Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Biased User Scheduling for Heterogeneous Wireless Federate Learning Network

Changxiang Wu, Yijing Ren, Daniel K. C. So, Jie Tang

TL;DR

The paper tackles the challenge of reducing convergence latency in wireless federated learning with heterogeneous devices and energy constraints. It proposes a unified framework that blends adaptive biased user scheduling (via PPO-based DRL) with two local-resource allocation strategies (LDRA and LCRA) to mitigate stragglers and exploit informative updates under Non-IID data. A convergence bound under non-IID data is derived to quantify the impact of scheduling bias, and the framework is validated through MNIST and CIFAR-10 experiments, showing substantial reductions in total convergence time compared with baselines. The results demonstrate robustness to dynamic channels and energy harvesting, providing a practical approach to deploy FL in real-world heterogeneous wireless networks with energy constraints and stragglers.

Abstract

Federated Learning (FL) has revolutionized collaborative model training in distributed networks, prioritizing data privacy and communication efficiency. This paper investigates efficient deployment of FL in wireless heterogeneous networks, focusing on strategies to accelerate convergence despite stragglers. The primary objective is to minimize long-term convergence wall-clock time through optimized user scheduling and resource allocation. While stragglers may introduce delays in a single round, their inclusion can expedite subsequent rounds, particularly when they possess critical information. Moreover, balancing single-round duration with the number of cumulative rounds, compounded by dynamic training and transmission conditions, necessitates a novel approach beyond conventional optimization solutions. To tackle these challenges, convergence analysis with respect to adaptive and biased scheduling is derived. Then, by factoring in real-time system and statistical information, including diverse energy constraints and users' energy harvesting capabilities, a deep reinforcement learning approach, empowered by proximal policy optimization, is employed to adaptively select user sets. For the scheduled users, Lagrangian decomposition is applied to optimize local resource utilization, further enhancing system efficiency. Simulation results validate the effectiveness and robustness of the proposed framework for various FL tasks, demonstrating reduced task time compared to existing benchmarks under various settings.

Adaptive Biased User Scheduling for Heterogeneous Wireless Federate Learning Network

TL;DR

The paper tackles the challenge of reducing convergence latency in wireless federated learning with heterogeneous devices and energy constraints. It proposes a unified framework that blends adaptive biased user scheduling (via PPO-based DRL) with two local-resource allocation strategies (LDRA and LCRA) to mitigate stragglers and exploit informative updates under Non-IID data. A convergence bound under non-IID data is derived to quantify the impact of scheduling bias, and the framework is validated through MNIST and CIFAR-10 experiments, showing substantial reductions in total convergence time compared with baselines. The results demonstrate robustness to dynamic channels and energy harvesting, providing a practical approach to deploy FL in real-world heterogeneous wireless networks with energy constraints and stragglers.

Abstract

Federated Learning (FL) has revolutionized collaborative model training in distributed networks, prioritizing data privacy and communication efficiency. This paper investigates efficient deployment of FL in wireless heterogeneous networks, focusing on strategies to accelerate convergence despite stragglers. The primary objective is to minimize long-term convergence wall-clock time through optimized user scheduling and resource allocation. While stragglers may introduce delays in a single round, their inclusion can expedite subsequent rounds, particularly when they possess critical information. Moreover, balancing single-round duration with the number of cumulative rounds, compounded by dynamic training and transmission conditions, necessitates a novel approach beyond conventional optimization solutions. To tackle these challenges, convergence analysis with respect to adaptive and biased scheduling is derived. Then, by factoring in real-time system and statistical information, including diverse energy constraints and users' energy harvesting capabilities, a deep reinforcement learning approach, empowered by proximal policy optimization, is employed to adaptively select user sets. For the scheduled users, Lagrangian decomposition is applied to optimize local resource utilization, further enhancing system efficiency. Simulation results validate the effectiveness and robustness of the proposed framework for various FL tasks, demonstrating reduced task time compared to existing benchmarks under various settings.
Paper Structure (21 sections, 3 theorems, 56 equations, 5 figures, 4 algorithms)

This paper contains 21 sections, 3 theorems, 56 equations, 5 figures, 4 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. Contributions
  4. System Model
  5. Federated Learning Workflow
  6. Enabling Federated Learning over Wireless Network
  7. Convergence Analysis and Problem Formulation
  8. Convergence Analysis
  9. Problem Formulation
  10. Deep Reinforcement Learning based User Scheduling
  11. Markov Decision Process Design
  12. Training DRL Agent over Wireless FL Network
  13. Local Resources Allocation
  14. CPUs Frequency Optimization
  15. Communication Process Optimization
  16. ...and 6 more sections

Key Result

Theorem 1

Assuming a decaying learning rate $\eta_k \leq \frac{1}{\beta_k}$, and $G_k\,=\, \max_{n\in\mathcal{V}_k}\left\| \nabla F_n({\boldsymbol{w}_k^n}) \right\|^2$, the expected gradient norm of the adaptive biased scheduling is bounded by:

Figures (5)

  • Figure 1: Training Convergence of the PPO-based DRL agent.
  • Figure 2: FL convergence performance on (a) MNIST and (b) CIFAR-10.
  • Figure 3: FL convergence time verses the iteration round of one FL task on (a) MNIST and (b) CIFAR-10.
  • Figure 4: Scheduled user number verses the iteration round of one FL task on (a) MNIST and (b) CIFAR-10.
  • Figure 5: Convergence time versus (a) Non-IID ratio and (b) target accuracy.

Theorems & Definitions (4)

  • Definition 1
  • Theorem 1
  • Proposition 1
  • Proposition 2