Hi, Sarah Goslee (jn reply to? Basic optimization question (I'm a rookie)):? "R is quite good at optimization." I wonder what is the experience of the R user community with high dimensional problems, various objective functions and various numerical methods in R. In my experience with my package CatDyn (which depends on optimx), I have fitted nonlinear models with nearly 50 free parameters using normal, lognormal, gamma, Poisson and negative binomial exact loglikelihoods, and adjusted profile normal and adjusted profile lognormal approximate loglikelihoods. Most numerical methods crash, but CG and spg often, and BFGS, bobyqa, newuoa and Nelder-Mead sometimes, do yield good results (all numerical gradients less than 1)? after 1 day or more running in a normal 64 bit PC with Ubuntu 16.04 or Windows 7. Ruben -- Ruben H. Roa-Ureta, Ph. D. Consultant, ORCID ID 0000-0002-9620-5224 Marine Studies Section, Center for Environment and Water, Research Institute, King Fahd University of Petroleum and Minerals, KFUPM Box 1927, Dhahran 31261, Saudi Arabia Office Phone : 966-3-860-7850 Cellular Phone : 966-540026401
I fit also model with many variables (>100) and I get good result when I mix several method iteratively, for example: 500 iterations of Nelder-Mead followed by 500 iterations of BFGS followed by 500 iterations of Nelder-Mead followed by 500 iterations of BFGS etc. until it stabilized. It can take several days. I use or several rounds of optimx or simply succession of optim. Marc Le 28/11/2018 ? 09:29, Ruben a ?crit?:> Hi, > > Sarah Goslee (jn reply to? Basic optimization question (I'm a > rookie)):? "R is quite good at optimization." > > I wonder what is the experience of the R user community with high > dimensional problems, various objective functions and various > numerical methods in R. > > In my experience with my package CatDyn (which depends on optimx), I > have fitted nonlinear models with nearly 50 free parameters using > normal, lognormal, gamma, Poisson and negative binomial exact > loglikelihoods, and adjusted profile normal and adjusted profile > lognormal approximate loglikelihoods. > > Most numerical methods crash, but CG and spg often, and BFGS, bobyqa, > newuoa and Nelder-Mead sometimes, do yield good results (all numerical > gradients less than 1)? after 1 day or more running in a normal 64 bit > PC with Ubuntu 16.04 or Windows 7. > > Ruben >
Hello, Genetic algorithm can prove handy as well here. see for instance https://cran.r-project.org/web/packages/GA/vignettes/GA.html with non-convex objective functions I usually try a genetic algorithm for a few rounds then finish using nlminb Best regards, Jeremie Marc Girondot via R-help <r-help at r-project.org> writes:> I fit also model with many variables (>100) and I get good result when > I mix several method iteratively, for example: 500 iterations of > Nelder-Mead followed by 500 iterations of BFGS followed by 500 > iterations of Nelder-Mead followed by 500 iterations of BFGS > etc. until it stabilized. It can take several days. > I use or several rounds of optimx or simply succession of optim. > > Marc > > Le 28/11/2018 ? 09:29, Ruben a ?crit?: >> Hi, >> >> Sarah Goslee (jn reply to? Basic optimization question (I'm a >> rookie)):? "R is quite good at optimization." >> >> I wonder what is the experience of the R user community with high >> dimensional problems, various objective functions and various >> numerical methods in R. >> >> In my experience with my package CatDyn (which depends on optimx), I >> have fitted nonlinear models with nearly 50 free parameters using >> normal, lognormal, gamma, Poisson and negative binomial exact >> loglikelihoods, and adjusted profile normal and adjusted profile >> lognormal approximate loglikelihoods. >> >> Most numerical methods crash, but CG and spg often, and BFGS, >> bobyqa, newuoa and Nelder-Mead sometimes, do yield good results (all >> numerical gradients less than 1)? after 1 day or more running in a >> normal 64 bit PC with Ubuntu 16.04 or Windows 7. >> >> Ruben >> > > ______________________________________________ > R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code.