PaperlayerView exact revision on arXiv

Source-linked AI summary

Confidence-Guided Adaptive Beam Selection for Vision-Aided Beam Tracking in mmWave Systems

Sohani Munteha Hiam, Mohammad Soleymani, Rafael F. Schaefer

arXiv:2610.21880v1eess.SP

TL;DR

Fixed top-K selection assigns the same candidate count to every prediction despite uncertainty varying across samples and future horizons. The policy temperature-calibrates predictor logits, then retains the smallest top-ranked beam set reaching a cumulative-probability threshold under a maximum budget. At matched overhead versus fixed top-K, candidate set accuracy gains up to 3.80 percentage points, normalized received power increases 0.32 percentage points, and outage probability decreases 2.06 percentage points.

  • Problem

    Fixed top-K selection assigns the same candidate count to every prediction despite uncertainty varying across samples and future horizons.

    Source: §I · p. 1

  • Method

    The policy temperature-calibrates predictor logits, then retains the smallest top-ranked beam set reaching a cumulative-probability threshold under a maximum budget.

    Source: §I · p. 1

  • Results

    At matched overhead versus fixed top-K, candidate set accuracy gains up to 3.80 percentage points, normalized received power increases 0.32 percentage points, and outage probability decreases 2.06 percentage points.

    Source: 1) Budget Matched Performance · p. 4

  • Takeaways & Limitations

    The policy provides a more favorable balance between communication reliability and training overhead, with candidate count increasing with prediction uncertainty while Kmax controls the maximum overhead.

    Source: §V · p. 5

    Future work may investigate conformal prediction-set calibration for an explicit coverage guarantee and evaluate the approach across additional propagation scenarios and beam codebooks.

    Source: §V · p. 5

Abstract

from arXiv · show

Vision-aided beam tracking reduces mmWave beam-training overhead by predicting beam indices from camera observations. Fixed top-$K$ selection, however, assigns the same candidate beam budget to all predictions, ignoring uncertainty variations across samples and horizons. This letter proposes confidence-guided adaptive candidate selection: predictor logits are temperature-calibrated, and the smallest top-ranked beam set whose cumulative probability reaches a target threshold is retained under a maximum budget. On DeepSense~6G Scenario~8, over five random seeds and at matched overhead, the policy improves candidate set accuracy by up to 3.80 percentage points, increases normalized received power by 0.32 percentage points, and reduces outage probability by 2.06 percentage points without retraining the predictor.

I. INTRODUCTION

Vision-aided beam tracking uses camera observations to predict future mmWave beam indices, but fixed candidate counts do not reflect uncertainty differences across samples and horizons. This work studies confidence-guided beam-budget allocation alongside the system objective of retaining the optimal beam with few candidates.

I. INTRODUCTION

  • Fixed candidate counts ignore uncertainty differences across predictions, while exhaustive beam search incurs substantial training overhead and latency.Camera observations at the base station can predict future beam indices [1],.Source: §I · p. 1
  • Prior work addresses visual representations, sensing, fusion, robustness, blockage, and beam prediction, whereas this letter focuses on candidate-beam resource allocation after prediction,,,,,,,.The comparison is with representative vision-aided beam-management approaches.Source: §I · p. 1, Table I · p. 2
  • The method calibrates predictor logits and retains the smallest top-ranked beam set reaching a cumulative-probability threshold, allocating budgets per sample and horizon.The predictor architecture is adopted from, temperature scaling from, and the cumulative-mass principle is related to top-p selection.Source: §I · p. 1

II. SYSTEM MODEL AND PROBLEM FORMULATION

  • The system objective is to include the optimal beam while keeping each candidate set small, balancing tracking reliability against beam-training overhead.The modeled downlink uses a base station with an antenna array and co-located RGB camera serving a high-mobility user.Source: §II · p. 2

III. PROPOSED ADAPTIVE CANDIDATE BEAM SELECTION

The proposed selection layer operates after the adopted vision-aided predictor produces multi-horizon beam logits. It calibrates those logits and uses the resulting probabilities for adaptive candidate selection.

III. PROPOSED ADAPTIVE CANDIDATE BEAM SELECTION

  • Beam tracking is treated as multiclass classification over the 64-beam codebook in DeepSense 6G Scenario 8.Source: §III · p. 2
  • The adopted YOLOv4 and encoder–decoder GRU predictor generates beam logits for multiple future horizons from RGB image sequences.Its architecture and training procedure remain unchanged.Source: §III · p. 2, Figure 2 · p. 2
  • The contribution follows prediction: temperature scaling calibrates the logits, and calibrated probabilities drive adaptive candidate-beam selection.The framework is applied to the output of the adopted predictor.Source: §III · p. 2, Figure 3 · p. 2

A. Confidence Calibration

A single global temperature is fitted per seed on calibration samples across prediction steps. Temperature scaling changes probability concentration without changing beam ordering, so it can affect adaptive counts but not fixed top-K sets.

A. Confidence Calibration

  • Sequence samples are split into approximately 72% training, 8% calibration, and 20% test data.Source: A. Confidence Calibration · p. 2
  • For each seed, one temperature is fitted jointly over calibration samples and prediction steps by minimizing negative log likelihood.Calibration uses the ground-truth optimal beam for each sample and prediction step.Source: A. Confidence Calibration · pp. 2–3
  • Temperature scaling preserves beam ordering but changes probability concentration, leaving fixed top-K sets unchanged while potentially changing the count needed to reach a cumulative-mass threshold.Temperatures below one sharpen distributions and temperatures above one soften them.Source: A. Confidence Calibration · pp. 2–3

B. Adaptive Candidate Beam Selection

Adaptive selection ranks calibrated beam probabilities separately for each prediction and retains the smallest set reaching the chosen cumulative-mass threshold, subject to a maximum count. Fixed top-K instead assigns the same count to every prediction.

B. Adaptive Candidate Beam Selection

C. Beam Tracking Performance Metrics

The section defines candidate-set accuracy, normalized received power, and outage probability alongside candidate count and beam-training overhead. Outage is determined by whether normalized received power falls below a 0.95 threshold.

C. Beam Tracking Performance Metrics

  • Candidate count is the size of the selected beam set, and average candidate count determines beam-training overhead.The metrics are evaluated over test samples and prediction steps.Source: C. Beam Tracking Performance Metrics · p. 3
  • Candidate-set accuracy measures whether the optimal beam is included in the selected candidate sets.The passage defines ACC over test samples and prediction steps.Source: C. Beam Tracking Performance Metrics · p. 3
  • Normalized received power is bounded between 0 and 1, with 1 corresponding to the optimal beam.The section defines received power using linear values.Source: C. Beam Tracking Performance Metrics · p. 3
  • An outage occurs when normalized received power is below ρmin = 0.95, and the outage probability is reported as a percentage.This threshold corresponds to retaining at least 95% of optimal linear received power.Source: C. Beam Tracking Performance Metrics · p. 3

D. Budget Matched Fixed Top-K Comparison

The comparison matches adaptive and fixed top-K policies by average candidate count, which implies equal average beam-training overhead. Fixed-policy values are interpolated between evaluated budgets when needed.

D. Budget Matched Fixed Top-K Comparison

  • Fixed top-K results are linearly interpolated between evaluated budgets when the adaptive policy's average candidate count falls between them.The evaluated fixed budgets are Kfix = {1, 2, 3, 5, 8}.Source: D. Budget Matched Fixed Top-K Comparison · p. 3
  • The matched fixed-policy value is obtained from the adjacent evaluated budgets surrounding the adaptive average candidate count.If the adaptive count equals an evaluated fixed budget, that budget's result is used directly.Source: D. Budget Matched Fixed Top-K Comparison · p. 3
  • Matching is performed independently for each seed before aggregation, and equal average candidate counts imply equal average beam-training overhead.Source: D. Budget Matched Fixed Top-K Comparison · p. 3

E. Aggregation Across Repeated Runs

Experiments use five random seeds, and reported metrics are aggregated as means with sample standard deviations shown as error bars.

E. Aggregation Across Repeated Runs

F. Calibration Metric

Calibration quality is measured with expected calibration error (ECE), computed from confidence bins and aggregated across seeds and prediction steps. Lower ECE indicates better calibration.

F. Calibration Metric

  • ECE is computed separately for uncalibrated T = 1 and calibrated T = T⋆ probabilities.Source: F. Calibration Metric · pp. 3–4
  • Confidence values are divided into equal-width bins over [0, 1], and ECE compares each nonempty bin's empirical top-1 accuracy with its mean predicted confidence.Empty bins contribute zero.Source: F. Calibration Metric · pp. 3–4
  • ECE is computed independently for each seed and prediction step, then aggregated across five seeds; lower values indicate better calibration.Source: F. Calibration Metric · p. 4

IV. RESULTS AND DISCUSSION

At matched beam-training overhead, adaptive selection outperforms fixed top-K across the evaluated performance measures. It allocates more beams to predictions with greater uncertainty, with candidate counts also increasing as the confidence-mass threshold rises.

1) Budget Matched Performance:

  • Up to 3.80 percentage points higher candidate set accuracy, 0.32 percentage points higher average normalized received power, and 2.06 percentage points lower outage probability at matched overhead.The largest gains occur at low training overhead and decrease as budgets increase; trends are consistent across five seeds.Source: 1) Budget Matched Performance · p. 4

2) Adaptive Candidate Count Under Uncertainty:

  • At τ = 0.95, mean selected candidate count rises from approximately 1.75 to 3.81 beams between the lowest- and highest-uncertainty groups.For every evaluated τ, counts rise with uncertainty; increasing τ also raises candidate counts because more cumulative probability mass must be retained.Source: 2) Adaptive Candidate Count Under… · p. 4
  • At τ = 0.95, the median, 90th-percentile, and maximum candidate counts are 3, 4, and 6, while Kmax = 8 is reached only at τ ∈{0.98, 0.99}.Evaluation used DeepSense 6G Scenario 8, with budget-matched comparisons aggregated over five random seeds.Source: §IV · p. 4, 2) Adaptive Candidate Count Under… · p. 5

3) Calibration Quality:

Global temperature scaling reduces ECE and changes adaptive candidate counts without changing beam ranking. The conclusion reports reduced beam-training overhead and presents the method as a lightweight decision layer for existing predictors.

3) Calibration Quality:

  • Averaged across prediction horizons, temperature scaling reduces ECE from 5.55% to 4.98%, an absolute reduction of 0.57%.Scaling reduces ECE at every horizon; fitted temperatures below one sharpen the distributions relative to raw softmax.Source: 3) Calibration Quality · p. 5

4) Effect of Calibration on Adaptive Selection:

  • At τ = 0.95, calibration reduces mean overhead from 4.65% to 4.35%, alongside a reduction in average candidate set size from 2.98 to 2.79 beams.The overhead reduction appears across evaluated τ values and random seeds; Fig. 7a presents mean overhead across thresholds.Source: Figure 7 · p. 5, 4) Effect of Calibration on Adaptive… · p. 5

V. CONCLUSION

  • Temperature scaling improves predicted probabilities and reduces the adaptive rule’s beam budget even though beam ranking remains unchanged.The predictor is neither modified nor retrained, making the method a lightweight decision layer for existing beam prediction architectures.Source: §V · p. 5

Related papers

Loading 2610.21880v1…