On the Effects of Continuous Pruning on Symbolic Regression for Different Variants of Evolutionary Search

In machine learning, pruningpruningtechniques are often used in decision trees, rule-based learning, or neural networks to obtain simpler models. In addition to improving interpretabilityinterpretability, the use of pruningpruningis often motivated by the avoidance of overfitting. In the field of symbolic regressionsymbolic regressionbased on genetic programming, pruningpruningis used comparatively less frequently and, when it is used, it is more likely to be post-pruningpruningafter the learning process has been completed. In this article, we analyze a pruningpruningvariant that is designed for use during the execution of an evolutionary algorithm for symbolic regressionsymbolic regressionand combines multiple techniques to enable beneficial simplificationsimplificationduring execution without simultaneously unduly restricting genetic diversity that is so important for the convergence behavior of an evolutionary algorithm. The effects of some interesting and characteristic algorithm instances are analyzed using benchmark problems currently frequently used in the scientific domain and analyzed in terms of population dynamics, as well as compared with results published in the literature.