How to use OutputFcn with fitnlm

Hi all,

How do I execute an OutputFcn function in each iteration inside fitlm? I have tried to run the following code but it does not print anything or throw an exception.

  load carbig
    tbl = table(Horsepower,Weight,MPG);
    modelfun = @(b,x)b(1) + b(2)*x(:,1).^b(3) + ...
                        b(4)*x(:,2).^b(5);
    
    beta0 = [-50 500 -1 500 -1];
    opts = statset("fitnlm");
    opts.OutputFcn = @(x)disp(x);
    mdl = fitnlm(tbl,modelfun,beta0,"Options",opts);

[dpb]

fitnlm doesn't document the 'OutputFcn' field in the options list; I'm guessing it's being ignored.

Try nlinfit instead.

This penchant for multiple nearly identical functions is maddening, indeed...

which -all fitnlm

/MATLAB/toolbox/stats/classreg/fitnlm.m

which -all nlinfit

/MATLAB/toolbox/stats/stats/nlinfit.m

Answer 1

[Matt J]

Only Optimization Toolbox solvers have an OutputFcn option. Perhaps you are thinking of lsqcurvefit,

    load carbig
    tbl = table(Horsepower,Weight,MPG);
    modelfun = @(b,x)b(1) + b(2)*x(:,1).^b(3) + ...
                        b(4)*x(:,2).^b(5);
    
    beta0 = [-50 500 -1 500 -1];
    
    opts = optimoptions('lsqcurvefit',OutputFcn=@outfun);

    beta1 = lsqcurvefit(modelfun,tbl{:,1:2},tbl{:,3},beta0,"Options",opts);


function stop = outfun(x,optimValues,state)
    stop = false;
    % Check whether directional derivative norm is less than .01.
    if strcmp(state,'iter')
        disp(x)
    end
end

[Khadija]

HI Mr Matt J,

If i use lsqcurvefit in my code, can i have more statistical informations with it , like confidence interval, plot the shaded region?

[Torsten]

If statistics for the fitted parameter were accessible when using "lsqcurvefit", it would be mentionned on the documentation page as possible output from the solver:

https://uk.mathworks.com/help/optim/ug/lsqcurvefit.html

But it isn't.

[Matt J]

@Adrian Batista Planas,

once you have solved with lsqcurvefit, you can take the solution beta1 and use it for beta0 with fitnlm:

    load carbig
    tbl = table(Horsepower,Weight,MPG);
    modelfun = @(b,x)b(1) + b(2)*x(:,1).^b(3) + ...
                        b(4)*x(:,2).^b(5);
    
    beta0 = [-50 500 -1 500 -1];
    
    opts = optimoptions('lsqcurvefit',OutputFcn=@outfun);

    beta1 = lsqcurvefit(modelfun,tbl{:,1:2},tbl{:,3},beta0,"Options",opts);
 
    mdl = fitnlm(tbl,modelfun,beta1,"Options",statset("fitnlm"));

You can then use mdl for whatever statistical analysis you had originally planned.