Transdisciplinary approach¶
The proposed research relies on finding fertile ground on the intersection of several different domains of knowledge and practice which are adjacent, in some ways, both due to the problems they try to solve but also the ideas they use for solving them. A brief description of these domains follows below.
The first domain is Computational Architecture, a subdomain of Architectural Design that focuses, among other things, on developing computational tools, workflows, and algorithms, that can generate and evaluate design artifacts in an efficient and diverse manner. The field spans most areas of design, ranging from the urban level all the way down to the construction component level and relates to practical areas of design and construction such us manufacturing, robotic fabrication, and modular design.
The second domain is Machine Learning (ML) and its subdomain Deep Learning (DL). DL has had incredible growth, research activity, and practical successes in the last decade. DL models play an important role for this research, in different ways:
allowing real-time evaluation of performance through surrogate modelling,
enabling an encoding of the design space at different moments of the design process,
and assisting with design exploration, visualization and finally extracting Design Intelligence (DL).
The DL models that will be developed and used in this research come from the areas of Computer Vision (CV), specifically from the research domain of Domain Adaptation, of Graph Theory and Graph Neural Networks (GNNs), and the more diverse domain of Self-Supervised Learning (SSL) which includes new ideas and techniques for learning through representations without labels, an important attribute for a generative design system.
The third domain is that of Evolutionary Computation (EC) with a focus on the newly established field of Quality Diversity (QD). EC began in the late 1950s and 1960s along with the larger availability of computer programs that allowed the testing of models of evolutionary processes and includes important EC paradigms such as evolutionary programming (EP), evolution strategies (ES), and genetic algorithms (GAs). The main focus is on simulating various aspects of evolution and while the techniques developed have differences in their implementations and insights, they all have one fundamental commonality: “they each involve the reproduction, random variation, competition, and selection of contending individuals in a population” (ref: Handbook of EC). QD, a much younger sub-domain of EC which started with the seminal work on Novelty Search by Kenneth Stanley (ref: novelty search), focuses on the ability of natural evolution to produce a diverse collection of organisms, in our case designs, that are all well-performing in their niche. QD therefore “flips the script”, changing the focus from traditional evolutionary approaches that look to find a single high-performing solution towards design space exploration, identifying viable and well-performing solutions at every niche of the design space.
The fourth domain is Gaming AI, which includes a variety of sub-domains of interest: procedural content generation (PCG), mixed-initiative design (MID), player experience (PEX) and user preference learning (PL). The work done in Gaming AI is both highly interesting and applicable for the AEC, something which is becoming evident by the growing interest and number of applications in the field which take advantage of gaming technologies and tools (e.g. VR/AR, generative design, visualization, and simulation). An advantage of intersecting with the domain of Gaming AI is that a lot of the research being done not only has practical implications for design but also typically includes a practical implementation, case studies showcasing these new ideas in practical applications (in new games or gaming environments). Additionally, a lot of insights can be gained in areas important for design, such as generative design and its role in creating content of all kinds, evaluating expressivity and diversity of the design artifacts generated, understanding the behavioral space of designs and their relationship with designers and users, and developing interfaces that enable an intuitive MID process with a human designer.
In a true transdisciplinary approach, the above domains cannot be thought of in isolation from each other nor can they simply be connected at obvious points of intersection, where specific design challenges for the AEC are found. Instead, they need to be embedded and intertwined within the very structure and dynamics of the proposed generative design system and together create a tool that offers a seamless design workflow and user experience. For that reason, the next few sections will review relevant works and ideas from these domains, from the point of view of the system and its modules, and will detail both how these relate to each module and how they allow for relationships and information transfer to form between modules.