Machine Learning Models For Directed Curation Of Design Solution Space
DesComp Thesis Completed by Christian Sjoberg (2017)- Presented at ACADIA 2017
As the influence of computation on design continues to grow, designers have the ability to optimize and account for high numbers of variables. This results in a high dimension solution space and an exponentially increasing number of possible variations. Optimization strategies enable designers to navigate toward solutions which are quantitatively superior in pre-defined tests, but limit the designer’s ability to search the solution space for qualitative aspects. Fully accepting a quantitatively optimized solution poses an issue for designers. Calculating a top performing design based on limited or standardized criteria reduces the role of the designer. This has the potential to situate design as a secondary layer on calculated models. This workflow limits the ability of designers to holistically consider the aspects of design problems which are abstract or qualitative. To counteract this side-effect, designers may include quantitative descriptions of qualitative criteria into their computational models. In doing so, they are able to constrain the solution space based on their pre-conceived expectations of the possible outcomes. In either situation, the designer is limited in their exploration of the full solution space, and reduces the opportunity to discover emergent properties of designs resulting from unexpected combinations of parameters.
If we are to avoid this, we must re-evaluate the possible methods for navigating vast solution spaces and create models which can evaluate solutions based on both quantitative evaluations and qualitative properties. In the context of this project, the term qualitative is used to describe features of a computationally generated design which are visibly recognizable as favorable to the designer, but are not measured by the designer’s script as an evaluation of the design’s performance. These features can be understood to
be the emergent result of the combined state of many parameters of the design. Due to the abstract and subjective nature of qualitative criteria, designers can benefit from a process of selection of solutions based on their own expertise and intuition rather than the simplification of a problem and constraint of possible solutions.
We may imagine, for example, an artist in the process of making a sculpture. The artist does not explicitly or mathematically define the proper curvature of a surface of the sculpture before they begin to form it. They instead manipulate their design iteratively, and visually evaluate the current state of the form until they recognize that it has taken on the quality they find favorable. As humans, we are far better at visually recognizing a geometry we find to be favorable then we are at explicitly mathematically defining it. We may even find that our original assumptions or what we thought to be ideal changes in the process of creating a design.
The methods that computational designers currently employ to search for design options in a defined parameter space do not represent the inherent artistic processes of the designer, because they usually require the explicit definition of what is acceptable as a solution before the search process ever begins. If we can create software systems that more closely resemble the artist’s evaluation of design qualities we can begin to remove some of the limitations that result from the need to explicitly define criteria at the beginning of our search. To perform meaningful evaluation of qualities subjectively favorable to an individual designer, a system must first create a model of the designer’s evaluation process.
This project seeks to develop and test a method for searching vast solution spaces based on inexplicit evaluation criteria demonstrated by designers using the system.
Machine learning methods allow for the formation of evaluative criteria based on patterns learned from the user’s selection of design options. As designers are iteratively presented with design choices, they will choose the solutions that meet their personal aesthetic, abstract, or qualitative criteria for design. The encoded design parameters for the possible selections as well as the actual user selections are recorded as a training set for an artificial neural network. A network trained on this selection data acts as a model of the user’s evaluation of design options. Using this model as the fitness function for an evolutionary solver provides a mechanism for searching solution spaces containing more design possibilities than could possibly be evaluated individually by a human. Learning from the designer’s evaluation of a sample of the possible solutions, this system demonstrates the ability to identify additional high performing regions of the solution space based on the user’s inexplicit qualitative criteria. Allowing designers to visually identify the presence favorable qualitative features leverages the human’s inherent strengths, while reducing the limitations resulting from the need for explicit limitation of the solution space.
A MACHINE LEARNING APPROACH
The discipline of architecture has explored the computational generation of designs through rule based logics for many years. One of the principle applications of this research has been the task of automated space planning. Charles Eastman’s General Space Planner (1973) is an example of one such system for creating space planning designs using decision tree logic for evaluation. The application of multiple pre-defined constraints to the problem space is used as a method for searching for design solutions (Eastman 1973). These and other similar methods rely heavily on the user and system’s
anticipation of all possible design conditions, and the presence of a pre-programmed response. This top down logic ultimately falls short of representing a human’s process of design, which is one of nuanced reasoning, accumulated experience, and personal intuition. To enable a system to better represent the complex process of design, it must be developed to recognize subtle decisions as well as the greater patterns of designers. The application of a bottom-up approach to learning design constraints has the potential to remove the requirement of explicitly defined design constraints and better represent the designer’s own process. It must be able to form a model of design responses that are not directly related to one specific condition, but to the presence of subtle patterns or combinations of multiple conditions.
The explicit curation of the solution space based on rules which operate at the level of the design parameter dramatically reduces the possibility of discovery of new options. Design is often a process of experimentation in which the designer tries numerous alternative approaches while both formulating and solving a problem. If the tool used by the designer requires them to explicitly define each of the criteria at the onset of the search for a solution, this process of exploration, discovery, and redirection is lost. To create tools which more naturally resemble the designer’s process, we can begin to develop a bottom-up approach to the creation of constraints for the solution space. The collection of a designer’s choices in their own process of curating possibilities to a design problem contains potentially identifiable patterns if we can consistently describe the steps of their process mathematically. Based on this limitation, a system capable of mathematically describing each of the design iterations explored by the designer can potentially form a model of the underlying relationships or patterns present in the set.
Creating an environment for the recording of the designer’s process poses a difficult problem. We must make trade-offs between the freedom provided to the designer and the effectiveness of the system to measure the outputs of their process. Designers working with analog methods such as sketching and physical modeling produce artifacts which are difficult to comparatively evaluate, due to inconsistency in the relative meaning of any one determined variable between one artifact and the next. If, for example a series of drawings are compared which represent the designer’s evolving understanding of a design, there is no guarantee that a specific pixel value is necessarily comparable with the same pixel in another image drawn at another scale, from another angle, or in another medium. For the purposes of this project, it is important to establish a consistent and comparable definition of the objects considered by the system.
The computational processes of architectural designers today are commonly developed within visual scripting platforms. These software platforms allow users with little to no programming experience to create models of the logic they wish to use to produce a design. This practice of scripting design demonstrates a voluntary, explicit definition of the possible parameter space. This project utilizes the user defined parameter space as a means of capturing consistent and comparable descriptions of the design options being considered. In this way, any potential design option within the user defined parameter space can be fully represented by its unique set of parameter values. This allows for simple comparison of one design’s parameter sequence to another. The comparison of a series of designs selected by the designer for the presence of favorable qualitative traits provides the potential for pattern recognition between parameter values and user selection.
