[ADMB Users] my ADMB program is very sensitive to initial guess values in parameters, but different results look okay???

[James R. Bence]

Dear Saang-Yoon

As usual I replied all to your message so it would go immediately to you 
and eventually to the list after moderation.  But the list one bounced 
because you sent to google groups and apparently I am not subscribed to 
that version of users but I am to the list at admb-project.org.  I think 
the google list is archaic and forwards to the other.  Johnnoel is this 
something to be cleaned up?

You are asking a very generic question related to numerical searches for 
best fitting parameters that is not really specific to admb.  The result 
you are seeing reflects something about your specific model and data and 
specifics of how you did the search.  The basic results you seem to be 
reporting suggest that the admb program thinks you have found a local 
minimum in the sense that the gradient is nominally zero (absolute values 
all below a theshold) and estimated Hessian is positive definite.  If 
starting with different starting values leads to different numerical 
answers there seem to be two possibilities.  First that for your problem 
there really are multiple local minima or second you really are not at a 
local minima just something that numerically appears to be and each time 
you do a search you end up at a different false minima.   The simplest 
thing you could try is to make the criterion for gradients to be assumed to 
be zero more stringent and see if the problem goes away  (check manual on 
how to change default gradient values).  It could be the case the problem 
does not go away even if you make the criterion so small that it takes very 
long time (many iterations to meet).  Then you may still have either of the 
problems but there would seem to be a problem with your model given the 
data that are available.  It could be that the model is poorly 
parameterized (scale of parameters very different from one another or 
parameters highly correlated) or it could be still that there are really 
local minima.  It is hard to give any specific advice without actually 
seeing code and data.  I suspect that there are a number of people on the 
list who would jump in with suggestions if you provided that.

Another thing you could do to try to diagnose what is going on is to take a 
look at the different solutions you are getting and compare your objective 
functions at the solutions.  If you are getting almost identical values for 
the objective function but the parameters are varying enough that you are 
concerned this suggests there really is a model problem in the sense that 
there really does not appear to be a unique minimum to the objective function.

Hope this helps

At 07:23 PM 6/15/2012, Saang-Yoon wrote:
>Hi.
>With small changes in initial guess values of parameters, different 
>results are generated. Interestingly different results are satisfactory in 
>terms of Hessian matrix calculation, and stability (i.e., maximum gradient 
>= almost zero).  Although a difference between two results are small, the 
>episode indicates a failure to regenerate the same results.  Given the 
>same data, the same results must be generated.   Also it is not good that 
>the results are *very* sensitive to initial guess values.  Appreciated is 
>any suggestion/advice about how to avoid the problem?
>Saang-Yoon
>_______________________________________________
>Users mailing list
>Users at admb-project.org
>http://lists.admb-project.org/mailman/listinfo/users

Jim Bence
Dept. of Fisheries and Wildlife
Michigan State University
http://www.msu.edu/user/bence/  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20120616/fcbc4f91/attachment.html>