Reinforcement Learning with Heterogeneous Graph Attention Mechanism Assisted Transformer for Scheduling Amphibious Unmanned Surface Vehicles
Hui Yu, Kaizhou Gao, Rong Su, Peiyong Duan, Ponnuthurai Nagaratnam Suganthan
Inner Mongolia University Macau University of Science and Technology Nanyang Technological University Qilu University of Technology
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
With the marine resource development and environmental protection, the demand for amphibious unmanned surface vehicles (AUSVs) is increasing for executing diverse tasks in complex and dynamical marine environments. However, limited research about AUSV scheduling problems has been conducted. This work focuses on the multi-objective heterogeneous AUSV scheduling problems with obstacle and ocean current constraints. First, a mathematical model is established to minimize both maximum completion time and total cost. Second, an improved A* algorithm is employed to reprocess the routes between different task points by avoiding obstacles. Third, a reinforcement learning framework assisted by a Transformer with a heterogeneous graph attention mechanism is developed. Specifically, 1) an action mechanism incorporating multiple local search operators is devised to iteratively refine the current task sequence; 2) a problem-feature-based heterogeneous graph attention mechanism that integrates node, position, and edge features is constructed to handle complex heterogeneous environments; and 3) an actor-critic network for multi-objective multi-AUSV task optimization is constructed, with proximal policy optimization (PPO) incorporated for strategy optimization. Finally, the mathematical model is verified on the small-scale instances by using an exact solver. The effectiveness of the improved strategies is evaluated on 20 real scenarios of different scales. Our newly designed algorithm is compared with four heuristic rules, three meta-heuristics, and two learning-based algorithms. The experimental results and discussions demonstrate its competitiveness.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
工程Scheduling and Optimization Algorithms
Vehicle Routing Optimization Methods · Maritime Navigation and Safety