PaperlayerView exact revision on arXiv

Source-linked AI summary

Distributed Federated Learning for Ultra-Reliable Low-Latency Vehicular Communications

Sumudu Samarakoon, Mehdi Bennis, Walid Saad, Merouane Debbah

arXiv:1807.08127v3cs.IT

TL;DR

The paper addresses network-wide power minimization for vehicular URLLC while controlling probabilistic queueing delays and rare extreme queue lengths. It combines EVT-based tail modeling, federated learning, and Lyapunov optimization in a distributed JPRA framework. FL achieves tail-estimation accuracy close to centralized learning with up to 79% less exchanged data, while the proposed methods also reduce large queues and power consumption relative to the stated baseline.

  • Problem

    Vehicular URLLC requires low errors and delays, but conventional queue-reliability constraints do not capture rare extreme queue lengths that determine worst-case latency.

  • Method

    The paper combines EVT-based GPD modeling, distributed FL tail estimation, and Lyapunov optimization for distributed joint power and resource allocation.

  • Results

    Up to 79% reductions in exchanged data are achieved with async-FL compared with CEN, while proposed tail-aware methods yield nearly equal worst-case queuing latencies in the reported range.

  • Takeaways & Limitations

    FL enables local learning of network-wide queue tails without sharing actual queue-length samples, reducing unnecessary signaling overhead.

Abstract

from arXiv · show

In this paper, the problem of joint power and resource allocation (JPRA) for ultra-reliable low-latency communication (URLLC) in vehicular networks is studied. Therein, the network-wide power consumption of vehicular users (VUEs) is minimized subject to high reliability in terms of probabilistic queuing delays. Using extreme value theory, a new reliability measure is defined to characterize extreme events pertaining to vehicles' queue lengths exceeding a predefined threshold. To learn these extreme events, assuming they are independently and identically distributed over VUEs, a novel distributed approach based on federated learning (FL) is proposed to estimate the tail distribution of the queue lengths. Considering the communication delays incurred by FL over wireless links, Lyapunov optimization is used to derive the JPRA policies enabling URLLC for each VUE in a distributed manner. The proposed solution is then validated via extensive simulations using a Manhattan mobility model. Simulation results show that FL enables the proposed method to estimate the tail distribution of queues with an accuracy that is close to a centralized solution with up to 79% reductions in the amount of exchanged data. Furthermore, the proposed method yields up to 60% reductions of VUEs with large queue lengths, while reducing the average power consumption by two folds, compared to an average queue-based baseline.

I. INTRODUCTION

The paper targets URLLC in vehicular networks by minimizing network-wide power while controlling queue reliability, including rare extreme queue-length events. It proposes a distributed EVT- and FL-based framework with Lyapunov optimization to reduce coordination overhead.

  • Motivation: URLLC for V2V networks is needed to support intelligent transportation applications requiring extremely low errors and delays.The paper connects URLLC to applications including platooning, navigation, collision avoidance, and autonomous driving.
  • Motivation: Low queuing latency and efficient radio resource management are required because queuing and over-the-air latency jointly determine end-to-end latency.The paper identifies a target end-to-end latency of 1 ms and also highlights energy consumption as a vehicular-network concern.
  • Contribution: Existing vehicular RRM studies address rate, delay, energy, and clustering objectives, but this paper focuses on jointly allocating transmit power and resources for URLLC.The proposed framework is distributed and targets network-wide power minimization under latency and reliability requirements.
  • Reliability modeling: EVT models rare queue-length exceedances through a generalized Pareto distribution, whose scale and shape parameters are estimated from threshold-exceeding samples.These extreme events correspond to worst-case queuing and end-to-end latency that ordinary reliability constraints may miss.
  • Distributed learning: Federated learning lets each VUE estimate local GPD parameters and share two gradient values, while the RSU averages models instead of collecting queue samples.This design supports distributed tail-distribution estimation while avoiding direct exchange of actual queue-length samples.
  • System model: The network model groups VUE pairs that can reuse RBs with low-to-no interference, while the RSU allocates RBs orthogonally across groups.Each VUE may use only the RB subset assigned to its current group.

III. PROPOSED DISTRIBUTED FRAMEWORK USING EVT AND LYAPUNOV OPTIMIZATION

The distributed solution decouples the stochastic optimization across VUE pairs using statistics of queue lengths exceeding a threshold.

  • The optimization is decoupled over VUE pairs using statistics of queue lengths exceeding q0.

A. Modeling Extreme Queue Lengths Using Extreme Value Theory

Extreme queue lengths are modeled with a generalized Pareto distribution under an IID assumption, enabling local reliability constraints and Lyapunov-based distributed control.

  • Modeling Extreme Queue Lengths Using Extreme Value Theory: Under an IID assumption and a sufficiently large q0, excess queue lengths are modeled using a generalized Pareto distribution.The GPD is characterized by shape ξ and scale σ parameters.
  • Modeling Extreme Queue Lengths Using Extreme Value Theory: Each VUE can estimate ξ and σ locally with RSU assistance, avoiding QSI sharing and decoupling reliability constraints.
  • Modeling Extreme Queue Lengths Using Extreme Value Theory: EVT-based first and second tail moments recast the original optimization problem into an equivalent form.
  • Lyapunov Optimization for Power Allocation: Lyapunov optimization models time-average constraints with virtual queues and introduces a parameter V controlling the queue-length–solution-accuracy tradeoff.
  • Lyapunov Optimization for Power Allocation: Controlling the Lyapunov-drift upper bound ensures stability of both actual and virtual queues.
  • Lyapunov Optimization for Power Allocation: The resulting policy minimizes an upper bound on drift plus penalty and can be decoupled among VUEs using estimated interference.

IV. LEARNING THE PARAMETERS OF THE MAXIMUM QUEUE DISTRIBUTION

Accurate estimation of excess-queue distribution parameters is necessary for optimal power allocation, but centralized estimation introduces communication overhead and latency.

  • Optimal power allocation depends on the excess queue distribution G^d_M(m) and therefore requires accurate estimates of σ and ξ.
  • Centralized modeling requires the RSU to compute and communicate distribution information with all VUEs at each time t.

A. Queue Sampling via Block Maxima (BM)

The framework samples block maxima of excess queues and estimates GPD parameters through centralized or distributed gradient-based procedures, while reducing sample-uploading overhead through model exchange.

  • A. Queue Sampling via Block Maxima (BM): Each VUE uses a block length w to select at most one maximum queue sample per block when the queue exceeds q0.The block length balances reduced sample correlation against undersampling.
  • A. Queue Sampling via Block Maxima (BM): Queue samples can vary across VUEs because each VUE independently performs its own sampling process.
  • B. RSU-Centric GPD Parameter Estimation: Maximum-likelihood estimation fits the GPD parameters d = [σ, ξ] to the excess-queue samples.
  • B. RSU-Centric GPD Parameter Estimation: CEN uses SVRGD at the RSU to iteratively estimate GPD parameters, then shares the resulting parameters with all VUEs.
  • B. RSU-Centric GPD Parameter Estimation: In CEN, frequent uploading of local queue samples increases wireless overhead and can cause congestion and latency as VUE density grows.
  • B. RSU-Centric GPD Parameter Estimation: Figure 2 links excess queue sampling, GPD parameter estimation, transmit-power and RB allocation, and local-global model exchange.

C. FL-Based GPD Parameter Estimation

The paper estimates GPD parameters distributively by having VUEs compute local likelihood gradients and models, then sharing them with an RSU through synchronous or asynchronous FL.

  • FL decomposes the network likelihood into per-VUE likelihoods evaluated locally from each vehicle’s queue samples.
  • Each VUE uses SVRGD to update local GPD parameters and gradients, then uploads its learning model to the RSU.
  • The shared local model contains gradients, GPD parameters, sample counts, and each VUE’s maximum queue sample, which determines the domain needed by SVRGD.
  • The RSU averages received models to update global GPD parameters and gradients before sharing the global model with the network.
  • Sync-FL aggregates simultaneous uploads at fixed intervals, whereas async-FL aggregates independently arriving models after each VUE collects a predefined number of new samples.

D. Cost of Communication with the RSU

Communication for centralized and federated GPD estimation is modeled as additional delay on V2V links because VUEs exchange queue samples or learning models with the RSU.

  • The communication exchanges in CEN, sync-FL, and async-FL are illustrated as switching between V2RSU and V2V communication periods.
  • CEN uploads all new queue samples and downloads GPD parameters, while sync-FL exchanges learning models instead of the full sample set.
  • (JΥ + Jd + JQ)/ru0 and (JΥ + Jd + JQ)/r0u represent sync-FL’s uplink and downlink communication delays for a VUE.
  • Async-FL lets VUEs access the RSU independently, reducing simultaneous-access interference compared with synchronous communication.

V. SIMULATION RESULTS AND ANALYSIS

The simulations use a Manhattan vehicular mobility model and evaluate queue-tail estimation and communication-learning behavior under different VUE populations.

  • The simulation network is a 250 m×250 m Manhattan model with nine intersections, 60 RBs, and maximum transmit power p0 = 10 W.
  • Queue-length CCDFs exceeding q0 are evaluated for CEN and async-FL across different numbers of VUEs.
  • The MLE-based cost fd(Q) is evaluated against SVRGD iteration count for CEN and async-FL using scenarios with about 2U queue-length samples.
  • Simulation parameters assign 8 bits to gradient values and 16 bits each to GPD parameters and other specified model quantities.

A. Centralized vs distributed GPD parameter estimation

The results compare centralized estimation with async-FL on GPD accuracy, exchanged data, reliability, transmit power, and queue lengths as VUE density changes.

  • Async-FL’s GPD estimates are almost equivalent to CEN’s estimates for U = 20, 60, and 100.
  • For more than two SVRGD iterations, async-FL’s cost is about 0.5% higher than CEN at U = 20 and about 0.5% lower at U = 100.
  • Async-FL reliability is slightly lower than CEN for U < 72 but higher than CEN for U > 72 when maintaining queue length below q0.
  • 27% and up to 79% reductions in exchanged data are achieved by async-FL relative to CEN at U = 28 and U = 100, respectively.
  • As VUE density increases beyond 28, CEN exchanges more data than async-FL because its sample size grows with the network.

B. Performance Evaluation

The evaluation compares EVT-based proposed methods with queue-based and fixed-power baselines across VUE densities, distances, and tail-queue statistics. Async-FL generally improves reliability, worst-case latency, extreme-queue behavior, and power efficiency, especially as network size increases.

  • Power consumption: Async-FL uses less transmission power than CEN and sync-FL, with reductions reaching 31.6% at U = 42 and remaining around 35% for U > 42.The reductions are attributed to lower power requirements for asynchronous learning-model uploads and downloads.
  • Worst-case latency: Async-FL reduces worst-case latency by 7.5%, 21.7%, and 29% versus QSR, CEN, and sync-FL, respectively, at U = 100.Asynchronous model sharing reduces delays imposed by RSU-VUE communications.
  • Evaluation settings: The evaluation examines async-FL tail distributions using mean and standard deviation versus VUE-pair count, with 40 VUEs and speeds of 40, 60, and 80 km/h.The corresponding finite-block-length values are D = 800, 534, and 400 bits.
  • Reliability: At U = 100, async-FL reduces outages by about 60.9%, 36%, and 35.9% compared with QSR, CEN, and sync-FL, respectively.For U ≥84, asynchronous model sharing produces lower queue lengths and reduced outages than CEN and sync-FL.
  • Tail distribution: For U > 44, async-FL has the lowest mean and fluctuations of queue lengths exceeding q0, reducing average extreme queue lengths by 28.6%, 41.9%, and 19.5% versus QSR, CEN, and sync-FL.At U = 100, fluctuation reductions are about 33.2%, 38%, and 47.1% against the same methods.
  • Distance sensitivity: As vTx-vRx distance increases, queue lengths rise in QSR and async-FL because over-the-air data rates decrease.Async-FL reduces average transmit power by 3.2%, 18.2%, and 43.1% at distances of 20 m, 50 m, and 80 m, respectively.

C. Impact of finite block length

The finite-block-length evaluation examines how vehicle speed affects async-FL under mobility-dependent coherence times. Higher speeds shorten block lengths and reduce achievable rates, increasing worst-case latency and queue outages while higher power can lower average queuing latency.

  • Finite-block-length setting: Vehicle speeds of 40, 60, and 80 km/h correspond to block lengths D = 800, 534, and 400 bits.Block length is proportional to channel coherence time and inversely proportional to VUE speed.
  • Speed effects: For ε = .5, higher VUE speeds increase worst-case queuing latency and queue outages in async-FL.The paper attributes this to lower achievable rates from smaller coherence times and shorter block lengths.
  • Power-latency trade-off: For ε = .5, higher power consumption yields lower average queuing latency in the finite-block-length evaluation.The evaluation reports average queue lengths, maximum queue length, average transmit power, and outages for async-FL.

APPENDIX A PROOF OF PROPOSITION 1

The appendix proves Proposition 1 by bounding the one-slot Lyapunov drift and separating control-dependent from control-independent terms. The derivation uses queue-update inequalities and boundedness under queue stability.

  • Drift bound: The proof begins with the one-slot Lyapunov drift ΔL_t = L(Ξ(t + 1)) − L(Ξ(t)).This drift is the quantity subsequently upper-bounded in the proof.
  • Queue expansion: The inequality ([q+a−r]+)^2 ≤ q^2+(a−r)^2+2q(a−r) provides upper bounds for terms in the drift expansion.For nonempty queues, the queue update is written without the positive-part operator.
  • Term grouping: Queue stability bounds the quadratic terms, replacing their sum with the bounded value Δ0.Terms independent of control variables are grouped into Δu for each VUE.
  • Conclusion: Combining the derived bounds and applying them to the drift expression completes the proof.The appendix explicitly combines the results of equations (24)–(26) with equation (23).
Loading 1807.08127v3…