On-Device Deep Reinforcement Learning for Decentralized task offloading: Performance trade-offs in the training process

Task offloading enables resource-constrained devices to execute demanding computing tasks by leveraging remote edge or cloud servers. However, deciding where and why to run each task is complex. Therefore, in this work, we propose a decentralized Deep Reinforcement Learning (DRL) agent that enables each device to independently select the appropriate computing location while adapting its decisions in real time based on local observations. Unlike most existing studies based on simulations and/or centralized optimization, we evaluate the proposed approach in a real-world 5G testbed composed of heterogeneous edge devices, a Multi-access Edge Computing (MEC) server, and a Cloud server. Our focus is on assessing the practical impact and feasibility of deploying DRL models directly on devices for decentralized decision‑making. We examine not only the agent’s ability to meet tasks’ requirements in dynamic environments, but also the practical implications of running DRL locally on resource-constrained hardware. We demonstrate that a simple and lightweight DRL algorithm can run directly on edge devices while still delivering competitive performance. Our experiments show that local, on‑device training not only provides good performance but also outperforms remote training due to the absence of communication latency. Moreover, we show that the energy overhead of training and inference on the devices is not excessive, making the approach practical for real deployments. Finally, we analyze the trade-offs between local and remote training in terms of latency, energy consumption, and task completion performance, providing concrete insights into the feasibility of decentralized, on-device intelligence for task offloading in real networks.