Any recommendations for choice of a constrained optimization library suitable for my optimization function? I am minimizing a i) non-linear function with linear equality and inequality constraints, and ii) have available the gradient and the hessian of the function.

If it helps, the function I am minimizing is the Kullback-Liebler divergence.

constrOptim only deals with inequality constraints. Quadprog handles quadratics. Trust does not support constraints. So the KL divergence does not fit into these solutions.

There are quite a few solutions on the R Cran Task page for Optimization. Iam able to perform the optimization in MATLAB using the fmincon() function which seems to use an interior-point or a trust-region-reflective. Ideally there is a library that is well-suited to the problem defined.

  • Are the constraints linear? – cardinal Dec 14 '11 at 2:05
  • @cardinal - yes - the constraints are linear – Ram Ahluwalia Dec 14 '11 at 2:12
  • 2
    So, if your equality constraints are $A x = b$, at least a work-around for constrOptim would be to include both $A x \leq b$ and $A x \geq b$. No? – cardinal Dec 14 '11 at 2:14
  • That is very clever. I will explore this and see how it performs. You should consider posting this as an answer. I'll keep the question open for a couple days to see what other tools are available – Ram Ahluwalia Dec 14 '11 at 2:55
  • 1
    @cardinal trick doesn't help for interior point method, like the one of constrOptim, since the method need a starting point into the interior of the feasible region -not on the border – stackovergio Oct 2 '14 at 8:08
up vote 12 down vote accepted

Both packages, alabama and Rsolnp, contain "[i]mplementations of the augmented lagrange multiplier method for general nonlinear optimization" --- as the optimization task view says --- and are quite reliable and robust. The can handle equality and inequality constraints defined as (nonlinear) functions again.

I have worked with both packages. Sometimes, constraints are a bit easier to formulate with Rsolnp, whereas alabama appears to be a bit faster at times.

There is also the package Rdonlp2 that relies on an external and in the optimization community well-known software library. Unfortunately, its license status is a bit uncertain at the moment.

  • 1
    Just want to add one property of Alabama inequality constraints which I found today. If your inequality constraints creates an infeasible region, then the code runs without any warning/error messages and takes the mean value of boundaries as fixed value of the parameter. For example if you have x > 6 and x <4, then the solution will provide a solution with x = 5 without any warning. – Gaurav Singhal Apr 3 '17 at 11:13

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.