Delay-Driven Vehicle Scheduling and Resource Optimization for Multi-Task Federated Learning in Heterogeneous Vehicular Networks

Abstract

Federated learning (FL) has become a promising paradigm for privacy-preserving and communication-efficient model training in vehicular networks. With the continuous expansion of vehicular networks, multiple FL tasks are often initiated concurrently by moving vehicles, which poses substantial challenges to the conventional single-task FL framework that relies on round-robin execution. Thus, effective vehicle scheduling is required to support concurrent participation in multiple FL tasks. Moreover, the straggler effect caused by vehicle heterogeneity can lead to inefficient resource utilization, as faster-training vehicles remain idle while waiting for slower ones. To address these challenges, we propose a distributed multi-task FL framework that enables independent and parallel execution of multiple FL tasks. A delay-driven hierarchical vehicle scheduling strategy is developed, which preferentially assigns vehicles with stronger training capabilities to tasks with more stringent latency requirements. To balance training latency and accuracy, we design a dynamic accuracy function, based on which a task utility function is defined. Subject to vehicle waiting-time constraints, we formulate a joint optimization problem that determines the number of global training rounds for each task, uplink bandwidth allocation, and vehicle CPU frequency to maximize the overall system utility. To solve it, we design a novel multi-agent twin-delayed deep deterministic policy gradient (MATD3)-based round-aware and resource optimization algorithm. Simulation results demonstrate that the proposed scheme significantly improves system utility while reducing total energy consumption compared with baseline methods.