Evolutionary systems for quality and diversity

Computational generative systems are becoming increasingly popular for many design and creative applications.

Known for their ability to generate complex and interesting designs from a relatively simple specification, they shift the focus of design from objects to processes.

However, generative systems can be hard to control and manipulate when trying to fit to the designer’s particular goals or taste. Often the shear number of possibilities they present can be overwhelming, making it difficult or even impossible to find the best or most suitable designs.

Evolutionary techniques are a popular method to find interesting designs in large search spaces. However, traditionally such search methods focus on optimisation: trying to find the single design that best fits the criteria for success. But this is not how design works. Designers like to compare different, good designs. Even for a single criteria, such as aesthetics, there may be multiple, different designs that are appealing, not a single optimum.

Evolutionary methods use metaphors from biological evolution to optimise a population
To evolve a design we need a measure of suitability or fitness, and measuring aesthetics remains a challenging problem
Using MAP-Elites we can search for diverse designs that are of high quality
We tested the system on a generative line drawing system to find a wide variety of interesting designs

In the method we have developed, we are interested in finding the most diverse collection of good designs in any generative system. So we are selecting designs based on two criteria: firstly, that they are different from any of the designs the system has seen before. And secondly, if they are not different, then at least they are better than similar designs seen previously.

So we are searching for all possible high quality designs, not just a single best design.

To do this, we need to be able to measure two things: how different a design is from anything produced to date (known as a “diversity measure”). And secondly how good or suitable it is, what is termed “fitness” in standard evolutionary algorithms.

Our method presents techniques for measuring the aesthetic quality of a generative design system, based on an individual designer’s preferences. We use a Variational Autoencoder to measure the diversity of designs produced.

Together, these measures of quality and diversity are used in a MAP-Elites algorithm to find a diverse set of high-fitness designs.

This video explains our method of designing with quality diversity algorithms

The research is on-going as we seek to improve the measures of both quality and diversity, and evaluate the system on a wider set of generative design systems.

This research was supported by an Australian Research Council Future Fellowship, FT170100033.