.png){kind=icon}
.png){kind=icon}
Return
A dual-layer bilevel optimization model to complete algorithm evolution in agile satellite task scheduling problem using Large Language Models
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.115533.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The Agile Imaging Satellite Task Scheduling (AISTS) is an NP-hard optimization problem. To address diverse AISTS scenarios, the development of efficient and adaptive algorithms is imperative. Learning-driven algorithmic automation can assist manual heuristic engineering in this task setting, but existing LLM-based evolutionary studies have mainly focused on simpler traditional optimization problems rather than the specialized constraints of AISTS. To bridge the gap between the generalist nature of LLMs and the specialized domain of AISTS, we design a human-architected LLM-based Bilevel Optimization Model (BOM) to support offline heuristic synthesis for AISTS. This framework organizes the LLM’s generative process into two hierarchical levels. In the upper level, a human-devised evolutionary mechanism optimizes prompts and refines the instructions given to the LLM. In the lower level, guided by these optimized prompts, the LLM operates within a fitness-driven evolutionary framework to craft domain-adapted, high-performance heuristics by combining and modifying algorithmic components. Because AISTS is also naturally modeled as a BOM with task-assignment and single-satellite scheduling subproblems, the proposed dual-layer framework aligns algorithm synthesis with the problem structure. Through AAD for the two subproblems within the outer BOM, we devise high-performing algorithm pairings for the evaluated AISTS scenarios. Empirical evidence demonstrates that, within the evaluated AISTS domain, the synthesized composite heuristics improve completion rate over the tested human-designed baselines by up to 15.41% while retaining lightweight online execution, with an online runtime equal to only 1.64% of the slowest tested baseline in the runtime comparison.
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W