LLM multi-agent tool automates satellite scheduling algorithm design

A new multi-agent system named AgentAD, detailed in a recent paper published in Engineering, leverages large language models (LLMs) to autonomously generate executable optimization algorithms for the Earth Observation Satellite Scheduling Problem (EOSSP), reducing manual labor required to build scenario-specific scheduling logic.

EOSSP is classified as an NP-Hard optimization task that demands matching satellite observation windows, task priorities, onboard power and storage limits, and satellite attitude transition time constraints to maximize completed observation tasks. Traditional workflow requires aerospace and optimization experts to manually build, test, and revise standalone algorithms for every unique satellite-target scenario, a process marked by heavy time and resource costs. While prior LLM-powered automatic algorithm design (AAD) research exists, most frameworks target simple combinatorial benchmarks such as traveling salesman problems, and they rely on predefined baseline algorithm structures that still demand human algorithm expertise to operate. EOSSP further raises technical barriers due to its domain-specific professional terminology and layered constraint rules, which complicate direct natural language-to-code conversion with single LLMs.

To resolve these limitations, the research team constructed AgentAD, a collaborative multi-agent system with five specialized LLM-driven roles: Algorithm Designer (AD), Algorithm Programmer (AP), Algorithm Optimizer (AO), Code Tester (CT), and Project Manager (PM). Each agent operates under tailored prompting paradigms that define fixed roles, task boundaries, input-output standards, and cross-agent communication protocols, with the overall workflow mirroring real-world software development split into four sequential phases: design, optimization, testing, and documentation. The system converts unstructured natural language descriptions of EOSSP scenarios into functional Python code through repeated atomic dialogue cycles between paired agents. Three core supporting mechanisms underpin stable collaboration: role specialization that restricts each agent to narrow professional duties, a permanent memory stream that archives all historical dialogue to avoid repeated errors, and chain-of-coding that breaks code writing and debugging into sequential subtasks to mitigate LLM hallucinations.

Within the four-phase pipeline, the AD first interprets raw scenario input and drafts initial algorithm frameworks including greedy, dynamic programming, and genetic structures. The PM coordinates agent communication and passes preliminary algorithm outlines to the AO, which proposes structural and parametric improvements to raise task completion performance. The AP translates revised algorithm logic into runnable annotated code, which the CT validates on EOSSP datasets and reports bugs or performance feedback back to the AO for iterative refinement. This optimization-test loop repeats until convergence or predefined iteration limits, after which the PM compiles all dialogue records and algorithm iteration trails into standardized documentation. Ablation experiments in the paper confirm the AO serves as the most indispensable component for performance improvement, followed by the CT for accurate feedback provision, while the memory stream delivers greater performance gains than the chain-of-coding module.

The team ran comparative trials across basic EOSSP settings with static constraints and complex scenarios augmented by dynamic task priorities, satellite forbidden time windows, and minimum imaging quality thresholds, benchmarking AgentAD-generated algorithms against established human-designed heuristics and alternative LLM algorithm design platforms including LLM4AD. The paper notes that algorithms autonomously evolved by AgentAD achieve higher task completion rates across most test environments, though the performance margin narrows as satellite and task scales expand due to increased contextual comprehension burdens on LLMs. Convergence tracking shows AgentAD’s fitness metric F_i​(a) rises steadily through iterative modification after identifying viable optimization trajectories, and trials across GPT-3.5-Turbo, GPT-4o, DeepSeekV3 and Llama3.3 reveal more capable LLMs cut total atomic dialogue counts and token consumption without substantial shifts in final code length. For future expansion, the authors outline plans to integrate multimodal satellite cloud data, introduce human expert feedback loops, and build visual interactive dashboards to enhance real-world operability of the automated algorithm generation pipeline.

The paper “A Large Language Model-based Multi-Agent Framework to Autonomously Design Algorithms for Earth Observation Satellite Scheduling Problem,” is authored by Jiawei Chen, Yingguo Chen, Duc Truong Pham, Yanjie Song, Jian Wu, Lining Xing, Yingwu Chen. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.10.027. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.