- Main
Interactive Evolutionary Generative Design in Augmented Reality
- Bailey, Elnaz Tafrihi
- Advisor(s): Caldas, Luisa
Abstract
This dissertation explores immersive technologies in the architectural domain, aiming to answer an increased demand for new remote design platforms that promote interactive collaboration. It explores the existing space of architectural design processes, along with the design applications of genetic algorithms and participatory design. In response to existing gaps, this dissertation introduces InsightXR, an Augmented Reality (AR) platform that facilitates collaboration during the architectural design process and provides individualized feedback to designers while exploring generative design. The use case of InsightXR presented in this research focuses on developing massing studies during early stages of architectural design using a new application of the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to the field of Operative Design. As a method and a design platform, however, InsightXR can be applied to many other design problems and disciplinary domains. The proposed workflow for massing studies involves utilizing a site model to generate a population using the Operative Generative NSGA-II algorithm. During each generation, users are presented with the three most distinct geometries, based on shape similarity calculated by mesh similarity methods, and asked to provide a qualitative fitness value to be used in the proposed algorithm. To understand the benefits and limitations of this new interactive generative design process, three experiments are developed in this dissertation. The first experiment presents a novel Operative Generative NSGA-II using algorithm experiments by investigating Floor Area Ratio (FAR), Non-Passive Zone (NPZ), Roofs and Best Oriented Surfaces (RBOS), and Usable Open Space (UOS) as objective functions. The second experiment explores the integration of the Operative Generative NSGA-II with InsightXR’s application using RBOS, UOS, and Area Violation (AV) as objective functions. Finally, the third experiment is designed in two stages. In the first stage, a human subject experiment is performed with experts and non-experts in architectural design using AV, NPZ, and UOS as objective functions. In the second stage, a follow-up experiment uses only the algorithm without including users, utilizing the first stage initial populations as seeds. The resulting building metrics and genetic algorithm performance are analyzed across the various experiments, exploring the benefits of the platform as well as looking at the effects on the results from Expert versus Non-Expert participants and Human-Directed versus Algorithm-Only executions. In Experiment 3, both user groups’ feedback conflicted with the other objective functions, demonstrating the relevance of introducing user participation in generative design processes. This was due to the tendency of the algorithm to generate massing options with smaller footprints and taller typologies to improve UOS and NPZ values, which most users described as their least preferred building type. A high percentage of both experts and non-experts rated NPZ as the most important objective function, demonstrating a valorization of design aspects that contribute to climate change mitigation. The NPZ values had a higher impact on experts compared to non-experts, signaling the importance that experts currently attribute to those performance criteria. Comparative results between the two user groups also highlighted the relevance of expertise in the generative design process. While providing complex spatial information to experts helped improve their decision-making process, in contrast, it did not significantly improve the decision-making process for non-experts. As an overall conclusion, while the AI system alone could often generate better-performing solutions in a quantitative-only optimization space, it tended to create solutions that were qualitatively unsatisfactory to many users. The absence of user preference in generative design processes has arguably been a major obstacle for the adoption of generative design in real world situations. Introducing user preference ratings as an additional objective function often generated a relative reduction in some quantitative metrics, but it simultaneously increased the likelihood that solutions were qualitatively appreciated by users, and thus more likely to be adopted. Furthermore, when expert users of our AR-driven participatory design interface were asked to rate alternative designs purely based on aesthetic preferences, results were typically inferior to when users were additionally given some simplified quantitative information to inform their rating preferences. This suggests that the successful integration of informed user preference with AI-based support for complex multi-dimensional design tasks may prove a successful and promising avenue for future design methods research.
Main Content
Enter the password to open this PDF file:
-
-
-
-
-
-
-
-
-
-
-
-
-
-