大语言模型驱动的关键基础设施网络恢复决策的智能体构建OA

To promote the intelligentization of post-disaster recovery decision-making(RDM)for critical in-frastructure networks(CIN),a CIN-RDM agent driven by large language models(LLMs)was proposed.First,a CIN-RDM toolbox and the corresponding tool knowledge graph were constructed.A tool filtering-rea-soning and acting(TF-ReAct)agent architecture was put forward.Then,the CIN-RDM agents based on 6 LLMs were developed,whose performance were evaluated and compared.The results show that compared with the traditional ReAct architecture,the TF-ReAct architecture improves the task completion rate of the agents by 41.1%and reduces the redundant action rate by 86.2%on average.The TF-ReAct agents driven by GPT-4 and GPT-4o achieve a task completion rate of 1.0 while eliminating redundant actions.This study con-tributes to enhancing the tool-use capabilities of agents,as well as facilitating the efficient application of CIN-RDM tools by infrastructure managers.