A deep reinforcement learning agent for distributed task offloading based on Graph Neural Networks

In this work, we propose a distributed task-offloading framework based on deep reinforcement learning (DRL) and graph neural networks (GNNs). Each edge node hosts an agent that autonomously decides whether to execute incoming tasks locally or forward them to neighboring nodes using local and two-hop network information. The problem is formulated as a Markov decision process, and a tailored reward function is designed to encourage deadline-compliant execution while reducing CPU usage and energy consumption. The agent is trained with Proximal Policy Optimization (PPO) over realistic MEC topologies and diverse task-generation patterns. Simulation results show that the proposed approach outperforms shortest-path-based greedy baselines in the most demanding scenarios while achieving comparable performance elsewhere.