PaperlayerView exact revision on arXiv

Source-linked AI summary

A 3GPP-Compliant Benchmark Dataset for RIS-Aided Beyond 5G Networks

Pujitha Mamillapalli, Pankaj Singh Rathour, Abhinav Kumar

arXiv:2609.39058v1cs.DBcs.AI

TL;DR

RIS research lacks standardized, high-fidelity datasets for data-driven evaluation, while large-scale phase optimization is computationally challenging. This paper builds a 3GPP-compliant dataset and uses CSI-to-CQI classification to benchmark models; the evaluation finds generalization remains difficult across simulated and real-world shifts.

  • Problem

    Data-driven RIS research lacks realistic, standardized, large-scale datasets with 3GPP compliance, scenario diversity, and quality labels for evaluation.

    Source: §1 · p. 2

  • Method

    The paper constructs a 3GPP TR 38.901-compliant RIS-aided mmWave dataset with oracle phase labels and benchmarks a CSI-to-CQI classification task.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    Models struggle to generalize across simulated and real-world distribution shifts, with degradation even within simulated settings and only partial mitigation from fine-tuning.

    Source: §7 · p. 9

  • Takeaways & Limitations

    The dataset offers a multi-scenario benchmark for evaluating robustness and domain generalization in RIS systems.

    Source: §7 · p. 9

    The dataset is simulation-based and omits hardware impairments, mobility, and time-varying propagation dynamics.

    Source: §4 · p. 6

Abstract

from arXiv · show

Reconfigurable Intelligent Surfaces (RIS) are emerging as a key technology for programmable wireless environments in the beyond the fifth generation (B5G) networks. However, data-driven RIS research remains bottleneck by the lack of standardized, high-fidelity and open-source datasets. In this paper, we introduce a large-scale 3GPP TR 38.901-compliant dataset for RIS-aided millimeter wave (mmWave) networks, that considers severe path loss, blockage sensitivity, and spatial channel sparsity make the RIS assistance more impactful. The dataset spans various canonical 3GPP deployment scenarios across 20 controlled variants, capturing diverse user densities, fading conditions, and blockage regimes. Uniquely, every sample includes oracle RIS phase configurations obtained via a globally optimal brute-force codebook search, providing gold-standard supervision labels that are absent from any existing public dataset. Rich multi-task annotations comprising full channel state information (CSI), per-link channel decomposition, optimal phase matrices, and channel quality index (CQI) labels support a broad range of machine learning paradigms and downstream tasks, including phase optimization, channel estimation, and interference management. As the primary benchmark task, we introduce a novel CSI-to-CQI mapping that frames RIS-aided link-quality prediction as a scalable scalar classification problem, thereby avoiding the exponential output complexity of the direct phase vector prediction. We have evaluated this mapping against state-of-the-art architectures under in-distribution, out-of-distribution, and real-world hardware measurement conditions. Our dataset provides a reproducible, extensible, and community-ready foundation to accelerate data-driven research in RIS-aided B5G networks.

1 Introduction

B5G demands are pushing conventional wireless infrastructure toward its limits, and RIS offer programmable control of signal propagation. Yet RIS optimization and data-driven research face computational and dataset barriers that motivate a standardized benchmark.

1 Introduction

  • RIS can program signal paths to improve coverage, reduce inter-user interference, and strengthen signals without power amplifiers or other complex radio hardware.The paper describes this as turning a stochastic channel into a deterministic, programmable resource at lower energy cost than active alternatives.Source: §1 · p. 1
  • Large-scale RIS optimization is difficult because its high-dimensional, non-convex formulation raises computational complexity, making iterative solvers impractical under blockage.The cited solver families include alternating optimization, successive convex approximation, and manifold optimization.Source: §1 · p. 2
  • RIS machine-learning research lacks realistic, standardized, large-scale datasets with 3GPP-aligned propagation, scenario diversity, and quality labels for rigorous evaluation.Existing work predominantly uses private, scenario-specific datasets based on simplified channel models.Source: §1 · p. 2
  • The proposed benchmark spans UMa, UMi, and InH scenarios across 20 controlled variants, with CSI, per-link decomposition, oracle phase labels, and CQI mapping.Oracle phases come from a globally optimal brute-force codebook search; the scalar CSI-to-CQI task avoids the exponential output complexity of predicting phase vectors.Source: §1 · p. 2

2 Related Work

Prior RIS optimization research spans iterative solvers and learning methods, but both face practical trade-offs. Existing datasets vary in realism and scope, limiting fair comparison and assessment of generalization.

2 Related Work

  • Classical iterative solvers have high computational cost and latency, while learning methods trade these costs for challenges such as sample inefficiency, instability, or dependence on oracle labels.Unsupervised methods avoid labels but lack guarantees of globally optimal solutions in non-convex settings.Source: §2 · pp. 2–3
  • Most prior learning studies rely on private, single-scenario datasets lacking standardized channel models, 3GPP compliance, or oracle labels, weakening fair comparison and generalization conclusions..Source: §2 · p. 3
  • Existing datasets and simulators have gaps in multi-scenario coverage, explicit transmitter–RIS and RIS–receiver decomposition, 3GPP alignment, and oracle phase labels.Real-world measurement datasets are typically small, cover few scenarios and frequencies, and lack oracle phase labels.Source: §2 · p. 3

3 Preliminaries

The section models a downlink RIS-aided MIMO system and defines oracle phase optimization by maximizing aggregate user throughput. It recasts link-quality prediction as CSI-to-CQI classification, with labels derived from oracle-optimal configurations.

System Model

  • The system comprises a base station with N antennas, a passive RIS with L² reflecting elements, and K single-antenna users.Its channel state includes BS-to-RIS, RIS-to-user, and direct BS-to-user links; the performance metric is α-mean throughput.Source: §3 · pp. 3–4

RIS Phase Optimization Task and Label Generation

  • The oracle phase vector maximizes summed per-user α-mean throughput over the discrete RIS phase codebook.Exhaustive codebook search guarantees a global optimum but has complexity O(|S|^φ), which grows exponentially with the number of RIS elements.Source: §3 · p. 4
  • CSI-to-CQI mapping quantizes throughput under the oracle-optimal phase configuration into one of C discrete CQI classes.The quantization levels align with 3GPP CQI definitions.Source: §3 · p. 4

Learning Framework

  • The mapping reduces the output space from |S|^L² phase configurations to C classes, enabling scalar classification with labels actionable by the BS scheduler.The classification objective is defined over CSI samples and oracle-generated CQI labels.Source: §3 · p. 5
  • The predictor is trained by empirical risk minimization on independently and identically distributed CSI samples paired with oracle CQI labels.Residual risk reflects model capacity or distribution mismatch because the labels are globally optimal by construction.Source: §3 · p. 5

4 Dataset Generation Pipeline

The pipeline applies the same generation procedure across 20 variants, changing only the channel configuration to isolate controlled propagation differences. It generates channels, searches codebook phase configurations for the optimum, and stores dataset samples; the dataset is simulation-based and excludes hardware impairments, mobility, and time-varying blockage.

4 Dataset Generation Pipeline

  • Across all 20 variants, only the channel configuration changes, so inter-variant differences arise from controlled propagation changes.UE positions and blockage states are sampled according to that configuration, and large-scale fading follows 3GPP TR 38.901.Source: §4 · p. 5
  • For each codebook phase configuration, the pipeline forms the effective channel, evaluates throughput, and stores the throughput-maximizing phase vector and its CQI label.It also stores the optimal aggregate alpha-mean throughput per sample.Source: §4 · p. 5
  • The dataset contains 200,000 channel realizations across 20 variants, with each sample represented by a 250,000-dimensional feature vector.The system targets 28 GHz mmWave deployments with a 36-antenna BS, a 30×30 RIS, and 100 users.Source: §4 · pp. 5–6
  • The simulation dataset omits hardware impairments and assumes static user locations and fixed blockage.The authors identify mobility and time-varying propagation as outside the current dataset scope.Source: §4 · p. 6

5 Dataset Analysis

The analysis examines the dataset’s variance distribution using principal component analysis. Variance accumulates across many components without a clear elbow, indicating that truncation-based dimensionality reduction is lossy.

5 Dataset Analysis

  • The dataset contains 200,000 channel realizations across 20 deployment variants, represented by 250,000-dimensional feature vectors.The analysis uses eigenvalue spectra to characterize variance concentration across principal components.Source: §5 · p. 6

5.1 Feature Variance Distribution via PCA

  • 35 principal components capture approximately 50% of variance, while 65, 86, and 93 components are required for 75%, 90%, and 95%, respectively.The cumulative explained-variance curve rises smoothly without a clear elbow.Source: §5.1 · pp. 6–7
  • The gradual eigenvalue decay indicates no spectral gap, so truncation-based dimensionality reduction is inherently lossy.The authors suggest models that preserve high-dimensional structure or learn task-aware compression.Source: §5.1 · p. 7

6 Evaluation Experiments

The evaluation benchmarks diverse models on CSI-to-CQI classification and tests generalization across simulated shifts and hardware-measured channels. Fine-tuning the FNN performs best across conditions, while distribution shifts expose substantial generalization challenges.

6 Evaluation Experiments

  • The benchmark compares classical, neural, and sequence models, using 36 standardized statistical features to make high-dimensional CSI more tractable.The features are 12 descriptors extracted from each of three sub-channels; soft-label supervision uses KL-divergence loss.Source: §6 · p. 7
  • XGBoost has the highest validation accuracy and lowest MAE but takes significantly longer to train; FNN offers a more favorable accuracy–efficiency balance.The CNN trains fastest but has lower accuracy, while Transformer and Mamba incur higher parameter counts and GPU memory use.Source: §6 · pp. 7–8

6.1 Test Datasets

  • T1 varies proportional-fairness α and noise variance; T2 tests unseen Bernoulli-Gaussian impulsive noise; T3 adds temporally correlated AR(1) noise; T4 evaluates BRISC hardware-measured RIS-aided channels at 5 GHz.T4 represents a joint shift in frequency, channel statistics, and hardware-noise characteristics, and is the most severe sim-to-real evaluation.Source: §6.1 · p. 8
  • T4 has the largest DKL, W1, and MMD shift, exceeding the shift of any simulated test set by more than 2.6×.The reported difficulty ordering is T1 ≈ T2 ≈ T3 ≪ T4.Source: §6.1 · p. 8, Table 3 · p. 8
  • FNN+FT achieves the best performance across all conditions, with significant T4 gains using only 100 labeled samples.The authors attribute the gains to pretrained FNN representations that adapt to target-domain statistics.Source: §6.1 · p. 8, Table 4 · p. 8
  • Models degrade under shifts even within simulation, and fine-tuning improves performance across T2, T3, and the sim-to-real T4 condition.The authors connect this challenge to the dataset’s multi-scenario diversity, which exposes variability beyond narrow single-scenario settings.Source: §6.1 · pp. 8–9

7 Conclusion and Future Directions

The paper introduces a reproducible, multi-scenario 3GPP-compliant RIS dataset with oracle phase labels and a CSI-to-CQI classification task. Benchmarks show generalization difficulties under simulated and real-world shifts, while the dataset remains simulation-based and excludes hardware impairments and temporal mobility.

7 Conclusion and Future Directions

  • The dataset combines 3GPP TR 38.901 channel modeling, per-link decomposition, oracle phase labels, multi-scenario coverage, and reproducible splits.The CSI-to-CQI mapping frames RIS phase optimization as a scalable classification task.Source: §7 · p. 9
  • Models struggle to generalize across simulated and real-world shifts, and fine-tuning provides only partial mitigation.The dataset is simulation-based and does not capture hardware impairments or temporal mobility; future extensions are planned for hardware-in-the-loop measurements and time-evolving channels.Source: §7 · p. 9
  • The dataset supports reproducible RIS-learning research, while the authors identify possible misuse for surveillance or sensing and unequal access to advanced communications.The stated potential benefits include wireless coverage, energy efficiency, reliability, and applications such as smart cities and autonomous systems.Source: §7 · p. 9

NeurIPS Paper Checklist

The checklist reports that the paper’s claims match its contributions and that the information needed to reproduce its main results is provided. It also records affirmative responses on experimental details, statistical significance, limitations, ethics, and asset documentation.

NeurIPS Paper Checklist

  • The authors report that the paper’s abstract and introduction accurately reflect its contributions and scope.They also report that the information needed to reproduce the main experimental results is described and that the dataset, metadata, and code are openly accessible.Source: NeurIPS Paper Checklist · pp. 13–15
  • The authors state that the paper discusses dataset limitations and that the research complies with the NeurIPS Code of Ethics.They identify Section 4 as the location of the dataset limitations discussion and say the datasets were generated through simulation.Source: NeurIPS Paper Checklist · pp. 13, 16
  • The authors report that training and test details are specified and that appropriate measures of statistical significance are reported.They direct readers to Section 6 and supplementary material for experimental details and statistical-significance measures.Source: NeurIPS Paper Checklist · p. 15
  • The authors report that the conclusion discusses both potential positive and negative societal impacts..Source: NeurIPS Paper Checklist · p. 16
  • The authors report that datasheets and detailed documentation for new assets are provided in the supplementary material, while human-subjects and crowdsourcing questions are marked not applicable.The asset documentation is described as accompanying the newly introduced assets.Source: NeurIPS Paper Checklist · p. 18

Related papers

Loading 2609.39058v1…