This thesis aims to formulate the challenging nature of design problems, and suggest a methodology through which data based methods can be used to solve design problems that require the negotiation of multiple criteria.

 

For the first part, we begin with an in depth study of design theory as it relates to: the nature of design problems, how designers have used computational tools to approach complex design problems, and the potentials of deep learning in the context of design. We use the information gathered in the first part, to propose a framework whereby a design problem can be approximated by a computer in a way that holistically combines multiple classes of information in a generative system using data-driven multi-objective generative adversarial networks (MOGAN). We demonstrate through two case studies (generation of digits, generation of chairs) that multiple design criteria can be negotiated and resolved using our MOGAN framework. Through this combined effort of theory and application from discourse in design and computer science, we demonstrate a new exciting approach to computational design.