Experience-based Computation Offloading by Deep Reinforcement Learning for Multi-access Edge Computing Network

Abstract

Recently, researchers have focused on a new paradigm called Multi-access Edge Computing (MEC), which ensures reduce the execution time for computation constrained edge devices through computation offloading techniques. Despite the exciting research on optimizing centralized task scheduling problems in the MEC architecture bringing several advantages, a few challenges, such as efficiently optimizing algorithms in dynamic and large-scale networks, are still a conundrum. This article investigates a computation offloading scheduling problem in a dynamic network. The tasks generated by the edge devices can be executed locally or sent to edge servers for execution based on the time-varying MEC network. Moreover, we formulate the computation offloading optimization problem as a Markov Decision Process (MDP) model. Furthermore, to improve the learning and convergence efficiency, we propose an Experience-Based replay Reinforcement Learning algorithm (EBRL) by collecting significant transformations and leveraging the most valuable knowledge from the experience pool. Experimental results show that our proposed algorithm effectively achieves faster convergence speed and reduces the system delay than other benchmarks in a dynamic MEC network.