Multi-Objective Decomposition Evolutionary DRL for UAV-Assisted MEC in Internet of Vehicles

Dynamic multi-objective optimization in Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) for Internet of Vehicles (IoV) faces significant challenges, due to complex operational environments and conflicting objectives. While Deep Reinforcement Learning (DRL) enables real-time optimization, conventional weighted-sum approaches fail to balance these objectives effectively. To address this, we propose a Multi-Objective Decomposition Evolutionary DRL (MODE-DRL) framework, which include the following three innovative aspects. Firstly, a multi-objective optimization model is developed, aiming to minimize delay and energy consumption while maximizing the number of completed tasks, thus ensuring overall network performance. Secondly, a novel MODE strategy that dynamically associates weight vectors with learning agents to optimize policy distribution and enhance population diversity. Lastly, two integrated algorithms, called MODE with Proximal Policy Optimization (MODE-PPO) and MODE with Deep Deterministic Policy Gradient (MODE-DDPG), are developed to combine DRL's dynamic decision-making with MODE's global optimization capabilities, enabling agents to rapidly adapt strategies based on different weights. Experimental results demonstrate that the MODE-DRL achieves a 33.2% improvement in hypervolume, along with a 16.3% reduction in average delay, a 15.5% decrease in average energy consumption, and a 34.4% increase in average number of completed tasks. These results confirm that MODE-DRL exhibits significant advantages in both convergence and diversity, while enhancing overall network performance. This work provides a scalable paradigm for real-time multi-objective decision-making in UAV-assisted MEC for IoV systems.