Can I see the parameters that fminunc generate in each try of the optimization search?

Hi,

I am using fminunc to optimize a likelihood function which inside has other functions which makes the problem a bit complicated.

When I try the values of all the matrices inside this likelihood with the initial parameters all is correct. With this I mean that I have correct covariance matrices, no zero determinants etc.

However, when I start the optimization process at a certain parameter values I get a singular matrix in one of the equations inside the likelihood.

Therefore, what I would like to do is to observe the vector of parameters being chosen by the fminunc in the iterations.

Is there a way to do this?

Thanks in advance.

Barbara

Answer 1

[Alan Weiss]

You can write an output function to obtain almost everything you want. However, fminunc takes finite difference steps that the output function will not report; output functions get called at each iteration, not at each step in the process.

If what you really want is to have the solver continue past a singularity, you might try using a different solver. The fmincon interior-point and sqp algorithms are robust to some failures in function evaluation. You need a constraint to run fmincon; feel free to use lb = -Inf, which fmincon accepts as a constraint but which has no bearing on the solution.

Good luck,

Alan Weiss

MATLAB mathematical toolbox documentation

Answer 2

[Barbara]

Thanks a lot! I will dig into this.

Answer 3

[Barbara]

Dear Alan,

so I have dived into this issue and I found that it is quite simple to get the outputs of each iteration. For the case of the likelihood values at each iteration inside the fminsearch (for instance) (it is a simple change of definition for the parameters case) things goes like this:

//create empty matrix global_likelihood_vector = [];

//Settings and call to "fminsearch" ... ...

function [current_likelihood] = myLikelihood(x);

   //Calculate likelihood
   current_likelihood = ...

   //Concatenate current likelihood to vector
   //of all likelihoods
   global_likelihood_vector = [global_likelihood_vector ; current_likelihood];

end;

And thats it :)