Maximizing Accuracy-Per-Cost in Vehicular Federated Learning via Active Inference and Two-Stage Resource Optimization

Abstract

Federated Learning (FL) enables collaborative model training in the Internet of Vehicles (IoV), helping vehicles overcome data and computational limitations while preserving privacy. However, its performance is constrained by both data heterogeneity and system heterogeneity. Additionally, different FL tasks exhibit diverse preferences regarding energy consumption and end-to-end latency. To address the diverse latency and energy preferences in IoV and improve model accuracy and convergence, this paper proposes a unified framework to optimize accuracy per unit of utility in vehicular FL. Utility is modeled as a weighted sum of energy consumption and latency. We analyze key factors influencing FL model accuracy in IoV environments and establish a joint optimization problem that balances accuracy with utility cost. To match the information asymmetry between the server and clients and reduce computational complexity, we decompose the problem into two coordinated stages: server-side client selection and bandwidth allocation, and client-side computation and transmission power control. On the server side, we model decision-making as a Partially Observable Markov Decision Process (POMDP), and develop an active-inference-based actor–critic method for adaptive client selection and resource management. On the client side, we derive closed-form optimal solutions for resource allocation under heterogeneous latency and energy preferences. By combining learning-based global coordination with lightweight analytical local control, the proposed provides a practical and scalable tradeoff between adaptability, efficiency, and implementation complexity. Extensive simulations validate the effectiveness and superiority of our algorithm.