Data-Efficient Design Exploration through Surrogate-Assisted Illumination

Abstract

Design optimization techniques are often used at the beginning of the design process to explore the space of possible designs. In these domains illumination algorithms, such as MAP-Elites, are promising alternatives to classic optimization algorithms because they produce diverse, high-quality solutions in a single run, instead of only a single near-optimal solution. Unfortunately, these algorithms currently require a large number of function evaluations, limiting their applicability. In this article, we introduce a new illumination algorithm, Surrogate-Assisted Illumination (SAIL), that leverages surrogate modeling techniques to create a map of the design space according to user-defined features while minimizing the number of fitness evaluations. On a two-dimensional airfoil optimization problem, SAIL produces hundreds of diverse but high-performing designs with several orders of magnitude fewer evaluations than MAP-Elites or CMA-ES. We demonstrate that SAIL is also capable of producing maps of high-performing designs in realistic three-dimensional aerodynamic tasks with an accurate flow simulation. Data-efficient design exploration with SAIL can help designers understand what is possible, beyond what is optimal, by considering more than pure objective-based optimization.

Procedural Content Generation through Quality Diversity

“…Numerous and long evaluations of content are usually required to test the quality of the generated content both in SBPCG and PCG-QD. However, we can imagine leveraging the abstraction capabilities of the machine-learned model to create surrogate of the simulations; this approach can reach comparable results in fewer evaluations [45].…”

Evolving the Behavior of Machines: From Micro to Macroevolution

“…The color denotes the fitness (best fitness values are in yellow). (C) Two-dimensional map of airfoils, adapted from Gaier et al (2018). We can understand that the feature on the vertical axis (X up ) has less influence on the fitness than the horizontal one (Z up ).…”

“…Most interestingly, this approach avoids the need to encode all the requirements in the fitness function, which is required when using traditional optimization algorithms. So far, promising results have been published for designing three-dimensional shapes of aerodynamic bikes (Gaier et al, 2018) ( Figure 2C), to generate content for different games (Fontaine et al, 2020;Gravina et al, 2019), to design molecules (Verhellen and Van den Abeele, 2020), to solve workforce scheduling and routing problems (Urquhart and Hart, 2018), and to find adversarial examples for deep neural networks (Nguyen et al, 2015) or malware (Babaagba et al, 2020).…”

Modeling user selection in quality diversity

“…Quality diversity (QD) algorithms, which combine optimization and novelty discovery by keeping track of solutions in an archive that is defined by feature based or behavioral diversity measures, are right at that forefront of current research in divergent optimization. QD searches for diversity in terms of a solution's expressed behavior [5] or features describing the expressed genome [8]. This makes QD an excellent search algorithm for design processes.…”