input data to output channel parameters, with applications including path loss, angle information, and channel impulse response.The cited examples include SVM-based AoA estimation from actual array measurements and an Extended Kalman Filter for tracking multipaths over time.Source: A. Channel Parameter Estimation · p. 4
  • Emerging higher-frequency systems, diverse scenarios, and more abundant devices create further challenges for parameter-extraction algorithms.The paper introduces these challenges as research concerns for ML-based channel parameter extraction.Source: A. Channel Parameter Estimation · p. 4

1) Performance in a non-stationary environment.:

  • Fast-moving vehicles, trains, and satellites produce strongly non-stationary channels, so ML-based parameter estimators' real-time ability and accuracy require verification in those conditions.In such channels, both realizations and channel statistics can change within a short time.Source: 1) Performance in a non-stationary… · p. 4

2) Selection of input and output parameters.:

The paper emphasizes choosing ML inputs and outputs to fit diverse estimation needs and obtaining adequate measurement data. It also reviews MPC clustering approaches, including shape-based identification and its accuracy trade-off.

2) Selection of input and output parameters.:

  • Input selection matters because ML parameter extractors map input data to target outputs, and system requirements call for different estimates, including multipath, scenario, and angle information.Existing input features include environmental information, signal strength, and channel impulse response; the paper calls for optimizing parameter selection for different requirements.Source: 2) Selection of input and output… · pp. 4–5

3) Acquisition of big data of measurement data.:

  • ML-based parameter extraction needs extensive real channel measurements, which are expensive and time-consuming to collect, especially for mmWave/THz and high-mobility scenarios.The paper suggests denoising and combining measurement with simulation as data-utilization approaches; simulation data still require measurement-based evaluation and verification.Source: 3) Acquisition of big data of… · p. 5

B. MPCs’ Cluster Identification

  • MPC clustering groups components with similar channel characteristics and can serve as preprocessing for ML-based channel modeling.The paper categorizes existing cluster-identification algorithms by clustering methodology.Source: B. MPCs’ Cluster Identification · p. 5

1) Shape-Based Cluster Identification:

  • Shape-based identification fits the MPC envelope in the power-delay profile to a distribution; approaches include HMM/Viterbi fitting, straight-line fitting on a semi-log scale, and kurtosis with Region Competition.The method needs little prior knowledge and uses relatively low-dimensional PDP information, but its accuracy is limited by the lack of angular information.Source: 1) Shape-Based Cluster Identification · p. 5

2) Optimization-Based Cluster Identification:

Optimization-based methods account for channel impulse response (CIR) behavior, while distance-based methods group multipath components (MPCs) by parameter similarity. Their trade-offs include angular information, prior-parameter needs, and sensitivity to initialization.

2) Optimization-Based Cluster Identification:

  • Optimization-based identification uses sparsity optimization to recover noise-reduced CIRs or angle-delay GMMs refined with EM and Variational Bayesian algorithms.The GMM approach uses EM for initial parameters and Variational Bayesian optimization to refine the number of Gaussian distributions.Source: 2) Optimization-Based Cluster… · p. 5
  • Both shape- and optimization-based methods identify PDP clusters with little prior knowledge, while optimization can incorporate angular characteristics to improve accuracy.Shape-based methods can also extend into the angular domain, but require assuming a PAS shape that is difficult to specify generally.Source: 2) Optimization-Based Cluster… · pp. 5–6

3) Distance-Based Cluster Identification:

  • MCD normalizes delay and angle before distance calculation, outperforms SED for measuring MPC similarity, and supports hierarchical, power-weighted KPM, and fuzzy-C-means clustering.KPM iteratively groups MPCs using power-weighted MCD; a later variant initializes centroids with the highest-power MPC.Source: 3) Distance-Based Cluster Identification · p. 6
  • Distance-based methods can identify clusters across power, delay, and angle with high accuracy and acceptable complexity, but often require prior cluster counts or centroid positions and are initialization-sensitive.The best initial settings depend on the processed data, with no general “golden rule.”Source: 3) Distance-Based Cluster Identification · p. 6

4) Density-Based Cluster Identification:

Density-based methods identify MPC clusters from local density, while computer-vision methods interpret visual patterns in channel data. Their approaches differ in whether clustering relies on density or image-processing operations.

4) Density-Based Cluster Identification:

  • KPD uses a modeled kernel density of MPCs, weighted by MPC power, to identify clusters; DBSCAN is another density-based approach.The density-based methods can identify MPC clusters across multiple domains without prior cluster counts or centroid positions.Source: 4) Density-Based Cluster Identification · pp. 6–7

5) Computer-Vision-Based Cluster Identification:

  • Computer-vision methods use Hough trajectories, PAS-based segmentation, or image denoising and segmentation to identify MPC clusters from channel representations.PASCT separates potential cluster groups from background noise, then divides clusters using density-peak search.Source: 5) Computer-Vision-Based Cluster… · p. 7
  • Vision-based identification follows an intuitive approach that can produce results conforming to human observation and draws on developments in computer vision.PASCT obtains the power angular spectrum using the Bartlett beamformer without high-resolution parameter extraction.Source: 5) Computer-Vision-Based Cluster… · p. 7

6) Evolution-Based Cluster Identification:

Evolution-based identification uses MPC behavior over time to distinguish dynamic clusters, including clusters that overlap in individual snapshots. Time-varying channel analysis also requires tracking MPCs or cluster trajectories across snapshots.

6) Evolution-Based Cluster Identification:

  • Clusters with distinct evolving patterns can be separated even when they overlap in a snapshot, and evolution-based identification can improve accuracy over independent snapshot clustering.The approach relies on MPCs in a cluster sharing similar evolving behavior; capturing that behavior is a critical challenge.Source: 6) Evolution-Based Cluster… · pp. 7–8
  • Methods for dynamic identification include moving-probability clustering with Kuhn-Munkres, data-stream clustering, and evolving similarity based on trajectory shape and MPC distance.The evolving-similarity approach applies KPD to identify dynamic clusters.Source: 6) Evolution-Based Cluster… · p. 8
  • Because measurement data lack cluster-identification ground truth, evaluation remains challenging; synthetic data and clustering indices are alternatives, with XB and GD generally performing well in comparisons.The cited comparison assessed XB, GD, and other measures for evaluating MPC cluster identification.Source: 6) Evolution-Based Cluster… · p. 8

C. Multipath Component/Cluster Tracking

  • Time-varying channel characterization requires tracking MPCs or clusters across consecutive snapshots, using methods such as Kalman filters, Hungarian assignment, or Kuhn-Munkres.Tracking approaches either associate MPCs first and then identify cluster trajectories, or identify clusters per snapshot and then track their centroids.Source: C. Multipath Component/Cluster Tracking · p. 8

1) Distance-Based Tracking:

Distance-based tracking associates MPCs across snapshots by proximity, while matching-based methods formulate tracking as an assignment problem. The approaches differ in global optimality and their handling of cluster and MPC changes.

1) Distance-Based Tracking:

  • Distance-based tracking links MPC pairs with the smallest distance between snapshots, using MCD, but these links provide a locally rather than globally optimal solution.Some methods instead initialize clustering by associating current-snapshot MPCs with prior-snapshot cluster centroids.Source: 1) Distance-Based Tracking · p. 8
  • Reflection-point tracking uses AoA and AoD to locate the reflecting object and follows the reflection point along its surface, but is mainly applicable to single-bounce reflections.The specific propagation process of a multi-bounce reflection is difficult to acquire.Source: 1) Distance-Based Tracking · p. 8

2) Matching-Based Tracking:

  • Matching-based tracking formulates MPC association as a weighted bipartite matching problem solved with Kuhn-Munkres; Hungarian assignment is also used, with all links assigned equal weight.The method is motivated by the difficulty of obtaining a globally optimal solution to the nonconvex tracking problem.Source: 2) Matching-Based Tracking · p. 8
  • Matching-based tracking has higher reported accuracy and robustness than distance-based tracking, but does not model cluster splitting or merging or MPC births and deaths.The comparison is reported in.Source: 2) Matching-Based Tracking · p. 9

3) Threshold-Based Tracking:

The section reviews threshold- and prediction-based tracking for evolving multipath clusters, then surveys ML antenna-selection methods that balance link performance, feedback, and computational complexity.

3) Threshold-Based Tracking:

  • Threshold tracking links MPCs or clusters using a preset threshold, enabling one-to-many or many-to-one associations for splitting and merging at low computational complexity.Threshold settings can substantially affect results and may need to vary across measurement data.Source: 3) Threshold-Based Tracking · p. 9

4) Prediction-Based Tracking:

  • Prediction-based tracking uses long-term evolution to estimate future MPC or cluster positions, which can be paired with matching to associate predictions with the next snapshot.Kalman filters track one target; particle filters can track multiple targets.Source: 4) Prediction-Based Tracking · p. 9
  • Evolution-based cluster identification jointly considers clustering and tracking and is reported as more robust and accurate on time-varying channels than either sequential ordering.Computational complexity remains a key implementation consideration for clustering algorithms.Source: 4) Prediction-Based Tracking · p. 9

A. ML-Based Transmit Antenna Selection

  • For two-stream TAS, KNN and SVM do not match max-min eigenvalue selection, but offer polynomial rather than exhaustive-search complexity; a new HHH-based feature improves SVM BER.For single-stream transmission, the compared methods have equivalent BER performance.Source: A. ML-Based Transmit Antenna Selection · pp. 9–10
  • In TAS, Naive Bayes and SVM nearly match exhaustive search for secrecy rate with half the feedback; CNN outperforms SVM and KNN on capacity loss and variance at higher complexity.The relative performance of SVM and Naive Bayes depends slightly on eavesdropper-channel knowledge.Source: A. ML-Based Transmit Antenna Selection · p. 10
  • ANN and GBDT can approach near-optimal TAS performance, with GBDT offering a better BER-performance-efficiency tradeoff; diagonal HHH features are twice as important as off-diagonal features.Source: A. ML-Based Transmit Antenna Selection · p. 10
  • The multi-user MCTS antenna-selection method achieves near-optimal BER with 50% of branch-and-bound complexity and a 1-dB BER gain over greedy selection.It uses channel capacity as its key performance indicator.Source: A. ML-Based Transmit Antenna Selection · p. 10

B. ML-Based Beam Selection and Antenna Tilt Optimization

ML-based beam selection uses channel data to reduce beam-sweeping overhead, while reinforcement learning has been applied to optimize antenna tilt for coverage, interference, and network performance.

B. ML-Based Beam Selection and Antenna Tilt Optimization

  • Sub-6 GHz channel information reduces exhaustive beam-sweeping overhead during millimeter-wave initial access by as much as 80%.The reported DNN approach typically uses the channel power-delay profile as a user-position fingerprint.Source: B. ML-Based Beam Selection and Antenna… · p. 10
  • For a 5G V2I setting, deep reinforcement learning beam-selection accuracy approaches dynamic programming’s optimum time-slot and beam allocation.The reviewed work focuses on generating realistic channel-evolution data by combining ray-tracing and traffic simulators.Source: B. ML-Based Beam Selection and Antenna… · p. 10
  • Distributed reinforcement learning for antenna tilt improves throughput fairness by 45-56% and energy efficiency by 21-47% over fixed strategies; urban sum rate rises 30% in another study.The latter result is reported for optimized tilt in an urban scenario.Source: B. ML-Based Beam Selection and Antenna… · p. 10

C. Data-Driven Design of Antenna Patterns

The section contrasts channel-informed antenna-pattern design with simpler antenna-property adaptation, and identifies unresolved challenges in AI-based channel parameter estimation.

C. Data-Driven Design of Antenna Patterns

IV. IMPORTANT ISSUES AND CHALLENGES

  • AI-based channel analysis challenges are framed as guidelines for future ML-based channel analysis and modeling research.Source: §IV · p. 11

A. DL-Based Channel Parameter Estimation

  • ANN-based deep learning could use spectrum results as prior information to estimate accurate channel parameters with fewer iterations, but training complexity, accuracy, and overfitting remain challenges.The target is to improve on high-resolution estimation accuracy while balancing training performance and overfitting.Source: A. DL-Based Channel Parameter Estimation · p. 11

B. DL-Based Cluster Identification

The paper identifies open challenges in cluster recognition, AI-based antenna design, and mmWave/THz applications, and presents its review of AI-enabled channel processing and antenna-channel optimization.

B. DL-Based Cluster Identification

  • Deep-learning cluster recognition has not improved accuracy as expected over unsupervised clustering, while existing methods require data-specific presets and manual adjustment for non-stationary channels.Joint tracking and clustering of time-varying MPCs also requires further investigation.Source: B. DL-Based Cluster Identification · p. 11

C. AI-Based Antenna Design for Coverage Optimization or Energy Saving

  • AI-based antenna-pattern synthesis remains underexplored, and translating a desired pattern into a reconfigurable antenna structure poses unresolved design and computational challenges.The proposed direction includes AI modeling to predict patterns across configurations, with impedance matching also requiring investigation.Source: C. AI-Based Antenna Design for… · pp. 11–12

D. AI-Based Applications for mmWave/THz Band Communications

  • For mmWave and THz channel applications, costly, time-consuming measurements constrain training-data acquisition, while newer channel properties remain underused in existing AI applications.The paper calls for investigating how sparsity, channel hardening, and non-stationarity could improve application efficiency, accuracy, and robustness.Source: D. AI-Based Applications for… · p. 12

V. CONCLUSION

  • The paper reviews AI-enabled channel data processing, parameter estimation, characterization, and antenna-channel optimization, and identifies future challenges for channel-data methods.Scenario identification and channel modeling are covered in Part II.Source: §V · p. 12

Related papers

Loading 2111.12227v1…