Prev: Plotting raw spectrogram
Next: rgb2gray not working
From: TS on 19 May 2010 16:43
I am using fmincon for estimating 6 different variables stored in 2 * 3 matrix C. These are actually probability values of 3 elements across 2 different classes. (row -> class no, column -> data instance no)
I have only linear inequality constraints i.e. sum across each column of C should be 1(the constraint we have for probability values). So, I should only specify A & b matrices in fmincon. I have defined A & b as 1*2 and 1*3 vectors. But an error is occurring saying

A must have 6 column(s).

[which is the total no of variables in the matrix(2*3)]

To correct this, I have specified A & b as 1*6 vectors but now I am getting another error

??? Error using ==> times
Matrix dimensions must agree.
Error in ==> fmincon at 759
[X,FVAL,EXITFLAG,OUTPUT,LAMBDA,GRAD,HESSIAN] =
barrier(funfcn,X,A,B,Aeq,Beq,l,u,confcn,options_ip.HessFcn, ...


Can somebody please give an example on how to use matrix variables with fmincon, given linear inequality constraints in the form of A & b?
I had been searching for such demo but couldnt get any.

Thanks,
From: Matt J on 19 May 2010 19:18


Don't pose the problem in terms of the matrix C. Pose it instead in terms of the vector x=C(:)

In this case your constraints on x are

x(1)+x(2)=1
x(3)+x(4)=1
x(5)+x(6)=1

which are expressed in the matrix form A*x=b with

A =

1 1 0 0 0 0
0 0 1 1 0 0
0 0 0 0 1 1

and

b=ones(3,1)
From: TS on 21 May 2010 09:41
Thanks Matt for the solution. But the problem is that this method is feasible if I am optimizing a few variables(stored in the form of vector).. But my real optimization problem involves estimating the probability values of half million variables so I guess specifying A in the way you have mentioned is probably not a good option..

Can someone point out any other way for specifying the linear constraints in case C is a huge matrix of thousands of variables?

Thanks a lot,


"Matt J " <mattjacREMOVE(a)THISieee.spam> wrote in message <ht1rju$37l$1(a)fred.mathworks.com>...
>
>
> Don't pose the problem in terms of the matrix C. Pose it instead in terms of the vector x=C(:)
>
> In this case your constraints on x are
>
> x(1)+x(2)=1
> x(3)+x(4)=1
> x(5)+x(6)=1
>
> which are expressed in the matrix form A*x=b with
>
> A =
>
> 1 1 0 0 0 0
> 0 0 1 1 0 0
> 0 0 0 0 1 1
>
> and
>
> b=ones(3,1)
From: Marcus M. Edvall on 24 May 2010 13:24
Hi,

If you have half a million variables you need to look at TOMLAB.

Just create the linear constraints in a sparse matrix and you should
be fine.

You can get a demo from here as needed: http://tomopt.com/scripts/register.php

Best wishes, Marcus
http://tomsym.com/
http://tomdyn.com/

From: Bruno Luong on 24 May 2010 13:51
"TS " <tanwistha4u(a)yahoo.com> wrote in message <ht62hj$lp2$1(a)fred.mathworks.com>...
> Thanks Matt for the solution. But the problem is that this method is feasible if I am optimizing a few variables(stored in the form of vector).. But my real optimization problem involves estimating the probability values of half million variables so I guess specifying A in the way you have mentioned is probably not a good option..
>
> Can someone point out any other way for specifying the linear constraints in case C is a huge matrix of thousands of variables?
>
> Thanks a lot,

SPARSE matrix won't help?

Bruno
 | 
Pages: 1
Prev: Plotting raw spectrogram
Next: rgb2gray not working