Continuous Pruning for Symbolic Regression

The generation of compact regression models is crucial for inter-pretability in symbolic regression. Pruning is a well-known technique to compactify models and is usually used in a post hoc fashion as a way to combat bloat and overfitting. This paper introduces a continuous pruning mechanism that simplifies symbolic regression trees during evolution, rather than as a post-processing step. Ex-periments on regression datasets from the Penn Machine Learning Benchmark show that while this approach reduces overfitting as expected, it also influences the search behavior itself. The impact of pruning depends on the selection procedure, particularly whether children compete with their siblings or their parents. Notably, while pruning is often viewed as a method to combat bloat, our results indicate that its most pronounced effects occur in already depth-constrained contexts, where selection pressure is high, and tree size is inherently limited.