COMPGROUPS.NET | Search | Post Question | Groups | Stream | About | Register

• Follow

```Hi all,

I just came across a problem regarding to the nonlinear constrained regression. There is a proposed equation like:

A>Ap; A<Ad and L>Lg are constrains. There are 9 unknown variables in this equation. i.e. a1;a2;a3;k1;k2;k3;Ap;Ad;Lg. Tor, ma and L are given. I want to run a regression problem to find out the optimal value of all the unknowns which could best fit the equation. Anyone has some idea about this??? Many thanx....
```
 0
Reply maoyi.tian (1) 11/22/2007 7:40:03 AM

```MAOYI TIAN <maoyi.tian@student.unsw.edu.au> wrote in message
<10372620.1195706433974.JavaMail.jakarta@nitrogen.mathforum.org>...
> Hi all,
>
> I just came across a problem regarding to the nonlinear constrained
regression. There is a proposed equation like:
>
>
> A>Ap; A<Ad and L>Lg are constrains. There are 9 unknown variables in
this equation. i.e. a1;a2;a3;k1;k2;k3;Ap;Ad;Lg. Tor, ma and L are given. I
want to run a regression problem to find out the optimal value of all the
unknowns which could best fit the equation. Anyone has some idea about
this??? Many thanx....

You want our opinions?

Its a very difficult problem to optimize.

The presence of inequalities inside the function
will introduce discontinuitues in the derivatives.
In turn, expect to find that most optimizers that
rely on differentiability to sometimes fail. At the
very least, expect to find many local solutions, so
you will need very good starting values.

Optimizers that are slightly more robust on
non-differentiable objectives, like fminsearch,
also tend to be poor on problems with as many
as 9 parameters. You might consider looking
into genetic algorithms, simulated annealing,
or particle swarm optimizers. They can be
far less sensitive to the above problems. Look
on the FEX - there is a good implementation
of particle swarms to be found there.

objectId=7506&objectType=file

Or look into the GADS toolbox:

HTH,
John

```
 0
Reply woodchips (7921) 11/22/2007 10:25:12 AM

1 Replies
59 Views

6/19/2013 6:29:35 PM