[SciPy-User] optimize.fmin_cobyla giving nan to objective function

Ralf Gommers ralf.gommers@googlemail....
Wed Aug 3 11:05:33 CDT 2011


On Tue, Aug 2, 2011 at 10:55 PM, Gustavo Goretkin <
gustavo.goretkin@gmail.com> wrote:

>
>>> Your function always returns inf, so it's not very surprising that you
>> get a nan after a few iterations. Could happen for example if the code
>> determines a derivative numerically, resulting in inf / inf = nan.
>>
>> It would be helpful if you had a realistic, self-contained example.
>>
>> Raln
>>
>
> In scikit-learn, fmin_cobyla is used to optimize some parameters of a
> Gaussian Process. The objective function returns inf when the parameters are
> such that the matrix calculations are unstable and NumPy throws a LinAlg
> exception. What would be a better way to handle this?
>

Let the objective function do something sensible? Like figure out what the
unstable region is and returning values that steer the optimizer away from
it.

With a slight modification to your last test script I see that fmin_cobyla
doesn't choke on receiving a first inf from the objective function (see
below). If it receives infs not for a single x, but several or a range, then
I'd expect it to fail.

from scipy.optimize import fmin_cobyla
import numpy as np

def objective(x):
    print 'Input: ', x, '  return value: ', x + 1./x
    return x + 1./x

def constraint1(x):
    return 0

xstar = fmin_cobyla(func=objective, x0=0, cons=[constraint1])

Cheers,
Ralf



> My gut feeling is that an optimizer should not pass nan to the objective
> function, since it cannot possibly be informative. Maybe checking for nan
> would be inefficient.
>
> _______________________________________________
> SciPy-User mailing list
> SciPy-User@scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/scipy-user/attachments/20110803/31274407/attachment.html 


More information about the SciPy-User mailing list