Deep Reinforcement Learning with Local Attention for Single Agile Optical Satellite Scheduling Problem.

Area of Science:

• Computer Science
• Aerospace Engineering
• Operations Research

Background:

• The single agile optical satellite scheduling problem is critical for earth observation due to increasing data demands.
• Traditional methods struggle with complex constraints and large solution spaces, leading to high computational costs.
• An efficient and scalable solution is needed to optimize satellite scheduling.

Purpose of the Study:

• To develop an efficient deep reinforcement learning algorithm for the single agile optical satellite scheduling problem.
• To address complex constraints and optimize the profit ratio of completed tasks.
• To improve solution quality and computational efficiency compared to existing methods.

Main Methods:

• A mathematical model was established to define the satellite scheduling problem and its constraints.
• A deep reinforcement learning framework with an encoder-decoder structure and a local attention mechanism was proposed.
• An adaptive learning rate strategy was integrated to enhance actor-critic training effectiveness.

Main Results:

• The proposed deep reinforcement learning algorithm demonstrated superior performance in solution quality.
• The algorithm showed improved generalization capabilities across different problem instances.
• Significant gains in computation efficiency were observed compared to conventional methods.

Conclusions:

• The developed deep reinforcement learning algorithm offers an effective solution for the single agile optical satellite scheduling problem.
• The local attention mechanism and adaptive learning rate strategy enhance the algorithm's performance and training.
• This approach provides a promising direction for optimizing future earth observation missions.