What Quadratic Programming Package?

[dudefromdayton]

@jbaker, make sure you're not creating test cases with all points on the same side of the line (see instructions). Maybe then you won't get the positive definite complaint.

[epsilon]

I use a built-in command qp in octave.
http://www.gnu.org/software/octave/d...ogramming.html

It seems to work, but I am not sure it can do well like cvxopt and other packages.

[alfansome]

Originally Posted by lucag

Thanks for the input ... I found the CVX library for Matlab and I will (try to) use that.
@Marcello I don't think any of those libraries are system specific, i.e., they can be used on a Mac as well as on Windows, or Linux, or BSD, or ...

I've had problems trying to install cvxopt on mac os x 10.7 (with Apple supplied python), so looking for alternate approaches.

[lucag]

Originally Posted by alfansome

I've had problems trying to install cvxopt on mac os x 10.7 (with Apple supplied python), so looking for alternate approaches.

what kind of problems? I have a Mac running Lion and installed cvxopt without a glitch...

[jbaker]

Yes, I caught the not-all-on-one-side requirement.

Now I've tried both in R and also using the CGAL library with C++, but the latter is also complaining that Q is not positive semidefinite!

Which makes me think that I'm just being silly and constructing Q wrong. If you have points x_(d,i) where d = {1,2} is the dimension and i = {1..N}, with response y_(i) in {-1, 1}, then the matrix elements are

Q_(ij) = y_(i) * y_(j) * sum_d [ x_(d,i) * x_(d,j) ]

Right?

[alfansome]

Originally Posted by lucag

what kind of problems? I have a Mac running Lion and installed cvxopt without a glitch...

I downloaded the tar.gz file cvxopt-1.1.5.tar.gz, unpacked it, and then ran the python setup program to build it; generated hundreds of errors and no usable cvxopt module for python. I also tried using easy_install and it too failed. Again, I am just using the python that comes bundled with mac os x (its python 2.7.1).

How did you install it?

[dudefromdayton]

@jbaker, I didn't have to go that far into the weeds, as I used an SVM package as opposed to a QP package. If you ignore the warnings, are you getting a sensible number of support vectors out? If you aren't, there definitely is a problem.

I don't see a problem with your formulation for Q, but again I didn't go that route. Someone else might see a problem that I missed.

[dudefromdayton]

@alfansome, are you okay on the second paragraph from the cvxopt installation instructions? I haven't done this myself, but I looked it up.

"The package requires version 2.7 or 3.x of Python, and is built from source, so the header files and libraries for Python must be installed, as well as the core binaries.

"The installation requires either ATLAS or BLAS + LAPACK. Using architecture optimized ATLAS libraries is recommended and gives a large performance improvement over standard BLAS & LAPACK libraries. Both header files and libraries must be installed."

[lucag]

Originally Posted by alfansome

I downloaded the tar.gz file cvxopt-1.1.5.tar.gz, unpacked it, and then ran the python setup program to build it; generated hundreds of errors and no usable cvxopt module for python. I also tried using easy_install and it too failed. Again, I am just using the python that comes bundled with mac os x (its python 2.7.1).

How did you install it?

mmm that's weird: I do get tons of warnings but at the end it works just fine (tried with one of the examples provided). The version installed is 2.7.1
Did you install it globally or only for a specific user? I don't know if it changes anything, but I installed it as root -- sudo python setup.py install -- on my machine.

[jbaker]

After way too much sanity checking, I convinced myself that one should indeed plow ahead despite the positive-definiteness alarms. However in order to get CGAL not to abort, you have to #define CGAL_QP_NO_ASSERTIONS before including the headers! Not sure how one would get the same behavior in R.

CGAL seems pretty slick, btw.