Genetic programming symbolic regression (GPSR) is a flexible way to discover equations from data, but the expressions it returns are often larger than they need to be. A new preprint on arXiv argues that these bloated models contain dormant branches and finely tuned algebraic configurations that mainly fit sampling noise rather than the underlying relationship.
To address this, the authors introduce a pruning approach based on the Fisher information geometry of the model. Rather than relying only on standard complexity penalties, the method identifies and removes parts of the expression that carry little statistical information, steering the search toward parsimonious models.
The work is framed as a response to existing complexity-control techniques, which the authors suggest do not fully solve the overfitting problem. The preprint does not present experimental comparisons in the abstract, so the practical gains are not yet clear from the announcement alone.