[Scipy-tickets] [SciPy] #1588: scipy.optimize.cobyla not consistant in Windows

SciPy Trac scipy-tickets@scipy....
Tue Jan 24 16:12:56 CST 2012


#1588: scipy.optimize.cobyla not consistant in Windows
-----------------------------+----------------------------------------------
  Reporter:  casperskovby    |       Owner:  somebody   
      Type:  defect          |      Status:  closed     
  Priority:  normal          |   Milestone:  Unscheduled
 Component:  scipy.optimize  |     Version:  0.9.0      
Resolution:  invalid         |    Keywords:             
-----------------------------+----------------------------------------------
Changes (by pv):

  * status:  needs_review => closed
  * resolution:  => invalid


Comment:

 Ok, EPD is compiled with Intel's Fortran compiler AFAIK, which can do
 optimizations producing issues like this. What you show is consistent with
 the issue being different code paths being chosen for different data
 alignments, so I guess that is the issue.

 Note that the requested tolerance of `fmin_cobyla` can be chosen by giving
 it the `rhoend` argument, see the documentation. You don't specify it
 above, so you get the default accuracy 1e-4.

 That's the reason why the differences are so big --- because of the
 alignment-dependent rounding error, the minimization algorithm probably
 does one additional iteration in one of the two cases, before reaching the
 requested tolerance. If you specify a smaller tolerance, the difference
 will not matter as much. If you want to double check, you can count the
 number of times the objective function is called.

 I'm closing this ticket as "not a bug", as the conclusion seems clear.
 Please reopen if specifying a smaller `rhoend` does not help.

-- 
Ticket URL: <http://projects.scipy.org/scipy/ticket/1588#comment:5>
SciPy <http://www.scipy.org>
SciPy is open-source software for mathematics, science, and engineering.


More information about the Scipy-tickets mailing list