Deep reinforcement learning-based edge computing offloading algorithm for software-defined IoT

Abstract

Edge computing offloading can effectively solve the problem of insufficient computing resources for terminal devices and improve the performance and efficiency of the system. When network states and tasks change rapidly, data-driven intelligent algorithms have difficulty obtaining comprehensive statistics for accurate prediction, resulting in degraded performance of computational offloading and difficulty in adaptive adjustment. It is a current challenge to improve the environment-aware, intelligent optimization so that the computational offloading algorithm can adapt to the dynamic changes in network state and task demands, thus achieving global multi-objective optimization. This paper presents optimized edge computing offloading algorithm for software-defined IoT. First, to provide global state for making decisions, a software defined edge computing (SDEC) architecture is proposed. The edge layer is integrated into the control layer of software-defined IoT, and multiple controllers share the global network state information via east–west message exchange. Moreover, an edge computing offloading algorithm in software-defined IoT (ECO-SDIoT) based on deep reinforcement learning is proposed. It enables the controllers to offload the computing task to the most appropriate edge server according to the global states, task requirements, and reward. Finally, the performance metrics for edge computing offloading were evaluated in terms of unit task processing latency, load balancing of edge servers, task processing energy consumption, and task completion rate, respectively. Simulation results show that ECO-SDIoT can effectively reduce task completion time and energy consumption compared with other strategies.