[prev in list] [next in list] [prev in thread] [next in thread] 

List:       scilab-users
Subject:    Re: [Scilab-users] Some questions about Scilab Optim function
From:       Serge Steer <serge.steer () inria ! fr>
Date:       2013-01-21 17:33:04
Message-ID: 438507643.28117966.1358789584409.JavaMail.root () inria ! fr
[Download RAW message or body]



----- Mail original -----
> De: "Rafael Guerra" <jrafaelbguerra@hotmail.com>
> À: users@lists.scilab.org
> Envoyé: Lundi 21 Janvier 2013 14:33:03
> Objet: [Scilab-users] Some questions about Scilab Optim function
> 
> Hello,
> 
> I am using Scilab 5.4 Windows7 optim function to find parameters in a
> non-linear model to fit experimental data and I have some questions
> about
> the method used in Scilab.
> 
> I am running optim with very good results despite a bit slowly, with
> the
> following function call (x is a vector of length 3 in my case):
> 
> [f,xopt]= optim(list(NDcost,costfun,param1,..,paramN),x0)
> ...
> function f = costfun(x,param1,...,paramN)
> ...
> 
> The questions I have regarding optim are:
> 
> 1.  When using the NDcost option to compute numerically the
> derivatives is
> optim using the quasi-Newton method without constraints?

The NDcost functio is just a way to give optim a function able to compute the cost \
and/or the gradient. So all optim options are available in this context. The quality \
and efficiency will depend on the quality of the estimated derivative

> 2. Is this NDcost option equivalent in terms of run time and accuracy
> to
> estimating numerically the derivatives with the dedicate Scilab
> functions
> (derivative or numdiff)?
Yes
> 
> 3. Can the NDcost option be run with constraints?
> 
Yes, under the assumption the numericall derivative is precise enough

> 4. How can we have an idea of the accuracy of the parameters
> estimated by
> optim, or at least of the stability of  the solution found?

You can look at the gradient or projected gradient norm


Serge Steer
_______________________________________________
users mailing list
users@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/users


[prev in list] [next in list] [prev in thread] [next in thread]