Lexicase Selection (tut125, Advanced Tutorials)
Thomas Helmuth, William La Cava
Lexicase parent selection decides which individuals in an EC population are selected not based on a single, aggregated fitness value, but instead based on random sequences of test cases (or objectives) used to filter the population down to a single individual. By never aggregating performance on separate cases, lexicase selection effectively makes use of additional information for parent selection. In Genetic Programming, this results in a semantic-aware parent selection, although lexicase selection has proven useful in other EC systems without a genotypic/semantic distinction. Since lexicase selection places importance on performing well on subsets of the test cases rather than generally on all cases at once, it often selects specialist individuals who perform well on some parts of a problem while poorly on others. These specialists contribute to the maintenance of increased diversity in populations evolved under lexicase selection. However, unlike some methods that focus entirely on diversity, lexicase selection also provides strong selection pressure toward individuals that perform well on many combinations of test cases, which can lead to excellent overall performance. This tutorial will give an introduction to lexicase selection and will discuss major research directions involving both advances in its use and explanations for why it performs well. We will discuss its effects on population diversity, its ability to select specialist individuals, and its propensity toward hyperselection, where it selects the same individual many times in one generation. Participants will come away with a strong understanding of lexicase selection and how it can be applied to any test case-based evolutionary computation system. In addition, they will gain insight into the research frontiers in lexicase selection.
GECCO 2021
The Genetic and Evolutionary Computation Conference
July 10-14, 2021 — Lille, France (online)
https://gecco-2021.sigevo.org
0 Comments