An Overview of Machine Learning Techniques for Radiowave Propagation Modeling

I. INTRODUCTION

• Empirical models are computationally efficient but may miss complex wave effects, while deterministic models can predict a given site reliably at high computational cost.The paper presents ML-driven models as promising for addressing this accuracy-efficiency dichotomy.Source: §I · p. 1
• ML-driven propagation models can be trained on measured or synthetic data and offer flexibility for site-specific or general-purpose prediction.Their nonlinear nature makes them candidates for predicting propagation parameters such as multipath fading.Source: §I · p. 1
• The overview discusses how propagation scenarios affect model decisions and groups the challenges around inputs, the ML model, and outputs.It also presents the key ideas from reviewed papers and includes a representative model-construction case study in the appendix.Source: §I · p. 1

A. ML propagation models: goal

• Predicting propagation parameters such as path loss or received signal strength from measured or synthetic data is a supervised learning problem.Supervised learning maps input-output pairs; training computes an approximation of the unknown mapping from the available pairs.Source: A. ML propagation models: goal · pp. 1–2
• Generalization requires low error on data not used for training, balanced against model complexity to avoid underfitting or overfitting.Training, validation, and test sets support training, model selection, and final evaluation, respectively.Source: A. ML propagation models: goal · p. 2

B. ML propagation models: challenges

• The flowchart identifies the input, ML model, and output as the three building blocks, alongside their main challenges.Source: B. ML propagation models: challenges · p. 2

1) Input:

• The input should represent propagation-relevant information compactly, using measured data or synthetic data generated by site-specific or empirical models.Source: 1) Input · p. 2

2) ML model:

• Model choice determines the function to be learned, and both the model’s hyperparameters and architecture require careful selection.The paper names ANNs, regression decision trees, and SVMs as examples of available models.Source: 2) ML model · p. 2

3) Output:

• Outputs can be scalar, vector-valued, or probabilistic, and predictions should have small out-of-sample error.Source: 3) Output · p. 2
• The review categorizes papers by their input, ML model, and output approaches, and includes a representative model-formulation case study in the appendix.Source: 3) Output · pp. 2–3

III. INPUT SPECIFICATION FOR ML MODELS

• Input features describing geometry and electromagnetic properties are processed before forming training pairs, after which the trained model predicts outputs for new inputs.Processing can include scaling, normalization, or dimensionality reduction; targets may come from measurements or a model.Source: §III · p. 3

A. Modeling environment

• Reviewed propagation-modeling studies span urban, rural, mixed outdoor, and specialized environments such as roads, mines, and subway tunnels.Source: A. Modeling environment · p. 3

1) Environmental and topographical features:

• Urban inputs can describe streets and buildings, forest inputs can represent vegetation and canopy, and terrain or obstructed urban paths can use sampled path profiles.A path profile samples ground elevation along the transmitter-receiver line to capture terrain variation.Source: 1) Environmental and topographical… · p. 3

2) Propagation features:

• Urban models may distinguish LOS from NLOS cases and include diffraction losses, which are associated with improved accuracy in the cited studies.The Paris ANN accounting for diffraction losses was more accurate than one that did not.Source: 2) Propagation features · p. 3

1) Type of input features:

• Inputs may be continuous, discrete, binary, or visual, while dependent samples can be passed as sequences whose length must be tuned to the problem.In one cited example, a sequence of 200 RSS samples performed better than shorter or longer sequences.Source: 1) Type of input features · p. 3

2) Number of input features:

• Adding uncorrelated input features generally improves predictions, and the cited urban ANN using both global and local building information was more accurate than one using local information alone.Source: 2) Number of input features · p. 3

3) Dimensionality reduction:

• Dimensionality reduction can improve computational performance while preserving comparable accuracy; PCA, nonlinear PCA, and grayscale conversion are among the approaches described.PCA reduced 9 environmental input features to 4 principal components in, while and reported up to 30% training-time savings for comparable accuracy.Source: 3) Dimensionality reduction · pp. 3–4

4) Impact of input features:

• Street orientation improved RBF-ANN accuracy more than building-height difference, while environmental features and reduced path-profile representations also affected prediction accuracy.Adding a highly correlated diffraction-loss feature produced no substantial improvement in.Source: 4) Impact of input features · p. 4

1) Size of training dataset:

• Larger training sets can improve predictions when they add representative propagation cases, but the model must be complex enough to avoid underfitting.For classification, balancing class samples improved localization accuracy by about 1% in.Source: 1) Size of training dataset · p. 4

2) Dataset augmentation:

• Augmentation can expand measured datasets through transformed images, simulated samples, source fusion, or generative models.In, rotating each image by 1° increased the training set by a factor of 180; in, combining empirical and measured data improved accuracy further.Source: 2) Dataset augmentation · p. 5

3) Impact of training data:

• Uniformly sampling measurements across airplane rows produced a more accurate model than concentrating them at the front or back, and clustering can help distribute representative samples.The k-means approach in adds the cluster count k as a hyperparameter.Source: 3) Impact of training data · p. 5

4) Type of training data:

• Synthetic training data can reduce reliance on infeasible large measurement collections, but the solver choice affects computational efficiency and environment construction.RT is commonly used for outdoor environments, while VPE is often used for waveguiding enclosures or terrestrial propagation.Source: 4) Type of training data · p. 5
• Solver-generated data can require solver-specific inputs, including ray counts, grid-cell attributes, or reflecting-surface information.Examples include indoor/outdoor cell labels in and maximum obstacle height per cell in.Source: 4) Type of training data · p. 5

IV. ML MODEL

• The reviewed propagation models are grouped as ANN-based, non-ANN-based, or hybrid, and model complexity and hyperparameter choices depend on task and data scale.Larger or more complex models may suit large datasets or many inputs, while simpler models may be faster for small-scale tasks; hyperparameters can be selected with heuristic search [11].Source: §IV · p. 6, Figure 3 · p. 6

A. ANN-based models

• Increasing ANN depth or neuron count has often improved prediction accuracy, but larger networks are more prone to overfitting when training data are limited.The reviewed countermeasures include early stopping and L1 or L2 regularization,.Source: A. ANN-based models · p. 6
• RBF-ANNs are widely used variants whose radial basis functions matter alongside activation-function choice; wavelet neural networks use wavelets as activation functions.Spline functions were used with good results in.Source: A. ANN-based models · p. 6
• CNNs can take grid-like propagation inputs with channels encoding features such as building height or transmitter and receiver distances, and their shared filters and sparse connections improve efficiency over standard MLPs.CNN inputs can also represent nonspatial information, such as RSS values arranged as an image in.Source: A. ANN-based models · pp. 6–7
• RNNs process spatial or temporal input sequences, and LSTMs can capture longer input dependencies than other RNN types.Their ability to learn temporal sequences makes them suitable for modeling time variations in a communication channel.Source: A. ANN-based models · p. 7
• GANs use a generator and discriminator in an adversarial setup, and their generative capacity has been used for dataset augmentation and settings with few labeled data.Conditional GANs can constrain generated data; deep belief networks are another generative-network option.Source: A. ANN-based models · p. 7

B. Non-ANN-based models

• Non-ANN options include SVR, genetic algorithms, k-nearest neighbors, and ensembles such as random forests and AdaBoost, with ensembles requiring a weighting scheme and potentially longer training.In, a Laplacian kernel gave more accurate test-set predictions than polynomial and Gaussian kernels.Source: B. Non-ANN-based models · p. 7

C. Hybrid models

• In error-correction hybrids, an ML model learns the baseline model’s error and adds its predicted correction to the baseline output.The paper describes this as a starting-point approach that is expected to make training more efficient.Source: C. Hybrid models · p. 7, Figure 4 · p. 8
• Reviewed error-correction baselines include empirical models and an Adaline network, while an ANN correction to a four-reflection RT prediction approached seven-reflection accuracy with smaller runtime.The ML model may use baseline inputs or additional parameters not modeled by the baseline.Source: C. Hybrid models · pp. 7–8
• Cascaded models combine networks for distinct tasks, such as CNN feature extraction followed by ANN prediction, or sequential building and room classification with two RNNs.Other examples combine ResNet spatial features with LSTM temporal processing, or use GAN-based models with classifiers and LSTMs.Source: C. Hybrid models · p. 8
• A baseline can also be integrated into an ANN itself, as in, where a node implements Hata’s model and contributes directly to the network output.Source: C. Hybrid models · p. 8

D. Input-constrained models

• Broader input ranges make models more flexible but require greater model complexity, more training data, and more computing power than specialized models.Examples of specialization include frequency-, environment-, route-, and distance-specific models; multi-frequency, multi-environment models are also described.Source: D. Input-constrained models · p. 8

V. OUTPUT

• Regression accuracy is commonly evaluated with MAE, MSE, or correlation factor, while classification accuracy is the probability of correctly predicting the class.The paper also discusses classification, probabilistic, sequential, and other output forms.Source: §V · p. 8

A. Type of output

• Classification can label environments as urban or rural and LOS or NLOS, but overlapping environmental features can make labels ambiguous.In, the optimal k-means cluster count matched the number of classes.Source: A. Type of output · pp. 8–9
• Probabilistic outputs can represent channel-model parameters or path-loss distributions, with the number of distribution bins requiring care,.The ANN in generated valid parameters for the path graph and Saleh-Valenzuela models, while estimated path-loss distributions across urban and suburban environments.Source: A. Type of output · p. 9
• MIMO LSTMs were more accurate than MISO or SISO LSTMs in the reported comparisons,.Source: A. Type of output · p. 9
• RSS and path-loss outputs can support localization, tracking, and higher-level environmental analysis; user location is estimated by coordinates or classified subspace,,,.Other discussed outputs include ground conductivity and permittivity and large-scale fading such as PLE.Source: A. Type of output · p. 9

B. Understanding output errors

• Channel-impulse-response models become more error-prone as multipath mechanisms grow more complex.Source: B. Understanding output errors · p. 9

1) Type of environment:

• Urban and NLOS scenarios generally produce lower accuracy than rural or semi-urban and LOS cases; local receiver information and higher transmitter height improved accuracy in some difficult cases,,,,,,,.More buildings in an urban grid were also associated with increased error.Source: 1) Type of environment · p. 9

2) Distance and frequency:

• Errors are generally higher near the transmitter and often increase with frequency as small-scale fading becomes more pronounced,,,,,.In, errors showed no specific frequency pattern after measurements were averaged over 40 wavelengths.Source: 2) Distance and frequency · p. 9

1) Comparison to empirical models:

• ML models, including hybrid approaches, were reported more accurate than multiple empirical baselines, and a combined Hata-ANN model outperformed each constituent model,,,,.Other comparisons include Walfish-Bertoni, single-slope, CWI, ARMA, and ITU-R P.1546 baselines.Source: 1) Comparison to empirical models · p. 10

2) Comparison to other ML models:

• Across model comparisons, RF often ranked highly, while the best approach varied; in, a voting regressor was slightly more accurate than five individually close learners.Other findings include CNN advantages in selected tasks and mixed rankings among recurrent architectures, –,.Source: 2) Comparison to other ML models · p. 10

1) Composition of the test set:

• A reliable generalizability test uses samples excluded from training, chosen to reflect the model's goal and to test interpolation or extrapolation as appropriate.Examples include testing on different transmitter locations, routes, grids, distances, or frequencies,,,,,.Source: 1) Composition of the test set · p. 10, D. Generalizability and test set · p. 10
• When training uses only synthetic data, model accuracy is bounded by the data-generating solver's accuracy; measurements can also serve as test cases.An ANN trained on ray-tracing tunnel data was tested on both simulated configurations and sections of the London Underground.Source: 1) Composition of the test set · p. 10

2) Test set distribution:

• Cross-route testing deteriorated when route 1's narrower urban range was used to predict route 2, and a model trained on a uniform urban grid had considerably higher error on a non-uniform grid,.Both examples illustrate the importance of training and test data distributions being representative of one another.Source: 2) Test set distribution · pp. 10–11

• Input features should be informative and weakly correlated, while dimensionality reduction can identify dominant features by removing redundant ones.Source: §VI · p. 11
• More training data across propagation scenarios improves accuracy; high-fidelity solver or empirical-model data and augmentation can expand training sets and refine accuracy.Source: §VI · p. 11
• RF was most accurate in several reviewed papers, but accuracy differences among models were generally small and implementation-dependent.Source: §VI · p. 11
• More general models covering broad frequency and environment ranges require more training data, as do models for complex scenarios such as urban propagation.Source: §VI · p. 11
• ML models can be connected into hybrid models for more complex propagation problems.Source: §VI · p. 11
• Evaluation requires a test set covering all propagation mechanisms, with samples drawn from the same distribution as training data.The overview notes that further questions remain open.Source: §VI · p. 11
• An open direction is connecting channel propagation mechanisms to model architecture and required training-data volume; ML may also simplify complex-environment inputs, such as replacing CAD models with image sequences.Source: §VI · p. 11
• Further research could examine GANs for accuracy with low-volume training data and investigate reinforcement learning for electromagnetic wave propagation modeling.Source: §VI · p. 11

APPENDIX - A CASE STUDY OF AN ML PROPAGATION

• The circular-tunnel example uses three inputs—tunnel geometry, axial Tx–Rx distance, and operating frequency—with separate ANNs for 900 MHz and 2.4 GHz.Table I lists the input features and their ranges.Source: APPENDIX - A CASE STUDY OF AN ML… · pp. 11–12, Table I · p. 12
• For each frequency band, 189 tunnel configurations supplied training and validation data, while 24 separately generated configurations formed the test set.Each configuration contributed 501 receiver-position input vectors; 20 configurations were randomly selected for validation.Source: APPENDIX - A CASE STUDY OF AN ML… · p. 12
• Inputs were normalized to zero mean and unit standard deviation, and the ANN used squared-error training with L2 regularization and ADAM optimization.The network used an identity output and tanh activations elsewhere.Source: APPENDIX - A CASE STUDY OF AN ML… · p. 12
• The checked four-layer ANN had slightly lower validation MAE than the three-layer alternative, so the authors selected four layers.For the three-layer model, 300-200-75 nodes seemed optimal; larger architectures added training time without substantial accuracy improvement.Source: Figure 7 · p. 12, Figure 6 · p. 12, APPENDIX - A CASE STUDY OF AN ML… · p. 12
• The evaluation figures plot predicted against actual received signal strength for two selected tunnel configurations at 900 MHz and 2.4 GHz.Source: Figure 8 · p. 13, Figure 9 · p. 13