GP-GOMEA gets simultaneous constant and structure optimization for symbolic regression
New merging of real-valued GOMEA with GP-GOMEA boosts accuracy by optimizing constants and expression structure together.
Johannes Koch and colleagues present a new method for symbolic regression that addresses a key limitation of GP-GOMEA—its reliance on ephemeral random constants without dedicated constant optimization. By merging GP-GOMEA with the real-valued variant of GOMEA, the approach optimizes expression structure and constants in a single, simultaneous evolutionary process. This integrated mixed discrete-continuous optimization is shown to outperform alternative strategies such as linear scaling, restarts, and constant tuning after GP optimization, achieving better accuracy on benchmark problems.
The work builds on prior research indicating that simultaneous optimization of both discrete and continuous parts leads to superior results, especially when interactions exist between them. The proposed approach generally performs best across comparisons, confirming the importance of co-optimizing constants and structure during evolution. The paper was presented at PPSN 2024 and is available on arXiv.
- Merges real-valued GOMEA with GP-GOMEA for simultaneous optimization of constants and expression structure.
- Outperforms linear scaling, restarts, and post-hoc constant tuning across benchmark symbolic regression problems.
- Confirms that well-integrated mixed discrete-continuous evolution yields more accurate and compact symbolic expressions.
Why It Matters
This advancement could make symbolic regression tools more accurate for scientific discovery and data modeling applications.