Can I implement lsqnonlin (or equivalent) in a way that handles NaNs?

I am trying to solve a non-linear least squares problem using Matlab's optimisation tool lsqnonlin. Specifically, I am trying to minimise the residual between my model and the data. However, my data, and hence the output of my anonymous function I use as input to lsqnonlin contains NaNs; causing lsqnonlin to fall over.

"Objective function is returning undefined values at initial point. lsqnonlin cannot continue."

Is there a way in which I can implement lsqnonlin (or equivalent) that can handle NaNs?

I have looked into giving my NaN values a zero weight, but I don't know whether this would work in this case? I have also thought about including logic such that the lsqnonlin only applies to values of my anonymous function that are non-NaN, but this seems improper, and, I suspect, very slow.

What does the community recommend?

Answer 1

[Matt J]

Specifically, I am trying to minimise the residual between my model and the data

Since your residual is of the form F(x,xdata)-ydata, it is better to use lsqcurvefit.

Remove the NaNs from your ydata before feeding it to lsqcurvefit. Write a model function F(x,xdata) that considers only the NaN-reduced ydata.