Source-linked AI summary
Artificial intelligence enabled radio propagation for communications-Part II: Scenario identification and channel modeling
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
TL;DR
The paper addresses how AI and ML can support wireless-propagation scenario identification and channel modeling or prediction. It reviews and compares existing methods, finding that feature selection and model configuration are central, with accuracies above 98% reported in some typical scenarios. The review also identifies challenges involving training-data collection, synthetic-data fidelity, and dynamic channel prediction.
Problem
Wireless channels depend on physical environments, so reliable scenario identification and channel modeling are needed across localization, communication design, and unmeasured propagation conditions.
Method
The paper reviews state-of-the-art ML methods for propagation scenario identification, channel modeling, and prediction, analyzing their characteristics, comparisons, and future challenges.
Results
Identification accuracies above 98% are reported in some typical scenarios, while feature selection and model configuration are emphasized as important performance factors.
Takeaways & Limitations
ML-based propagation-channel processing offers typical solutions for scenario identification and channel modeling or prediction, but application design must match features and data structure to the task.
Abstract
from arXiv · showhide
This two-part paper investigates the application of artificial intelligence (AI) and in particular machine learning (ML) to the study of wireless propagation channels. In Part I, we introduced AI and ML as well as provided a comprehensive survey on ML enabled channel characterization and antenna-channel optimization, and in this part (Part II) we review state-of-the-art literature on scenario identification and channel modeling here. In particular, the key ideas of ML for scenario identification and channel modeling/prediction are presented, and the widely used ML methods for propagation scenario identification and channel modeling and prediction are analyzed and compared. Based on the state-of-art, the future challenges of AI/ML-based channel data processing techniques are given as well.
I. INTRODUCTION
The paper motivates ML for wireless propagation by linking channel behavior to physical environments and reviews AI-enabled scenario identification, channel modeling, and prediction.
- Propagation channels are fundamental to wireless-system design and performance evaluation, while ML-based AI addresses bottlenecks in conventional channel and antenna methods.
- Different physical environments produce different channel characteristics, making scenario identification useful for localization, channel modeling, transportation, and network applications.
- Channel modeling relates multipath parameters such as delay, received power, and angle to physical environments, enabling prediction at unmeasured locations, frequencies, or directions.
- The review covers ML-based communication scenario identification, channel modeling and prediction, and associated challenges and future research directions.
A. LoS/NLoS Identification
LoS/NLoS identification has progressed from statistical and threshold-based methods toward AI-based classification, especially for dynamic channels where fixed boundaries are unreliable.
- LoS/NLoS identification supports localization and channel modeling, but dynamic, time-varying channels make reliable identification difficult.
- AI-based automatic identification is divided into supervised methods requiring classified training data and unsupervised methods using unclassified characteristics.
- Statistical approaches use received signal strength, ToA, RMS-delay spread, propagation distance, or kurtosis to distinguish LoS from NLoS conditions.
- Fixed thresholds work well in static scenarios but cannot perfectly separate overlapping LoS and NLoS channel data in dynamic environments.
- Hypothesis-testing methods are usually treated as statistical rather than AI methods in some research.
2) Supervised Identification:
Supervised identification learns nonlinear LoS/NLoS boundaries from labeled channel data, with model and feature choices strongly affecting performance.
- 2) Supervised Identification:: Supervised methods address nonlinear channel-data distributions using SVM, RVM, Decision Tree, Random Forest, and ANN classifiers.
- 2) Supervised Identification:: SVM classification projects features into a higher-dimensional space through kernels such as Gaussian radial basis or Polynomial kernels.
- 2) Supervised Identification:: RVM uses Bayesian learning, requiring fewer training samples than SVM but longer training time with the same amount of data.
- 2) Supervised Identification:: In one comparison, KNN achieved the best accuracy, while SVM and Decision Tree showed similar performance; Random Forest addresses single-tree overfitting.
- 2) Supervised Identification:: Random Forest outperformed competing methods in most cases, ANN performed best on PAS training samples, and time-varying or angular features improved identification.
- 2) Supervised Identification:: Supervised identification improves accuracy significantly but requires preclassified training data, and feature selection remains critical to final performance.
B. Communication Scenario Identification
Communication scenario identification uses channel fingerprints to classify environments for system and modeling applications, with feature selection and model configuration central to performance.
- B. Communication Scenario Identification: Changing environments, especially for dynamic devices, affect channel models, physical-layer algorithms, and network-layer design.
- B. Communication Scenario Identification: GIS-based identification can be unsuitable for indoor or dynamic environments because unexpected situations complicate localization.
- B. Communication Scenario Identification: Model configuration involving channel features, training-data size, and neural-network structure produced identification accuracies above 98% in some typical scenarios.
- B. Communication Scenario Identification: LoS/NLoS and communication scenario identification share a classification formulation based on propagation-channel fingerprints but serve different application environments.
- B. Communication Scenario Identification: Effective ML scenario identification requires both an appropriate feature set and an accurate identification algorithm, with feature selection often having greater impact.
III. CHANNEL MODELING
As cellular generations evolve, channel modeling becomes increasingly multi-dimensional. 5G and beyond add constraints involving nearby mobile terminals and physically large base-station antenna arrays.
- Channel modeling becomes increasingly multi-dimensional as cellular network generations evolve.
- 5G and beyond introduce multiple nearby mobile terminals as an added channel-model constraint.
- 5G and beyond also introduce physically large base-station antenna arrays as an added channel-model constraint.
1) Statistical Modeling:
Statistical channel models use empirical observations and physically justifiable parameter ranges, but they lack implicit spatial consistency and base-station non-stationarity for 5G and beyond. Geometrical models incorporate environmental geometry and multipath interactions, with different model types varying in their reliance on scatterer coordinates.
- Statistical Modeling: Traditional stochastic models describe path loss, shadowing, and small-scale fading using formulas derived from extensive empirical observations.
- Statistical Modeling: Pure statistical models lack implicit modeling of user spatial consistency and base-station non-stationarity for 5G and beyond.They are therefore used mainly for link-level analyses with independent spatial, temporal, and frequency properties.
- Geometrical Modeling: Geometrical modeling relates channel responses to device and scattering-object locations, movement, and wave interactions.It accounts for reflection, scattering, diffraction, and penetration when deriving multipath parameters.
- Geometrical Modeling: Site-specific geometrical modeling mathematically characterizes multipath components arising from radio-wave interactions with physical objects.Ray tracing provides an approximate wave-propagation solution using geometrical optics and ray concepts.
- Geometrical Modeling: 3GPP TR38.901 derives geometry only for the line-of-sight path, whereas COST2100 uses scatterer coordinates fully to derive multipath parameters.The 3GPP model is straightforward to implement statistically, while COST2100 implicitly addresses spatial consistency through visibility regions.
B. ML-Enabled Channel Modeling and Prediction
Because measurement campaigns cannot cover all scenarios, channel prediction is needed for unknown environments. The review distinguishes ML-based characterization/modeling from ML-based prediction according to whether models learn environment-to-channel relationships or channel histories.
- Limited channel measurements make prediction in unknown environments essential.The paper highlights ML prediction methods for their accuracy and flexibility in complicated environments.
- The review separates ML studies into channel characterization/modeling and channel prediction.Characterization/modeling provides channel statistical parameters, while prediction uses acquired channel information to estimate characteristics across locations, frequencies, or times.
1) ML-Based Characterization/Modeling:
ML-based characterization and modeling map environmental or system inputs to channel properties or emulate measured data, while related prediction methods learn channel histories to extrapolate future or unknown channels. The reviewed studies use diverse neural and non-neural methods, with training-data quality often strongly affecting accuracy.
- ML-Based Characterization/Modeling: ML characterization directly captures channel properties such as path loss or delay spread from selected inputs.Neural networks provide a theoretical basis for approximating channel behavior as functions of physical, geometrical, or bias parameters.
- ML-Based Characterization/Modeling: RBF networks can improve 60 GHz path-loss accuracy relative to generalized MLP networks, but require higher computational complexity.Random Forest, KNN, and GA+SVM also outperform generalized neural networks in reported path-loss or signal-strength comparisons.
- ML-Based Characterization/Modeling: ML channel simulation maps system layout inputs, including frequency points and propagation distance, to channel impulse responses.Reported applications include AWGN channels with Doppler shift and generalized wireless channels.
- ML-Based Characterization/Modeling: Insufficient measurement data limits trained-model accuracy, motivating GAN-based generation of synthetic wireless-channel samples.The described framework compares generated fake samples with real channel samples using a channel-data discriminator.
- ML-Based Channel Prediction: Channel prediction focuses on specific conditions, including locations, times, or frequency bands, rather than only generalized channel characteristics.History can span both time and frequency dimensions, and RNNs and Kalman filters are used for future channel-matrix prediction.
- ML-Based Channel Prediction: Coverage prediction uses physical positions and environmental factors as inputs to estimate received power or field strength.Examples include outdoor terrain features, indoor propagation effects, and weather information for satellite channels.
- ML-Based Channel Prediction: Training-data quality can affect modeling and classification accuracy more than the choice between specific network models.FNN and RBF networks show similar modeling performance in one reported comparison.
- ML-Based Channel Prediction: Prediction methods generally learn mappings from channel history to future or unknown channels, with ANN designs and input-output choices remaining major challenges.The review notes offline training and low-complexity online implementations among many ANN-based predictors.
A. Training Data Collection for AI-based Applications
Training data volume and type strongly affect AI-based wireless applications, while synthetic data can supplement measurements when comprehensive channel campaigns are impractical.
- Channel measurements are commonly used for practical deployment but are time-consuming, costly, and difficult to collect across massive scenarios.
- Synthetic training data can be generated from conventional or ML-based models, ray-tracing simulators, or GAN-based data reproduction.
- Model- and simulator-based synthetic channels can supplement measurements for extension scenarios, but trained performance depends on their accuracy.
- GAN-based simulators reproduce inner patterns from measurement data and cannot provide new information for extension scenarios.
B. AI-Based Channel Characterization for Positioning
Future IoT positioning depends on channel information to support moving devices in complex environments affected by people and surrounding objects.
- Channel information contributes to localization and positioning for moving devices such as smartphones and vehicles in complex environments.
- Multiantenna techniques and increased device computing power enable more channel information to be obtained during communications.
C. AI-Based Channel Prediction for Complicated/Combined Scenario
AI-based channel models may improve prediction across changing environments, but generalization, scenario transitions, rapid channel changes, and processing efficiency remain challenges.
- AI-based models can incorporate physical environmental factors to improve accuracy while maintaining a general structure across different scenarios.
- Most existing studies still separate propagation environments into representative scenarios before modeling.
- Rapid wireless-channel changes during transitions between scenarios remain difficult to model and predict.
- High-speed movement and ultra-low-latency B5G and 6G applications challenge data-processing efficiency and conventional resource allocation.
- The paper reviews AI-enabled scenario identification and channel modeling or prediction and identifies future challenges in channel-data processing.