Study finds GP initialization method has negligible effect on SR results

A new study published on arXiv (2606.31990) by Lukas Kammerer and colleagues systematically evaluates how the initial population affects genetic programming (GP) for symbolic regression (SR). The researchers compared three standard random initialization methods against a more sophisticated approach: seeding the population with small, optimized solutions from exhaustive symbolic regression (ESR). The experiments used a multi-objective evolutionary algorithm (NSGA-II) across twelve synthetic problems of varying complexity and one real-world dataset.

Contrary to expectations, the team found that after just a few generations, any initial advantage from the ESR-based initialization disappeared. The final Pareto fronts—showing the trade-off between accuracy and model complexity—were statistically indistinguishable across all initialization methods. The key insight: as long as initial diversity is similar, the choice of initialization method has negligible impact on the final quality of GP-based symbolic regression. This negative result is valuable for practitioners, suggesting they can use simpler random initialization without sacrificing performance.