Multi-objective task offloading optimization using deep reinforcement learning with resource distribution clustering

Task offloading in multi-access edge computing (MEC) systems is critical for managing computational tasks in dynamic urban environments. Existing strategies face challenges such as high communication overheads and regional performance deviations including centralized and distributed methods. Clustering approaches have been explored to address these issues, yet they often rely on physical proximity to form clusters, overlooking the variability in task rate distributions across edges. To overcome these limitations, this paper proposes a graph-driven inter-cluster resource distribution (GIRD) clustering scheme that clusters edge nodes based on task request distribution and computing resource status, ensuring similar resource utilization across clusters. Building on this, a proximal policy optimization (PPO)-enabled intra-cluster task offloading algorithm (PITO) is introduced to determine one execution server for task offloading—either an edge server within a cluster or a cloud server—using various network state information. This dynamic decision-making process optimizes a multi-objective function that includes task processing delay, consumed energy, success rate, and cloud cost. Simulation results demonstrate the proposed GIRD-PITO framework achieves superior task success rates, reduced delays, and improved regional performance fairness, making it a promising solution for large-scale MEC systems. 2018 The Korean Institute of Communications and Information Sciences. Publishing Services by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).