A Study of On-Device Deep Reinforcement Learning for Task Offloading under Dynamic 5G Channel Conditions

Multi-Access Edge Computing (MEC) is a paradigm that enables Internet-of-Things (IoT) applications and devices to run tasks in different locations, from IoT devices to servers in the Cloud. This way, less capable devices can offload computation loads to more powerful or available servers. However, choosing the optimal location for a particular task can be complex due to the features of each location and restrictions of the task. For this, numerous approaches in the literature adopt a centralized strategy for the computation offloading decision, which introduces a single point of failure and can be a bottleneck for resource-demanding applications. In this work, we propose a decentralized Deep Reinforcement Learning (DRL) agent to solve the choice of computing locations, and its assessment in a real testbed. This testbed is formed of an end-user device running the agent, which connects to a MEC server and a Cloud server through 5G. We compare the algorithm against four alternatives, one based on another DRL approach, and analyze their performance in terms of meeting the computing tasks’ requirements and energy consumption in the User Equipment (UE), where synthetically generated tasks are executed or offloaded. DRL algorithms are shown to provide the best tradeoff between performance and energy consumption in changing conditions.