Tree Structure Choice Affects Symbolic Regression Success
A new paper on arXiv (2604.23256) investigates how the fixed tree structure in symbolic regression impacts recovery of target formulas. Three structures, sharing operators and target language but differing in variable entry, were tested across over 12,700 training runs. One structure achieved 100% recovery on one target while another scored 0%, with rankings reversing on a different target. Expressiveness does not guarantee gradient descent success; the most expressive structure failed on targets a restricted alternative solved reliably. Switching operators also changed which targets succeeded.
Key facts
- Paper arXiv:2604.23256 tests tree structure choice in symbolic regression.
- Three structures compared, differing in how variables enter the tree.
- Over 12,700 training runs conducted.
- One structure recovers a target at 100%, another at 0%.
- Ranking reverses on a different target.
- Most expressive structure fails on targets a restricted alternative solves.
- Switching operators changes which targets succeed.
- Expressiveness does not guarantee gradient descent finds the solution.
Entities
Institutions
- arXiv