Abstract: In vehicular edge-computing environments with limited communication and computation resources， improper client selection in federated learning （FL） for vehicle-to-everything （V2X） often causes poor performance and high training delays in the aggregated global model. In this study， we propose an optimization method for client selection based on soft actor-critic （SAC）. First， this method establishes a system model that integrates the vehicle channel state， vehicle location， and local computing resources to characterize the key factors affecting the FL performance in the vehicular edge-computing environment. Next， the vehicle-selection process is modeled as a Markov decision process to achieve decision control in the continuous action space. Optimization of vehicle-selection decisions is also established to enhance the long-term goal of high FL performance. Finally， a reward function is designed to balance communication， computing cost， and FL loss， and a corresponding client-selection algorithm is presented for V2X FL based on SAC. This algorithm incorporates a maximum-entropy reinforcement learning framework and dual-critic network architecture and outputs vehicle-participation decisions through the actor network. Furthermore， the proposed algorithm realizes collaborative optimization of the model performance and delay by leveraging environmental reward guidance strategies. The numerical simulation results demonstrate that the proposed method fully considers the practical characteristics of vehicle mobility， dynamic channel conditions， and heterogeneous computational capabilities in V2X environments. The proposed method can effectively select high-quality clients to participate in FL. Compared with the traditional FL， FedCS， and DDPG algorithms， the proposed algorithm achieves faster convergence， significantly improves the model accuracy， and reduces the average training delay by approximately 11.56%. Meanwhile， the proposed algorithm maintains client-selection diversity and enhances the generalization capability of FL models in V2X scenarios.