[scilab-Users] Academic question

[Nicolas FEZANS]

Hi Paul!

Before trying such approaches, if you cannot (or don't want to)
compute gradients and hessian, you may let Scilab approximate them
numerically thanks to the "NDcost" function. It generally works very
well so I strongly recommend it before even thinking using some
stochastic methods (which are not bad but should not be used when not
necessary and I guess you don't need them in your case).

Note that they are plenty of fully deterministic heuristics to
optimize without gradients (not even numerically approximated), which
are too often forgotten and often even not included in optimization
packages. As for all methods they are always some problems, but it
solved quite efficiently many optimization problem in the past and not
all optimization problems are tough or are worth writing 30 pages of
calculus to get an analytical expression for the gradient of the cost
function...

Anyway... have fun with NDcost!

Regards
Nicolas


On Tue, Jan 25, 2011 at 4:49 PM, Charles Basenga Kiyanda <cbk at lanl.gov> wrote:
> You could try a genetic-type algorithm, such as differential evolution. I
> don't think there's a scilab implementation (at the very least, I don't
> think there's one distributed with scilab).
>
> See information (and code for several ports) here:
> http://www.icsi.berkeley.edu/~storn/code.html
>
> Implementing the basic algorithm in scilab (in any language for that matter)
> is easy and quick.
>
> There are several references on that page. I recommend the first few
> chapters of "Differential Evolution - A Practical Approach to Global
> Optimization" by Ken Price, Rainer Storn, and Jouni Lampinen.
>
> Otherwise, search for algorithms such as 'simulated annealing'. I've heard
> good things although I've lately been helping colleagues, who relied on that
> algorithm for some minimization problems, migrate to DE as a potentially
> more robust and efficient algorithm.
>
> Good luck,
>
> Charles
>
>
> On 01/25/2011 04:38 AM, Carrico, Paul wrote:
>>
>> Dear All,
>>
>>
>>
>> For some times I'm testing Scilab macros for optimization purposes ;
>> "optimization" has to be intended here as an inverse method for fitting
>> parameters (i.e. so that finite element analysis fit test measurements).
>>
>>
>>
>> I tested basic macros such as fminsearch for unbounded fitting as well
>> as Nelder-Mead one for bracketed parameters fitting. Basically the "cost
>> function" is the normalized Sum of the Square errors SSE from the FEA to
>> the measurements.
>>
>>
>>
>> In my FEA's the variables can be material parameters, spring
>> stiffness's, damping ratio and so on : the simplex method is well
>> adapted since the former variables are not analytically described in the
>> cost function and since it's not necessary to calculate the gradient
>> vector nor the Hessian matrix ...
>>
>>
>>
>>
>>
>> If this method is robust nevertheless it remains rather slow !
>>
>>
>>
>>
>>
>> Is there a way or another macro I can use to reduce the number of loops
>> and the time consuming consequently ?
>>
>>
>>
>> Nota : I was thinking in OPTIM macro but from my understanding the cost
>> function need to be twice differentiable at least  (for the Gradient and
>> the Hessian ) i.e. analytically link to the parameters ... isn't it ?
>>
>>
>>
>> Thanks in advance for any advice (for a better understanding)
>>
>>
>>
>> Paul
>>
>>
>>
>>
>> --------------------------------------------------------------------------------
>>
>>
>> Le présent mail et ses pièces jointes sont confidentiels et destinés à la
>> personne ou aux personnes visée(s) ci-dessus. Si vous avez reçu cet e-mail
>> par erreur, veuillez contacter immédiatement l'expéditeur et effacer le
>> message de votre système. Toute divulgation, copie ou distribution de cet
>> e-mail est strictement interdite.
>>
>> This email and any files transmitted with it are confidential and intended
>> solely for the use of the individual or entity to whom they are addressed.
>> If you have received this email in error, please contact the sender and
>> delete the email from your system. If you are not the named addressee you
>> should not disseminate, distribute or copy this email.
>>
>
>
>