Asynchronous Federated Learning for Commercial Vehicle Fleets
| Title | Asynchronous Federated Learning for Commercial Vehicle Fleets |
|---|---|
| Summary | explore and design Asynchronous Federated Learning strategies for commercial vehicle fleets in AI-driven digital services |
| Keywords | |
| TimeFrame | Fall 2025 or Spring 2026 |
| References | [[References::[1] Xu, C., Qu, Y., Xiang, Y., & Gao, L. (2023). Asynchronous federated learning on heterogeneous devices: A survey. Computer Science Review, 50, 100595.
[2] Chen, Z., Liao, W., Hua, K., Lu, C., & Yu, W. (2021). Towards asynchronous federated learning for heterogeneous edge-powered internet of things. Digital Communications and Networks, 7(3), 317-326. [3] Imteaj, A., & Amini, M. H. (2020, December). Fedar: Activity and resource-aware federated learning model for distributed mobile robots. In 2020 19th IEEE International Conference on Machine Learning and Applications (ICMLA) (pp. 1153-1160). IEEE. [4] Wu, W., He, L., Lin, W., Mao, R., Maple, C., & Jarvis, S. (2020). SAFA: A semi-asynchronous protocol for fast federated learning with low overhead. IEEE Transactions on Computers, 70(5), 655-668. [5] Chen, Y., Sun, X., & Jin, Y. (2019). Communication-efficient federated deep learning with layerwise asynchronous model update and temporally weighted aggregation. IEEE transactions on neural networks and learning systems, 31(10), 4229-4238. [6] Xie, C., Koyejo, S., & Gupta, I. (2019). Asynchronous federated optimization. arXiv preprint arXiv:1903.03934. [7] Sun, W., Lei, S., Wang, L., Liu, Z., & Zhang, Y. (2020). Adaptive federated learning and digital twin for industrial internet of things. IEEE Transactions on Industrial Informatics, 17(8), 5605-5614. [8] Zhang, Y., Duan, M., Liu, D., Li, L., Ren, A., Chen, X., ... & Wang, C. (2021). CSAFL: A clustered semi-asynchronous federated learning framework. IJCNN, pp. 1-10.]] |
| Prerequisites | |
| Author | |
| Supervisor | Yuantao Fan, Zahra Taghiyarrenani |
| Level | Master |
| Status | Open |
Traditional machine learning workflows often depend on centralized data aggregation, which creates high communication overhead and raises privacy concerns. While federated learning (FL) seeks to address these issues, it faces scaling difficulties in large, heterogeneous fleets due to differences in device capabilities and connectivity. Synchronization requirements in standard FL can cause idle waiting for slower devices, reducing overall efficiency. In addition, repetitive computations on previously learned data lead to unnecessary resource
consumption in large-scale deployments. These obstacles emphasize the need for more flexible, asynchronous approaches that accommodate changing device connectivity, varying computation resources and data availability, and diverse fleet configurations.
This project explores and aims to design scalable Asynchronous Federated Learning (AFL) solutions that account for fleet heterogeneity, storage efficiency, and computational constraints. By utilising multiple aggregated models and personalized learning, the research aims to enhance model accuracy, adaptability and efficiency for AI-driven digital services in commercial vehicle fleets. Target application areas include energy consumption forecasting, representation learning, anomaly detection, and related intelligent mobility tasks.
Existing research on AFL spans several key areas: i) node selection: while classical FL selects nodes based primarily on their data volume, AFL prioritizes nodes with heightened resilience, network and computational capacity. For example, work presented in [2] employs a heuristic greedy node selection strategy to iteratively involve nodes in global learning based on local computing and communication resources, whereas other works use trust scores [3], or crashing probability [4] to guide the selection process; ii) weighted aggregation: increasing the weight of recently updated local models [5] or adjusting a hyperparameter to balance convergence speed and variance due to staleness [6]; iii) gradient compression, to reduce communication expenses. Fourth, employing semi-asynchronous FL as a hybrid approach, aggregating local models that arrive early, and involving slow devices based on the magnitude of stabledness; iv) cluster-based FL, which groups nodes based on aggregation frequency [7], gradient direction, or latency [8] to optimize overall performance. These research directions address key challenges related to heterogeneity, communication constraints, and model convergence, and are of interest to explore.
