[SciPy-user] Typo in SciPy code

Travis Oliphant oliphant at ee.byu.edu
Mon Mar 27 15:48:13 CST 2006


Armando Serrano Lombillo wrote:

>In http://projects.scipy.org/scipy/scipy/browser/trunk/Lib/optimize/info.py
>where it says:
>
>min_cobyla   -- Contrained Optimization BY Linear Approximation
>
>it should say [...] -- Constrained [...]
>
>This also appears when you use help(scipy.optimize.fmin_cobyla).
>
>
>BTW I've been testing some of the optimization routines in scipy
>(Nelder-Mead, Powell and COBYLA) and COBYLA was 10 times faster than
>the others and it showed much better precision and stability. Is this
>due to the implementation (fortran instead of python) or is it just
>that the algorithm works better for my problem?
>
>  
>
Try the BFGS method (fmin_bfgs) instead of Powell.  The Powell 
implementation in SciPy needs some work to be truly representative.  The 
BFGS methods should be fairly decent (it uses a good line search).  The 
Nelder-Mead algorithm is usually always going to use more function 
evaluations.

It's always hard to predict which optimization algorithm is going to 
work "best" for a particular problem (that's why there are so many of 
them).  I suspect it's the number of function evaluations that's most 
important rather than whether or not the iteration is written in FORTRAN 
or Python.

-Travis



More information about the SciPy-user mailing list