Multi-agent deep reinforcement learning based multi-task partial computation offloading in mobile edge computing

Abstract

Mobile edge computing (MEC) can enhance the computation performance of end-devices by providing computation offloading service at the network edge. However, given that both end-devices and edge servers have finite computation resources, inefficient offloading policies may lead to overload, thereby increasing the computation delays of tasks. In this paper, we investigate a multi-task partial computation offloading problem combined with a queue model. Based on achieving load-balancing across the MEC system, our objective is to minimize the long-standing average task-processing cost of the end-devices while ensuring the delay thresholds of tasks. For this purpose, a distributed offloading algorithm utilizing the multi-agent deep reinforcement learning (MADRL) method is proposed. Specifically, through interacting with the MEC environment and accumulating experience data, the device agents can collaborate to optimize their local offloading decisions over continuous time-slots, which includes adjusting the transmission power and determining the tasks’ offloading ratios under the dynamic wireless channel conditions. Exhaustive experimental results demonstrate that in contrast with the baseline algorithms, the proposed offloading algorithm can not only better balance the computation loads between the end-devices and the MEC server, but also more effectively reduce the task-processing cost of the end-devices, as well as the percentage of timeout tasks.