On the use of feature-maps for improved quality-diversity meta-evolution

Abstract

Quality-Diversity (QD) algorithms evolve a behaviourally diverse archive of high-performing solutions. In QD meta-evolution, one evolves a population of QD algorithms by modifying algorithmic components (e.g., the behaviour space) to optimise an archive-level objective, the meta-fitness. This paper investigates which feature-map is best for defining the behaviour space for an 8-joint robot arm. Meta-evolution with non-linear feature-maps yields a 15-fold meta-fitness improvement over linear feature-maps. On a damage recovery test, archives evolved with non-linear feature-maps outperform traditional MAP-Elites variants.