Quality Evolvability ES: Evolving Individuals With a Distribution of Well Performing and Diverse Offspring

One of the most important lessons from the success of deep learning is that\nlearned representations tend to perform much better at any task compared to\nrepresentations we design by hand. Yet evolution of evolvability algorithms,\nwhich aim to automatically learn good genetic representations, have received\nrelatively little attention, perhaps because of the large amount of\ncomputational power they require. The recent method Evolvability ES allows\ndirect selection for evolvability with little computation. However, it can only\nbe used to solve problems where evolvability and task performance are aligned.\nWe propose Quality Evolvability ES, a method that simultaneously optimizes for\ntask performance and evolvability and without this restriction. Our proposed\napproach Quality Evolvability has similar motivation to Quality Diversity\nalgorithms, but with some important differences. While Quality Diversity aims\nto find an archive of diverse and well-performing, but potentially genetically\ndistant individuals, Quality Evolvability aims to find a single individual with\na diverse and well-performing distribution of offspring. By doing so Quality\nEvolvability is forced to discover more evolvable representations. We\ndemonstrate on robotic locomotion control tasks that Quality Evolvability ES,\nsimilarly to Quality Diversity methods, can learn faster than objective-based\nmethods and can handle deceptive problems.\n