Article ID: 2025EAL2009
The Agile Earth Observation Satellite Constellation Mission Planning (AEOSCMP) problem focuses on optimizing target selection and scheduling for multiple satellites to maximize global observation rewards while adhering to resource constraints. To tackle this challenging task, this letter employs the Multi-Agent Transformer (MAT) to convert the joint policy search problem into a sequential decision-making process, optimizing observation policies through the attention mechanism. This approach could provide a theoretical guarantee of monotonic improvement during online training, ensuring consistent and reliable performance enhancements. Experimental results demonstrate that MAT achieves superior observation efficiency compared to state-of-the-art Multi-Agent Reinforcement Learning (MARL) methods.
