I think I've seen this somewhere out there on the webs... http://dirk.eddelbuettel.com/blog/2011/07/14/
------------------------------------------------------------------------- J.P. Olmsted 029 Corwin (Office) 130 Corwin Hall (Mail) Politics Department Princeton University Princeton, NJ 08544 t: 609.258.6202 f: 609.258.1110 jolms...@princeton.edu http://about.me/olmjo ------------------------------------------------------------------------- On Thu, Aug 1, 2013 at 11:31 AM, Jonathan Olmsted <jolms...@princeton.edu>wrote: > Just my quick thoughts: > > Bayesian MCMC is what brought me to Rcpp. So, I have always found those > examples the most compelling. This is precisely because good R coding can't > improve performance on these problems and the gain in computational > performance really justified the development cost as opposed to cases where > you come out with a time profit only in the *very* long-run. > > In particular, I usually work with models that are conceptually similar to > factor analysis (i.e., nothing on the "right-hand side is observed"). This > is relevant because the next place one might go if R is slow for these > problems is JAGS. Or maybe it's the first place you go. I'm not sure. > Anyway, JAGS (as of when I last looked into it) doesn't handle the things I > was doing efficiently. It used a less efficient sampler because it didn't > detect that conditional on every other parameter, I was updating something > that came from a Normal dist. So, with just a vanilla Gibbs step, I could > out perform JAGS' slice sampler (where efficiency = increase in Effective > Sample Size / time). > > I settled on Rcpp for these problems before Stan was really out there, so > I haven't doubled-back to see how Stan would factor in to someone's > decision-making process. Taken together: > > 1. If you are doing MCMC in R for a model that isn't packaged, you > have to code it by hand > 2. Even if your R code is efficient, your runtime will be slow > 3. Outside options don't get you too much improvement (e.g. JAGS), but > they get you some > 4. For some problems, that just isn't fast enough and Rcpp-based C++ > code can help you get you a lot more speedup with a reduced learning curve > > Particularly nice about this class of problems is that it > "straightforward" on the implementation side. It focuses on computation > (not advanced features of C++ or the nuances of R objects), the algorithm > is conceptually simple, and you don't have to rely on boost or armadillo or > anything beyond basic C++. Lastly, with all the recent sugar, MCMC Rcpp > code is veryR-like. > > Also, while it's never a complaint of mine, there is often some conceptual > hurdle for folks looking at an example like simulating pi and seeing how it > applies to their work. At least with MCMC, it is a bit closer to a real > application. I'm not saying it's anyone else's job to jump that hurdle for > them, but if it can be removed for free, why not? > > Apologies for long-windedness. > > -Jonathan > > > > > ------------------------------------------------------------------------- > J.P. Olmsted > > 029 Corwin (Office) > 130 Corwin Hall (Mail) > > Politics Department > Princeton University > Princeton, NJ 08544 > > > t: 609.258.6202 > f: 609.258.1110 > jolms...@princeton.edu > http://about.me/olmjo > ------------------------------------------------------------------------- > > > On Thu, Aug 1, 2013 at 10:19 AM, Dirk Eddelbuettel <e...@debian.org> wrote: > >> >> I'm giving a talk to the Chicago RUG that is limited to 30 mins, and I >> would >> like to include some nice examples (besides standards like Fibonacci and >> SimulatingPi). The other talk is on ggplot(2), so the house may be full >> with >> new users / non-C++ hackers. >> >> What examples should I talk about? Bonus points for links for list >> questions, StackOverflow questions, or Rcpp Gallery posts. >> >> Dirk >> >> -- >> Dirk Eddelbuettel | e...@debian.org | http://dirk.eddelbuettel.com >> _______________________________________________ >> Rcpp-devel mailing list >> Rcpp-devel@lists.r-forge.r-project.org >> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel >> > >
_______________________________________________ Rcpp-devel mailing list Rcpp-devel@lists.r-forge.r-project.org https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
