Source-linked AI summary
Age-Based Scheduling Policy for Federated Learning in Mobile Edge Networks
Howard H. Yang, Ahmed Arafa, Tony Q. S. Quek, H. Vincent Poor
TL;DR
Federated learning must schedule a subset of devices over limited wireless resources while preserving timely model updates. This paper proposes a low-complexity AoU-based policy that jointly accounts for update staleness and instantaneous channel quality, and simulations show improved FL convergence over MaxPack.
Problem
Limited wireless resources allow only a subset of UEs to update each round, while existing scheduling approaches overlook update staleness despite its effect on FL convergence.
Method
The paper defines age-of-update and uses it with instantaneous channel quality in a greedy scheduling protocol for federated learning.
Results
The ABS policy achieves marked test-accuracy improvement over MaxPack, especially during the initial training stage below 40 communication rounds.
Takeaways & Limitations
Accounting for update staleness alongside channel quality accelerates FL convergence in the evaluated mobile edge network.
Takeaways & Limitations
The evaluation assumes independently and identically distributed quasi-static block-fading channels that remain constant during each transmission block.
Abstract
from arXiv · showhide
Federated learning (FL) is a machine learning model that preserves data privacy in the training process. Specifically, FL brings the model directly to the user equipments (UEs) for local training, where an edge server periodically collects the trained parameters to produce an improved model and sends it back to the UEs. However, since communication usually occurs through a limited spectrum, only a portion of the UEs can update their parameters upon each global aggregation. As such, new scheduling algorithms have to be engineered to facilitate the full implementation of FL. In this paper, based on a metric termed the age of update (AoU), we propose a scheduling policy by jointly accounting for the staleness of the received parameters and the instantaneous channel qualities to improve the running efficiency of FL. The proposed algorithm has low complexity and its effectiveness is demonstrated by Monte Carol simulations.
1. INTRODUCTION
FL moves training to user devices to preserve privacy and reduce communication overhead, but limited wireless resources make UE scheduling important. Existing approaches emphasize communication efficiency while neglecting update staleness, motivating AoU-based scheduling that jointly considers staleness and channel quality.
- Federated learning trains models on user devices and sends only resultant parameters to edge servers, reducing communication overhead and preserving data privacy.
- Stragglers are more critical in federated learning than in conventional data-center training.
- Prior FL scheduling protocols target transmission latency, spectral utility, or channel selection, but largely ignore update staleness.
- Because update staleness affects distributed-learning convergence, the paper introduces age-of-update to schedule timely updates using both staleness and communication quality.
2. SYSTEM MODEL
The system comprises an access point and K single-antenna UEs sharing N subchannels under block-fading channels. FL alternates local computation, aggregation, and model distribution, while limited spectrum forces the AP to schedule only a subset of UEs each round.
- The mobile edge network contains one access point and K single-antenna UEs, each with a local dataset, communicating through a common spectrum divided into N equal-length subchannels.
- The model assumes N < K and independently and identically distributed quasi-static block-fading channels that remain constant during each transmission block.
- Federated training uses local device computation, AP aggregation of received updates, and redistribution of the improved global model over repeated communication rounds.
- Because the wireless medium is resource-constrained, the AP selects only a subgroup of UEs for updates in each round, and selection affects FL convergence.
- The AoU example depicts its time evolution under transmissions occurring at timestamps t1, t2, and t3.
3. SCHEDULING POLICY DESIGN
The paper designs age-based scheduling around AoU to balance update freshness, channel quality, fairness, and spectrum use. A greedy ABS procedure selects feasible UEs and assigns subchannels with relatively low complexity.
- 3.1. Design Metric: AoU measures update staleness and increases when a UE is not selected, allowing the AP to prioritize older updates.The metric resets through selection and otherwise grows with each communication round.
- 3.2. Problem Formulation: The AP selects UEs by minimizing functions of AoU while enforcing rate, transmit-power, and orthogonal-access constraints.The function fα represents the server’s sensitivity to local-update staleness.
- 3.2. Problem Formulation: The objective balances fairness and radio-channel quality, reducing to maximum UE packing for constant fα and N-Round-Robin under ideal communication.These special cases show how the formulation connects spectrum utilization with fairness.
- 3.3. Algorithm: Because the mixed-integer problem lacks generally low-complexity solutions, the paper solves it with a greedy algorithm.ABS repeatedly forms feasible candidates, selects UEs using AoU relative to subchannel needs, and recalculates candidates after selections.
- 3.3. Algorithm: ABS allocates subchannels and schedules UEs recursively, selecting k* = arg maxk∈Is fα(Tk[t])/∥Ck∥0 among candidates.The procedure first constructs a candidate list satisfying the target rate, then updates available subchannels after each selection.
- 3.3. Algorithm: The resulting wireless FL procedure iterates AP scheduling, local UE updates, parameter collection, AoU updates, and global-model aggregation.Algorithm 3 repeats these steps across communication rounds.
4. NUMERICAL RESULTS
Simulations evaluate ABS against MaxPack for SVM training on MNIST. ABS achieves higher test accuracy, with the clearest convergence gain during the first 40 communication rounds.
- 4. Numerical Results: ABS attains a marked test-accuracy improvement over MaxPack as communication rounds increase.The evaluation uses SVM training on MNIST with data divided among 100 UEs.
- 4. Numerical Results: Before 40 communication rounds, ABS shows especially pronounced gains, indicating accelerated FL convergence and improved learning efficiency.The reported advantage is attributed to balancing aggressive channel use with short-term fairness and gradient diversity.
5. CONCLUSION
The paper proposes an FL scheduling policy for mobile edge networks that uses AoU to account for update staleness and instantaneous channel quality, accelerating FL convergence.
- The paper proposes a scheduling policy for federated learning in mobile edge networks.
- AoU jointly accounts for update staleness and instantaneous channel qualities in the scheduling scheme.
- The proposed scheme is intended to accelerate the convergence rate of federated learning.