QoS-Aware Multi-AIGC Service Orchestration at Edges: An Attention-Diffusion-Aided DRL Method

Yaju Liu, Siyuan Li, Xi Lin, Xiuzhen Chen, Gaolei Li, Yuchen Liu, Bolin Liao, Jianhua Li

Abstract: AI-Generated Content (AIGC) services have gained significant popularity among mobile users due to their automated, high-quality, and diverse content creation capabilities. The evolution of edge networks has further accelerated the adoption of ubiquitous AIGC edge services. However, there is still a lack of research on the collaboration of multiple AIGC services in multi-user and multi-edge environments. Additionally, AIGC services continue to face limitations imposed by constrained resources in edge networks. To improve the Quality of Service (QoS) from the generated content, this paper proposes an innovative attention-diffusion-aided Deep Reinforcement Learning (DRL) method to achieve the QoS-aware multi-AIGC service collaborative orchestration under resource-constrained edge networks. Specifically, we model a mobile user utility function to comprehensively evaluate orchestration decisions based on the inherent capabilities and real-time performance of AIGC services. The proposed Attention-based Diffusion Soft Actor-Critic (ADSAC) algorithm presents the attention-based diffusion model as a policy network in the off-policy reinforcement learning framework to extract probability distributions for complex edge networks and diverse user tasks. The introduction of the attention mechanism captures important features that affect the user utility by selecting the relevant contextual information. Extensive experiments demonstrate the effectiveness of our algorithm in prompting QoS-aware AIGC service at edges. Compared to the existing methods, our proposed ADSAC algorithm improves the overall user utility by at least 30.4% and reduces the server crash rate by at least 17.2%.

External IDs:dblp:journals/tccn/LiuLLCLLLL25