Service Association and Subcarrier Allocation in Vehicular Slicing Networks With Dynamic Multi-Service

The vehicular slicing network with dynamic multi-service (VSN-DMS) enables vehicles to support multiple services with dynamically changing types. In this architecture, vehicles can access multiple base stations (BSs) and leverage different network slices (NSs) to meet the requirements of various services. However, frequent service variations and vehicle mobility introduce high handover (HO) costs and inefficient decision making. This letter proposes a novel two-timescale optimization framework to jointly optimize vehicle service (VS) association and subcarrier (SC) allocation. Specifically, in the long-timescale, we propose a federated deep reinforcement learning (FDRL) algorithm to optimize VS association. Unlike conventional reinforcement learning approaches that require centralized data collection, FDRL enables distributed model training across vehicles while aggregating parameters in the cloud. This mitigates the issue of insufficient training data caused by frequent VS changes, while also ensuring vehicle data privacy. In the short-timescale, we propose a versatile alternating coalitional game and swap-matching game algorithm (ACGSM) to optimize SC allocation. By iteratively executing coalitional formation and swap-matching mechanisms, ACGSM efficiently improves the quality of the solution while adapting to dynamic network conditions. Simulation results demonstrate the convergence of FDRL and show that the proposal outperforms benchmarks in improving long-term QoS and reducing the number of HOs.