Length-niching Selection and Spatial Crossover in Variable-length Evolutionary Rule Set Learning
David Pätzel, Richard Nordsieck, Jörg Hähner · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
We explore Metaheuristic Rule Set Learners with variable-length encodings for regression tasks by performing a first benchmark of two variable-length operators that have not yet seen use in this context: Spatial crossover is a well-known operator which seems to be a natural fit for these systems since it is meant to promote building blocks in continuous parameter spaces like the ones defined by interval-based rule conditions. Length-niching selection is a recently proposed operator that promotes population diversity with respect to solution length which is meant to prevent some forms of premature convergence. We perform comparisons with other established operators (i. e. tournament selection and cut-and-splice crossover) within a simplistic GA that uses the corrected Akaike Information Criterion as a fitness measure. The 54 learning tasks considered are synthetic and highly likely to be learnable by the algorithms considered. While all variants of the GA tested perform similarly in terms of Mean Absolute Error after a fixed number of iterations, with respect to solution complexity (i. e. number of rules in the best found solution), not using any crossover seems to outperform both crossover operators tested.