應用大型語言模型多代理人模擬建築專案風險之初探

Abstract

營建專案中複雜的分包商結構與交互行為，形成了工作定義模糊的營建生態，進而使工程相關利益者面臨工程遲延、成本超支或其他不可歸責之風險。目前藉由大型語言模型的開發，我們可透過自然語言指導大型語言模型(LLM) 完成目標明確之任務，並且亦能建立大型語言模型(LLM) 代理人系統來拆解模糊定義之任務以達到更好的成果。本研究旨在通過開發一個基於大型語言模型的多代理系統（LLM-MA）框架模擬營建專案中利益相關者的行為，從而加深對營建專案中契約關係的理解。本研究利用賽局理論整合於大型語言模型多代理（LLM-MA），建立了架構地圖去描繪如欲建立大型語言模型多代理系統建築專案(LLM-MA) 所需要的元素，並透過角色扮演遊戲測試LLM 代理的能力。實驗過程允許觀察微觀行為，例如模擬角色決定接受或協商提案，以及專案的宏觀狀態，包括利害關係人關係與合約付款。結果證明，LLM 代理產生的類似利益關係人的行為，可以符合賽局中的策略行為，此實驗架構也展現未來潛力於增強建築專案中合約風險分析技術。

Construction projects involve complex cooperation structures due to contracting, and ambiguous job definitions can lead to delays and unforeseen risks. Recent advancements in natural language processing enable the use of Large Language Models (LLMs) to guide task decomposition and create LLM-based agent systems capable of handling complex tasks. This study aims to enhance our understanding of contracting relationships in construction projects through developing a Multi-Agent System (MAS) framework that utilizes LLM-based agents to simulate stakeholder behaviors. By integrating game theory into Large Language Model based Multi-Agents (LLM-MA), this work identify a roadmap for building LLM-MA of construction project and test the abilities of LLM agent by roleplaying game. The experiment process allows for observing micro-behaviors, such as simulated roles deciding to accept or negotiate proposals, as well as macro-states of a project, including stakeholder relationships and contract payments. The results demonstrate that human-like behaviors generated by LLM agents can align with strategic approaches to cooperative problems. Furthermore, this experimental framework shows potential for enhancing contractual risk analysis techniques in construction projects.