Benchmarking parent selection for program synthesis by genetic programming

Abstract

In genetic programming, the parent selection method determines which individuals in the population are selected to be parents for the next generation, and how many children they create. This process directly impacts the search performance by determining on which areas of the search space genetic programming focuses its attention and how it balances exploration and exploitation. Many parent selection methods have been proposed in the literature, with aims of improving problem-solving performance or other characteristics of the GP system. This paper aims to benchmark many recent and common parent selection methods by comparing them within a single system and set of benchmark problems. We specifically focus on the domain of general program synthesis, where solution programs must make use of multiple data types and control flow structures, and use an existing benchmark suite within the domain. We find that a few methods, all variants of lexicase selection, rise to the top and demand further study, both within the field of program synthesis and in other domains.