Client Selection and Model Pruning for Energy-Efficient Federated Learning on Resource-Constrained Edge Devices
Year: 2026
Type: conference-paper
Abstract: Federated learning on heterogeneous edge devices is constrained by differences in computation, memory, bandwidth, and energy. Existing client selection strategies mainly optimize training time, energy efficiency, or model performance, often overlooking memory constraints on edge devices. Sub-model based techniques address this by allowing devices to train pruned versions of the global model according to their resources. In this work, we propose a sub-model aware client selection strategy using Multi-agent Reinforcement Learning (MARL) that jointly considers device heterogeneity, memory constraints, energy efficiency, and model accuracy. The approach dynamically selects clients and assigns suitable sub-models based on device conditions. Experimental results show superior energy efficiency over heuristic and reinforcement learning baselines.
Institution: Karlsruhe Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7225752039
