Optimizing AIGC Services Using Learning-Based Stackelberg Game in Vehicular Metaverses

Bingkun Lai, Xiaofeng Luo, Jiawen Kang, Xiaozheng Gao, Zuyuan Yang, Dusit Niyato, Shiwen Mao

Abstract: The emerging vehicular metaverse embodies the next-generation vehicular networking paradigm. In the vehicular metaverses, Artificial Intelligence-Generated Content (AIGC) technology as a powerful content generation tool, is capable of providing an immersive experience for Vehicular Metaverse Users (VMUs). Due to limited computational resources within vehicles, VMUs rely on AIGC Service Providers (ASPs) to execute resource-intensive AIGC tasks within vehicular metaverses. However, large-scale AIGC service requests can lead to resource scarcity within the ASP, ultimately leading to declining service quality for VMUs. To tackle this challenge, we introduce a novel Stackelberg game framework utilizing the Generative Diffusion Model (GDM) for AIGC services, in which we experimentally reveal a relationship between image quality and diffusion steps. A Transformer-based Deep Reinforcement Learning (TDRL) algorithm is employed to find the optimal Stackelberg equilibrium under incomplete information. Numerical results indicate that our method converges to equilibrium efficiently, with superior utilities compared to baseline approaches.

External IDs:dblp:journals/tvt/LaiLKGYNM25