Abstract In vehicular networks, federated learning faces significant challenges due to resource heterogeneity, dynamic participation patterns, and intermittent connectivity among vehicles. Traditional client selection mechanisms often fail to consider the two-tier decision-making process inherent in vehicular network environments, where both central servers and individual vehicles must make participation decisions based on their respective constraints. Moreover, existing model aggregation algorithms typically assume fixed client participation and cannot adapt to the highly variable participation patterns unique to vehicular networks. This paper proposes a comprehensive vehicular federated learning framework with three key innovations. First, we introduce a strategy-driven adaptive client participation mechanism with a two-tier decision-making process that combines server-side reinforcement learning-based client selection with client-side autonomous participation decisions based on local resource thresholds. Second, we develop an incremental online policy learning algorithm based on Proximal Policy Optimization (IO-PPO) to address the data scarcity challenge in federated learning environments by enabling continuous learning from limited trajectory data. Third, we propose a dynamic client size-adaptive optimized model aggregation algorithm that adapts to different participation patterns while considering both current and historical client contributions. Our approach leverages a synergistic combination of reinforcement learning for adaptive decision-making, asynchronous federated learning principles for flexible participation, and graph-based modeling for capturing network topology effects. Extensive experimental results demonstrate that compared to existing methods, the proposed framework significantly improves learning efficiency, convergence stability, and model performance in realistic vehicular network scenarios.
Lin et al. (Wed,) studied this question.