基于多Agent协作的智能应急决策支持系统
Intelligent Emergency Decision Support System Based on Multi-Agent Collaboration
摘要
针对传统应急决策系统存在的数据来源单一、智能化程度低、决策路径固化等问题,我们提出一种基于多Agent协作的智能应急决策支持系统。系统采用了我们提出的Plan-Execute-Monitor循环架构,集成Web搜索、知识图谱查询和地理信息服务等多源信息融合模块,构建了思维树推理驱动的多智能体协作机制。通过引入计划导向路径生成、多维度进展评估和自适应执行监控等关键技术,解决了传统多智能体系统指令遵循失败、步骤重复和上下文丢失等问题。我们参考GAIA评估框架,基于政府开源文件和网络爬虫自主构建了包含135个任务的GAIA-应急管理领域数据集。在该数据集上的实验结果表明,PEM架构准确率达到48.7\%,比传统迭代搜索架构的28.6\%高出20.1个百分点,平均执行时间降低63.7\%,验证了系统的有效性。Abstract
To address the limitations of traditional emergency decision systems including single data sources, low intelligence levels, and rigid decision pathways, we propose an intelligent emergency decision support system based on multi-agent collaboration. The system employs a Plan-Execute-Monitor (PEM) cyclic architecture, integrating multi-source information fusion modules including Web search, knowledge graph querying, and geographic information services, and constructs a Tree-of-Thoughts-driven multi-agent collaboration mechanism. By introducing key techniques such as plan-oriented path generation, multi-dimensional progress evaluation, and adaptive execution monitoring, the system resolves critical issues in traditional multi-agent systems including instruction-following failures, step repetition, and context loss. We constructed a GAIA-Emergency Management domain dataset containing 135 tasks based on government open-source files and web crawling, referencing the GAIA evaluation framework. Experimental results on this dataset show that the PEM architecture achieves an accuracy of 48.7\%, outperforming the traditional iterative search architecture's 28.6\% by 20.1 percentage points, with an average execution time reduction of 63.7\%, validating the system's effectiveness. 
评论