Convex optimization with linear constraint?
I'm using PETSc + TAO's LMVM method for a convex optimization problem. As the project progresses, it's clear that some linear constraints are also needed, the problem now looks like: minimize f(vec_x) (f is neither linear or quadratic, but is convex) subject to A * vec_x = vec_b As LMVM does not support linear constraints, I'm looking for another solver. TAO lists several functions dealing with constraints, but they're all in the developer section, and in the samples linked from the manual I haven't found one that's linearly constrained. Is there a suitable one in TAO? Liu Chang
Liu, Unless I'm missing something, I don't think you will directly find what you are looking for. You will prob have to solve A * vec_x = vec_b directly in your FormObjective function then use an adjoint method or something to compute your gradient directly in your FormGradient routine. On Thu, 30 Apr 2009, liu chang wrote:
I'm using PETSc + TAO's LMVM method for a convex optimization problem. As the project progresses, it's clear that some linear constraints are also needed, the problem now looks like:
minimize f(vec_x) (f is neither linear or quadratic, but is convex) subject to A * vec_x = vec_b
As LMVM does not support linear constraints, I'm looking for another solver. TAO lists several functions dealing with constraints, but they're all in the developer section, and in the samples linked from the manual I haven't found one that's linearly constrained. Is there a suitable one in TAO?
Liu Chang
Thanks. I made LMVM optimize over a reduced number of variables and solve the rest from the linear equations. On Thu, Apr 30, 2009 at 3:37 AM, David Fuentes <[email protected]> wrote:
Liu,
Unless I'm missing something, I don't think you will directly find what you are looking for. You will prob have to solve
A * vec_x = vec_b
directly in your FormObjective function then use an adjoint method or something to compute your gradient directly in your FormGradient routine.
On Thu, 30 Apr 2009, liu chang wrote:
I'm using PETSc + TAO's LMVM method for a convex optimization problem. As the project progresses, it's clear that some linear constraints are also needed, the problem now looks like:
minimize f(vec_x) (f is neither linear or quadratic, but is convex) subject to A * vec_x = vec_b
As LMVM does not support linear constraints, I'm looking for another solver. TAO lists several functions dealing with constraints, but they're all in the developer section, and in the samples linked from the manual I haven't found one that's linearly constrained. Is there a suitable one in TAO?
Liu Chang
participants (2)
-
David Fuentes -
liu chang