PaperlayerView exact revision on arXiv

Source-linked AI summary

FedQoS: Federated QoS-Risk Learning for Heterogeneous Indoor-Outdoor Access Selection

Nguyen Van Thieu, Ti Ti Nguyen, Ons Aouedi, Zerihun Huruy, Vu Nguyen Ha, Symeon Chatzinotas

arXiv:2608.25496v1cs.LG

TL;DR

Heterogeneous indoor-outdoor access selection cannot reliably use instantaneous radio measurements alone because future QoS also depends on mobility, blockage, load, and resource competition. FedQoS learns candidate-link QoS-failure risk from distributed access-node logs through federated aggregation, then uses those risks for access decisions. Simulations show lower QoS-failure rates than signal-based and historical-QoS heuristics, with near-centralized prediction performance and gains under mild non-IID data.

  • Problem

    Instantaneous radio strength alone may select overloaded or otherwise unreliable links in dynamic heterogeneous networks.

  • Method

    FedQoS trains a federated supervised model that predicts future QoS-failure probabilities from distributed access-node observations and uses them for access selection.

  • Results

    QoS-failure rate is reduced by approximately 13.9% relative to FedAvg and 96.6% relative to historical QoS in the mild non-IID setting.

  • Takeaways & Limitations

    FedQoS supports reliability-aware access decisions without centralizing user-level network data and remains effective under mild and severe non-IID conditions.

Abstract

from arXiv · show

Reliable access selection in dynamic and heterogeneous indoor-outdoor environments is challenging because instantaneous radio measurements alone cannot capture future QoS degradation caused by mobility, blockage, traffic load, and resource competition. This paper proposes FedQoS, a federated QoS-risk learning framework for predicting the future reliability of candidate access links and supporting access-node selection without centralizing user-level network data. In FedQoS, each access node locally learns from its observed network logs, including radio, traffic, load, and service-context features, while a global QoS-risk predictor is trained through federated aggregation. The learned model estimates the probability of QoS failure for each candidate link, and the controller uses these risk scores to select reliable access nodes under dynamic network conditions. To evaluate the framework, we construct physics-based synthetic indoor-outdoor wireless datasets using the Sionna framework, covering normal traffic, mobility, event-driven congestion, and non-IID client observations. Simulation results show that learning-based access selection substantially reduces the QoS-failure rate compared with signal-based and historical-QoS heuristic methods. FedQoS achieves near-centralized predictive performance and provides clear reliability gains under mild non-IID data while remaining competitive under the more challenging severe non-IID condition. These results demonstrate the potential of federated QoS-risk learning for reliable, data-local access selection in dynamic wireless environments.

I. INTRODUCTION

FedQoS addresses access selection in heterogeneous indoor-outdoor networks by predicting future QoS-failure risk rather than relying on instantaneous signal strength alone. It uses federated learning to train on distributed access-node observations while supporting reliability-oriented decisions.

  • Indoor-outdoor access selection affects throughput, load balancing, delay, handover behavior, and quality of service as users move across heterogeneous environments.
  • Strongest-signal association can fail under heavy load, while weaker-signal alternatives may provide higher throughput or lower latency.
  • FedQoS predicts the probability that each candidate AP, BS, or aerial node will violate a user’s QoS requirement within a future prediction window.
  • FedQoS combines stabilized local training with QoS-aware aggregation to address non-IID observations across access nodes.
  • The controller converts predicted QoS-failure probabilities into access decisions while explicitly accounting for access-node load and handover cost.
  • The evaluation uses realistic dynamic wireless datasets and compares FedQoS with centralized, federated, and representative non-learning baselines.

A. Heterogeneous Indoor-Outdoor Access Setting

The system models a campus network with indoor APs, terrestrial BSs, and static aerial nodes serving mobile users. Each access node stores local observations, while the controller selects candidate nodes using contextual, link, load, service, and handover features.

  • The network contains indoor APs, terrestrial BSs, and static aerial access nodes serving varying users within a campus area.
  • A federated controller selects the most suitable next access node for each active user.
  • Candidate nodes are drawn from the current serving node, top-K logging-score candidates, and occasional probing candidates.
  • The strongest instantaneous signal is not necessarily associated with the lowest future QoS-failure risk under blockage, mobility, density, and load variation.
  • The feature vector includes user context, access-node attributes, link measurements, node load, service requirements, and handover indicators.
  • Access nodes retain logs for served or probed users, and the controller receives only model parameters or updates during federated training.

C. Future QoS-Failure Label

FedQoS labels candidate decisions by their future throughput and BLER over a prediction horizon. This target captures combined radio, contention, service-demand, and mobility effects rather than instantaneous SINR alone.

  • The future QoS outcome is measured over a prediction window using average throughput and BLER.
  • A candidate receives a failure label when its future performance violates the user’s minimum-rate or maximum-BLER requirement.
  • The label represents whether a candidate node is expected to violate or satisfy the user’s QoS requirement over the prediction horizon.
  • The target incorporates radio quality, resource contention, service demand, and mobility, with SINR serving only as an explanatory feature.
  • Links with similar SINR can still have different QoS outcomes because of load, scheduling, or handover conditions.

D. QoS-Risk Learning

FedQoS trains a probabilistic QoS-risk model from access-node-local datasets under naturally non-IID observations. Unlike centralized training, it exchanges model parameters while keeping local access logs at each node.

  • Access nodes observe naturally non-IID data because locations, structures, user densities, channel conditions, services, and mobility patterns differ.
  • The QoS-risk model estimates each candidate’s probability of failing the user’s requirement over the prediction horizon.
  • Each client evaluates a local empirical loss using binary cross-entropy for its stored observations.
  • FedQoS keeps each local dataset at its access node and exchanges only model parameters with an edge controller.

E. FedQoS Local Update

FedQoS stabilizes local training under heterogeneous client data and aggregates updates using both dataset size and observed QoS-failure risk.

  • Local update: Each participating client initializes its local model from the broadcast global model and trains locally for E epochs before uploading the update.A proximal objective penalizes deviation from the current global model, stabilizing learning under heterogeneous local data.
  • QoS-aware aggregation: FedQoS balances sample size and local QoS-failure exposure when weighting client updates.The aggregation gives greater influence to clients observing QoS-critical conditions while retaining information from larger datasets.
  • Training procedure: The training procedure repeats client selection, global-model broadcast, parallel local updates, aggregation, and final deployment of the global QoS-risk model.Algorithm 1 deploys the resulting model to support later access decisions.
  • QoS-aware aggregation: q controls dependence on client dataset size, while λ controls emphasis on clients observing more QoS failures.q = 1 and λ = 0 recover standard sample-size weighting; smaller q reduces large-client domination and positive λ increases high-risk-client influence.

G. Handover-Aware Access Selection

FedQoS selects among candidate access nodes using predicted QoS-failure risk while accounting for handover cost and stability constraints.

  • Risk-based selection: The deployed model estimates each candidate access node’s probability of QoS failure.These probabilities feed the access-selection score.
  • Risk-based selection: The controller minimizes a handover-aware risk score that includes a handover cost coefficient.Selecting a candidate that triggers handover incurs the modeled cost.
  • Handover stability: A handover occurs only when the candidate improves the serving score by at least ξ_h and the user satisfies the minimum dwell-time requirement T_dwell.Otherwise, the user remains connected to the previous serving node, limiting ping-pong handovers.

III. EXPERIMENTAL SETUP

The evaluation uses Sionna RT and Sionna SYS to generate physics-based network logs for heterogeneous indoor-outdoor environments.

  • Experimental setup: Sionna RT and Sionna SYS generate physics-based network logs for heterogeneous indoor-outdoor environments.Logs are partitioned by observing access node, with local training and model-parameter exchange through the edge controller.
  • Experimental setup: The resulting models are evaluated both as QoS-failure predictors and as access-selection policies.This evaluates predictive performance and its use in access decisions.

A. Sionna-Based Dataset Generation

The dataset generator models a campus-scale heterogeneous wireless environment and produces locally partitioned samples across varied mobility, traffic, blockage, and non-IID conditions.

  • Scenario construction: The simulated campus contains multi-floor buildings, indoor APs, an outdoor BS, quasi-static UAV access nodes, walkways, and an event area.Materials, furniture, vegetation, vehicles, and other obstacles create blockage and non-line-of-sight conditions.
  • Scenario construction: Sionna RT computes geometry-aware propagation, while Sionna SYS generates received power, SINR, BLER, throughput, load, and resource-utilization quantities.Users follow short indoor-outdoor trajectories across occupancy, transition, event, stress, and access-node-degradation scenarios.
  • Sample construction: Each log entry is a user-candidate tuple (u, a, t) associated with a serving node and top-K feasible measurement candidates.Future throughput and BLER are probed over the prediction window to derive QoS-failure labels.
  • Sample construction: Samples remain at their observing access nodes and are split locally into 70%/15%/15% training, validation, and test sets.Unselected links require probing, dual connectivity, or controlled exploration to obtain outcomes.
  • Non-IID conditions: The evaluation compares mild and severe non-IID conditions using different service rates, event traffic, outages, room skew, churn, probing, and candidate-set reductions.The severe condition increases traffic stress and client heterogeneity relative to the mild condition.

B. Baselines and Metrics

The evaluation compares federated and non-learning access-selection baselines using prediction metrics and system-level reliability, throughput, and BLER metrics.

  • Learning baselines: Centralized, Local-only, FedAvg, and FedProx provide learning baselines with different data-sharing and aggregation strategies.Centralized pools all training data; Local-only trains independently; FedAvg uses sample-size weighting; FedProx adds proximal regularization.
  • Non-learning baselines: Current serving, Strongest SINR, Load-aware RSRP, AP-first, Historical QoS, and Oracle represent practical or reference access-selection policies.Oracle uses realized future QoS outcomes and is included only as a non-implementable reference.
  • Prediction metrics: Balanced accuracy, recall, and F1-score measure QoS-risk prediction performance, with recall targeting imminent failure detection.Balanced accuracy accounts for class imbalance, while F1 summarizes failure detection and false alarms.
  • Access-selection metrics: QoS-failure rate is the primary access-selection metric, complemented by mean throughput and mean BLER.Lower failure rate and BLER indicate greater reliability, whereas higher throughput indicates better data-rate performance.

IV. RESULTS AND DISCUSSION

FedQoS improves failure-oriented prediction and downstream access reliability across mild and severe non-IID settings. Its strongest reliability gains occur under mild heterogeneity, while severe heterogeneity narrows differences among federated methods and increases overall difficulty.

  • Prediction and access-selection performance: Learning-based methods outperform conventional heuristics in both non-IID settings, while centralized learning generally provides the strongest prediction reference.The comparison uses ten random seeds, with results reported as mean ± standard deviation.
  • Prediction and access-selection performance: FedQoS achieves the highest balanced accuracy and recall among federated methods under mild non-IID data, with F1 nearly matching FedAvg and FedProx.FedAvg and FedProx have slightly higher displayed F1-scores, but differences are small relative to seed variability.
  • Prediction and access-selection performance: FedQoS achieves the highest mean balanced accuracy, recall, and F1 among federated methods under severe non-IID data.The advantage remains modest numerically but persists under stronger topology-induced heterogeneity, increased user density, and partial AP unavailability.
  • Prediction and access-selection performance: FedQoS reduces mild-setting QoS-failure rate by approximately 13.9% versus FedAvg, 6.1% versus FedProx, 43.6% versus Local-only, and 96.6% versus Historical QoS.It also achieves the lowest mean QoS-failure rate and BLER among federated methods.
  • Prediction and access-selection performance: FedQoS achieves reliability improvements with average throughput approximately 0.49% below FedAvg and 0.27% below FedProx in the mild setting.The policy favors candidates with lower predicted future QoS-violation risk rather than only higher instantaneous data rate.
  • Prediction and access-selection performance: FedQoS maintains similar throughput and BLER to federated baselines under severe non-IID data, with throughput less than 0.2% below FedAvg and BLER differing by 10^-4.FedQoS ties FedAvg at displayed precision for the lowest federated failure rate.
  • Prediction and access-selection performance: Overall, FedQoS provides the clearest reliability gains under mild non-IID data and remains robust and competitive under severe non-IID conditions.The oracle remains an unattainable performance reference because it uses future QoS outcomes.

B. Sensitivity to q and λ

The sensitivity analysis shows that FedQoS remains stable across a broad grid of client-size exponent q and QoS-risk emphasis λ values. Client-size weighting has greater influence than risk emphasis, while λ mainly fine-tunes performance after reasonable q selection.

  • B. Sensitivity to q and λ: The analysis varies the policy QoS-failure rate over the client-size exponent q and QoS-risk emphasis parameter λ.The QoS-failure rate is the primary system-level objective reported for the grid.
  • B. Sensitivity to q and λ: In mild non-IID data, the QoS-failure rate ranges from approximately 0.00302 to 0.00334, with the minimum near (q, λ) = (0.75, 1.5).The main setting (0.75, 0.5) also lies within the low-failure region.
  • B. Sensitivity to q and λ: In severe non-IID data, the QoS-failure rate remains within approximately 0.13553–0.13594 across the complete grid.The minimum occurs near (q, λ) = (0.5, 4), but neighboring combinations are nearly indistinguishable.
  • B. Sensitivity to q and λ: Across both settings, q influences performance more visibly than λ, especially when q is close to zero.Once reasonable client-size weighting is selected, λ mainly fine-tunes the policy.
  • B. Sensitivity to q and λ: FedQoS remains stable over a broad parameter range, so its reported policy gains do not depend on narrowly tuned q and λ values.The observed stability is consistent across the mild and severe non-IID settings.
Loading 2608.25496v1…