1: \begin{abstract}

2: This paper proposes a vehicular edge federated learning (VEFL) solution, where an edge server leverages highly mobile connected vehicles' (CVs') onboard central processing units (CPUs) and local datasets to train a global model. 

3: Convergence analysis reveals that the VEFL training loss depends on the successful receptions of the CVs' trained models over the intermittent vehicle-to-infrastructure (V$2$I) wireless links.

4: Owing to high mobility, in the full device participation case (FDPC), the edge server aggregates client model parameters based on a weighted combination according to

5: the CVs' dataset sizes and sojourn periods, while it selects a subset of CVs in the partial device participation case (PDPC).

6: We then devise joint VEFL and radio access technology (RAT) parameters optimization problems under delay, energy and cost constraints to maximize the probability of successful reception of the locally trained models.

7: Considering that the optimization problem is NP-hard, we decompose it into a VEFL parameter optimization sub-problem, given the estimated worst-case sojourn period, delay and energy expense, and an online RAT parameter optimization sub-problem.

8: Finally, extensive simulations are conducted to validate the effectiveness of the proposed solutions with a practical $5$G new radio ($5$G-NR) RAT under a realistic microscopic mobility model.

9: \end{abstract}

10: