Dynamic Offloading Strategy in Vehicular Networks via Federated Continuous Reinforcement Learning

The dynamic nature of vehicular ad-hoc networks (VANETs) presents significant challenges, including computational constraints, decision conflicts, and performance degradation under environmental dynamics. Traditional offloading strategies often struggle to balance real-time responsiveness, energy efficiency, and knowledge retention in heterogeneous vehicle clusters. To address these issues, this paper proposes a federated continual reinforcement learning framework with asynchronous updates and spatiotemporal knowledge fusion. An optimized multi-agent deep deterministic policy gradient algorithm with lazy agent updates (OMADDPG-LA) is introduced to reduce policy conflicts while improving communication efficiency. Additionally, a spatiotemporally decoupled knowledge retention mechanism adjusts parameter importance weights based on vehicle-specific behaviors and incorporates generative replay to alleviate storage constraints. A federated aggregation paradigm with credibility-aware weighting further enhances adaptability to dynamic tasks. Experimental results show that the proposed framework significantly improves information freshness, energy efficiency, and communication overhead compared to existing methods. It also achieves better convergence stability under dynamic conditions, providing a privacy-preserving and scalable solution for collaborative intelligence in large-scale VANETs.