On the interaction between lexicase selection, modularity and data subsets
Benjamin Portman, Malcolm Iain Heywood · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Lexicase selection represents a framework for maintaining population diversity, and has therefore demonstrated a capacity for discovering solutions to a wide range of tasks. We note, however, that the benchmarking studies used to demonstrate these properties also tend assume some form of modularity in the representation. With that in mind, an empirical study using five classification datasets is performed in which we compare: 1) Lexicase and a canonical 'breeder' model of selection, and 2) modularity is explicitly enabled and disabled. A clear preference is demonstrated for including modularity with both forms of selection. However, the effect is most pronounced under Lexicase selection with significantly simpler solutions also resulting when modularity is enabled.