Here’s Bob’s talk from the NYC machine learning meetup. And here’s Stan himself:
Four pages dense with problems that had to be solved in order to get this to work — wow, I hope you guys are being paid by the hour! (smile) Seriously, it looks great; how long until it will be usable by us mere mortals? Say g’day to Bob from me.
Typo bottom of page 19, I think.
I need this Software because I am a Professor of Animal Science in Camagüey University, in Animal
[...] Stan: A (Bayesian) Directed Graphical Model Compiler [...]
Your source code is currently on code.google. After you release it, will it be moved to github/bitbucket? In case others want to submit pull requests to add to it, or fork it for their own use?
In my experience, when you transform parameters, like variance parameters, by taking the log, that gives you some very steep gradient functions in some parts of the parameter space. Because HMC is a discrete approximation, this can lead to nasty behavior of the sampler. Does Stan/NUTS handle that well? I typically just integrate the constrained parameters out analytically, which solves the problem – but with a general-purpose solution like Stan, that’s not really an option.
(By the way, I can’t wait to try this software.)