Constrained optimization

[sloisel]

How to implement constrained optimization?

• Which algorithm to use? Barrier method?
• How much derivative information is needed for constraint function(s)? How to pass constraint function(s)? Is it an Array of functions, or a function returning an Array? (Latter case: is the gradient a Jacobian? Will it need an order 3 Hessian as well? Does not seem reasonable since the penalty function ends up being scalar-valued.)
• How competitive should it be compared to numeric.solveLP, when used on a linear program?
• What is a suitable battery of tests for this? This is important, untested optimization programs usually don't work.

[albertosantini]

Hey @sloisel.

I developed a module for solving quadratic programming problems with constraints: https://github.com/albertosantini/node-quadprog

Hope that helps and available to changes, Alberto

[sloisel]

Dear Alberto,

Thanks for your note. Your program is already included in numeric.js!

Thanks,

Sébastien Loisel

[albertosantini]

I see... Thanks for the reference.

I didn't implement the lagrangian, but it is matter of a few lines of code, just in case. I will link numericjs in quadprog readme.

Regards, Alberto

[sloisel]

Hi,

Yes, I have included your License in the resources/ subfolder. I've slightly patched quadprog.js but otherwise left it untouched. I have also mentioned that you did this in the documentation.

Thanks,

Sébastien Loisel

[albertosantini]

Please, update to 1.4.0 (available in a few hours), because I fixed an important bug in the constraints and added input arguments validation.

Please, update also, LICENSE.quadprog in resources, because I removed the email address for the spam.

Available to any change, if you think there is a better way to include quadprog in numeric without copy&paste (more changes) when the package is updated.