簡易檢索 / 詳目顯示

回結果列表
研究生： 巫政和
Wu, Cheng-Ho
論文名稱： 建築設計的閉環模式：具自我修正與工作流程重組之人工智慧系統
A Closed-Loop Model for Architectural Design: An Artificial Intelligence System with Self-Revision and Workflow Reconfiguration
指導教授： 鄭泰昇
Jeng, Tay-Sheng
宋立文
Sung, Li-Wen
學位類別： 碩士
Master
系所名稱： 規劃與設計學院 - 建築學系
Department of Architecture
論文出版年： 2026
畢業學年度： 114
語文別： 中文
論文頁數： 172
中文關鍵詞： 建築人工智慧 、閉環設計系統 、多代理人協作 、自我修正 、工作流程重組
外文關鍵詞： Architectural Artificial Intelligence, Closed-Loop Design System, Multi-Agent Collaboration, Self-Revision, Workflow Reconfiguration
相關次數： 點閱：5  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 當代建築設計是一項高度複雜且反覆迭代的專業活動，設計者需在空間機能、形式表現、環境回應、結構合理性與使用需求之間持續協調。近年生成式人工智慧（Generative Artificial Intelligence）雖已能加速概念發想、圖像生成與文字整理，但多數工具仍停留於單向輸入與單次產出的開環（Open-Loop）模式，缺乏對生成條件、成果偏差、修正依據與工具失敗的系統性回收機制。
    本研究以建築設計的閉環（Closed-Loop）模式為核心，建構一套具自我修正與工作流程重組能力之人工智慧系統。OpenClaw 與 Codex 在本研究中構成受人類監督的協作與跨平台操作層，負責任務編排、工具調用、本機資料處理與操作紀錄；xFigura 作為節點化多模態生成平台，Rhino 則作為幾何延伸與檢視環境。系統將可重複的操作、檢核與失敗恢復整理為可追蹤流程。
    流程由設計意圖轉譯、多模態生成、設計延伸建模與設計價值評估四個功能代理人構成。四者依序處理業主需求與跨媒材手法、節點化圖像與模型生成、Rhino 多方案延伸，以及圖像與幾何資料的多指標評估。系統將來源檔案、RAG 檢索索引、Memories、Skills 與 Skills Router 分層管理，使知識來源、操作技能、任務經驗與分配規則不被混用，並讓輸入、成果、評估證據與修正條件可跨階段追溯。
    研究以單一住宅案例進行三輪探索性比較，測試系統能否依據基地照片、固定需求與不同跨媒材手法，完成需求轉譯、圖像生成、設計延伸、九方案建模與七項指標評估。結果顯示，第一輪方案分布較廣，第二輪在 LEGO 模矩約束下平均提高且離散縮小，第三輪因音樂方法引入新空間語法而再次擴張；三輪最高分並未單調提升。此結果支持系統能記錄並重組設計搜尋條件，但不證明閉環必然提高建築品質。研究成果應被視為受人類監督、可追蹤且可回饋之建築設計工作流程原型。

    Contemporary architectural design depends on continuous negotiation among spatial, functional, environmental, formal, and material requirements. Generative AI accelerates text, image, and model production, but most tools remain open-loop: they generate from prompts without evaluating failure, revising strategy, or reorganizing workflows. The problem is not generation itself, but the absence of a mechanism that can transform evaluation into the next design action. This thesis proposes a closed-loop model for architectural design with self-revision and workflow reconfiguration, aiming to make design iteration more traceable, evaluable, and reusable.
    The system integrates OpenClaw and Codex as a human-supervised collaboration and cross-platform operation layer, xFigura for node-based multimodal generation, and Rhino for three-dimensional design development. Four functional agents translate design intent, organize multimodal generation, extend alternatives as Rhino models, and evaluate design value. The architecture records inputs, intermediate artifacts, tool states, and evaluation evidence so that failures can be localized and subsequent steps can be reconfigured without claiming fully autonomous architectural design.
    Source files, retrieval-augmented generation (RAG) indexes, memories, skills, and routing rules are managed as distinct knowledge resources. A single narrow-lot housing case was examined through three exploratory rounds using hand drawing, LEGO composition, and music-derived spatial sequencing. The second round showed higher consistency and a narrower score distribution, whereas the third round reopened the search space after introducing a less conventional method. The results indicate conditional convergence followed by re-expansion rather than monotonic improvement.

    摘要 I 誌謝 V 目錄 VI 表目錄 XII 圖目錄 XIV 第一章 緒論 1 1.1 研究背景 1 1.1.1 從數位輔助到生成式設計的工具角色轉變 1 1.1.2 開環生成下的回饋斷裂與設計認知負荷 2 1.1.3 邁向閉環設計智慧的多代理人協作與工作流程重組 4 1.2 研究動機 5 1.2.1 生成式人工智慧在建築設計中的開環生成限制 5 1.2.2 建築設計對閉環回饋與自我修正的需求 6 1.2.3 多工具環境下工作流程重組的必要性 7 1.2.4 設計概念轉譯與性能最佳化之界線 8 1.3 研究目標 9 1.3.1 建立建築設計閉環系統之理論架構 10 1.3.2 建構多代理人分工之設計工作流程 11 1.3.3 整合跨平台工具操作與資料流重組機制 11 1.3.4 驗證閉環流程於建築設計案例中的可行性 11 1.4 研究方法 12 1.4.1 文獻分析與知識轉譯 12 1.4.2 多代理人閉環系統建構 13 1.4.3 跨平台工作流程實作 14 1.4.4 三輪迭代案例驗證 14 1.5 研究架構 16 第二章 文獻回顧 19 2.1 人工智慧輔助建築設計流程之角色演化 19 2.1.1 生成式人工智慧於早期設計探索之應用 19 2.1.2 人機代理協同與建築設計工作流程重組 20 2.1.3 從單次生成到流程型設計智慧 22 2.2 閉環建築設計系統 23 2.2.1 參數化最佳化與設計知識發現 23 2.2.2 數位孿生架構下的虛實雙向同步 24 2.2.3 強化學習於空間配置的動態回饋 25 2.2.4 生成式人工智慧與系統性評估機制 27 2.2.5 代理式迴圈工程與自主執行機制 28 2.3 多代理人協作系統 32 2.3.1 多代理人協作系統之拓樸結構演進 32 2.3.2 閉環協作系統於建築設計流程之重要性 33 2.3.3 代理人擴展定律與專業分工之綜效 35 2.4 多目標評估 36 2.4.1 建築設計最佳化與多目標決策 36 2.4.2 性能導向之衍生式設計與空間探索 38 2.4.3 人工智慧與多代理人系統於多目標評估之整合 40 2.4.4 文獻總結與本研究之啟發 42 2.5 基於人工智慧之既有設計系統 43 2.5.1 生成式設計與性能驅動的最佳化系統 44 2.5.2 基於深度學習與多模態模型之概念生成系統 46 2.5.3 大型語言模型與多代理人協作系統 50 2.5.4 既有系統之侷限與「開環」瓶頸 53 第三章 建築設計閉環系統之架構與代理人機制 54 3.1 建築設計閉環系統架構 55 3.1.1 設計需求轉譯與可執行工作流程 55 3.1.2 四代理人協作與雙層閉環機制 57 3.1.3 共享知識記憶技能與路由架構 59 3.1.4 工作流程重組與人機協作修正 61 3.2 設計意圖轉譯代理人 63 3.2.1 需求與跨媒材設計手法解讀 63 3.2.2 設計知識工程與意圖語言輸出 64 3.3 多模態生成代理人 66 3.3.1 xFigura節點化生成與跨平台執行 66 3.3.2 多模態成果輸出與工作流程保存 68 3.4 設計延伸建模代理人 68 3.4.1 設計解讀與Rhino幾何轉譯 69 3.4.2 建築知識工程與建模規則 69 3.4.3 方案變體生成與品質檢核 71 3.5 設計價值評估代理人 72 3.5.1 圖像模型解析與評估基準 72 3.5.2 多目標評分與手法追溯性 74 3.5.3 評估回饋決策與跨代理人修正 75 第四章 閉環系統實作與迭代驗證 76 4.1 平台整合與跨工具操作測試 76 4.1.1 系統環境與工具配置 76 4.1.2 OpenClaw與Codex跨瀏覽器操作 78 4.1.3 Rhino MCP檔案相容與失敗恢復 80 4.2 四代理人端到端工作流程實作 81 4.2.1 設計意圖轉譯代理人之跨媒材意圖生成 82 4.2.2 多模態生成代理人之節點工作流程與成果輸出 85 4.2.3 設計延伸建模代理人之多方案建模 89 4.2.4 設計價值評估代理人之多指標評估 94 4.2.5 代理人交接與成果完整性檢核 97 4.3 三輪設計生成與方案演化 98 4.3.1 案例條件與比較變項 99 4.3.2 第一輪手繪輸入與探索性展開 100 4.3.3 第二輪積木輸入與約束導向收斂 104 4.3.4 第三輪音樂輸入與方法性再擴張 108 4.4 三輪評估與閉環成效比較 112 4.4.1 評估架構與跨輪比較界線 112 4.4.2 三輪評估結果與方案排序 113 4.4.3 修正紀錄與閉環條件性成效 119 4.5 綜合討論 120 4.5.1 系統整合與工作流程智慧發現 120 4.5.2 研究有效性限制與適用邊界 121 第五章 研究結論與建議 122 5.1 系統研究成果 122 5.1.1 四代理人閉環協作成果 123 5.1.2 三輪迭代之收斂與再擴張結果 124 5.2 研究貢獻 125 5.2.1 理論與研究方法貢獻 125 5.2.2 系統實作與建築設計應用貢獻 126 5.3 技術瓶頸 128 5.3.1 系統自主性與工具操作限制 128 5.3.2 評估方法與實驗設計限制 128 5.3.3 外部效度與知識更新限制 130 5.4 後續研究建議 131 5.4.1 自主工作流程與回饋學習 131 5.4.2 BIM與幾何資訊整合 131 5.4.3 法規及AEC知識持續擴充 132 5.4.4 專家評估與多層驗證 132 5.4.5 長期案例與產業部署 133 5.4.6 從生成結果走向流程智慧 133 參考文獻 135 學位論文 135 期刊論文 135 會議論文與預印本 137 專書 139 網路與軟體資料 139 附錄 141 附錄A 代理交接資料契約與範例 141 附錄B 設計意圖語言與建模操作代號 142 附錄C 七項評估指標原始分數與計算方式 143 附錄D 三輪工作流程與成果檔案索引 146 附錄E 軟體版本、錯誤恢復與人工介入紀錄 147 附錄F 人工智慧使用與資料來源說明 148 附錄G 三輪設計延伸建模代理人之方案資訊與建築圖說 149

    學位論文
    1. 吳宛霖（2022）。《衍生式設計結合建築性能最佳化於建築初期階段草案自動生成研究》〔碩士論文，國立成功大學建築學系〕。
    2. 傅乙晟（2025）。《CoEvo：生成式人工智慧之多代理系統在建築設計創新流程的探討》〔碩士論文，國立成功大學建築學系〕。
    期刊論文
    1. Abou Ali, M., Dornaika, F., & Charafeddine, J. (2025). Agentic AI: A comprehensive survey of architectures, applications, and future directions. Artificial Intelligence Review. https://doi.org/10.1007/s10462-025-11422-4
    2. As, I., Pal, S., & Basu, P. (2018). Artificial intelligence in architecture: Generating conceptual design via deep learning. International Journal of Architectural Computing, 16(4), 306–327. https://doi.org/10.1177/1478077118800982
    3. Bailey, E. T., & Caldas, L. (2023). Operative generative design using non-dominated sorting genetic algorithm II (NSGA-II). Automation in Construction, 155, 105026. https://doi.org/10.1016/j.autcon.2023.105026
    4. Cheung, L. H., Wang, L., & Lei, D. (2025). Conversational, agentic AI-enhanced architectural design process: Three approaches to multimodal AI-enhanced early-stage performative design exploration. Architectural Intelligence. https://doi.org/10.1007/s44223-025-00092-5
    5. Chen, N., Lin, X., Jiang, H., & An, Y. (2024). Automated building information modeling compliance check through a large language model combined with deep learning and ontology. Buildings, 14(7), 1983. https://doi.org/10.3390/buildings14071983
    6. Danhaive, R., & Mueller, C. T. (2021). Design subspace learning: Structural design space exploration using performance-conditioned generative modeling. Automation in Construction, 127, 103664. https://doi.org/10.1016/j.autcon.2021.103664
    7. Du, C., Esser, S., Nousias, S., & Borrmann, A. (2026). Text2BIM: Generating building models using a large language model-based multiagent framework. Journal of Computing in Civil Engineering. https://doi.org/10.1061/JCCEE5.CPENG-6386
    8. Evins, R. (2013). A review of computational optimisation methods applied to sustainable building design. Renewable and Sustainable Energy Reviews, 22, 230–245. https://doi.org/10.1016/j.rser.2013.02.004
    9. Fan, Z., Liu, J., Wang, L., Cheng, G., Liao, M., Liu, P., & Chen, Y. F. (2023). Automated layout of modular high-rise residential buildings based on genetic algorithm. Automation in Construction, 152, 104943. https://doi.org/10.1016/j.autcon.2023.104943
    10. Jang, S., Roh, H., & Lee, G. (2025). Generative AI in architectural design: Application, data, and evaluation methods. Automation in Construction, 174, 106174. https://doi.org/10.1016/j.autcon.2025.106174
    11. Kakooee, R., & Dillenburger, B. (2024). Enhancing architectural space layout design by pretraining deep reinforcement learning agents. Journal of Computational Design and Engineering. https://doi.org/10.1093/jcde/qwae109
    12. Koo, J., & Yoon, S. (2026). Digital twin synchronization in closed-loop HVAC control for building operations. Automation in Construction, 182, 106723. https://doi.org/10.1016/j.autcon.2025.106723
    13. Lee, J., Song, J., Koo, J., Choi, S., Hwang, J., Hasnain Saif, S. M., Li, Y., Li, J., Yoo, J., Lee, G., Seok, M., & Yoon, S. (2025). Agentic built environments: A review. Energy and Buildings, 346, 116159. https://doi.org/10.1016/j.enbuild.2025.116159
    14. Su, Z., & Yan, W. (2015). Creating and improving a closed loop: Design optimization and knowledge discovery in architecture. International Journal of Architectural Computing, 13(2), 123–142. https://doi.org/10.1260/1478-0771.13.2.123
    15. Wang, L., Janssen, P., & Ji, G. (2024). Optimization-based design exploration of building massing typologies—EvoMass and a typology-oriented computational design optimization method for early-stage performance-based building massing design. Frontiers of Architectural Research, 13(6), 1400–1422. https://doi.org/10.1016/j.foar.2024.06.003
    16. Wang, L., Liu, J., Zeng, Y., Cheng, G., Hu, H., Hu, J., & Huang, X. (2023). Automated building layout generation using deep learning and graph algorithms. Automation in Construction, 156, 105036. https://doi.org/10.1016/j.autcon.2023.105036
    17. Zhang, Z., Guo, Z., Zheng, H., Li, Z., & Yuan, P. F. (2024). Automated architectural spatial composition via multi-agent deep reinforcement learning for building renovation. Automation in Construction, 168, 105702. https://doi.org/10.1016/j.autcon.2024.105702
    會議論文與預印本
    1. Cai, Y., Gu, Z., Du, Z., Ye, Z., Cao, S., Xu, Y., Feng, H., & Chen, P. (2025). MIRAGE: Exploring how large language models perform in complex social interactive environments. arXiv. https://doi.org/10.48550/arXiv.2501.01652
    2. Fitriawijaya, A., & Jeng, T. (2024). Multimodal interaction and AI-supported architectural design workflow. CAADRIA 2024 Conference Proceedings.
    3. Hughes, N., Chang, Y., Hu, S., Talak, R., Abdulhai, R., Strader, J., & Carlone, L. (2023). Foundations of spatial perception for robotics: Hierarchical representations and real-time systems. arXiv. https://doi.org/10.48550/arXiv.2305.07154
    4. Jiang, H., Xiao, Y., Hurley, R., & Liu, S. (2026). The Impact of Generative AI on Architectural Conceptual Design: Performance, Creative Self-Efficacy and Cognitive Load. arXiv. https://doi.org/10.48550/ARXIV.2601.10696
    5. Ko, J., Ajibefun, J., & Yan, W. (2023). Experiments on generative AI-powered parametric modeling and BIM for architectural design. arXiv. https://doi.org/10.48550/arXiv.2308.00227
    6. Macedo, S. (2026). Stop hand-holding your coding agent: Engineering the loops that replace step-by-step prompting. arXiv preprint arXiv:2607.00038. https://doi.org/10.48550/arXiv.2607.00038
    7. Ponzio, A. P., Chornobai, S. R., Fagundes, C. V. M., Rodrigues, R. C., & Cunha Hafez José, G. (2023). Exploring AI-assisted thinking in the design process: The design intelligence strategy. Proceedings of CAADRIA 2023, 2, 49–58.
    8. Qian, C., Xie, Z., Wang, Y., Liu, W., Zhu, K., Xia, H., Dang, Y., Du, Z., Chen, W., Yang, C., Liu, Z., & Sun, M. (2024). Scaling large language model-based multi-agent collaboration. arXiv. https://doi.org/10.48550/arXiv.2406.07155
    9. Tang, C., Kam-Ming Mark, T., & Herr, C. M. (2025). Archi-Agents: Approximating the architectural design process with collaborative specialized multi-agent system LLM and building information modeling. Proceedings of CAADRIA 2025, 141–150. https://doi.org/10.52842/conf.caadria.2025.3.141
    10. Wortmann, T., & Fischer, T. (2020). Does architectural design optimization require multiple objectives? A critical analysis. Proceedings of the 25th Conference on Computer Aided Architectural Design Research in Asia (CAADRIA), 1, 365–374. https://doi.org/10.52842/conf.caadria.2020.1.365
    11. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., Zheng, R., Fan, X., Wang, X., Xiong, L., Zhou, Y., Wang, W., Jiang, C., Zou, Y., Liu, X., … Gui, T. (2023). The rise and potential of large language model based agents: A survey. arXiv. https://doi.org/10.48550/arXiv.2309.07864
    專書
    1. Chaillou, S. (2020). ArchiGAN: Artificial intelligence x architecture. In P. F. Yuan, M. Xie, N. Leach, J. Yao, & X. Wang (Eds.), Architectural Intelligence (pp. 117–127). Springer Nature Singapore. https://doi.org/10.1007/978-981-15-6568-7_8
    2. Ching, F. D. K. (2010). Design Drawing (2nd ed.). John Wiley & Sons.
    3. Ching, F. D. K. (2014). Architecture: Form, Space, and Order (4th ed.). John Wiley & Sons.
    4. Eastman, C., Teicholz, P., Sacks, R., & Liston, K. (2018). BIM Handbook: A Guide to Building Information Modeling for Owners, Designers, Engineers, Contractors, and Facility Managers (3rd ed.). John Wiley & Sons.
    網路與軟體資料
    1. 褚崇名（2026年7月12日）。Loop Engineering（迴圈工程）是什麼？別再手動下 prompt，學會設計會自己跑的 AI Agent 迴圈。WHOOPS SEO。檢索於2026年7月13日，https://seo.whoops.com.tw/loop-engineering/
    2. Anthropic. (2024). Building effective AI agents. Retrieved July 13, 2026, from https://www.anthropic.com/engineering/building-effective-agents
    3. buildingSMART International. (n.d.). Industry Foundation Classes (IFC). Retrieved July 7, 2026, from https://www.buildingsmart.org/standards/bsi-standards/industry-foundation-classes/
    4. h30190. (2026). HJPLUS_Taiwan_Architect_KB: Taiwan AEC industry skill open knowledge base [GitHub repository]. GitHub. Commit f66aa1fa605b8f6cf7e895f67cdf946ce63e3817. https://github.com/h30190/HJPLUS_Taiwan_Architect_KB
    5. McNeel. (n.d.). Rhinoceros 7 documentation. Retrieved July 7, 2026, from https://docs.mcneel.com/rhino/7/help/en-us/
    6. OpenAI. (n.d.). Codex. Retrieved July 7, 2026, from https://openai.com/codex/
    7. Osmani, A. (2026, June 7). Loop engineering. Retrieved July 13, 2026, from https://addyosmani.com/blog/loop-engineering/
    8. xFigura. (n.d.). xFigura. Retrieved July 7, 2026, from https://xfigura.ai/

    QR CODE