Prev: ? remove dicom info
Next: filter design in matlab
From: Amar on 16 Apr 2010 11:20
Hello Peter,
I tried using statset('robust','on') in my code, but it gave me an error

I used the following code:

options = statset('FunValCheck','off','DerivStep',10^-12,'robust','on');

beta = nlinfit([parameter_values(3,:);parameter_values(4,:)],Y,@NTCPmodel, beta_seed,options);

And this the error:

??? Error using ==> times
Matrix dimensions must agree.

Error in ==> nlinfit>nlrobustfit at 417
radj = r .* adjfactor;

Error in ==> nlinfit at 188
[beta,J,sig,cause] =
nlrobustfit(X,y,beta,model,J,ols_s,options,verbose,maxiter);

Any pointers as to whats going on there?

Thanks
Amar

Peter Perkins <Peter.Perkins(a)MathRemoveThisWorks.com> wrote in message <guf4j2$hsr$1(a)fred.mathworks.com>...
> Bendik Mjaaland wrote:
> > Hi!
> >
> > I have worked with a prediction task for some time, now I have decided to use nonlinear regression. I have little experience with matlab, so bear with me.
> >
> > so, I have a set of data points YDATA, about 60 values.
> > I define the X-axis
> > X1 = 1:60;
> >
> > I have two models:
> > y = a + b*log(1 + c/x)
> > y = a + b*exp(c/x)
> > implemented in the following functions
> > function [F] = funlog(a,data)
> > F=a(1)+a(2)*log(1 + a(3)./data);
> > function [F] = funexp(a,data)
> > F=a(1) + a(2)*exp(a(3)./data);
> >
> > initial beta0 = [10 5 5];
> > To generate the beta values
> > [beta,r,J,COVB,mse] = nlinfit(X1,YDATA,@funexp,beta0);
> > and the same for @funlog
> >
> > Issue 1:
> > Both functions give me this error
> > Warning: The Jacobian at the solution is ill-conditioned, and some
> > model parameters may not be estimated well (they are not identifiable).
> > - why does this happen?
> > - how can I make sure my results are valid?
>
> In general this means one of two things:
>
> 1) two or more parameters in the model are aliased, e.g., b and c in y = a + b*exp(c+x), or
>
> 2) the solver has wandered off to a region in the parameter space that is far from the "true" MLE, and the log-likelihood surface has a very bad shape. It sounds from your followup post like that might be the case. Good starting values are pretty important in non-linear least squares, and in general it's impossible to pick them automatically.
>
> > Issue 2:
> > The exp-function also gives me this error:
> > Warning: Iteration limit exceeded. Returning results from final iteration.
> >> In nlinfit at 193
> > - Is this something I can modify in the options? How? Unlike LSQCURVEFIT, there is no options parameter.
>
> It isn't, but if NLINFIT doesn't converge in 100 iterations, that's a pretty good indication of something gone wrong. You can restart at the last value if you want, but given the above, it sounds like that would be pointless.
>
> > Issue 3:
> > The log-function returns complex values. Not being a statician, i hardly understand what this complex results really mean, I understand the cause - the logarithm of something negative produces an imaginary number, but as far as I know I do not need anything this sophisticated for an answer. I am interested in the plots, and plotting the log-results was really weird, the graph was not contineous.
> > - Can I merely ignore the complex addition so that if " a = 10.424 + 1.2325i " I can use "a = 10.424"?
> > - Can I tell matlab only to return real numbers?
>
> No, iy's not something to ignore. It's kind of hard to diagnose this without specifics. What are the 60 values?
>
> > And final question
> > - The documentation says something about outliers, which I would like to ignore. But I cannot figure out how this works.
>
> You probably wan to look at using the 'Robust' option. First call statset('robust','on'), then pass the resulting options structure into NLINFIT.
>
> Hope this helps.
From: Amar on 16 Apr 2010 13:34
Hello Peter,
I tried using statset('robust','on') in my code, but it gave me an error

I used the following code:

options = statset('FunValCheck','off','DerivStep',10^-12,'robust','on');

beta = nlinfit([parameter_values(3,:);parameter_values(4,:)],Y,@NTCPmodel, beta_seed,options);

And this the error:

??? Error using ==> times
Matrix dimensions must agree.

Error in ==> nlinfit>nlrobustfit at 417
radj = r .* adjfactor;

Error in ==> nlinfit at 188
[beta,J,sig,cause] =
nlrobustfit(X,y,beta,model,J,ols_s,options,verbose,maxiter);

Any pointers as to whats going on there?

Thanks
Amar

Peter Perkins <Peter.Perkins(a)MathRemoveThisWorks.com> wrote in message <guf4j2$hsr$1(a)fred.mathworks.com>...
> Bendik Mjaaland wrote:
> > Hi!
> >
> > I have worked with a prediction task for some time, now I have decided to use nonlinear regression. I have little experience with matlab, so bear with me.
> >
> > so, I have a set of data points YDATA, about 60 values.
> > I define the X-axis
> > X1 = 1:60;
> >
> > I have two models:
> > y = a + b*log(1 + c/x)
> > y = a + b*exp(c/x)
> > implemented in the following functions
> > function [F] = funlog(a,data)
> > F=a(1)+a(2)*log(1 + a(3)./data);
> > function [F] = funexp(a,data)
> > F=a(1) + a(2)*exp(a(3)./data);
> >
> > initial beta0 = [10 5 5];
> > To generate the beta values
> > [beta,r,J,COVB,mse] = nlinfit(X1,YDATA,@funexp,beta0);
> > and the same for @funlog
> >
> > Issue 1:
> > Both functions give me this error
> > Warning: The Jacobian at the solution is ill-conditioned, and some
> > model parameters may not be estimated well (they are not identifiable).
> > - why does this happen?
> > - how can I make sure my results are valid?
>
> In general this means one of two things:
>
> 1) two or more parameters in the model are aliased, e.g., b and c in y = a + b*exp(c+x), or
>
> 2) the solver has wandered off to a region in the parameter space that is far from the "true" MLE, and the log-likelihood surface has a very bad shape. It sounds from your followup post like that might be the case. Good starting values are pretty important in non-linear least squares, and in general it's impossible to pick them automatically.
>
> > Issue 2:
> > The exp-function also gives me this error:
> > Warning: Iteration limit exceeded. Returning results from final iteration.
> >> In nlinfit at 193
> > - Is this something I can modify in the options? How? Unlike LSQCURVEFIT, there is no options parameter.
>
> It isn't, but if NLINFIT doesn't converge in 100 iterations, that's a pretty good indication of something gone wrong. You can restart at the last value if you want, but given the above, it sounds like that would be pointless.
>
> > Issue 3:
> > The log-function returns complex values. Not being a statician, i hardly understand what this complex results really mean, I understand the cause - the logarithm of something negative produces an imaginary number, but as far as I know I do not need anything this sophisticated for an answer. I am interested in the plots, and plotting the log-results was really weird, the graph was not contineous.
> > - Can I merely ignore the complex addition so that if " a = 10.424 + 1.2325i " I can use "a = 10.424"?
> > - Can I tell matlab only to return real numbers?
>
> No, iy's not something to ignore. It's kind of hard to diagnose this without specifics. What are the 60 values?
>
> > And final question
> > - The documentation says something about outliers, which I would like to ignore. But I cannot figure out how this works.
>
> You probably wan to look at using the 'Robust' option. First call statset('robust','on'), then pass the resulting options structure into NLINFIT.
>
> Hope this helps.
From: Tom Lane on 19 Apr 2010 18:31
> options = statset('FunValCheck','off','DerivStep',10^-12,'robust','on');
>
> beta = nlinfit([parameter_values(3,:);parameter_values(4,:)],Y,@NTCPmodel,
> beta_seed,options);
>
> And this the error:
>
> ??? Error using ==> times
> Matrix dimensions must agree.

Amar, this is hard to figure out without the data. If you are comfortable
debugging, you may want to set a breakpoint at the location of the error and
see what's going on.

Also, is there are reason you turn off FunValCheck? That might help here. As
best I can tell looking at the code, the function is finding an NaN in an
unexpected place, removing it, and then getting an array of the wrong size.

-- Tom


 | 
Pages: 1
Prev: ? remove dicom info
Next: filter design in matlab