Generalization of mutation-only GP benefits beyond polynomial benchmarks

Investigate whether the performance advantage of mutation-only genetic programming ((1+λ)-GP) observed on Large-Scale Polynomial benchmarks generalizes to other symbolic regression problem classes, and determine the conditions under which mutation-only variation operators outperform recombination-based approaches.

Background

In experiments on the Large-Scale Polynomial benchmark, the authors observed that mutation-only GP ((1+λ)-GP) suffers less from bloat and converges faster than canonical GP while achieving similar final loss and span. This suggests a potential advantage of mutation-only variation for polynomials.

The authors explicitly defer studying whether this advantage extends to other problem types, making the generalization question an open problem.