Diffusion Model Based Preference-Guided Learning for Edge Server Placement

Abstract

Edge server placement (ESP) is critical in mobile edge computing (MEC) by enabling low-latency services and cost-efficient operation through effective resource allocation. However, with the rapid growth of candidate nodes and the integration of heterogeneous server types, ESP becomes a complex large-scale multi-objective optimization problem. The resulting exponential expansion of the decision space, combined with stringent engineering constraints, leads to sparse feasible regions. Consequently, conventional multi-objective evolutionary algorithms face significant challenges in locating feasible solutions, often getting trapped in infeasible regions and struggling to maintain a trade-off between convergence and diversity. To address these challenges, a diffusion model-based preference-guided learning algorithm, called DMPGL, is proposed. Specifically, a preference-based partitioning strategy is designed to divide deployment solutions into positive and negative sets based on solution feasibility, service latency, energy consumption, and load balancing. Then, a diffusion model is trained on the positive set to learn its distribution and generate high-quality ESP solutions. Furthermore, a classifier-guided fine-tuning strategy is introduced, which leverages discriminative information from both positive and negative samples to guide the generative process toward feasible regions. Experimental evaluations on Shanghai Telecom network traces and the Australian edge-computing dataset demonstrate the superiority of DMPGL over state-of-the-art algorithms in solving high-dimensional ESP problems under different deployment scenarios.