# Quality-Diversity Meta-Evolution: customising behaviour spaces to a meta-objective

**Authors:** David M. Bossens, Danesh Tarapore  
**arXiv:** [2109.03918](https://arxiv.org/abs/2109.03918)  
**Date:** 2021-09-08  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2109.03918v1/

## Key points

- Quality-Diversity (QD) algorithms aim to evolve behaviorally diverse and high-performing solutions.
- Definition of a suitable behavior space is crucial in QD algorithms.
- Dimensionality reduction techniques are necessary for dealing with high-dimensional behavior spaces.
- Customizing behavioral diversity according to specific meta-objectives set by end-users is essential.
- Quality-Diversity Meta-Evolution (QD-Meta) framework evolves populations of QD algorithms to optimize both the algorithms themselves and resulting archives towards user-defined meta-objectives.
- QD-Meta has shown promising results compared to traditional QD algorithms, but needs further comparison against leading behavior space automation techniques such as CVT-MAP-Elites and AURORA.
- Empirical analysis demonstrated that QD-Meta archives exhibited improved average performance and faster adaptation to environmental changes compared to CVT-MAP-Elites and AURORA.
- Resulting archives from QD-Meta were finely tailored to meet specific meta-objectives outlined by end-users.

## Summary

In the realm of Quality-Diversity (QD) algorithms, which aim to evolve behaviorally diverse and high-performing solutions, the definition of a suitable behavior space is crucial. When dealing with high-dimensional behavior spaces, it becomes necessary to employ dimensionality reduction techniques to maintain a manageable number of behavioral niches. While current automated methods for defining behavior spaces often focus on altering geometry or utilizing unsupervised learning, there is a growing need for customizing behavioral diversity according to specific meta-objectives set by end-users. Enter the innovative framework known as Quality-Diversity Meta-Evolution (QD-Meta), where populations of QD algorithms with varying algorithmic and representational characteristics are evolved to optimize both the algorithms themselves and the resulting archives towards user-defined meta-objectives. Despite showing promising results compared to traditional QD algorithms, QD-Meta has yet to be thoroughly compared against leading behavior space automation techniques such as Centroidal Voronoi Tessellations Multi-dimensional Archive of Phenotypic Elites Algorithm (CVT-MAP-Elites) and Autonomous Robots Realising their Abilities (AURORA). To address this gap in research, a recent study conducted an empirical analysis of QD-Meta using function optimization and multilegged robot locomotion benchmarks. The results demonstrated that QD-Meta archives not only exhibited improved average performance but also showcased faster adaptation to unforeseen environmental changes when contrasted with CVT-MAP-Elites and AURORA. Additionally, a qualitative examination revealed how the resulting archives from QD-Meta were finely tailored to meet the specific meta-objectives outlined by end-users. Authored by David M. Bossens and Danesh Tarapore, this study sheds light on the potential of Quality-Diversity Meta-Evolution in customizing behavior spaces towards user-defined meta-objectives, offering valuable insights into the field of evolutionary algorithms and optimization strategies.