Dynamic Computation Offloading With Energy Harvesting Devices: A Graph-Based Deep Reinforcement Learning Approach

We study a joint partial offloading and resource allocations (JPORA) problem for the mobile edge computing (MEC) with energy harvesting (EH), where the number of MDs, computation task, and harvested energy of mobile device (MD) are highly dynamic. It is critical to acquire an algorithm that optimally adapts JPORA decisions. Recent studies employ deep deterministic policy gradient (DDPG) agent to tackle the JPORA problem. However, traditional DDPG cannot generalize well to different MEC network scale, due to the deep neural networks in DDPG can only extract latent representation from Euclidean data, with the characteristics of MEC network structural information ignored. To this end, by taking advantage of the graph-based relationship deduction ability form graph convolutional networks (GCN) and the self-evolution ability from the experience training of DDPG, we propose a centralized GCN-DDPG agent to learn making decisions for MDs, including offloading ratio, local computation capacity, and uplink transmission power. Experimental results show that the proposed GCN-DDPG provides significant performance improvement over a number of state-of-the-art DRL agents.