Prototype Discovery Using Quality-Diversity

Abstract

An iterative computer-aided ideation procedure is introduced, building on recent quality-diversity algorithms, which search for diverse as well as high-performing solutions. Dimensionality reduction is used to define a similarity space, in which solutions are clustered into classes. These classes are represented by prototypes, which are presented to the user for selection. In the next iteration, quality-diversity focuses on searching within the selected class. A quantitative analysis is performed on a 2D airfoil, and a more complex 3D side view mirror domain shows how computer-aided ideation can help to enhance engineers' intuition while allowing their design decisions to influence the design process.

Modeling user selection in quality diversity

“…In [9] the initial intuition of the design space produced by QD was analyzed by clustering QD's resulting genomes in an unsupervised manner, forming design classes and prototypical representatives. The discovery that the behavior based niches in the QD archive contain a relatively small set of design classes is in line with the analysis in [24].…”

“…That work showed that the elite hypervolume, the part of the genotype space that contains the archive's elites, is often less spread out than the elites are in behavior space. The concise representation in [9] lets a user select interesting classes without being overwhelmed by the large number of solutions produced by QD. To influence the continuation of the QD algorithm, the authors reseeded its archive with the selected results, similar to [25], thereby forcing QD to start searching around the selection.…”

“…QD has to be constrained by modeling the user's selection decisions. The genotype based selection process from [9] is subject to another problem. When a non-linear genotype to phenotype mapping is used or the phenotype's behavior is reactive, e.g.…”

“…However, the genotype often has many parameters, causing distance metrics like Euclidean distance to become meaningless, which can happen for as few as 10-15 dimensions [2]. [9,21] showed that when the dimensionality of genotypic space is reduced with t-SNE [23], more structure in the objective landscape can be discovered by clustering techniques that are based on the notion of distance. This provides evidence that distances measured in the similarity space resulting from t-SNE are more effective in describing genotypic similarity than those measured in the original genotypic space.…”

“…Computer aided ideation has the potential to reduce the number of iterations in a design process by automated generation of design candidates [3]. In principle, QD was shown to be usable in the design by shopping paradigm [9].…”