Computing Assignment and Radio Resource Allocation for Edge-End Collaborative Inference in Mobile Edge Computing

Abstract

The growing usage of AI and large models has driven increasing demand for edge-end collaborative inference in mobile edge computing. However, inefficient resource utilization still prevents effective support for such services, especially reflected in difficulty of coordinating numerous computation requests with limited computational resources of edge servers. To address this issue, we consider an edge-end collaborative inference framework which integrates bottom-up computation offloading with top-down computing assignment, and utilizes computational resources of idle end devices (EDs) to avoid task timeout. In this framework, we propose a Multi-Agent Actor-Critic (MA-AC) algorithm to solve the problem of jointly optimizing computing assignment and resource allocation of bandwidth and downlink transmission power, aiming to minimize the tasks' total completion time under energy constraints of EDs. Simulation results show that the MA-AC algorithm outperforms baselines under various communication environmental settings, and demonstrate its adaptability, stability and effectiveness.