Machine learning in housing design : exploration of generative adversarial network in site plan / floorplan generation

Abstract

Technology has always been an important factor that shapes the way we think about Architecture. In recent years, Machine Learning technology has been gaining more and more attention. Different from traditional types of programming that rely on explicit instructions, Machine Learning allows computers to learn to execute certain tasks "by themselves". This new technology has revolutionized many industries and showed much potential. Examples like AlphaGo and OpenAI Five had shown Machine Learning's capability in solving complex problems. The Architectural design industry is not an exception. Early-stage explorations of this technology are emerging and have shown potential in solving certain design problems. However, basic problems regarding the nature of Machine Learning and its role in Architecture design remain to be answered. What does Machine Learning mean to Architecture? What will be its role in Architectural design? Will it replace human architects? Will it merely be a design tool? Or is it relevant to Architecture at all? To answer these questions, this thesis explored with a specific type of Machine Learning algorithm called Pix2Pix to investigate what can and cannot be learned by a computer through Machine Learning, and to evaluate what Machine Learning means for architects. It concluded that Machine Learning cannot be a creative design agent, but can be a powerful tool in solving conventional design problems. On this basis, this thesis proposed a prototype pipeline of integrating the technology into the design process, which is a combination of Generative Adversarial Network (Pix2Pix), Bayesian Network and Evolutionary Algorithm.