Multi-Objective Qualitative Optimization (MOQO) in Architectural Design
Abstract
Architectural design problems are often multi-objective in nature, involving both qualitative and quantitative objectives. Previous research has focused exclusively on the development of multi-objective optimization algorithms that work with multiple quantitative objectives. No previous research has looked at the topic of multi-objective qualitative optimization (MOQO), in which multiple qualitative objectives are optimized simultaneously. This research addresses MOQO through the development of a unique multi-objective optimization algorithm for the conceptual design phase that uses three-dimensional convolutional neural networks (3D CNNs) to measure user-defined qualities in architectural massing models.
Introduction
The process of architectural design often involves challenging optimization problems in which there are multiple and often conflicting objectives that must be simultaneously satisfied. To further complicate matters, these multi-objective problems (MOPs) often involve both quantitative (e.g., useful daylight, deflection, cost, etc.) and qualitative (e.g., publicness, slowness, etc.) objectives. Multi-objective optimization (MOO) algorithms have been studied extensively to solve MOPs in architecture and engineering. In the field of architecture, one MOO approach that has received significant attention are multi-objective evolutionary algorithms (MOEAs) (Von Buelow, 2012; Mueller and Ochsendorf, 2011; Turrin et al., 2011). MOEAs have been popular because of their generalizability to many different types of architectural MOPs. Their use in the field, however, has largely been cut-and-paste without significantly modifying the standard algorithms provided from the fields of optimization, operations research, and computer science to the specificities of architectural design. This has led to MOEA-based digital tools which are biased toward optimizing quantitative objectives and not qualitative ones. How can MOEAs be modified for the specificities of architectural design to be biased toward qualitative optimization?
In order to address this issue, we propose a new sub-category of MOO focused on qualities. We call this new area of optimization research multi-objective qualitative optimization (MOQO). MOQO involves the optimization of more than one qualitative objective simultaneously. A MOQO process may also involve a mixture of quantitative and qualitative objectives, as long as more than one qualitative objective is present. MOQO presents several challenges to researchers developing algorithms for this type of optimization. The chief problem is developing qualitative objective functions that can mathematically represent qualities.
Previous work in architecture has approached this problem through either solution ranking (Mueller and Ochsendorf, 2015)or function approximation (Beorkrem and Ellinger, 2017). Solution ranking involves having the user rank, or score, solutions during or after the optimization based on a perceived quality (e.g., degree of beauty, privacy, etc.). This approach causes user fatigue (Takagi, 2001)and restricts the extent of the search significantly. Function approximation attempts to address the user fatigue problem by finding a function to represent a user-defined quality. Examples of function approximation techniques include linear regression, support vector machines (SVMs), neural networks (NNs), deep neural networks (DNNs), and convolutional neural networks (CNNs). Of these approaches, three-dimensional CNNs (3D CNNs) offer the most accurate method to recognize qualities in three dimensional objects (Maturana and Scherer, 2015). Previous research in the field of architecture, however, has not used 3D CNNs for approximating qualitative objective functions.
This research addresses this gap in current research through the development of a unique MOEA-based optimization platform to address MOQO problems in the conceptual design phase. We call this platform the multi-objective qualitative optimization toolkit (MOQOT) and test its effectiveness on a qualitative optimization problem in the conceptual design phase. In summary the research presented here makes the following contributions:
This research is the first to deal with the optimization of multiple qualitative objective functions and contributes a new sub-field of study under MOO focused on qualitative phenomena that we call MOQO.
MOQOT uses 3D CNNs for the first time in the field of architecture to learn multiple user-defined qualitative objective functions (e.g., cellular quality; mat quality; tower quality).
This research contributes a novel approach to deal with the problem of generating large training sets to train 3D CNNs. We propose the use of generative algorithms to quickly create large training sets that exemplify specific qualities.
Figure 1. The architecture of the 3D CNN used for qualitative assessment is shown.
Figure 2. The image shows examples from the three generative algorithms used to produce the training set for the 3D CNN. The three categories of shapes relate to the three qualitative objectives being optimized (e.g., cellular quality; mat quality; tower quality).
Figure 3 The samples from MOQOT’s combined Pareto front for 10 different runs are shown in the image.
Figure 4. The image shows the combined Pareto front created during MOQOT’s 10 runs.