Evolutionary Multi-Objective Deep Reinforcement Learning for Autonomous UAV Navigation in Large-Scale Complex Environments

Autonomous navigation of Unmanned Aerial Vehicles (UAVs) in large-scale complex environments presents a significant challenge in modern aerospace engineering, as it requires effective decision-making in an environment with limited sensing capacity, dynamic changes, and dense obstacles. Reinforcement Learning (RL) has been applied in sequential control problems, but the manual setting of hyperparameters, including reward functions, often results in suboptimal solutions and inadequate training. To address these limitations, we propose a framework that combines Multi-Objective Evolutionary Algorithms (MOEAs) with RL algorithms. The proposed framework generates a set of non-dominating parameters for the reward function using MOEAs, leading to diverse decision-making preferences, efficient convergence, and improved performance. The framework was tested on the autonomous navigation of UAVs and demonstrated significant improvement compared to traditional RL methods. This work offers a novel perspective on the problem of autonomous UAV navigation in large-scale complex environments and highlights the potential for further improvement through the integration of RL and MOEAs.