摘要
对多智能体系统(Multi-Agent System,MAS)在建筑数字技术领域的研究进展进行综述性研究。选取近百个相关研究进行归纳总结,将研究原型根据智能体代理的主体不同分为:多智能体代理人、多智能体代理建筑空间单元、多智能体代理城市空间单元、其他代理类型等类别。从研究内容、智能体的控制指标、具体运行规则、研究所对应的建筑学问题以及研究所属的建筑学本体维度等角度分别分析各研究案例特征并进行统计。通过统计分析总结多智能体系统在建筑学本体维度研究中的优势和局限性。希望统计分析的结果能为未来多智能体系统在建筑及规划领域的研究提供建议和指南,即面临相应问题时可选择何种智能体代理模式、控制指标及运行规则。此外,还分析比较了MAS的不同研究平台和研究工具。最后,提出结合了深度强化学习的多智能体系统是其在建筑学本体维度研究值得探索的方向。
This study reviews the research progress on the Multi-Agent System(MAS) in the field of architecture's digital technologies.The first part of this paper contains three elements: first, the Citespace keyword time map was analyzed to understand recent research issues and trends of MAS in the architecture and planning fields. Secondly, seven main ontology dimensions of architectural generative design were introduced briefly, including spatial combination, plane layout, flow, form, structure, building facade,and pipeline equipment. Since MAS is widely used in the first three dimensions, these are the focus of this study. Thirdly, this paper gives a brief introduction to MAS.The second part of this paper summarized nearly 100 related studies. The research prototypes were classified into four categories according to intelligent agents, including multi-agents, multi-agent architectural space units, multi-agent urban space units, and others.Among them, multi-agent architectural space units and multi-agent urban space units were unified as "space units". The multi-agent and agent space units were analyzed and summarized, respectively. A statistical analysis on research cases of different types was carried out from perspectives of research contents, control indicators of intelligent agents, specific running rules, relevant architectural problems, and architectural ontology dimension of the study.The third part of this paper summarized the advantages and limitations of multi-agent system studying architectural ontology dimension through statistical analysis. According to analysis and statistical results, it was found that research on categories of agents mainly focuses on the flow problem in the architectural ontology dimension, and the corresponding agent control indicators are mainly pheromone accumulation, site accessibility, and destination setting. Research on categories of agent space unit focuses on plane layout, space combination, and planning in the architectural ontology dimension. The corresponding agent control indicators include proximity control(topological control), quantity control,spacing control, etc. This type of research emphasizes architectural generative design and involves many changes to control rules according to research problems. This phenomenon illustrates the flexibility of MAS tools in solving problems of different architectural dimensions. In addition, the expression forms of MAS model were also summarized. The categories of agents are mainly expressed by moving points, while categories of agent urban space units are mainly expressed by grids. There are diversified expressions under other categories, including but not limited to within bubbles, moving points, Voronoi diagrams, grids, polygons, graphs, etc.This study is expected to provide guidance for future studies on MAS in architecture and planning, such as which agent mode should be chosen according to relevant problems, control indicators, and running rules. Additionally, different research platforms and research tools of MAS were compared. Finally, it is proposed that the multiagent system combined with deep reinforcement learning is a direction worthy of exploration in the ontology dimension of architecture.
作者
马成也
宋明星
MA Chengye;SONG Mingxing
出处
《南方建筑》
CSCD
北大核心
2023年第5期56-67,共12页
South Architecture
基金
国家自然科学基金委员会青年项目(52008033):有流条件下建筑通风管道-共鸣器系统的声学特性及流声耦合作用机理研究
湖南省自然科学基金(2023JJ30148):气候适应下长株潭城市群城市空间韧性评估与优化研究。
关键词
多智能体系统
建筑数字技术
生成设计
人流仿真模拟
multi-agent system
architecture's digital technologies
generative design
flow simulation