Foundation Model Enhanced Joint Multi-Hop Task Offloading in Dynamic R2X/V2X-based Edge Computing Networks

Recent popularization of the Internet of Vehicles (IoV) and vehicle-to-everything (V2X) enables the emergence of real-time vehicular applications, posing challenges for resource-limited vehicles. Toward this end, vehicular edge computing (VEC) has been proposed to alleviate the computational burden on vehicles by leveraging resources from roadside units (RSUs) and VEC servers. While existing works mainly focus on the task requirements for either vehicles or RSUs, the joint task offloading for both V2X and RSU-to-everything (R2X) has not been fully studied. In this paper, we aim to optimize task offloading strategies for both vehicles and RSUs by adopting a multi-hop task offloading manner to fully utilize VEC network resources. This problem introduces a severe state-action space shift issue with varying dimensions and representations, posing challenges for conventional Deep Reinforcement Learning (DRL) approaches. To address this, we propose the Bidirectional Encoder Representations from Transformers (Bert)-based Matching Q-Network (BMQN) algorithm. First, we design the BMQN model to efficiently capture correlations among all vehicles and RSUs through bi-directional attention. Then, we introduce type-embedded grouped attention and available action embedding to mitigate sequence-length overfitting, thereby enhancing generalization capacity. Moreover, we address the state-action space shift through a matching-based manner, which significantly enhances offloading performance by matching states among devices. Simulation results demonstrate that BMQN achieves superior performance compared to existing approaches in scenarios with varying numbers of vehicles and RSUs, and exhibits robust generalization capacity when adapting to unseen scenarios.