機器學習在建築初步設計之應用

本研究的啟發來自人工智慧與機器學習演算法技術在當代社會的普及與重要性。在可見的未來，此種藉由統計科學為基礎的程序決策方式或將成為掌握社會發展的關鍵技術，而在建築設計領域勢必也與其緊密相關。本研究藉由整理「機器學習」技術從1950年代以來在科技與文化上的概略發展歷程，思考其成為建築數位設計工具的潛力與未來趨勢，試圖了解其與近代參數化建築設計的關聯，針對不同的機器學習演算法試執行數個實驗性的小設計並加以整理，最後藉由一個真實基地嘗試處理真實設計議題與建築量體的規劃。　　本研究由三個部分所組成，首先整理機器學習在近代建築與藝術設計交互影響下的發展歷程、常見演算法簡介、Rhino與Grasshopper平台上的插件整理、以及整理相關的設計應用案例。第二部分聚焦在Rhino與Grasshopper平台上現存的機器學習插件功能：Back-Propagation、SOM、CPPN、K-Means演算法等為基礎下，以多個設計案例探索此技術在建築設計應用上的可能性。最後歸納出幾個常見的設計應用方向：擬和曲面、預測資料、形態找尋、無監督分群、與資料拓樸。第三部分藉由前部分的研究歸納出幾個可行的演算流程，並且真實應用在真實基地與設計議題上—基於機器學習演算法應用於都市噪音之的分析與初步建築量體設計。　　藉由本研究在機器學習領域的整理與應用探討，可以初步了解到此技術在建築設計的發展近況與應用方向；並且在應用實作的過程中，藉由不同參數的設定與程序的設計，了解設計者的主觀意識如何介入設計流程，以達到人機協作的的效果。希望本研究可供後續相關研究者參考，延續發展出更多藉由機器學習與資料處理方式在設計中運用的可能性；並且利用參數化設計發展數位設計，除了生產更多不同的建築樣式之外，同時也能夠處理更多實際議題。

關鍵字

機器學習；參數化設計；神經網路；演算法；人工智慧

並列摘要

This research approach is derived from artificial intelligence and machine learning, which is a kind of computational technology that has been widely used and profoundly discussed in modern society. In the foreseeable future, such a procedure-based decision-making method through statistical science may become a key technology to social development where the realm of architectural design will be inevitably involved. In this research, the “machine learning” technology is based on consolidating the overall technical and cultural development since the 1950s to study current development status and future trends of the building-related digital design tools and to understand its impact on the digital-based architectural design in modern days. In this regard, several smaller-scale experimental designs and summarizations are conducted with different types of machine-learning algorithms. Finally, a physical base subject is established to deal with real issues and execute the overall architectural planning. This research is composed of three portions. The first portion deals with the summarization of the development process of machine learning under the interaction of modern architectural and artistic design, the introduction of algorithms normally used, the plug-in processing for the Rhino-Grasshopper Platform and the consolidation of historical application cases for the front-end design. The second portion is focusing on the machine learning plug-in functions (Backpropagation, SOM, CPPN, K-means Algorithm) being used by the Rhino-Grasshopper Platform nowadays. As such, several smaller-scale design cases are based to study the possibility of applying such technology in the architectural design to conclude the following design application approaches normally seen and they are fitting curve, data forecasting, pattern searching, unsupervised groups, and information topology. In the third portion, the aforesaid study is based to obtain several practical computation processes for applying in the physical base and issues, i.e. “Preliminary Testing for Urban Analysis and Design in Applying Machine Learning-Based Computation Method in Urban Noise - Taking Danshui Area as an Example.” Through the summarization and the application study of the machine learning realm that has been conducted in this research, we will be able to understand current development status and application direction by applying such kind of technology in the architectural design. With different parameters and procedures, we can even know how the designer will apply his subjective consciousness in the design process to achieve a harmonic relationship between people and machines. It is hoped that this research will provide a referential basis for the subsequent researchers to develop more processing methods through machine learning so that they can be used in the design. In the meantime, the parameter-based design will also be based to develop the digital-based design so that more physical issues can be processed in addition to presenting other types of architectural patterns.