Hi Federico,

Have you checked your gradients using finite differences in both cases?

Regards,
James


On 3 Oct 2013, at 20:27, Jaap Kroes <[email protected]> wrote:

> Hi,
> 
> Most likely there is a problem in the calculation of your gradients.
> Did you perhaps forget to initialize them to zero before every update?
> From the numbers the python gradient is very small on the second step
> while the nlopt gradients look like the old ones + something small.
> 
> Best,
> Jaap
> 
> 2013/10/3 federico vaggi <[email protected]>:
>> Hi everyone,
>> 
>> I've been using NLopt to estimate the parameters of a system of differential
>> equations by minimizing the least squares between my simulations and a set
>> of experimental data.
>> 
>> The non-gradient based optimizations have worked very well.  I started
>> experimenting with using gradient based methods and estimating the jacobian
>> from the sensitivity equations calculated by sundials, and I found that
>> using the SciPy fmin_l_bfgs_b I'm able to obtain a local minima very
>> quickly.  All NLopt algorithms, however, quickly throw an NLopt error.
>> 
>> The system I am working with is quite gnarly - some parameters are very,
>> very sensitive, while others are completely robust.
>> 
>> SciPy bfgs:
>> 
>> http://pastie.org/8374252
>> 
>> NLopt bfgs:
>> 
>> http://pastie.org/8374254
>> 
>> I've tried changing NLopt algorithms, but every single gradient based method
>> I tried eventually throws an error.  Any clue what might be causing it?  I
>> am happy to share the code, but this is a fairly large wrapper around SciPy
>> and Assimulo ODEINT solvers, so posting a minimal example is not very easy.
>> 
>> Federico
>> 
>> _______________________________________________
>> NLopt-discuss mailing list
>> [email protected]
>> http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
>> 
> 
> _______________________________________________
> NLopt-discuss mailing list
> [email protected]
> http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
> 


_______________________________________________
NLopt-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss

Reply via email to