[SciPy-user] fmin_tnc changes since scipy 0.6.0

dmitrey dmitrey.kroshko@scipy....
Fri Jun 20 16:42:23 CDT 2008


Nils Wagner wrote:
> On Fri, 20 Jun 2008 19:26:48 +0000 (UTC)
>   Yannick Copin <yannick.copin@laposte.net> wrote:
>   
>> Yannick Copin <y.copin <at> ipnl.in2p3.fr> writes:
>>     
>>> I'm a bit puzzled: I had a code which worked fine using 
>>> scipy.optimize.fmin_tnc with scipy 0.5.2. Unfortunately, 
>>> since I upgraded 
>>> scipy to 0.6.0, I hit some convergence issues...
>>>
>>> Does it ring a bell to somebody? I understand I'm a bit 
>>> vague for the moment 
>>> (I did not look for a simple test case yet), but maybe 
>>> somebody had this 
>>> experience too. Besides the reverted output order, was 
>>> there some significant 
>>> algorithmic changes in fmin_tnc between 0.5.2 and 0.6.0?
>>>       
>> To complete my previous email, it seems my convergence 
>> problems is related to
>> the new 'offsets' in fmin_tnc, which allows for 
>> fine-tuning of parameter
>> scaling. If I force these offsets to be all null 
>> (instead of their actual
>> defaults), I find results very similar to scipy 0.5.2 
>> ones. 
>>
>> PS: I suspect the TNC algo within OpenOpt is actually 
>> the one from Scipy, right?
>>     
>
> Dmitrey should reply here since he is the author of 
> openopt.
>   
yes, indeed, versions are same since tnc is called from scipy.
>   
>> So this would not solve my problem. Furthermore, I try 
>> to limit the number of
>> external packages to be needed by my code, so I restrict 
>> myself to routines from
>> numpy/scipy.
>>
>>     
>
> IMHO, one should compare the results obtained by different 
> optimizations tools, wrt. efficiency, convergence 
> behaviour, etc.
>   
Maybe Yannick's demands are completely covered by scipy solvers, so 
making additional dependencies really are not appreciated here. First of 
I write  openopt for some classes of problems beyond scipy could solve.
D.
> Nils
>   



More information about the SciPy-user mailing list