Pith. sign in

REVIEW 4 major objections 5 minor 47 references

A single measured path-loss value, inverted through a cos^n beam model, gives sub-degree angle-of-arrival estimates that track the Cramér–Rao lower bound.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

AoA can be estimated from a single path loss value by fitting a cos^n inversion with symbolic regression, but the reported accuracy is only demonstrated on the training data and one fixed angle.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful experimental demonstration undercut by in-sample evaluation; Stage II's near-zero MAE is a trivial constant fit, so the accuracy claims are not yet supported. the 4 major comments →

arxiv 2510.26340 v1 pith:MPN7KVC4 submitted 2025-10-30 eess.SP cs.LG

SABER: Symbolic Regression-based Angle of Arrival and Beam Pattern Estimator

Shih-Kai Chou , Mengran Zhao , Cheng-Nan Hu , Kuang-Chung Chou , Carolina Fortuna , Jernej Hribar This is my paper
classification eess.SP cs.LG
keywords Angle-of-arrival estimationSymbolic regressionBeam patternPath lossReconfigurable intelligent surfaceCramér-Rao lower boundcos^n beam modelMillimeter-wave
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that angle of arrival (AoA) can be recovered from a single scalar measured quantity—the path-loss coefficient—without antenna arrays, snapshot averaging, or a black-box neural network. It proposes SABER, a symbolic-regression framework that fits closed-form inversions of a cos^n(θ) beam-pattern model, and reports sub-0.5° mean absolute error in a free-space anechoic sweep and near-zero error in a fixed-angle RIS-aided indoor link. The interpretable inversions are shown to track the Cramér–Rao lower bound closely, with deviation only as the angle approaches 90°. If correct, this gives a transparent, physically grounded alternative to high-resolution subspace methods and opaque ML estimators for beam alignment and localization.

Core claim

The central claim is that the beam pattern's main lobe is well approximated by U(θ,φ)=cos^n(θ)cos^m(φ), and that this functional form can be inverted so that an unknown incident angle follows directly from the logarithm of the measured S-parameter: θ̂ = arccos(10^{ΔPL/(10n_R)}) + offset. Fitting the directivity exponents with symbolic regression (n_R = 28 for the free-space receiver), SABER recovers AoA with 0.42° MAE, compared to 0.396° for an unconstrained symbolic-regression fit; in the RIS-aided testbed both direct inversion and unconstrained SR converge to the same cosine model and recover the fixed 55° angle with MAE around 6.5×10^-7. Benchmarking the estimator variance against the Cra

What carries the argument

The load-bearing object is the normalized radiation model U(θ,φ)=cos^n(θ)cos^m(φ) (Eq. 5), where n and m set the main-lobe width. Path loss differences are built from this model, and the paper inverts it in two ways: a direct analytical inversion θ̂ = arccos(10^{ΔPL/(10n_R)}) + offset, and a quadratic-polynomial surrogate for the cosine fed through arccos. Symbolic regression is used to fit the exponents and offsets from measured S-parameters. The same model supplies h(θ) for the Cramér–Rao bound, so the near-bound comparison is internal to the chosen beam-pattern form.

Load-bearing premise

The inversion rests on the assumption that the measured main-lobe response is adequately represented by cos^n(θ) with a single scalar exponent across the whole measured angular range and frequency band; Section VI-A concedes this model cannot capture ripples, asymmetries, or null depths, and accuracy degrades near 90°.

What would settle it

Measure a held-out set of angles beyond 80° (or at a frequency not used in fitting) and compute the direct-inversion error against a Cramér–Rao bound computed from the actual measured h(θ). If SABER's RMSE is far above the true bound—or if its MAE jumps by orders of magnitude—while the cos^n model visibly mismatches the pattern, the near-optimality and sub-degree claims do not generalize.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • AoA estimation reduces to evaluating one closed-form expression on a single scalar measurement, enabling real-time beam alignment on low-power devices.
  • Because the estimator is interpretable, its failure modes (e.g., near 90°) are predictable from the derivative of the beam pattern, not hidden inside a network.
  • The cosine-inversion structure transfers to a RIS-aided two-hop link, so the approach is not limited to free-space line-of-sight.
  • Imposing the physics-informed structure costs only about 0.02° MAE in the multi-angle sweep compared with unconstrained symbolic regression, while keeping the model readable.
  • The derived Cramér–Rao bound is a closed-form, parameter-only curve that can be used to judge whether any future estimator is near-optimal in the same geometry.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the exponents n, m are fitted on the same data used to report accuracy, the 0.42° figure is an in-sample number; a held-out frequency or angular sweep would test true generalization.
  • In multipath environments the path-loss coefficient will mix several arrivals, so the single-cosine inversion will likely need a separation stage or an extended model with multiple lobes.
  • The polynomial-surrogate variant's larger error (5.96° in Stage I) suggests low-order polynomial approximations are only safe near boresight; the arccos branch becomes ambiguous once the true pattern has sidelobes.
  • The near-90° sharp rise in the Cramér–Rao bound is a real physical sensitivity loss for oblique incidence, so the model's limitation points toward a practical design rule: keep steering within the main lobe's valid angular sector.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper proposes SABER, a symbolic-regression framework that maps scalar path-loss/S21 measurements to angle-of-arrival (AoA). It assumes a cos^n(θ) beam-pattern model, fits directivity parameters with PySR, and then inverts the model directly or through a low-order polynomial surrogate. Validation is carried out in two stages: an anechoic chamber multi-angle sweep (Stage I) and a RIS-aided indoor setup with a single fixed AoA of 55° (Stage II). The paper reports MAE 0.42° for SABER direct inversion, 0.396° for unconstrained SR, near-zero MAE in Stage II, and closeness to a CRLB derived from the fitted model.

Significance. If the reported accuracies reflected predictive performance, the idea of interpretable, single-snapshot AoA estimation from a scalar power measurement would be of interest to mmWave/6G localization. The paper includes real measurements and a detailed CRLB derivation, which are strengths. However, as detailed below, the evaluation is entirely in-sample, the Stage II experiment cannot separate angle estimation from constant fitting, and the CRLB comparison is circular. These issues affect the central claims, so the current evidence does not establish the method's accuracy or generalization.

major comments (4)
  1. [Section IV-A and Table I] All Stage I MAE values (0.396°, 0.42°, 5.96°) are computed on the same 51-angle × 6-frequency data used to fit n_R, m_R, the direct-inversion offset, the polynomial coefficients, and the unconstrained SR expression. No train/test split, cross-validation, or separate validation set is described. The 'sub-0.5°' claim is therefore an in-sample fit metric; the 0.396° value for unconstrained SR may reflect memorization rather than estimation skill. A held-out evaluation is required before the method can be claimed as an AoA estimator.
  2. [Section IV-B and Table II] Stage II measures only one true AoA (55°). The expressions returned by both Unconstrained SR and SABER Direct Inversion are the constant 0.9599 rad (≈55°), so the reported MAE of 6.53×10^-7 is achieved by any constant predictor set to 55°. This experiment does not demonstrate recovery of AoA from path loss; it only shows the methods can fit one point. Multi-angle RIS measurements are needed to support the paper's AoA-estimation claims.
  3. [Section VI-C, Eq. (26)] The CRLB is computed from h(θ) in Eqs. (16) and (17), i.e., the same fitted cos^n model used to construct the estimator. Matching this bound shows the inversion of the assumed model is efficient for that model, not that SABER is accurate on real measured patterns. This is particularly problematic since Section VI-A explicitly admits cos^28(θ) cannot capture ripples, asymmetries, and null depths. The near-CRLB claim is therefore circular and should be reframed as a model-consistency check, or the CRLB should be computed from the measured beam pattern.
  4. [Section V-A, Eq. (12)] The direct inversion as written, θ̂ = arccos(10^{|S21|}) + offset, is inconsistent with Algorithm 1 (line 6), which uses arccos(10^{x/(10 n_R)}) + offset, and with Table I, which uses arccos(10^{ΔPL/(10 n_R)}) + offset. It is also dimensionally unclear whether |S21| is linear or dB. This obscures the actual estimator fitted and prevents reproducibility. Additionally, the abstract's claim that both the direct inversion and the polynomial surrogate achieve sub-0.5° MAE is contradicted by Table I, where the polynomial surrogate MAE is 5.96°.
minor comments (5)
  1. [Section V-B, Eq. (13)] Equation (13) is garbled; the exponent and denominator are unclear. Please restate the polynomial surrogate cleanly and align it with the expression in Table I.
  2. [Section VI-B and VI-C] The Monte Carlo CDF in Fig. 10 uses σ_PL = 3° with no sensitivity analysis or justification, and the CRLB comparison assumes σ^2 = 10^-3 and M = 1000 without explaining how these values are derived from the measurements. This makes the quantitative claims in Fig. 11 hard to verify.
  3. [General] No code or data release is mentioned. Given the small dataset and the in-sample evaluation, releasing data and code is essential for independent verification of the reported MAEs.
  4. [Abstract] The abstract states that 'the same SR-learned inversions' are applied in Stage II, but the fitted directivity parameters differ (Stage II reports n_T = 4, n_R = 1, while Stage I reports n_T = n_R = 28). The wording should be clarified to 'the same SR-based methods'.
  5. [Section VI-A and Section I-C] The paper does not compare against any classical (e.g., MUSIC/ESPRIT) or neural-network baseline on the same measurements, so the claim of being an 'alternative to state-of-the-art and black-box ML-based methods' is not directly supported by the experiments.

Circularity Check

3 steps flagged

Predictions reduce to calibration: Stage II's near-zero MAE is a constant fit at the single known angle; Stage I's sub-0.5° MAE is scored on the fitting grid; the CRLB is computed from the same fitted cos^n model.

specific steps
  1. fitted input called prediction [Section VI-B, Table II; Section IV-B (Stage II setup)]
    "For the sake of simplicity, we set the angle between both transmitter and receiver to70◦, i.e., θaxis = 35◦ ... we can obtain the AoA is55◦ with simple geometric manipulations ... As shown in Table II, both the unconstrained SR and SABER with Direct Inversion approach (row 1 and 2)converge to the same closed-form cosine model and achieve a virtually zero MAE (6.53×10−7) (column 2), effectively recovering the true55◦ angle with negligible error (column 3)."

    In Stage II the AoA is a single fixed value, 55° = 0.9599 rad, known a priori from the experimental geometry for both 2 m and 3 m. Both Table II 'expressions' are the constant 0.9599 rad, i.e., the learned estimator is a constant equal to the only label present in the dataset. Any predictor returning 0.9599 for every input attains the same 6.53e-7 MAE regardless of the path-loss value, so the near-zero error is a residual of fitting a constant to one label — equivalent to the input label by construction. It carries no information about general AoA estimation ability, as the paper itself concedes: 'prior beam-pattern knowledge is not strictly necessary when estimating a single, fixed AoA.'

  2. fitted input called prediction [Section V-A (Eq. 12, Algorithm 1), Section VI-A (Table I)]
    "Specifically, we fix the receive directivity (nR) and fit only a single additive offset ... Our findings from this method indicate that nT and mT are 1, while nR and mR are 28 ... we can then utilize SR to search for the AoA with respect to the path loss coefficients ... the performance (in MAE) and the final expression, are summarized in Table I."

    The direct-inversion estimator's parameters (n_R = 28, offset = 0.05813), the polynomial coefficients (a, b, c), and the unconstrained SR expression are all obtained by fitting the same anechoic-chamber data — 51 angles × 6 frequencies, Section IV-A — on which Table I computes the reported MAE. The paper says the data are used 'to train SR-based ML models and further validate their performance' but nowhere defines a train/test split, and Algorithm 1's claim of 'inference on unseen data' is never realized in the results. The headline 0.42° (and the unconstrained SR's 0.396°) MAE is therefore an in-sample goodness-of-fit of parameters calibrated on the very data being scored, i.e., a fitted input reported as prediction accuracy.

  3. fitted input called prediction [Section VI-C, Eqs. (24)–(26), Fig. 11]
    "Finally, by inserting the explicit form of the beam-pattern gainh(θ)and its derivativeh′(θ)given by our closed-form path loss expressions in Eqs. (8) and (16), we obtain a completely analytical curve p CRLB(θ)that depends only on known system parameters, the measurement noise varianceσ2, and the number of snapshotsM."

    Eq. (16) is the SR-fitted cos^28 model ('nR and mR are 28') obtained from the measured Stage I data, and the empirical RMSE in Fig. 11 comes from inverting that same fitted model on the same data. The CRLB is therefore the lower bound of the assumed, fitted model, and the near-CRLB agreement demonstrates only that the inverse of the fitted model is statistically efficient for the fitted model's own derivative — an internal consistency check, not an independent validation. The paper itself concedes that cos^28 'cannot capture the small amplitude ripples, asymmetries, and null depths observed in measurements,' so a bound derived from this fitted approximation cannot certify near-optimal accuracy on the true measured beam pattern.

full rationale

The symbolic-regression fitting itself (PySR discovering cos^28, a constant, or a polynomial) is a legitimate empirical procedure; the circularity lies in how the fitted quantities are then presented as validated predictions. (1) Stage II: the AoA is fixed at 55° by the experimental geometry for both 2 m and 3 m, and both 'recovered' expressions in Table II are the single constant 0.9599 rad — equal to the only label present. Any constant predictor at 55° attains the same 6.53e-7 MAE, so this headline result reduces by construction to the input label. (2) Stage I: n_R, m_R, the direct-inversion offset, the polynomial coefficients, and the unconstrained SR expression are all fitted on the same 51-angle × 6-frequency grid that Table I scores; no train/test split is described anywhere, and Algorithm 1's 'inference on unseen data' is never demonstrated. The sub-0.5° MAE is therefore an in-sample fit error, not an out-of-sample prediction; the paper's own nod to sacrificing 'out-of-sample behavior guarantees' confirms that no such guarantee is provided. (3) CRLB: the bound is computed by inserting the fitted cos^28 pattern (Eq. 16) into Eq. (26), and the empirical RMSE is produced by the inverse of that same fitted model on the same data; near-CRLB agreement is an internal consistency check for the assumed model, and the paper's admission that cos^28 misses ripples, asymmetries, and null depths means the model-bound comparison does not establish efficiency on the true measured patterns. On self-citation: the cos^n ansatz is cited to external [31]; the RIS path-loss equation cites the authors' own [32] but is checked against measured CDFs (Fig. 10); related-work self-citations [30], [33] and hardware references [39], [40] are not load-bearing to the estimator claims. There is no uniqueness-imported-from-authors or ansatz-via-self-citation pattern. Overall score 7: at least one advertised 'prediction' (Stage II near-zero MAE) reduces by construction, and the remaining headline numbers reduce to in-sample evaluation of fitted models. A genuine held-out evaluation or a multi-angle Stage II test would lower the score to 0–2.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central result relies on a small number of fitted beam-pattern exponents, offsets, and polynomial coefficients, plus standard propagation and signal-model assumptions. No new physical entities are introduced. The biggest burden is the implicit assumption that in-sample symbolic-regression fits generalize.

free parameters (7)
  • n_R receiver directivity exponent = 28 (Stage I), 1 (Stage II)
    Fitted by SR to beam-pattern data; used in direct inversion. The reported AoA MAE depends on this fit.
  • n_T transmitter directivity exponent = 1 (Stage I), 4 (Stage II)
    SR-fitted; enters free-space and RIS path loss models.
  • m_T/m_R elevation directivity exponents = 1 and 28 (Stage I)
    SR-fitted; part of the cos^n cos^m beam pattern model, though not used in azimuth inversion.
  • Direct-inversion offset = 0.05813 rad
    Fitted additive offset in θ = arccos(...) + offset; the 0.42° MAE depends on it.
  • Polynomial surrogate coefficients = a=0.03879, b=0.1165, c=0.8303
    Quadratic fit to cos(θ) as a function of ΔPL; used in polynomial-based inversion.
  • Unconstrained SR formula coefficients = 0.53076 and 0.12336 (Stage I); constant 0.9599 (Stage II)
    Discovered by PySR. Stage II reduces to a constant because only one angle is in the training set.
  • Monte Carlo angular spread σ_PL = 3°
    Chosen to make the analytical CDF match measurement; not independently measured.
axioms (6)
  • domain assumption Free-space path loss model PL_FS = 20log10(R)+20log10(f)+20log10(4π)-20log10(c)
    Standard FSPL model used in Eq. (2), assuming far-field and free-space propagation.
  • ad hoc to paper Antenna gain factorizes as G=Gmax U with U=cos^n(θ)cos^m(φ)
    Introduced in Eq. (5) as the beam pattern model; central to the direct inversion and CRLB derivation.
  • domain assumption RIS path loss expression Eq. (7) from [32] applies and ε_ap=1 in Stage II
    Borrowed from prior literature; the authors assume the aperture efficiency is unity for the deployed RIS.
  • domain assumption Steering angle of the RIS is θaxis = sin^-1(λ/2d) from a 180° progressive phase delay
    Standard phased-array relation used to compute the geometry in Stage II.
  • domain assumption Signal model y=αh(θ)x+n with n∼CN(0,σ²I), |α| known from calibration, and M=1000 snapshots
    Used for the CRLB derivation (Eqs. 18-26); noise variance and snapshot count are chosen, not measured.
  • ad hoc to paper In-sample MAE on the training data is representative of estimation accuracy
    The paper reports MAE without a train/test split or cross-validation, implicitly assuming no overfitting.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SABER: Symbolic Regression-based Angle of Arrival and Beam Pattern Estimator." pith.science (2026). https://pith.science/paper/MPN7KVC4

@misc{pith2026251026340,
  author       = {Pith},
  title        = {Pith review of: SABER: Symbolic Regression-based Angle of Arrival and Beam Pattern Estimator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MPN7KVC4}},
  note         = {Machine review of arXiv:2510.26340}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Accurate Angle-of-arrival (AoA) estimation is essential for next-generation wireless communication systems to enable reliable beamforming, high-precision localization, and integrated sensing. Unfortunately, classical high-resolution techniques require multi-element arrays and extensive snapshot collection, while generic Machine Learning (ML) approaches often yield black-box models that lack physical interpretability. To address these limitations, we propose a Symbolic Regression (SR)-based ML framework. Namely, Symbolic Regression-based Angle of Arrival and Beam Pattern Estimator (SABER), a constrained symbolic-regression framework that automatically discovers closed-form beam pattern and AoA models from path loss measurements with interpretability. SABER achieves high accuracy while bridging the gap between opaque ML methods and interpretable physics-driven estimators. First, we validate our approach in a controlled free-space anechoic chamber, showing that both direct inversion of the known $\cos^n$ beam and a low-order polynomial surrogate achieve sub-0.5 degree Mean Absolute Error (MAE). A purely unconstrained SR method can further reduce the error of the predicted angles, but produces complex formulas that lack physical insight. Then, we implement the same SR-learned inversions in a real-world, Reconfigurable Intelligent Surface (RIS)-aided indoor testbed. SABER and unconstrained SR models accurately recover the true AoA with near-zero error. Finally, we benchmark SABER against the Cram\'er-Rao Lower Bounds (CRLBs). Our results demonstrate that SABER is an interpretable and accurate alternative to state-of-the-art and black-box ML-based methods for AoA estimation.

Figures

Figures reproduced from arXiv: 2510.26340 by Carolina Fortuna, Cheng-Nan Hu, Jernej Hribar, Kuang-Chung Chou, Mengran Zhao, Shih-Kai Chou.

Figure 1
Figure 1. Figure 1: System model of free-space scenario. of light, respectively. As depicted in the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Beam pattern with different n and m in polar coordi￾nates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Flow chart of SABER: from the data collection to verification. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Experimental setup for Stage II (RIS-aided scenario). [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Measurement with Rohde & Schwarz ZVA40 VNA [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Measurement for the forward transmission coefficient, [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of magnitude between probe antenna [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: CDF of the measuring data in Stage II and the analyt [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Empirical RMSE of SABER: Direct Inversion com [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 2 canonical work pages

  1. [1]

    Wireless access in ultra- reliable low-latency communication (URLLC),

    P. Popovski, C. Stefanovic, J. J. Nielsen, E. de Carvalho, M. Angjelichi- noski, K. F. Trillingsgaard, and A.-S. Bana, “Wireless access in ultra- reliable low-latency communication (URLLC),”IEEE Transactions on Communications, vol. 67, no. 8, pp. 5783–5801, 2019

    2019

  2. [2]

    Exploring the 6G potentials: Immersive, hyper reliable, and low- latency communication,

    A. A. Shamsabadi, A. Yadav, Y . Gadallah, and H. Yanikomeroglu, “Exploring the 6G potentials: Immersive, hyper reliable, and low- latency communication,” 2025. [Online]. Available: https://arxiv.org/ abs/2407.11051

    Pith/arXiv arXiv 2025

  3. [3]

    AI-powered urban transportation digital twin: Methods and applications,

    X. Di, Y . Fu, M. K. Turkcan, M. Ghasemi, Z. Mo, C. Zang, A. Adhikari, Z. Kostic, and G. Zussman, “AI-powered urban transportation digital twin: Methods and applications,” 2024. [Online]. Available: https://arxiv.org/abs/2501.10396

    arXiv 2024

  4. [4]

    Millimeter wave mobile communications for 5G cellular: It will work!

    T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y . Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular: It will work!”IEEE Access, vol. 1, pp. 335–349, 2013

    2013

  5. [5]

    Terahertz-band ultra-massive spatial modulation MIMO,

    H. Sarieddeen, M.-S. Alouini, and T. Y . Al-Naffouri, “Terahertz-band ultra-massive spatial modulation MIMO,”IEEE Journal on Selected Areas in Communications, vol. 37, no. 9, pp. 2040–2052, 2019

    2040

  6. [6]

    Ultra-dense cell-free massive MIMO for 6G: Technical overview and open questions,

    H. Q. Ngo, G. Interdonato, E. G. Larsson, G. Caire, and J. G. Andrews, “Ultra-dense cell-free massive MIMO for 6G: Technical overview and open questions,” 2024. [Online]. Available: https://arxiv.org/abs/2401.03898

    Pith/arXiv arXiv 2024

  7. [7]

    RIS- assisted ISAC systems for robust secure transmission with imperfect sense estimation,

    C. Jiang, C. Zhang, C. Huang, J. Ge, D. Niyato, and C. Yuen, “RIS- assisted ISAC systems for robust secure transmission with imperfect sense estimation,”IEEE Transactions on Wireless Communications, vol. 24, no. 5, pp. 3979–3992, 2025. 12

    2025

  8. [8]

    AOA localization for vehicle-tracking systems using a dual-band sensor array,

    M. A. G. Al-Sadoon, R. Asif, Y . I. A. Al-Yasir, R. A. Abd-Alhameed, and P. S. Excell, “AOA localization for vehicle-tracking systems using a dual-band sensor array,”IEEE Transactions on Antennas and Propa- gation, vol. 68, no. 8, pp. 6330–6345, 2020

    2020

  9. [9]

    The angle of arrival estimation of frequency-hopping cooperative object based on software-defined radio,

    Rustamaji, S. Kliwati, and W. Widada, “The angle of arrival estimation of frequency-hopping cooperative object based on software-defined radio,”Scientific Reports, vol. 14, no. 1, p. 7732, 2024. [Online]. Available: https://doi.org/10.1038/s41598-024-58488-8

    work page doi:10.1038/s41598-024-58488-8 2024

  10. [10]

    Machine learning for 5G/B5G mobile and wireless communications: Potential, limitations, and future directions,

    M. E. Morocho-Cayamcela, H. Lee, and W. Lim, “Machine learning for 5G/B5G mobile and wireless communications: Potential, limitations, and future directions,”IEEE Access, vol. 7, pp. 137 184–137 206, 2019

    2019

  11. [11]

    Multiple emitter location and signal parameter estimation,

    R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Transactions on Antennas and Propagation, vol. 34, no. 3, pp. 276–280, 1986

    1986

  12. [12]

    ESPRIT-estimation of signal parameters via ro- tational invariance techniques,

    R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via ro- tational invariance techniques,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 37, no. 7, pp. 984–995, 1989

    1989

  13. [13]

    Stability and super-resolution of MUSIC and ESPRIT for multi-snapshot spectral estimation,

    W. Li, Z. Zhu, W. Gao, and W. Liao, “Stability and super-resolution of MUSIC and ESPRIT for multi-snapshot spectral estimation,”IEEE Transactions on Signal Processing, vol. 70, pp. 4555–4570, 2022

    2022

  14. [14]

    Optimal spatial filtering of single trial EEG during imagined hand movement,

    H. Ramoser, J. Muller-Gerking, and G. Pfurtscheller, “Optimal spatial filtering of single trial EEG during imagined hand movement,”IEEE Transactions on Rehabilitation Engineering, vol. 8, no. 4, pp. 441–446, 2000

    2000

  15. [15]

    Machine learning assisted phase- less millimeter-wave beam alignment in multipath channels,

    B. W. Domae, R. Li, and D. Cabric, “Machine learning assisted phase- less millimeter-wave beam alignment in multipath channels,” inProceed- ings 2021 IEEE Global Communications Conference (GLOBECOM), Dec 2021, pp. 1–7

    2021

  16. [16]

    Fast millimeter wave beam alignment,

    H. Hassanieh, O. Abari, M. Rodriguez, M. Abdelghany, D. Katabi, and P. Indyk, “Fast millimeter wave beam alignment,” inProceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, ser. SIGCOMM ’18. New York, NY , USA: Association for Computing Machinery, 2018, p. 432–445. [Online]. Available: https://doi.org/10.1145/3230543.3230581

    arXiv 2018

  17. [17]

    Deep learning based predictive beamforming design,

    J. Zhang, G. Zheng, Y . Zhang, I. Krikidis, and K.-K. Wong, “Deep learning based predictive beamforming design,”IEEE Transactions on Vehicular Technology, vol. 72, no. 6, pp. 8122–8127, 2023

    2023

  18. [18]

    5G-advanced AI/ML beam management: Performance evaluation with integrated ML models,

    N. Jayaweera, A. Bonfante, M. Schamberger, A. M. A. Tehrani, T. Sanguanpuak, P. Tilak, K. Jayasinghe, F. W. V ook, and N. Rajatheva, “5G-advanced AI/ML beam management: Performance evaluation with integrated ML models,” 2024. [Online]. Available: https://arxiv.org/abs/2404.15326

    Pith/arXiv arXiv 2024

  19. [19]

    Integrated probing-beam pattern learning and beam prediction for mmWave mas- sive MIMO,

    Q. Xue, A. Nordio, K. Niu, C. Dong, and C. F. Chiasserini, “Integrated probing-beam pattern learning and beam prediction for mmWave mas- sive MIMO,”IEEE Transactions on Communications, pp. 1–1, 2025

    2025

  20. [20]

    Deep learning approach in DOA estimation: A systematic literature review,

    S. Ge, K. Li, S. N. B. M. Rum, and S.-B. Tsai, “Deep learning approach in DOA estimation: A systematic literature review,”Mob. Inf. Syst., vol. 2021, Jan. 2021. [Online]. Available: https://doi.org/10.1155/2021/6392875

    work page doi:10.1155/2021/6392875 2021

  21. [21]

    Angle-of-arrival estimation using an adaptive machine learning framework,

    A. Khan, S. Wang, and Z. Zhu, “Angle-of-arrival estimation using an adaptive machine learning framework,”IEEE Communications Letters, vol. 23, no. 2, pp. 294–297, 2019

    2019

  22. [22]

    Deep networks for direction-of-arrival estimation in low SNR,

    G. K. Papageorgiou, M. Sellathurai, and Y . C. Eldar, “Deep networks for direction-of-arrival estimation in low SNR,”IEEE Transactions on Signal Processing, vol. 69, pp. 3714–3729, 2021

    2021

  23. [23]

    3D direction of arrival estimation: An innovative deep neural network approach,

    C. M. Mylonakis, P. Velanas, P. I. Lazaridis, P. Sarigiannidis, S. K. Goudos, and Z. D. Zaharis, “3D direction of arrival estimation: An innovative deep neural network approach,” in2024 13th International Conference on Modern Circuits and Systems Technologies (MOCAST), 2024, pp. 01–04

    2024

  24. [24]

    Direction of arrival (DOA) estimation using a deep unfolded learned iterative shrinkage thresholding algorithm (LISTA) network in a non-uniform metasurface,

    X. Niu, X. Su, L. He, and G. Chen, “Direction of arrival (DOA) estimation using a deep unfolded learned iterative shrinkage thresholding algorithm (LISTA) network in a non-uniform metasurface,” Remote Sensing, vol. 17, no. 7, 2025. [Online]. Available: https: //www.mdpi.com/2072-4292/17/7/1253

    2025

  25. [25]

    Neural network-based intelligent reflecting surface assisted direction of arrival estimation,

    Y . Azhdari and M. Farhang, “Neural network-based intelligent reflecting surface assisted direction of arrival estimation,” 2025. [Online]. Available: https://arxiv.org/abs/2406.18306

    Pith/arXiv arXiv 2025

  26. [26]

    ML-assisted RIS for ISAC systems: Initial results in the 6G study band,

    D. T. Phan, Q. D. Nguyen, N. Takanen, T. N. Nguyen, M. Juntti, and P. J. Soh, “ML-assisted RIS for ISAC systems: Initial results in the 6G study band,” inProceedings of 2025 19th European Conference on Antennas and Propagation (EuCAP), 2025, pp. 1–5

    2025

  27. [27]

    Deep learning based AoA and AoD estimation for millimeter wave MIMO systems,

    D. Lloria, S. Roger, C. Botella-Mascarell, and M. Cobos, “Deep learning based AoA and AoD estimation for millimeter wave MIMO systems,” inProceedings of the 12th Euro American Conference on Telematics and Information Systems, ser. EATIS 2024. New York, NY , USA: Association for Computing Machinery, 2025. [Online]. Available: https://doi.org/10.1145/368524...

    arXiv 2024

  28. [28]

    Machine learning-based chan- nel estimation for multi-RIS-assisted mmwave massive-MIMO OFDM system in a dynamic environment,

    T. C. Rapudu and O. O. Oyerinde, “Machine learning-based chan- nel estimation for multi-RIS-assisted mmwave massive-MIMO OFDM system in a dynamic environment,”IEEE Transactions on Wireless Communications, vol. 24, no. 6, pp. 5297–5309, 2025

    2025

  29. [29]

    Symbolic regression via genetic pro- gramming,

    D. Augusto and H. Barbosa, “Symbolic regression via genetic pro- gramming,” inProceedings. Vol.1. Sixth Brazilian Symposium on Neural Networks, 2000, pp. 173–178

    2000

  30. [30]

    Towards automated and interpretable pathloss approximation methods,

    A. Anaqreh, S.-K. Chou, I. Bara ˇsin, and C. Fortuna, “Towards automated and interpretable pathloss approximation methods,” in Proceedings of AAAI 2025 Workshop on Artificial Intelligence for Wireless Communications and Networking (AI4WCN), 2024. [Online]. Available: https://openreview.net/forum?id=M1WT5NZ4bj

    2025

  31. [31]

    A novel approximate antenna pattern for directional antenna arrays,

    N. Deng and M. Haenggi, “A novel approximate antenna pattern for directional antenna arrays,”IEEE Wireless Communications Letters, vol. 7, no. 5, pp. 832–835, 2018

    2018

  32. [32]

    On the aperture efficiency of intelligent reflecting surfaces,

    S.-K. Chou, O. Yurduseven, H. Q. Ngo, and M. Matthaiou, “On the aperture efficiency of intelligent reflecting surfaces,”IEEE Wireless Communications Letters, vol. 10, no. 3, pp. 599–603, 2021

    2021

  33. [33]

    Automated modeling method for pathloss model discovery,

    A. Anaqreh, S.-K. Chou, M. Mohor ˇciˇc, T. Lagkas, and C. Fortuna, “Automated modeling method for pathloss model discovery,” 2025. [Online]. Available: https://arxiv.org/abs/2505.23383

    Pith/arXiv arXiv 2025

  34. [34]

    Deep symbolic optimization: Reinforcement learning for symbolic mathematics,

    C. F. Hayes, F. L. D. Silva, J. Yang, T. N. Mundhenk, C. S. Lee, J. F. Pettit, C. Santiago, S. Kim, J. T. Kim, I. A. Solis, R. Glatt, A. R. Goncalves, A. Ladd, A. C. Solak, T. Desautels, D. Faissol, B. K. Petersen, and M. Landajuela, “Deep symbolic optimization: Reinforcement learning for symbolic mathematics,” 2025. [Online]. Available: https://arxiv.org...

    Pith/arXiv arXiv 2025

  35. [35]

    Symbolic regression via neural-guided genetic programming population seeding,

    T. N. Mundhenk, M. Landajuela, R. Glatt, C. P. Santiago, D. M. Faissol, and B. K. Petersen, “Symbolic regression via neural-guided genetic programming population seeding,” 2021. [Online]. Available: https://arxiv.org/abs/2111.00053

    Pith/arXiv arXiv 2021

  36. [36]

    Grammar variational autoencoder,

    M. J. Kusner, B. Paige, and J. M. Hern ´andez-Lobato, “Grammar variational autoencoder,” 2017. [Online]. Available: https://arxiv.org/ abs/1703.01925

    Pith/arXiv arXiv 2017

  37. [37]

    AI Feynman: a physics-inspired method for symbolic regression,

    S.-M. Udrescu and M. Tegmark, “AI Feynman: a physics-inspired method for symbolic regression,” 2020. [Online]. Available: https: //arxiv.org/abs/1905.11481

    Pith/arXiv arXiv 2020

  38. [38]

    KG,R&S® ZVA / ZVB / ZVT Vector Network Analyzers Operating Manual, Rohde & Schwarz GmbH & Co

    Rohde & Schwarz GmbH & Co. KG,R&S® ZVA / ZVB / ZVT Vector Network Analyzers Operating Manual, Rohde & Schwarz GmbH & Co. KG, M ¨unchen, Germany, 2020. [Online]. Available: https://scdn.rohde-schwarz.com/ur/pws/dl downloads/ dl common library/dl manuals/dl user manual/ZV AZVB ZVT OperatingManual en 33.pdf

    2020

  39. [39]

    Research on millimeter-wave phased antenna array for wireless communication applications,

    C.-N. Hu, Y .-X. Li, and K.-C. Chou, “Research on millimeter-wave phased antenna array for wireless communication applications,” in 2024 IEEE 12th Asia-Pacific Conference on Antennas and Propagation (APCAP), 2024, pp. 1–2

    2024

  40. [40]

    Passive reconfigurable intelligent surface (RIS) for enhanced wireless communications,

    C.-N. Hu, Q.-C. Li, P.-Y . Ho, F.-S. Hsu, K.-C. Chou, W.-J. Cheng, C.- H. Chang, and P.-S. Wan, “Passive reconfigurable intelligent surface (RIS) for enhanced wireless communications,” in2024 Photonics & Electromagnetics Research Symposium (PIERS), 2024, pp. 1–4

    2024

  41. [41]

    F6522 Short-Form Datasheet: 8-Ch Tx Active Beamforming IC 27.5–31.0 GHz,

    Renesas Technology Corp., “F6522 Short-Form Datasheet: 8-Ch Tx Active Beamforming IC 27.5–31.0 GHz,” https://www.renesas.com/en/ document/sds/f6522-short-form-datasheet, 2020

    2020

  42. [42]

    Reconfigurable intelligent surfaces: A signal pro- cessing perspective with wireless applications,

    E. Bj ¨ornson, H. Wymeersch, B. Matthiesen, P. Popovski, L. Sanguinetti, and E. de Carvalho, “Reconfigurable intelligent surfaces: A signal pro- cessing perspective with wireless applications,”IEEE Signal Processing Magazine, vol. 39, no. 2, pp. 135–158, 2022

    2022

  43. [43]

    Interpretable machine learning for science with PySR and symbolic regression,

    M. Cranmer, “Interpretable machine learning for science with PySR and symbolic regression,” 2023. [Online]. Available: https: //arxiv.org/abs/2305.01582

    Pith/arXiv arXiv 2023

  44. [44]

    DeepAoANet: Learning angle of arrival from software defined radios with deep neural networks,

    Z. Dai, Y . He, V . Tran, N. Trigoni, and A. Markham, “DeepAoANet: Learning angle of arrival from software defined radios with deep neural networks,”IEEE Access, vol. 10, pp. 3164–3176, 2022

    2022

  45. [45]

    H. L. V . Trees,Detection, Estimation, and Modulation Theory. John Wiley & Sons, Ltd, 2001

    2001

  46. [46]

    Cramer-rao bound for AOA-based VLP with an aperture-based receiver,

    H. Steendam, T. Q. Wang, and J. Armstrong, “Cramer-rao bound for AOA-based VLP with an aperture-based receiver,” in2017 IEEE International Conference on Communications (ICC), 2017, pp. 1–6

    2017

  47. [47]

    Demonstration of tunable steering and multiplexing of two 28 GHz data carrying orbital angular momentum beams using antenna array,

    G. Xie, Z. Zhao, and Y . Y . and, “Demonstration of tunable steering and multiplexing of two 28 GHz data carrying orbital angular momentum beams using antenna array,”Scientific Reports, vol. 6, p. 37078, 2016

    2016

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.