Re: non-linear fitting with bounds

[Olaf Till]

On Sat, Mar 13, 2010 at 08:37:40PM +0100, Stefan Neumann wrote:
> Hi all,
> 
> with leasqr() it is possible to fit data to an arbitrary non-linear
> function.
> 
> Is there a way to introduce bounds, for example limit a parameter to be less
> that zero?
> 
> Is there a way to introduce non-linear bounds?

In the last version (in package optim-1.0.12) bounds and linear
inequality constraints are implemented. Do not use an older version,
even if it should have a bounds implementation --- that implementation
was not good. But bounds are not implemented in such a way that you
can keep a parameter less than zero, you only can keep it less than or
equal to zero. You could keep it less than or equal to a small
negative number though. Bounds (but not general nonlinear constraints)
are also respected by the default function for the numerical
gradient. Nonlinear inequality constraints are probably a bit more
complicated. I intend to look into the latter in the future.

Olaf