Specialization and elitism in lexicase and tournament selection

Prior work has demonstrated that genetic programming systems often maintain higher levels of population diversity when using lexicase selection than when using other parent selection methods, and that the use of lexicase selection improves problem-solving performance in many circumstances. It has been suggested that it is not only the maintenance of diversity that is responsible for the performance of lexicase selection, but more specifically the production and maintenance of “specialists” that matters, where specialists are defined to be individuals with the lowest error, relative to the rest of the population, on a small number of training cases regardless of total error. Here we provide results of experiments that uphold this suggestion by tracking the numbers of specialists selected by lexicase selection and by tournament selection in a genetic programming system solving software synthesis problems. Our results also show that lexicase selection selects parents with poor total error far more frequently than tournament selection, even near the ends of successful runs, suggesting that such selections are integral to the improved problem-solving performance of lexicase selection.