fminunc on multiple variables

Hi,

I am new to using fminunc and have looked quite a bit into it but couldn't find what I was looking for. If I have a function like: myFun(x,y,z,a,b,c), how can i use fminunc to optimize over x,y and z simlutaneously.

REGARDS

[Shivani Chepuri]

Hi, Can you specify how to write lower and upper bound constraints for x,y,z variables here, using fminunc and fmincon? Also, all x,y,z are vectors. Thanks

Answer 1

[Walter Roberson]

@(X) myFun(X(1), X(2), X(3), a, b, c)

[Asad]

THANKS. What if x,y and z are matrices of different sizes. Will this method still work?

[Walter Roberson]

In such a case, make your initial point, x0, a vector whose total length is the sum of the number of elements in x, y, and z. Then in the anonymous call, extract and (if necessary) reshape the elements of the argument and pass them in to the appropriate location.

@(X) myFun( X(1:5), reshape(X(6:26), 3, 7), reshape(X(27:30), 2, 2), a, b, c)

for the case where the first argument should be 5 x 1, the second 3 x 7, the third 2 x 2.

[Louis Perriens]

Very solid!! thank you!!

[Shivani Chepuri]

Hi, Can you specify how to write lower and upper bound constraints for x,y,z variables here, using fminunc and fmincon? Also, all x,y,z are vectors. Thanks

Answer 2

[Sean de Wolski]

Objective functions expect all of the variables being optimized to be in "x"

myfun(x,a,b,c)

Where x is now a 1x3 vector. Then inside of myfun:

Use x(1) to refer to the first element, x(2) for y and x(3) for z. To pass in other variables, look at:

http://www.mathworks.com/help/releases/R2013b/optim/ug/passing-extra-parameters.html

[Asad]

THANKS. Can x be a cell in which each element is a matrix of different size?

[Sean de Wolski]

No, x must be a vector.

This would mean that you would be optimizing every value in the matrix to solve a numel(elements_in_matrix) dimensional problem - something completely different.

What are you trying to do?

Also, just to be clear, the additional parameters (a,b,c) that are not being perturbed for the minimization can be anything.

[Asad]

I am trying to do unconstrained optimization to calculate location of 3D points, planes and camera matrices.

[Sean de Wolski]

So you are basically trying to optimize over n dimensions where n is the total number of elements in all of your matrices.

You can do this the same way Walter and I have suggested: essentially pass in a 1xn vector x and then split it into its matrices etc. inside of the objective function.

Since I would expect fminunc to have trouble with a problem like this due to many local minima and discontinuities, you might want to try using patternsearch in the Global Optimization toolbox. This solver is much better able to handle the above. You might also choose to constrain you search wherever possible. The constraints might not be obvious but some things to think about: are some variables positive? are some within a range? Do some of their value depend on others or for example a row of your matric must sum to x? etc.

http://www.mathworks.com/help/releases/R2013b/gads/patternsearch.html