A comparison of semantic-based initialization methods for genetic programming

During the initialization step, a genetic programming (GP) system traditionally creates a population of completely random programs to populate the initial population. These programs almost always perform poorly in terms of their total error---some might not even output the correct data type. In this paper, we present new methods for initialization that attempt to generate programs that are somewhat relevant to the problem being solved and/or increase the initial diversity (both error and behavioral diversity) of the population prior to the GP run. By seeding the population---and thereby eliminating worthless programs and increasing the initial diversity of the population---we hope to improve the performance of the GP system. Here, we present two novel techniques for initialization (Lexicase Seeding and Pareto Seeding) and compare them to a previous method (Enforced Diverse Populations) and traditional, non-seeded initialization. Surprisingly, we found that none of the initialization methods result in significant differences in problem-solving performance or population diversity across five program synthesis benchmark problems. We attribute this lack of difference to our use of lexicase selection, which seems to rapidly converge on similar populations regardless of initialization method.