解構多代理人協作：架構設計方法論構築;Deconstructing Multi-Agent Collaboration: A Methodology for Architectural Design Construction

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98166

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98166

题名:	解構多代理人協作：架構設計方法論構築;Deconstructing Multi-Agent Collaboration: A Methodology for Architectural Design Construction
作者:	鄧祺文;Teng, Chi-Wen
贡献者:	資訊工程學系
关键词:	大型語言模型;多代理人系統;架構最佳化;Large Language Model (LLM);Multi Agent System (MAS);System Architecture Optimization
日期:	2025-06-24
上传时间:	2025-10-17 12:27:03 (UTC+8)
出版者:	國立中央大學
摘要:	隨著大型語言模型（LLM）的快速發展，智慧代理人（Agent）技術成為實現通用人工智慧（AGI）的關鍵路徑之一。然而，當前面對複雜真實世界問題時，單一代理人能力有限，多代理人系統（Multi-Agent System, MAS）雖具潛力，卻常因協作機制設計不足，面臨資訊傳遞失真、誤差累積放大、以及協作模式僵化等挑戰，限制了其處理複雜動態任務的效能與可靠性。為應對此挑戰，本研究提出一套新穎且系統化的多代理人系統設計方法論。此架構借鑒成熟之MIAT 方法論思想，創新性地整合了IDEF0 系統架構與GRAFCET 離散事件建模技術。在宏觀層次，運用IDEF0 的階層化與模組化原則，將複雜系統自頂向下分解為功能明確、介面清晰的代理人模組，精確定義各模組的輸入、輸出、控制與機制（ICOMs），從而降低設計模糊性，提升系統的模組化與可維護性。在微觀層次，引入GRAFCET 為各代理人進行精確的離散事件建模，透過定義原子行動序列與狀態轉移條件，細緻刻畫代理人的循序、平行與選擇邏輯。尤為關鍵的是，此方法能有效約束LLM 驅動之代理人的潛在隨機性，將其行為納入可控的離散框架內，提升系統行為的可預測性與可靠度。本研究闡述了整合IDEF0 宏觀架構與GRAFCET 微觀邏輯的系統合成方法，透過預先定義的介面與互動協議，使LLM 能專注於核心推理任務。本研究為驗證方法論的效能，執行了設計概念驗證與多代理人小說創作評測。在設計概念驗證階段，針對六個小型語言模型（參數≤ 30億）於複雜協同研究情境測試，相較於基準架構，本方法論在函數模型建構與離散事件建立的迭代次數上均顯著減少，錯誤修正週期亦大幅縮短（幅度達32.7% 至50.3%），且多數受測模型（五個）能成功完成任務。在以GPT-4o 為骨幹模型的小說創作評測中，本方法論與現有設計進行比較，於多個評估維度均展現領先，總體性能最優，初步驗證了本方法論在提升MAS 設計效率與穩健性方面的潛力。本研究貢獻在於提出整合IDEF0 與GRAFCET 的創新MAS 設計方法論，為應對當前LLM 驅動之MAS 在協作效率、可靠性及設計複雜度方面的挑戰提供解決思路，並為開發更高效、穩健的複雜AI 系統奠定理論與工程基礎。;The rapid development of Large Language Models (LLMs) positions agent technology as key to Artificial General Intelligence (AGI). However, single agents struggle with complex real-world problems. Multi-Agent Systems (MAS), despite their potential, suffer from challenges like information distortion, error amplification, and rigid collaboration patterns due to inadequate mechanism design, limiting their performance and reliability in dynamic tasks. This research proposes a novel, systematic MAS design methodology to overcome these limitations. Inspired by MIAT, it integrates IDEF0 system architecture and GRAFCET discrete event modeling. IDEF0’s macro-level hierarchical and modular principles guide top-down decomposition into agent modules with defined functions, interfaces, and ICOMs, reducing ambiguity and enhancing modularity. At the micro level, GRAFCET precisely models agent behavior (sequential, parallel, and selective logic) via defined atomic action sequences and state transitions. This crucially constrains LLM-driven agent randomness within a controllable discrete framework, improving system predictability and reliability. The study details a system synthesis method integrating IDEF0 macroarchitecture and GRAFCET micro-logic, allowing LLMs to focus on core reasoning via pre-defined interfaces. Validation involved proof-of-concept (PoC) experiments and multi-agent novel writing evaluations. PoC tests with six small LLMs (≤ 3B parameters) in collaborative research scenarios showed significantly reduced iterations for function modeling and discrete event establishment, and shortened error correction cycles (by 32.7%-50.3%), with five models succeeding. In novel writing evaluations using GPT-4o, our methodology outperformed existing designs in multiple evaluation dimensions. These results preliminarily validate its potential for enhancing MAS design efficiency and robustness. This research contributes an innovative MAS design methodology integrating IDEF0 and GRAFCET, addressing challenges in collaboration efficiency, reliability, and design complexity for LLM-driven MAS. It establishes a theoretical and engineering foundation for developing more efficient and robust complex AI systems.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	8	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....