Forum: developers
Monitor Forum | 
Start New Thread
| RE: constrained optimization [ Reply ] By: Arne Henningsen on 2009-08-31 16:44 | [forum:1869] |
|
Hi Ott! It is great that you added support for imposing (linear inequality) constraints on the parameters! This is even more general than box constraints (as in L-BFGS-B of optim() or in nlminb(), which uses the PORT routines). However, the latter approaches might be more efficient in case of (only) box constraints. I don't know an optimization package on CRAN that allows for equality constraints. Do you any? Anyway, equality constraints can be implemented by the *user* by changing his/her likelihood function (and removing one parameter per constraint). Hence, I don't think that this should have a high priority on our TODO list. |
|
| constrained optimization [ Reply ] By: Ott Toomet on 2009-08-30 15:36 | [forum:1866] |
|
I did a little work canning constrOpim to maxLik. Good side of the story is that it seems to be working. However, it may make the stuff more messy. * I added argument 'constraints' to max{BFGS,NM} and forward the value of it to constrOptim() (which uses optim() itself). Maybe use more meaningful names for the components, currently I use 'ui' and 'ci' as in constrOptim(). * It only supports inequality constraints. There is stuff available on cran using equality constraints. How to choose between them? * summary should tell something about constraints. Hence something should be included to 'maxim' structure itself. Ideas? |
|
