Prof J C Nash (U30A)
2014-Dec-18 14:10 UTC
[R] Maximum likelihood with analytical Hessian and
Of the tools I know (and things change every day!), only package trust uses the Hessian explicitly. It would not be too difficult to include explicit Hessian by modifying Rvmmin which is all in R -- I'm currently doing some cleanup on that, so ask offline if you choose that route. Given that some parameters are between 0 and 1, you could use the hyperbolic transformation (section 11.2 of my book Nonlinear parameter optimization using R tools) with trust, and I think I'd try that as a first attempt. You probably need to adjust the Hessian for the transformation carefully. Generally the work in computing the Hessian ( # obs * (# parameters)^2 in size) is not worth the effort, but there are problems for which it does make a lot of sense. JN On 14-12-18 06:00 AM, r-help-request at r-project.org wrote:> Message: 12 > Date: Wed, 17 Dec 2014 21:46:16 +0100 > From: Xavier Robin <robin at lindinglab.org> > To: r-help at r-project.org > Subject: [R] Maximum likelihood with analytical Hessian and > Message-ID: <5491EB98.6090606 at lindinglab.org> > Content-Type: text/plain; charset=utf-8 > > Dear list, > > I have an optimization problem that I would like to solve by Maximum > Likelihood. > I have analytical functions for the first and second derivatives of my > parameters. > In addition, some parameters are constrained between 0 and 1, while some > others can vary freely between -Inf and +Inf. > > I am looking for an optimization function to solve this problem. > > I understand that the base optim function doesn't take a Hessian > function, it only computes it numerically. > I found the maxLik package that takes the function as a "hess" parameter > but the maxNR method (the only one that uses the Hessian function) can't > be bounded. > Surprisingly I couldn't find a function doing both. > > Any suggestions for a function doing bounded optimization with an > analytical Hessian function? > > Thanks, > Xavier > >
Dear John, Thank you for your suggestions. I'll have a look at the trust package - the trust zone may be doing what I need. The tanh transformation could be a good alternative too. Best wishes Xavier On 18. 12. 14 15:10, Prof J C Nash (U30A) wrote:> Of the tools I know (and things change every day!), only package trust > uses the Hessian explicitly. > > It would not be too difficult to include explicit Hessian by modifying > Rvmmin which is all in R -- I'm currently doing some cleanup on that, so > ask offline if you choose that route. > > Given that some parameters are between 0 and 1, you could use the > hyperbolic transformation (section 11.2 of my book Nonlinear parameter > optimization using R tools) with trust, and I think I'd try that as a > first attempt. You probably need to adjust the Hessian for the > transformation carefully. > > Generally the work in computing the Hessian ( # obs * (# parameters)^2 > in size) is not worth the effort, but there are problems for which it > does make a lot of sense. > > JN > > On 14-12-18 06:00 AM, r-help-request at r-project.org wrote: >> Message: 12 >> Date: Wed, 17 Dec 2014 21:46:16 +0100 >> From: Xavier Robin <robin at lindinglab.org> >> To: r-help at r-project.org >> Subject: [R] Maximum likelihood with analytical Hessian and >> Message-ID: <5491EB98.6090606 at lindinglab.org> >> Content-Type: text/plain; charset=utf-8 >> >> Dear list, >> >> I have an optimization problem that I would like to solve by Maximum >> Likelihood. >> I have analytical functions for the first and second derivatives of my >> parameters. >> In addition, some parameters are constrained between 0 and 1, while some >> others can vary freely between -Inf and +Inf. >> >> I am looking for an optimization function to solve this problem. >> >> I understand that the base optim function doesn't take a Hessian >> function, it only computes it numerically. >> I found the maxLik package that takes the function as a "hess" parameter >> but the maxNR method (the only one that uses the Hessian function) can't >> be bounded. >> Surprisingly I couldn't find a function doing both. >> >> Any suggestions for a function doing bounded optimization with an >> analytical Hessian function? >> >> Thanks, >> Xavier >> >>-- Xavier Robin, PhD Cellular Signal Integration Group (C-SIG) - Linding Lab Biotech Research and Innovation Center (BRIC) - University of Copenhagen Anker Engelundsvej, DTU Campus, Building 301, DK-2800 Lyngby, DENMARK Mobile: +45 42 799 833 www.lindinglab.org - www.bric.ku.dk