[SciPy-User] optimize.fmin_cg terminates when w - grad*1e-10 yields lower obj & grad

Jason Rennie jrennie@gmail....
Mon Oct 5 13:38:48 CDT 2009


On Mon, Oct 5, 2009 at 1:48 PM, Sebastian Walter <sebastian.walter@gmail.com
> wrote:

> In 90% of all cases when an optimization fails the user provided the
> wrong gradient.
>  |g| = 2.644732e+07 at the termination point clearly points in that
> direction. Maybe you provided -gradient instead of gradient?
>

That is certainly a mistake I made many times before I had a solid obj/grad
checker, but I can't remember a single case since.  OTOH, I've seen problems
like this resulting from subtle issues in line search implementation.  My
objective is convex; it's basically sum(exp(dot(X,w))^2).  So, obj/grad
wouldn't go down by subtracting 1e-10*grad from the param. vector if it were
a sign error, no?


> It is also possible that you simply cannot identify your 12 parameters
> with the 50 measurements you made.
> Then you'll have to improve your modeling resp. get more measurements.


I threw out those numbers to give a sense of the size of the problem, but I
don't understand how they're as relevant as you seem to be suggesting.  The
function I'm trying to minimize is a least squares loss between target and
predicted values, but the function doesn't change other than via the 'x'
that fmin_cg manipulates.  fmin_cg should still find a local min even if I
have too many parameters, no?

FYI, this is the sequence of iterations I'm seeing (first line is calculated
before fmin_cg call; thereafter from callback).  2-norm of the gradient is
the "grad" number.

obj=2.710e+06 grad=9.200e+04
obj=1.928345e+06 grad=7.290656e+05
obj=2.483969e+06 grad=2.642844e+07
obj=2.484517e+06 grad=2.644731e+07
obj=2.484517e+06 grad=2.644732e+07
obj=2.484517e+06 grad=2.644732e+07
obj=2.484517e+06 grad=2.644732e+07
obj=2.484517e+06 grad=2.644732e+07

Jason

-- 
Jason Rennie
Research Scientist, ITA Software
617-714-2645
http://www.itasoftware.com/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/scipy-user/attachments/20091005/ddf81d3e/attachment.html 


More information about the SciPy-User mailing list