PaperlayerView exact revision on arXiv

Source-linked AI summary

Uncertainty-Aware Federated Learning for Joint Communication and Computation in XR Systems

Robert-Jeron Reifert, Hayssam Dahrouj, Alaa Alameer Ahmad, Haris Gacanin, Aydin Sezgin

arXiv:2610.21969v1cs.ITcs.LGmath.IT

TL;DR

Joint uncertainty in communication and computation remains insufficiently addressed in AI-driven resource allocation, despite its relevance to XR delay in multi-BS networks. The paper develops centralized and federated DNN frameworks that inject both uncertainty types during training, and reports improved robust delay with FL close to centralized performance.

  • Problem

    AI-driven resource allocation has limited systematic treatment of joint channel and computing uncertainty, although both affect delay and share constrained resources.

    Source: §I · p. 1, A. Related Works · p. 2

  • Method

    Centralized and federated DNNs map estimated channels and computing requirements to resource allocations, with sampled channel and computing errors injected after DNN outputs during training.

    Source: §I · p. 2, B. Contribution · pp. 2–3

  • Results

    Joint uncertainty injection improves network-wide robust delay across uncertainty levels and densities, while FL approaches the centralized upper bound and outperforms non-ML approaches.

    Source: B. Contribution · p. 3

  • Takeaways & Limitations

    The reported results support the feasibility of an AI-driven framework for joint communication and computing resource management in cellular XR networks.

    Source: B. Contribution · p. 3

    Limited-capacity BS–central-processor backhaul can bottleneck FL training, and aggressive compression may slow convergence or degrade allocation quality.

    Source: §I · p. 15

Abstract

from arXiv · show

The convergence of communication and computation, empowered by artificial intelligence (AI) enhanced capabilities, emerges as a key enabler for realizing the promising functionalities of the sixth-generation of communication systems (6G). To this end, this paper investigates an AI-native, robust network, where multiple base stations (BSs) simultaneously serve several extended reality (XR) devices. The paper then presents an innovative distributed deep learning-based approach to simultaneously manage the transmit and computing powers, and computation allocation, amidst uncertainties in both channel and computing states information. Under such an interference network scenario, the paper proposes a robust approach for minimizing the worst-case delay across the served XR devices, subject to computation and power constraints. The paper leverages a deep neural network (DNN) that can be trained either centrally or in a federated manner to map estimated channels and computation requirements to optimized resource allocations per BS. During training, uncertainty samples, reflecting both channel and computing estimation errors, are injected after the DNN output to account for such joint uncertainties. Our results validate the enhanced robust delay performance and scalability of joint uncertainty injection as compared to uncertainty-agnostic methods and state-of-the-art benchmarks, with federated learning-based solutions achieving near-centralized performance while enhancing efficiency, privacy, and adaptability.

I. INTRODUCTION

The paper targets joint communication and computation uncertainty in multi-BS XR networks, where worst-case delay depends on coupled resource decisions. It proposes centralized and federated DNN approaches with joint uncertainty injection; the reported evaluations show robust-delay gains and FL performance close to a centralized upper bound.

I. INTRODUCTION

  • Joint channel and computing uncertainty remains underexplored in AI-driven optimization, despite the importance of coordinating communication and computation in dense 6G networks.Prior work often treats the two uncertainty sources separately or assumes perfect channel and computing information.Source: §I · p. 1, A. Related Works · p. 2

A. Related Works

  • Prior work addresses XR latency, individual communication or computation uncertainty, and learning-based resource management, but the coupled uncertainty problem in multi-BS interference settings remains open.The paper extends the single-BS conference version to a multi-BS setting.Source: A. Related Works · p. 2, B. Contribution · p. 2

B. Contribution

  • The framework minimizes worst-case delay across XR devices by jointly allocating transmit power, computing power, and computation, using DNNs trained centrally or through FL.Training injects sampled channel and computing errors directly after DNN outputs.Source: B. Contribution · pp. 2–3, §I · p. 2
  • Federated solutions approach the centralized upper bound while outperforming non-ML approaches and improving robust delay across uncertainty levels and network densities.The paper also reports computational efficiency and reduced data exchange.Source: B. Contribution · p. 3

C. Notation and Organization

  • The paper is organized from system modeling and centralized learning through decentralized FL, then evaluates the approaches and concludes.The notation section specifies vector, matrix, set, and operator conventions.Source: C. Notation and Organization · p. 3

II. SYSTEM MODEL AND PROBLEM FORMULATION

  • The network comprises multiple edge-computing BSs serving XR devices, with each device associated with exactly one BS and centralized or federated model arrangements.In the federated arrangement, BSs train local models while the central processor maintains a global model.Source: §II · p. 3

A. XR Device Operation

  • Each device’s modeled delay has communication and computation components, while XR processing focuses on rendering and encoding at the BS.Control signaling and cache access delays are treated as negligible or omitted relative to the dominant delays.Source: A. XR Device Operation · pp. 3–4

B. Channel and Communication Delay Models

The communication model uses estimated aggregate channels while accounting for channel-estimation error and interference from devices served by the same or other base stations. The framework assumes fixed device association, with association and handover treated as slower-timescale processes.

B. Channel and Communication Delay Models

  • Each base station has estimated channels, while the true channels include estimation error whose distribution is not known in practice.The framework operates on sampled uncertainty realizations; simulations use a Gaussian assumption and also evaluate non-Gaussian models.Source: B. Channel and Communication Delay… · p. 4
  • Device association is fixed in the model because association and handover decisions occur more slowly than transmit-power and computation-resource allocation.Any valid association can be incorporated without fundamentally changing the framework.Source: B. Channel and Communication Delay… · p. 4
  • Communication slots are typically sub-ms, while channel statistics evolve over milliseconds to seconds and training or federated aggregation occurs offline or in the background.Only inference is latency-critical; dynamic association and variable-size architectures are left for future work.Source: B. Channel and Communication Delay… · pp. 4–5
  • Intra-base-station interference can be managed through power allocation, whereas inter-base-station interference is beyond the serving base station’s control.A device receives its desired signal, interference from other signals, and noise.Source: B. Channel and Communication Delay… · p. 5

C. Computing Framework and Computation Delay

The computing model allocates base-station computation cycles to device tasks whose required intensity may be uncertain. Total device delay combines computation and communication, while base-station power couples the resources used for both.

C. Computing Framework and Computation Delay

  • Computation intensity is the required cycles per input-data bit, and task requirements may be uncertain because computing environments vary with load and concurrent processes.The framework trains on sampled uncertainty realizations; Gaussian errors are used in simulations, although the algorithm is agnostic to the error distribution.Source: C. Computing Framework and Computation… · p. 5
  • Each base station’s computation allocation is bounded by its maximum capacity, and computation power is linked to allocated cycles through the CPU model.Allocated computation power is the total power used by that base station for all tasks.Source: C. Computing Framework and Computation… · p. 5

D. Joint Communication and Computation Considerations

  • A device’s total delay combines computation and communication components, and the worst-case delay is the maximum device delay to provide fair service.Computation uncertainty enters through ω_k, while channel conditions affect the achievable rate r_k.Source: D. Joint Communication and Computation… · p. 5
  • Communication and computation power consumptions share each base station’s power budget, making their allocation interdependent.The model abstracts channel fluctuations and workload variations as uncertain states that jointly affect end-to-end XR delay.Source: D. Joint Communication and Computation… · p. 5

E. Robust Delay Utility

The paper models robust delay using uncertain channel and computing states, then formulates worst-device-delay minimization under resource constraints. Its centralized learning approach maps estimated states to allocations and uses uncertainty-aware resource-allocation processing.

E. Robust Delay Utility

  • The DNN takes estimated channels and computing requirements as inputs and injects sampled channel and computing errors after its output to optimize a robust delay objective.The target is the γ-th quantile of the worst-case delay distribution.Source: E. Robust Delay Utility · p. 6

F. Problem Formulation

  • The optimization minimizes worst-case delay across devices subject to capacity, power, and quantile-robustness constraints.Its difficulty includes coupled communication and computation variables and unknown delay distributions, while centralized solving requires global estimated channel and workload information.Source: F. Problem Formulation · p. 6

III. UNCERTAINTY-AWARE CENTRALIZED LEARNING

  • The centralized DNN maps estimated channels and computing requirements to transmit-power, computation-power, and computation-capacity allocations.Its input has 2LK + K neurons, and its output has 2K + B entries.Source: §III · p. 6

A. DNN-Driven Resource Allocation Strategy

  • The resource-allocation strategy uses softmax outputs to allocate base-station transmit and computation powers, then derives computation cycles subject to power and capacity limits.The power allocation uses the full base-station power budget, and computation allocation uses the smaller of the power-limited and capacity-limited cycles.Source: A. DNN-Driven Resource Allocation… · pp. 6–7

B. Joint Uncertainty Injection Scheme

Joint uncertainty injection evaluates resource allocations under sampled channel and computing errors after the DNN output. The resulting delay samples estimate the robust quantile used to train the network.

B. Joint Uncertainty Injection Scheme

  • Uncertainty injection estimates the robust objective by sampling errors, computing multiple objective values after the DNN output, and using them to derive gradients for backpropagation.The paper’s joint scheme samples both channel and computing uncertainties because the true worst-case delay distribution depends on both unknown error distributions.Source: B. Joint Uncertainty Injection Scheme · p. 7
  • The sampled realizations produce N worst-case delays, which are sorted to estimate the γ-th quantile using the (γ · N)-th highest value.This estimate represents the empirical robust-delay objective.Source: B. Joint Uncertainty Injection Scheme · p. 7
  • Uncertainty sampling is used during training only; once trained, the DNN allocates resources from estimated inputs without further uncertainty sampling.The paper distinguishes training-related components from real-time operation and testing in Fig. 2.Source: B. Joint Uncertainty Injection Scheme · p. 7

C. Unsupervised Learning via Backpropagation

The centralized DNN is trained without labels by differentiating a high-quantile worst-case delay objective. Minibatch uncertainty realizations and validation guide updates and model selection before the paper turns to federated learning.

C. Unsupervised Learning via Backpropagation

  • Gradients of the γ-th highest computed worst-case delay update the DNN, which implicitly learns the channel and computing uncertainty distributions.The gradient is computed using PyTorch.Source: C. Unsupervised Learning via… · p. 8
  • Each epoch processes minibatches with newly sampled realizations and saves parameters when validation yields the lowest observed robust delay.Training ends at a convergence criterion or the maximum epoch count.Source: C. Unsupervised Learning via… · p. 8
  • The paper moves from the centralized solution to a practical distributed network architecture based on federated learning.Source: C. Unsupervised Learning via… · p. 8

IV. UNCERTAINTY-AWARE FEDERATED LEARNING

The section develops a federated architecture for robust resource allocation, in which BSs train and allocate locally while coordinating model parameters and inter-BS interference. It also specifies how channel knowledge and interference information are handled.

IV. UNCERTAINTY-AWARE FEDERATED LEARNING

  • The proposed federated approach addresses joint channel and computing uncertainty in distributed optimization, with the parameter server coordinating models rather than accessing global network state.Source: §IV · p. 8

A. Distributed Network Architecture

  • BSs perform training, inference, and resource allocation locally, while the CP aggregates and transmits models and facilitates network-wide training and interference-information exchange.This creates a hybrid RAN with distributed decisions and coordination.Source: A. Distributed Network Architecture · p. 8

B. Channel Knowledge and Inter-BS Interference

  • Because local BSs cannot directly manage inter-BS interference, the framework exchanges mean interference values rather than per-estimation, per-sample values.The mean is shared at each training epoch and during inference is updated only when channel statistics change significantly.Source: B. Channel Knowledge and Inter-BS… · p. 9

C. Decentralized DNN Architecture

Each BS uses a local DNN to allocate resources from its local estimates and mean inter-BS interference, with channel and computing uncertainty samples incorporated into robust-delay evaluation. Federated training aggregates local models and interference information at the CP.

C. Decentralized DNN Architecture

  • Each BS’s DNN maps local channel and computing estimates plus mean inter-BS interference to resource allocations for its associated devices.The distributed input adds K_b neurons for mean interference, and the DNN output has 2K_b + 1 neurons.Source: C. Decentralized DNN Architecture · p. 9
  • After mapping DNN outputs to transmit power and computation allocation, each BS injects channel and computing errors to evaluate the empirical γ-quantile worst-case delay.Source: C. Decentralized DNN Architecture · p. 9

D. Federated Deep Learning for Robust Delay Optimization

  • BSs update local models, upload parameters and interference terms, and receive the CP’s averaged global model and aggregated interference for the next round.FedAvg may not converge under highly heterogeneous data; formal convergence analysis is out of scope, though simulations show convergence within a practical number of epochs under uncertainty injection.Source: D. Federated Deep Learning for Robust… · p. 9

E. Data Exchange

The section contrasts centralized and federated learning overhead, then describes the simulation setup and benchmark schemes used to evaluate the proposed methods. The simulations use offline training based on channel and computing statistics.

E. Data Exchange

  • Centralized training sends estimated channels, computing requirements, and training uncertainty samples to the CP, while the CP broadcasts resource allocations.Its total exchange formula accounts for epochs, minibatches, validation size, BSs, antennas, and devices.Source: E. Data Exchange · p. 10
  • Federated exchange consists of broadcasting the global model and average per-device interference, then uploading local DNN parameters and locally computed interference values.The DNN parameter count depends on its hidden layers and input and output sizes.Source: E. Data Exchange · p. 10

V. SIMULATION RESULTS AND DISCUSSION

  • The simulations use three BSs with six antennas each and five XR devices per BS by default, Rayleigh fading, minimum mean-square channel estimation, and RZF beamforming.DNN training is offline and based on channel and computing statistics that vary relatively slowly compared with training time.Source: §V · p. 10

1) Benchmark Schemes:

  • Benchmarks include robust and non-robust centralized and federated learning, optimization-based FP and Uniform FP, and SAA/WMMSE schemes.The paper also compares joint uncertainty injection with channel-only and computing-only injection, FedProx, and local learning under heterogeneous uncertainty.Source: 1) Benchmark Schemes · pp. 10–11

2) Practical Channel Estimation Error:

Across channel and computing uncertainty levels, uncertainty-aware FL performs close to centralized learning and outperforms non-robust methods; joint uncertainty injection also remains effective as network density grows. The simulations examine practical estimation-error settings, uncertainty-specific comparisons, and scalability.

2) Practical Channel Estimation Error:

  • With γ = 0.2, β = 0.4, and N = L, the channel estimation error variance spans the range in Fig. 4 under moderate changes in pilot length, transmit power, and noise level.The paper describes these values of σ2 h as realistic operating regimes rather than isolated parameter choices.Source: 2) Practical Channel Estimation Error · p. 11, Figure 4 · p. 11

3) Practical Computing Estimation Error:

  • At σ2 ω = 3200, the simulations use a representative intermediate computing-uncertainty value and draw overall conclusions from trends across uncertainty levels.Workload variability ω reflects dynamic XR processing requirements such as field-of-view rendering and encoding.Source: 3) Practical Computing Estimation Error · p. 11, Figure 5 · p. 12

A. Impact of Communication and Computation Uncertainty

  • At σ2 h = 0.008, Robust FL and Robust Cen reduce robust delay by 18 ms and 12 ms, respectively, versus their non-robust counterparts; Robust FL remains close to centralized performance.Across tested channel uncertainties, ML-based schemes outperform baselines, and Robust Cen consistently achieves the lowest robust delay.Source: A. Impact of Communication and… · p. 11, Figure 4 · p. 11
  • As computing uncertainty rises, non-robust FL loses its advantage over Uniform FP at σ2 ω = 6400 and over SAA/WMMSE at σ2 ω = 12800; Robust FL maintains a 14 ms gap from Robust Cen.At fixed σ2 h = 0.02, computation uncertainty has a more moderate impact on Robust FL than communication uncertainty.Source: A. Impact of Communication and… · pp. 11–12, Figure 5 · p. 12

B. Advantages of Joint Uncertainty Injection

  • Across uncertainty levels, joint-injection Robust FL achieves the lowest robust delay; the better single-uncertainty scheme depends on which uncertainty dominates.Comm UI FL performs comparably to Robust FL when communication uncertainty dominates, while Comp UI FL and Robust FL are similar at σ2 ω = 800.Source: Figure 6 · p. 12, B. Advantages of Joint Uncertainty… · p. 12

C. Scalability and Density

  • At K = 20, Robust FL reduces robust delay by 69 ms and Robust Cen by 88 ms versus Uniform, while the gap between Robust FL and Robust Cen grows from 6 ms at K = 15 to 19 ms at K = 20.At K = 25, the centralized solution cannot determine a reliable solution within a reasonable time; Robust FL’s advantage over Non-robust FL reaches 13 ms.Source: Figure 7 · p. 12, C. Scalability and Density · pp. 12–13
  • In dense networks, robust delay rises from 32 ms to 113 ms for Robust FL and from 44 ms to 145 ms for Non-robust FL; DNN inference remains around 4–5 ms while optimization takes several seconds.The Robust FL versus Non-robust FL gap grows from 12 ms in sparse networks to 32 ms in dense networks; the DNN inference time scales sublinearly with device count.Source: Figure 8 · p. 13, C. Scalability and Density · p. 13, Table II · p. 13

D. Heterogeneous Uncertainties

The section examines how cross-BS uncertainty heterogeneity and reliability requirements affect robust delay, then tests whether the framework generalizes to non-Gaussian errors. Robust performance depends on the uncertainty regime, while Gaussian-trained models show only minor degradation under Laplacian uncertainty.

D. Heterogeneous Uncertainties

  • Increasing the spread of BS-specific error variances creates more diverse local uncertainty conditions while preserving the same average uncertainty.The heterogeneity analysis retains Gaussian error realizations but assigns their variances by BS.Source: D. Heterogeneous Uncertainties · p. 13
  • At high heterogeneity, Robust Local beats Robust FL and Robust FedProx by about 2 ms, while Robust Cen retains the lowest delay.With similar uncertainty levels across BSs, Robust FL outperforms Robust Local and Robust FedProx; fully local adaptation becomes advantageous under high heterogeneity.Source: D. Heterogeneous Uncertainties · pp. 13–14

E. Impact of Reliability Parameter γ

  • Lower γ imposes stricter reliability and higher delay, widening the robust–non-robust gap; Robust Cen is less sensitive to γ than Robust FL.The comparison treats γ as a tunable trade-off between delay and the probability of deadline violations.Source: Figure 10 · p. 14, E. Impact of Reliability Parameter γ · p. 14

F. Impact of Non-Gaussian Uncertainty Assumptions

  • Increasing uncertainty raises robust delay, and Laplacian errors are evaluated at variance-matched levels to the Gaussian case.The Laplacian distribution has heavier tails and represents rare but significant deviations.Source: F. Impact of Non-Gaussian Uncertainty… · p. 14
  • A Gaussian-trained model incurs only 0.5 ms of degradation under Laplacian uncertainty, supporting cross-distribution generalization.The Laplacian model itself yields slightly higher delay than Gaussian uncertainty because of its heavier tails.Source: F. Impact of Non-Gaussian Uncertainty… · p. 14

G. Convergence Behavior

The paper reports near-centralized robust performance alongside lower data exchange and millisecond-scale edge inference for federated solutions. Its evaluations also examine convergence, uncertainty-sample and DNN-size trade-offs, and a backhaul limitation that can affect FL training.

G. Convergence Behavior

  • Robust Cen and Robust FL converge within 500 epochs, but FL’s lower per-BS final objectives do not necessarily mean a lower global robust worst-case delay.Robust FL shows marked objective drops at epochs 10–40 and 200–400.Source: G. Convergence Behavior · p. 14

H. Impact of Uncertainty Samples

  • Increasing uncertainty samples from N = 500 to N = 1000 yields lower delay across epochs but approximately doubles training time.The performance gap is relatively small, so moderate sample sizes can approximate the robust objective sufficiently accurately.Source: Figure 13 · p. 15, H. Impact of Uncertainty Samples · p. 15

I. Training and Inference Data Exchange

  • At B = 8 and Kb = 6, centralized inference requires 19 kB versus 1.5 kB for the federated approach.For training, Algorithm 2 requires 800–4000 MB, while the centralized approach requires up to 495 TB at B = 8 and Kb = 6. FL exchanges model parameters and inter-BS interference levels, keeping excessive estimates local.Source: Figure 15 · p. 15, §I · pp. 15–16, Figure 14 · p. 15
  • Limited BS-CP backhaul capacity may bottleneck FL training; aggressive compression can introduce update errors, slow convergence, or increase robust delay.Communication-efficient techniques reduce transmitted bits per aggregation round, but the paper leaves this communication–robustness trade-off for future work.Source: §I · p. 15

J. Ablation Study

  • Larger hidden layers generally improve delay until gains saturate; deeper FL networks can hurt when training time is insufficient.Robust Cen benefits more strongly from larger DNNs, while the effect of FL depth becomes less significant with longer training.Source: Figure 16 · p. 16, J. Ablation Study · p. 16

K. Edge Deployment Feasibility

  • On an NVIDIA Jetson Orin Nano, Robust Cen inference averages 9.578 ms and Robust FL averages 4.587 ms at batch size one.The authors attribute FL’s lower latency to reduced input dimension and a smaller local DNN; training is offline or background, while inference is latency-critical.Source: K. Edge Deployment Feasibility · p. 16

VI. CONCLUSION

  • FL-based solutions achieve near-centralized performance while improving scalability, efficiency, and robustness in joint communication and computation resource management.The conclusion identifies other objectives, online or hybrid learning, non-terrestrial network models, and stale interference information as future directions.Source: §VI · p. 16

APPENDIX

The appendix defines the benchmarking schemes used in the simulations, including non-robust learning, numerical optimization, robust SAA/WMMSE, FedProx, and fully local learning.

APPENDIX

  • The appendix benchmarks data-driven and numerical-optimization methods under specific optimization formulations.The listed schemes are Non-robust Cen, Non-robust FL, Non-robust FP, Uniform FP, SAA/WMMSE, FedProx, and Local.Source: APPENDIX · p. 16

A. Non-Robust Centralized Learning

  • Non-robust centralized learning omits channel and computing uncertainty and the γ-quantile constraint, minimizing nominal worst-case delay instead.Its benchmark applies Algorithm 1 without uncertainty injection.Source: A. Non-Robust Centralized Learning · pp. 16–17

B. Non-Robust Federated Learning

  • Non-robust FL omits channel and computing uncertainty and minimizes nominal worst-case delay among devices associated with each BS.Its benchmark applies Algorithm 2 without uncertainty injection while retaining the stated computation, power, and inter-BS interference constraints.Source: B. Non-Robust Federated Learning · p. 17

C. Non-Robust Numerical Optimization

D. Uniform Numerical Optimization

  • Uniform FP adds equal-transmit-power and equal-computation-allocation constraints across devices, optimizing only the power split between transmission and computation.Source: D. Uniform Numerical Optimization · p. 17

E. Robust Numerical Optimization

  • SAA/WMMSE samples channel and computing errors, replaces the γ-quantile constraint with an SAA constraint, and solves the resulting problem using SAA and WMMSE.Source: E. Robust Numerical Optimization · p. 17

F. Robust FedProx-based Learning

  • FedProx adds a proximal penalty that discourages local-model drift toward the global model; inference uses the final aggregated global model.The implementation sets δ = 10^-3, between FedAvg-like behavior at small δ and more constrained local updates at large δ.Source: F. Robust FedProx-based Learning · p. 17

G. Fully Local Learning

  • Fully local learning removes global aggregation and maintains separate BS models, but still exchanges inter-BS interference information.Source: G. Fully Local Learning · p. 17

Related papers

Loading 2610.21969v1…