Using machine learning techniques for architectural design tracking: An experimental study of the design of a shelter

In this paper, we present a study aimed at tracking and analysing the design process. More concretely, we intend to explore whether some elements of the conceptual design stage in architecture might have an influence on the quality of the final project and to find and assess common solution pathways in problem-solving behaviour. In this sense, we propose a new methodology for design tracking, based on the application of data analysis and machine learning techniques to data obtained in snapshots of selected design instants. This methodology has been applied in an experimental study, in which fifty-two novice designers were required to design a shelter with the help of a specifically developed computer tool that allowed collecting snapshots of the project at six selected design instants. The snapshots were described according to nine variables. Data analysis and machine learning techniques were then used to extract the knowledge contained in the data. More concretely, supervised learning techniques (decision trees) were used to find strategies employed in higher-quality designs, while unsupervised learning techniques (clustering) were used to find common solution pathways. Results provide evidence that supervised learning techniques allow elucidating the class of the best projects by considering the order of some of the decisions taken. Also, unsupervised learning techniques can find several common problem-solving pathways by grouping projects into clusters that use similar strategies. In this way, our work suggests a novel approach to design tracking, using quantitative analysis methods that can complement and enrich the traditional qualitative approach.