We’re happy to announce that Stan 2.9.0 is fully available(1) for CmdStan, RStan, and PyStan — it should also work for Stan.jl (Julia), MatlabStan, and StataStan. As usual, you can find everything you need on the
The main new features are:
- R/MATLAB-like slicing of matrices. There’s a new chapter in the user’s guide part of manual up front explaining how it all works (and more in the language reference on the nitty-gritty details). This means you can write foo[xs] where xs is an array of integers and use explicit slicing, as with bar[1:3, 2] and baz[:3, , xs] and so on.
- Variational inference is available on an experimental basis in RStan and PyStan, and the adaptation has been improved; we still don’t have a good handle on when variational inference will work and when it won’t, so we would strongly advise only using it for rough work and then verifying with MCMC.
- Better-behaved unit-vector transform; alas, this is broken already due to a dimensionality mismatch and you’ll have to wait for Stan 2.9.1 or Stan 2.10 before the unit_vector type will actually work (it never worked in the past, either—our bad in both the past and now for not having enough tests around it).
We also fixed some minor bugs and cleaned up quite a bit of the code and build process.
We also would like to welcome two new developers: Krzysztof Sakrejda and Aki Vehtari. Aki’s been instrumental in many of our design discussions and Krzysztof’s first major code contribution was sparse matrix multiplication, which leads to our next topic.
We have also released the first version of RStanARM package. The short story on RStanARM is that it’s an MCMC and VB-based replacement for lm() and glm() from core R, and to some extent, lmer() and glmer() from lme4. I believe there’s also a new version of ShinyStan (2.1) available.
We also wrote up a paper on Stan’s reverse-mode automatic differentiation, the cornerstone of the Stan Math Library:
- Bob Carpenter, Matthew D. Hoffman, Marcus Brubaker, Daniel Lee, Peter Li, and Michael Betancourt. 2015. The Stan Math Library: Reverse-Mode Automatic Differentiation in C++. arXiv 1509.07164.
The Stan Development Team
(1) Apologies to those of you who tried to download and install RStan as it was trickling through the CRAN process. The problem is that the managers of CRAN felt a single RStan package was too large (4MB or so) and forced us to import existing packages and break RStan down (BH for the Boost headers, RcppEigen for the Eigen headers, StanHeaders for the Stan header files, and RStan itself for RStan itself). Alas, they provide no foolproof way to synchronize releases. We can insist on a particular version, but R always tries to download the latest or just fails. In the future, we’ll be more proactive and let people know ahead of time when things are in an unsettled state on CRAN and how to install through GitHub. Thanks for your patience.