How can I avoid the negative results from nlinfit function?

How can I avoid the negative results from the nlinfit function? Some articles advice to use lsqcurvefit but how and where, I don't understand?

thankx a lot

Answer 1

[Matt J]

LSQCURVEFIT allow you to specify upper and lower bounds on the parameters via the ub and lb input arguments. I can't tell if NLINFIT lets you do the same. You could also try a transformation of your parameters that ensures positivity, i.e., instead of making your model a function of a parameter x, make it a function of x=z^2 and fit with respect to z.

[Amjad]

Thanks for answer, but I'm a beginner in using matlab, and I dont understand how to use LSQCURVEFIT, should I use it on nlinfit results or instead of it??

thanks a lot,

[Matt J]

Instead of it. For instructions on its use,

 >> doc lsqcurvefit

[mwer]

Is there any possibility to specify lower and upper bonds for the resultig value y (y = f(x1,x2,...,xn))?

[Matt J]

Not with lsqcurvefit. You can pose the fitting as a non-linearly constrained optimization problem and solve with FMINCON. However, bound constraints lb<=f()<=ub on a nonlinear function f can give feasible regions that are not connected and therefore, difficult to optimize over.

[John D'Errico]

Matt is completely correct here. I'll add that one common reason for needing a positive result is that your model is really not subject to additive Gaussian noise. This is the presumption for nlinfit, as well as many other tools like lsqnonlin, etc.

When you have the above problem, it may indicate that you have proportional (multiplicative) noise. The trick then is to log your model. So instead of trying to fit a model in the form

y = f(X,P)

you fit a model of the form

log(Y) = G(X,P)

where G is simply log(F).

The nice thing is now your predictions are entirely positive, ALWAYS. This is true because at the end, you will predict Y as exp(G(X,P)). You can still use standard tools like nlinfit, lsqcurvefit, etc. Effectively you have a noise structure that is lognormal, instead of the traditional additive normal noise.