R help - Jun 2005 - Bounding or constraining parameters in non-linear regressions

Dear R-Users,

Being an engineer and not a statistician, my desired course of action may 
either be impossible or very simple.

I am attempting to fit a non-linear model to some measured data.  One term in 
the model contains a square-root, but in the course of regression, this term 
turns negative and an error occurs.  I started using Micrsoft's Excel
Solver,
and then I turned to NIST's Datplot statistical package.  I can constrain in
Solver, but it violates those constraints. :)  Dataplot does not have the 
capability to constrain parameters.

Does R have the capability to constrain or bound parameters in non-linear 
regressions?


Thanks,
Mark Davis

On 6/7/05, mwdavis at nist.gov <mwdavis at nist.gov>
wrote:
> Dear R-Users,
> 
> Being an engineer and not a statistician, my desired course of action may
> either be impossible or very simple.
> 
> I am attempting to fit a non-linear model to some measured data.  One term
in
> the model contains a square-root, but in the course of regression, this
term
> turns negative and an error occurs.  I started using Micrsoft's Excel
Solver,
> and then I turned to NIST's Datplot statistical package.  I can
constrain in
> Solver, but it violates those constraints. :)  Dataplot does not have the
> capability to constrain parameters.
> 
> Does R have the capability to constrain or bound parameters in non-linear
> regressions?
Sort of.  If you look at the stats package in r-devel you will see
that a function called nlminb has been added.  This function calls
optimization software from the Port package
(http://www.netlib.com/port/).  The Fortran code for constrained
nonlinear least squares problems is included in the package but the
interface code for R has not yet been written.  The energetic could
create such interface code by emulating that for nlminb - it's not
that long.

Alternatively you could use either optim or nlminb on the function
which is the residual sum of squares from your model.