You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Something I have noticed is that many users expect the result of .fit(X, y) to be converged for the default parameters, and are surprised that running it a second time results in a different equation. This is a reasonable assumption given that converged solutions are the typical behavior for other ML models or optimization routines!
So, I think the default behavior for PySR should be a bit more clear that "convergence" is not really well-defined for symbolic regression. Perhaps we could leave the same behavior, but basically show a warning to the user if niterations is not set explicitly, such as:
No stopping criteria given, so will exiting after 40 iterations. Increase `niterations` to search for longer.
This seems like a simple solution which would make it clear to users that they shouldn't expect the equations to be converged.
Something I have noticed is that many users expect the result of
.fit(X, y)
to be converged for the default parameters, and are surprised that running it a second time results in a different equation. This is a reasonable assumption given that converged solutions are the typical behavior for other ML models or optimization routines!So, I think the default behavior for PySR should be a bit more clear that "convergence" is not really well-defined for symbolic regression. Perhaps we could leave the same behavior, but basically show a warning to the user if
niterations
is not set explicitly, such as:This seems like a simple solution which would make it clear to users that they shouldn't expect the equations to be converged.
We might also think about renaming
niterations
tomax_iterations
to be more clear about this. We should make this in line with the other stopping criteria: https://astroautomata.com/PySR/api/#stopping-criteriaThe text was updated successfully, but these errors were encountered: