PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning for Ultra-Reliable Low-Latency V2V Communications

Sumudu Samarakoon, Mehdi Bennis, Walid Saad, Merouane Debbah

arXiv:1805.09253v1cs.NIcs.LGstat.ML

TL;DR

The paper addresses distributed V2V power and resource allocation for URLLC while controlling probabilistic queue-length reliability and avoiding centralized exchange of queue information. It combines EVT, FL, and Lyapunov optimization, achieving similar tail-estimation reliability with lower exchange overhead and reducing extreme queue events without additional transmit power.

  • Problem

    The paper seeks to minimize network-wide power consumption while ensuring queue stability and keeping threshold-exceeding outages below a specified probability in mission-critical V2V networks.

  • Method

    The approach models queue-length tails with EVT, estimates generalized Pareto parameters through distributed FL without sharing actual queue samples, and uses Lyapunov optimization for joint power and resource allocation.

  • Results

    The proposed method reduces threshold-exceeding VUEs by up to 60% versus Baseline 2 without additional transmit power, while reducing exchanged data versus centralized estimation by up to 79%.

  • Takeaways & Limitations

    FL provides a distributed way to estimate queue-tail reliability with performance similar to centralized estimation while reducing communication overhead and extreme queue events.

Abstract

from arXiv · show

In this paper, a novel joint transmit power and resource allocation approach for enabling ultra-reliable low-latency communication (URLLC) in vehicular networks is proposed. The objective is to minimize the network-wide power consumption of vehicular users (VUEs) while ensuring high reliability in terms of probabilistic queuing delays. In particular, a reliability measure is defined to characterize extreme events (i.e., when vehicles' queue lengths exceed a predefined threshold with non-negligible probability) using extreme value theory (EVT). Leveraging principles from federated learning (FL), the distribution of these extreme events corresponding to the tail distribution of queues is estimated by VUEs in a decentralized manner. Finally, Lyapunov optimization is used to find the joint transmit power and resource allocation policies for each VUE in a distributed manner. The proposed solution is validated via extensive simulations using a Manhattan mobility model. It is shown that FL enables the proposed distributed method to estimate the tail distribution of queues with an accuracy that is very close to a centralized solution with up to 79\% reductions in the amount of data that need to be exchanged. Furthermore, the proposed method yields up to 60\% reductions of VUEs with large queue lengths, without an additional power consumption, compared to an average queue-based baseline. Compared to systems with fixed power consumption and focusing on queue stability while minimizing average power consumption, the reduction in extreme events of the proposed method is about two orders of magnitude.

I. INTRODUCTION

The paper targets URLLC in V2V networks by controlling rare large-queue events that conventional probabilistic queue constraints do not adequately address. It proposes decentralized tail estimation and joint power-resource allocation to reduce signaling overhead while minimizing network-wide power.

  • Motivation: URLLC is required because vehicular applications depend on extremely low communication errors and delays, including a target end-to-end latency of 1 ms.Queuing and over-the-air latency are coupled, making efficient radio resource management necessary.
  • Motivation: Probabilistic queue-length constraints improve reliability but fail to control rare events involving large queues and potentially unacceptable latencies.Extreme value theory models these tail events using location, scale, and shape parameters.
  • Problem formulation: The framework formulates distributed joint transmit-power and resource allocation as network-wide power minimization subject to low-latency and high-reliability queue constraints.Reliability is characterized through excess queues above a predefined threshold.
  • Proposed approach: Federated learning enables VUEs to estimate queue-tail distributions without sharing individual queue-length samples, using local GPD models aggregated by an RSU.Vehicles learn locally on a short time scale, while global model averaging occurs on a longer time scale.
  • System model: The system partitions VUE pairs into zones and allocates resource blocks orthogonally across zones to reduce interference among nearby pairs.Each VUE may use only the RBs assigned to its corresponding zone.
  • Problem formulation: A centralized solution would require exchanging CSI and QSI across the network, creating unacceptable signaling overhead and motivating distributed analytical tools.The optimization is also challenging because decisions depend on future network states and the excess-queue distribution is initially unavailable.

III. PROPOSED DISTRIBUTED FRAMEWORK USING EVT AND LYAPUNOV OPTIMIZATION

The distributed solution decouples the network-wide power-minimization problem across VUE pairs using excess-queue statistics modeled over the network.

  • The network-wide power-minimization problem is decoupled over VUE pairs.The decoupling relies on the statistics of excess queues over the network.
  • Excess-queue statistics are used to decouple the objective function.
  • Excess-queue statistics are also used to decouple the reliability constraint.

A. Modeling Excess Queues Using Extreme Value Theory

The paper models queue excesses above a high threshold with a generalized Pareto distribution and uses its parameters to impose local reliability constraints.

  • A. Modeling Excess Queues Using Extreme Value Theory: Excess queues above a sufficiently high threshold are modeled with a generalized Pareto distribution.This modeling uses EVT under an i.i.d. queue assumption as the threshold approaches the upper endpoint.
  • A. Modeling Excess Queues Using Extreme Value Theory: The generalized Pareto distribution uses scale and shape parameters, collected as d = [σ, ξ].The scale parameter satisfies σ > 0, while the shape parameter is ξ.
  • A. Modeling Excess Queues Using Extreme Value Theory: The expected excess queue is bounded and equals σ/(1−ξ) only when ξ < 1.
  • A. Modeling Excess Queues Using Extreme Value Theory: The reliability constraint is rewritten using the time-average indicator of queues exceeding q0 and the bound σ/(1−ξ).
  • A. Modeling Excess Queues Using Extreme Value Theory: Each VUE pair estimates ξ and σ locally without sharing queue-state information, enabling local enforcement of the constraint.

B. Lyapunov Optimization

Lyapunov optimization converts the modified stochastic problem into drift-plus-penalty control, yielding distributed per-VUE power and resource decisions.

  • B. Lyapunov Optimization: Lyapunov optimization recasts the time-average reliability constraint using virtual queues.The constraint becomes E[1_{q_u}] ≤ ϵ for each VUE.
  • B. Lyapunov Optimization: The combined queue state is Ξ_u(t) = [q_u(t), Υ_u(t), A_u(t)].
  • B. Lyapunov Optimization: The quadratic Lyapunov function is L(Ξ(t)) = 1/2 Ξ†(t)Ξ(t), with one-slot drift defined by L(Ξ(t + 1)) − L(Ξ(t)).
  • B. Lyapunov Optimization: A penalty term V E[P_u^†p_u | Ξ(t)] is added to the expected drift to form the drift-plus-penalty objective.
  • B. Lyapunov Optimization: The drift bound contains queue, reliability, arrival, threshold, and GPD-parameter terms.
  • B. Lyapunov Optimization: Each VUE estimates interference from past channel-quality indicators, enabling decentralized minimization of the drift-plus-penalty bound.
  • B. Lyapunov Optimization: The convex per-VUE power problem is solved by water-filling with a Lagrangian dual coefficient λ_u(t) ≥ 0.

IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION

The proposed learning mechanism estimates the excess-queue distribution parameters locally with FL, reducing reliance on RSU-centered data collection and communication.

  • IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION: Accurate estimation of the GPD parameters σ and ξ from local queue-state samples is crucial to the distributed control mechanism.
  • IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION: An RSU-centric approach is impractical because frequent communication with highly dynamic VUEs creates overhead and vehicles may not share queue-state information.
  • IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION: Federated learning lets each VUE estimate GPD parameters from local queue-state observations with minimal communication with the RSU.
  • IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION: The RSU averages received local parameters into a global model and sends that model back to the VUEs.

A. Queue sampling via block maxima (BM)

Each VUE samples queue behavior through block maxima, balancing reduced sample correlation against sampling overhead. These local samples support the paper’s distributed queue-tail estimation workflow with the RSU.

  • Queue sampling via block maxima (BM): Each VUE uses a time window w to draw one maximum queue-length sample.The window is chosen large enough to reduce correlation while remaining small enough to limit sampling overhead.
  • Queue sampling via block maxima (BM): Fig. 2 links excess queue sampling, GPD parameter estimation, resource allocation, and local/global model exchange with the RSU.The figure presents these four processes as interrelated components between VUEs and the RSU.
  • Queue sampling via block maxima (BM): The kth sample is the maximum queue length observed within the kth block of w time slots.The sampling blocks span consecutive intervals indexed by k.
  • Queue sampling via block maxima (BM): VUEs can sample independently, so the number of collected samples may differ across vehicles.The resulting per-VUE sample sets are used in the queue-sampling procedure illustrated in Fig. 2.

B. Estimating the GPD parameters

The excess queue distribution is modeled by a GPD whose scale and shape parameters are estimated by maximum likelihood. Federated learning distributes this estimation across VUEs while using the RSU for model aggregation.

  • Estimating the GPD parameters: The GPD parameter vector d = [σ, ξ] is estimated by fitting the distribution to excess queue samples.The feasible set requires σ > 0, ξ < 1, and 1 + ξQ/σ ≥ 0 for all samples Q.
  • Estimating the GPD parameters: MLE determines d by maximizing the log-likelihood, equivalently minimizing its negative.The network likelihood is formed from the VUE-specific likelihood contributions.
  • Estimating the GPD parameters: The derivative of the negative GPD log-likelihood provides the gradient needed for iterative parameter optimization.The paper states that gradient-descent-based algorithms can determine the optimal parameter vector.
  • Estimating the GPD parameters: The RSU aggregates distributed model information because sharing all queue samples is impractical with many VUEs and reluctant data sharing.The paper identifies communication overhead and VUE unwillingness to share QSI as the motivating constraints.
  • Estimating the GPD parameters: FL lets each VUE update local GPD estimates using its own samples, gradients, and sample size before uploading the model to the RSU.The local estimation uses SVRG with a predefined positive step size δ.

V. NUMERICAL RESULTS

The numerical evaluation uses a Manhattan mobility scenario with VUEs sharing radio resources. Queue-length CCDFs are evaluated for the proposed FL method and centralized SVRG across different VUE densities.

  • NUMERICAL RESULTS: The evaluation uses a 250 m × 250 m Manhattan mobility model containing nine intersections.Each road has two lanes, each 4 m wide, with VUE pairs spaced 50 m apart and receivers following transmitters at 60 kmph.
  • NUMERICAL RESULTS: Fig. 3 compares queue-length CCDFs from the proposed FL approach and centralized SVRG for different VUE densities.The comparison concerns the estimated queue-tail behavior under the two GPD-estimation approaches.
  • NUMERICAL RESULTS: VUEs share 60 RBs and have a maximum transmit power of p0 = 10 W.The RB allocation per zone is adopted from prior referenced designs.

A. Centralized vs distributed GPD parameter estimation

The proposed FL estimator produces GPD estimates and reliability nearly equivalent to centralized SVRG while reducing data exchange as VUE density grows. The centralized method is more efficient at the smallest tested density because its sample volume is small.

  • Centralized vs distributed GPD parameter estimation: FL GPD estimations are almost equivalent to centralized SVRG estimations for U = 20, 60, and 100 VUEs.Fig. 3 compares the estimated GPDs and corresponding excess queue samples across these densities.
  • Centralized vs distributed GPD parameter estimation: FL and centralized SVRG achieve nearly equal reliability for maintaining queue length below q0 across different VUE densities.Reliability decreases as the number of VUEs increases for both approaches.
  • Centralized vs distributed GPD parameter estimation: Fig. 5 evaluates the transmit-power versus excess-queue-length tradeoff for worst-case VUEs whose queues exceed q0.The figure focuses on the relationship between power expenditure and queue excess beyond the threshold.
  • Centralized vs distributed GPD parameter estimation: At U = 20, centralized estimation exchanges less data because the network sample size is small.FL incurs higher exchange at this density because VUEs upload and download parameters and gradients.

B. Performance evaluation

The evaluation compares the proposed approach with fixed-power and queue-based baselines across power, latency, excess queues, and unreliability. The proposed method particularly reduces extreme queue-threshold violations without additional transmit power.

  • Evaluation setup: The study evaluates a fixed-power model and two additional baselines using average power, queuing latency, excess queue lengths, and unreliability.
  • Performance comparison: For U = 100 VUEs, the proposed method reduces average latency by 3% relative to Baseline 2.
  • Performance comparison: Baseline 2 achieves lower average latency than the proposed method for U = 20 and 60, while its latency is slightly higher when U = 100.
  • Performance comparison: 60%, 8%, and 9% reductions in the fraction of VUEs exceeding the queue threshold are achieved versus Baseline 2 for U = 20, 60, and 100, respectively, without additional transmit power.Compared with fixed-power and Baseline 1 models, reductions are as high as 97-99%.

APPENDIX A PROOF OF PROPOSITION 1

The appendix derives gradients for functions built from the generalized tail expression g_d(Q). It introduces X(Q) and proceeds by first differentiating g_d(Q).

  • Proof of Proposition 1: X(Q) is defined as 1/σ multiplied by g_d(Q) raised to the power ξ + 1.
  • Proof of Proposition 1: The proof holds the stated summation term fixed while differentiating g_d(Q) first.
  • Proof of Proposition 1: The gradient of f_d(Q) is then calculated using the derivative of g_d(Q).
Loading 1805.09253v1…