Hi AlL, I ahve this problem that my objective function is discontinous in the paramaters and I need to use methods such as nelder-mead to get around this. My question is: How do i compute standard errors to a problem that does not have a gradient? Any literature on this is greatly appreciated. Jean,
Spencer Graves
2005-Mar-24 05:11 UTC
[R] non-derivative based optimization and standard errors.
Have you considered bootstrap or Monte Carlo?
spencer graves
Jean Eid wrote:
>Hi AlL,
>
>I ahve this problem that my objective function is discontinous in the
>paramaters and I need to use methods such as nelder-mead to get around
>this. My question is: How do i compute standard errors to a problem that
>does not have a gradient?
>
>
>Any literature on this is greatly appreciated.
>
>
>Jean,
>
>______________________________________________
>R-help at stat.math.ethz.ch mailing list
>https://stat.ethz.ch/mailman/listinfo/r-help
>PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
>
>
The problem is that it is a very complicated model and bootstrap will probably take months. The objective function itself is making use of Monte Carlo simulation because it is next to impossible to get at a closed form solution (of the objective function itself). So I simulate this function and get its expectation and match that to data. I thought of doing a bootstrap but it will take so much time. I guess if this is the only way, then it has to be done. Jean On Wed, 23 Mar 2005, Spencer Graves wrote:> Have you considered bootstrap or Monte Carlo? > > spencer graves > > Jean Eid wrote: > > >Hi AlL, > > > >I ahve this problem that my objective function is discontinous in the > >paramaters and I need to use methods such as nelder-mead to get around > >this. My question is: How do i compute standard errors to a problem that > >does not have a gradient? > > > > > >Any literature on this is greatly appreciated. > > > > > >Jean, > > > >______________________________________________ > >R-help at stat.math.ethz.ch mailing list > >https://stat.ethz.ch/mailman/listinfo/r-help > >PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html > > > > > >
Huntsinger, Reid
2005-Mar-24 15:23 UTC
[R] non-derivative based optimization and standard errors.
You'll really need to give some details if you want anything like a relevant answer. There aren't really general methods for dealing with discontinuous functions you can't compute. A few things come to mind. 1) You might have a look at the literature on segmented regression. Non-differentiable and even discontinuous objective functions arise there. 2) Monte Carlo: you may be able to adapt one of the Monte Carlo optimization approaches to your situation, avoiding having to do Monte Carlo within Monte Carlo. I'd be happy to be more specific if you'll supply details. Reid Huntsinger -----Original Message----- From: r-help-bounces at stat.math.ethz.ch [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Jean Eid Sent: Thursday, March 24, 2005 9:12 AM To: Spencer Graves Cc: r-help at stat.math.ethz.ch Subject: Re: [R] non-derivative based optimization and standard errors. The problem is that it is a very complicated model and bootstrap will probably take months. The objective function itself is making use of Monte Carlo simulation because it is next to impossible to get at a closed form solution (of the objective function itself). So I simulate this function and get its expectation and match that to data. I thought of doing a bootstrap but it will take so much time. I guess if this is the only way, then it has to be done. Jean On Wed, 23 Mar 2005, Spencer Graves wrote:> Have you considered bootstrap or Monte Carlo? > > spencer graves > > Jean Eid wrote: > > >Hi AlL, > > > >I ahve this problem that my objective function is discontinous in the > >paramaters and I need to use methods such as nelder-mead to get around > >this. My question is: How do i compute standard errors to a problem that > >does not have a gradient? > > > > > >Any literature on this is greatly appreciated. > > > > > >Jean, > > > >______________________________________________ > >R-help at stat.math.ethz.ch mailing list > >https://stat.ethz.ch/mailman/listinfo/r-help > >PLEASE do read the posting guide!http://www.R-project.org/posting-guide.html> > > > > >______________________________________________ R-help at stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Ingmar Visser
2005-Mar-24 15:23 UTC
[R] non-derivative based optimization and standard errors.
Hi Jean, Profiling may be another option and/or finite difference gradients. In any case, if your objective function is discontinuous at some point close to the optimal parameter values, standard errors may not make much sense. Best, Ingmar On 3/24/05 9:12 AM, "Jean Eid" <jeaneid at chass.utoronto.ca> wrote:> The problem is that it is a very complicated model and bootstrap will > probably take months. The objective function itself is making use of Monte > Carlo simulation because it is next to impossible to get at a closed form > solution (of the objective function itself). So I simulate this function > and get its expectation and match that to data. I thought of doing a > bootstrap but it will take so much time. I guess if this is the only way, > then it has to be done. > > > Jean > > On Wed, 23 Mar 2005, Spencer Graves wrote: > >> Have you considered bootstrap or Monte Carlo? >> >> spencer graves >> >> Jean Eid wrote: >> >>> Hi AlL, >>> >>> I ahve this problem that my objective function is discontinous in the >>> paramaters and I need to use methods such as nelder-mead to get around >>> this. My question is: How do i compute standard errors to a problem that >>> does not have a gradient? >>> >>> >>> Any literature on this is greatly appreciated. >>> >>> >>> Jean, >>> >>> ______________________________________________ >>> R-help at stat.math.ethz.ch mailing list >>> https://stat.ethz.ch/mailman/listinfo/r-help >>> PLEASE do read the posting guide! >>> http://www.R-project.org/posting-guide.html >>> >>> >> >> > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html-- Ingmar Visser Department of Psychology, University of Amsterdam Roetersstraat 15, 1018 WB Amsterdam The Netherlands http://users.fmg.uva.nl/ivisser/ tel: +31-20-5256735