Benchmarks
Ufinq is measured in two ways: against other methods on the datasets the field uses, and on the data people bring to it.
Against other methods
A controlled comparison against PySR, Operon and gplearn on datasets from the SRBench corpora: PMLB, Feynman and Strogatz. Same data, same time, several seeds. The weekly sweep runs ten datasets from PMLB and the UCI repository.
- Latest sweep
- —
On the data people bring
Every public analysis on Diafunc, as its owner ran it. Nothing is re-run or selected: the page shows where Ufinq did well and where it did badly.
- Public analyses
- —
Why two
The standardized benchmark answers how Ufinq compares with the alternatives: every method gets the same data and the same time, so a difference is a result. It cannot say how Ufinq does on the data people have. The real-world benchmark shows that, but each analysis is run once, by its owner, so it is an observation and not a comparison. How both are run is described under Methodology. The public analyses themselves are on Diafunc: browse them, or run one on your own data.