ReSO: A Real-Time Scalable Offloading Approach for Dynamic Vehicle-to-Everything Systems

Abstract

Smart vehicles increasingly rely on a growing range of nearby devices, such as parked cars and roadside units, as service nodes (SNs) for offloading latency-sensitive tasks in Vehicle-to-Everything (V2X) networks. Deep reinforcement learning (DRL) has emerged as a powerful scheduler for this setting, yet dimensional mismatch at runtime renders a pre-trained agent incompatible whenever system scale or task count changes, and retraining from scratch is prohibitively costly and often impractical. To resolve this issue, we present ReSO, a DRL-based Real-time Scalable Offloading approach built on two targeted mechanisms. To eliminate task-count-induced dimension variation, we adopt sequential single-task dispatching in place of batch scheduling, casting the offloading problem as an event-driven Semi-Markov Decision Process (Semi-MDP). A homotopy reward then corrects the resulting temporal credit mis-assignment. For scalability across varying system scales, we leverage a heterogeneous graph attention encoder to produce scale-invariant embeddings, and a pointer output module to generate actions that dynamically adapt to the number of reachable SNs. Together, these mechanisms make ReSO transferable across dynamic V2X systems. Experiments confirm that, after zero-shot transfer to systems of varying scale, ReSO consistently reduces system cost by up to 35% over all baselines, even those afforded full retraining.