Hi Arvind,

I'm inclined to agree with Grey Gordon in that the problem is likely
on your end. The documentation for NLOPT_LN_NEWUOA[1] makes no
reference to the usage of random numbers.

My recommendation would be twofold:

1) First, test that you have NLopt along set up correctly along with
your development environment, by building a much simpler program to
test the optimisation of some dummy function[2], using the same
driver. If this also fails, you're going to have to work out what
isn't configured right. Which language/platform are you using NLopt
with?

2) Second, write some to evaluate your function at a repeating
sequence of known inputs, and check that the output repeats, and
agrees between separate fresh executions. You're probably going to
have to either use a debugger or lots of debug-I/O to work out where
the issue is.

Good luck! If you discover anything interesting, or can't make any
progress, I'm sure someone here can give you some more pokes in the
right direction!

Best regards,
Adam

[1]
http://ab-initio.mit.edu/wiki/index.php/NLopt_Algorithms#NEWUOA_.2B_bound_constraints

[2] https://en.wikipedia.org/wiki/Test_functions_for_optimization

----- Original Message -----
From:
 "Iyer Arvind S" <[email protected]>

To:
"[email protected]" <[email protected]>
Cc:

Sent:
Thu, 15 Dec 2016 13:05:15 +0000
Subject:
Re: [NLopt-discuss] Welcome to the "NLopt-discuss" mailing list

        Hi, 

        I am a new user of the nl-opt library. Thanks to the developers for
such a wonderful collection. 

        I recently observed that NLOPT_LN_NEWUOA based optimisation 

        yields different results (sometimes drastically different) on
multiple invocation over
 the same problem starting from the same guess.

        The problem itself can have multiple local minima and its
understandable that 
 this algorithm clings to the 
nearest one. However, I did not expect different 

        results for the same problem with same starting point.

This seems like the presence of a random number initialization of some
parameter.

        I am myself not aware of the details of the algorithm and its
implementation.

        Can someone 
confirm that this is indeed so? And perhaps a way to seed the 

        random number if this is the case.

        Thanks again 

        Arvind


_______________________________________________
NLopt-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss

Reply via email to