This is the mail archive of the gsl-discuss@sources.redhat.com mailing list for the GSL project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: multimin convergence problem


Roland Roberts writes:
 > What I'm finding is that the ability to get a solution at all is very
 > sensitive to how I initialize the line search parameters.  The setting
 > of TOL can vary a lot, but the value of STEP has to fall in a fairly
 > narrow range or GSL starts spitting out NaNs.  What is frustrating is
 > that neither the function nor the derivative is returning NaN at any
 > point; I've verified this by printout out all the function
 > evaluations.

Hi,

Are you doing least-squares fitting of the residuals between the two
curves?  If so the nonlinear 'multifit' functions should handle that
better.


 > Is there some magic to picking a good value for STEP?  Shouldn't GSL
 > return an error rather than setting all my parameters to NaN?

Yes, it should return an error not NaN.  There must be an unchecked
division by zero that we need to catch in there somewhere.   Can
send your program to the list.

Thanks

-- 
Brian Gough

----------------------------------------------------------------------
Network Theory Ltd            Phone: 0117 3179309 (+44 117 3179309)
15 Royal Park                 WWW: http://www.network-theory.co.uk/
Bristol BS8 3AL               Email: bjg@network-theory.co.uk     
United Kingdom                
----------------------------------------------------------------------


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]