Adaptive Protection boosts symbolic regression on 98 benchmarks

Evolutionary feature construction automatically discovers informative transformations of raw features to boost simple base learners, but existing methods often throw away valuable genetic material when genetic operators disrupt effective features. To fix this, Hengzhe Zhang and colleagues from Victoria University of Wellington and partnering institutions developed an adaptive protection mechanism that leverages feature importance scores to decide which constructed features get protected during evolution. More important features receive stronger protection, while less important ones remain open to modification and can incorporate building blocks from their stronger counterparts. This balances exploitation of high-value features with exploration of new ones, making the search process more efficient.

The team tested the approach on 98 regression benchmark datasets using multiple feature importance calculation methods and different base learners, and found it consistently improved solution quality over baseline symbolic regression methods. They also applied it to two credit classification datasets, showing the mechanism generalizes beyond regression to improve search effectiveness for classification tasks. The work was accepted to PPSN 2026, a top conference in parallel problem solving from nature. The paper is available on arXiv (2608.14209).