A new open-source tool targets one of symbolic regression's persistent weak spots: datasets with far more variables than rows.
Researchers released Evolutional Math, a genetic programming system built around four specific design choices. Fitness is scored on held-out cross-validation folds — not training-set correlation — so the search can't cheat by finding variables that correlate but don't predict. Independent "island" populations, each seeded with a different operator set (algebraic, logarithmic, trigonometric, or all of the above), run in parallel and exchange candidates periodically, keeping the search from collapsing into one corner of formula space. A deduplication layer strips near-identical expressions that differ only in constants, so the system's elite archive stays structurally diverse. Finally, the top candidates get their numeric constants tuned by a standard numerical optimizer after each migration cycle, separating the job of finding a formula's shape from fitting its numbers.
The target regime — small row counts, wide feature sets — shows up constantly in clinical-trial monitoring, biostatistics, and engineering pilot studies, where collecting more rows is expensive or impossible. Standard symbolic regression tends to overfit badly here, producing bloated expressions that describe the training noise rather than the underlying relationship. On a real anonymized clinical dataset with 24 rows and roughly 290 candidate features, the system recovered compact formulas with R-squared at or above 0.99.
The code is out under a noncommercial source-available license, which puts it closer to "open for inspection" than fully open-source — a distinction worth noting before anyone builds a product on top of it.