Dynamic task scheduling in wireless edge computing using deep reinforcement learning with ordinal optimization

Abstract In latency-critical Internet of Things (IoT) applications, multi-access edge computing (MEC) in wireless networks reduces core network strain by pushing computation and data resources to the edge. However, the limited computing power of edge servers struggles to handle numerous computation-intensive tasks. To improve the quality of service (QoS) for IoT users, task scheduling in wireless edge computing to effectively allocate the computing resource becomes particularly urgent. The traditional game theory-based or auction-based scheduling approaches cannot adjust policies in dynamically changing environments. Moreover, the high-dimensional action space of the current deep reinforcement learning (DRL)-based methods causes them to converge slowly. To address this, we propose a novel deep reinforcement learning-based task scheduling algorithm that leverages ordinal optimization to enhance convergence speed and learning efficiency. The proposed method enables dynamically adaptive scheduling of tasks and computing resources across cloud centers and edge servers, aiming to minimize average service latency and energy consumption in IoT wireless edge networks. By integrating ordinal optimization with DRL, we achieve an improved tradeoff between exploration and exploitation, maximizing long-term utility in dynamic environments. Simulation results demonstrate that our approach significantly reduces latency and energy use compared to conventional and state-of-the-art techniques, with rapid convergence suitable for real-time applications.