SECVC: Service-Enhanced Computation Offloading in Complex VEC Environments via Continual Reinforcement Learning

The rapid development of information and communication technologies has accelerated the deployment of Internet of Vehicles applications and sharply increased the volume and complexity of vehicular tasks. Because individual vehicles have limited onboard computation, storage, and energy resources, local execution alone often cannot satisfy the latency requirements of large-scale applications. Mobile edge computing provides a promising remedy by offloading tasks to nearby edge nodes. In practical vehicular edge computing (VEC) systems, however, heterogeneous task dependencies and high vehicle mobility cause frequent topology changes, which complicate reliable offloading. To model these coupled dynamics, we represent heterogeneous workflow tasks with a directed acyclic graph and develop a time-varying topology model for dynamic VEC environments. On this basis, we propose SECVC, a service-enhanced computation offloading method based on continual reinforcement learning. SECVC integrates continual learning into a policy-gradient deep reinforcement learning framework to reduce catastrophic forgetting as the environment changes and to derive adaptive offloading policies. Experiments in simple and complex VEC environments show that SECVC outperforms representative baselines in delay, energy consumption, and task completion rate. The results indicate that SECVC improves policy robustness and adaptability in dynamic VEC systems and provides a viable approach for computation offloading under heterogeneous tasks and time-varying network conditions.