Deep Reinforcement Learning-Empowered Trajectory and Resource Allocation Optimization for UAV-Assisted MEC Systems

In this letter, we address the energy minimization problem for the UAV-assisted MEC system under the long-term dynamic environment by jointly optimizing UAV trajectory, computation resource allocation and offloading decisions. The formulated optimization problem is modeled as a constrained Markov decision process (CMDP) to obtain a sequential optimization decision, where the optimization variables are coupled over multiple time slots. A double parametrized deep Q-network (DPDQN)-based algorithm is proposed for trajectory planning and computation resource allocation. We incorporate penalty and prioritized experience replay (PER) mechanisms to handle the large action space and multi-slot coupling. Simulation results validate that the proposed algorithm significantly reduces energy consumption.