Stan meeting notes Nov 1, 2018


#1

Andrew: Worked on Rsquared.

Mitzi: HTMLizing function reference done. Adding JSON support.

Breck: Stan Governing Body work, admin. Did not have quorum at SGB meeting.

Jonah: Revising royal stat society paper, fixining unit tests in Bayes Plot, tweaks to RStanArm for Andrew’s book.

Charles: More working about quadratic optimizer, prototyping optimizer for dense covariates.

Ben G: Working on getting CRAN working, got pull request from Imad on spatial work in RStanArm. Sam Brillman is adding some additional survival models in RStanArm. Have pull req for RStanTools.

Lauren: Got tutorial out.

Sean: More work on lower triangle inverse. Considering adding pointer to eigenmatrix to do some bookkeeping. Thread on discourse a few months back. Jenkins is still choking on big AWS servers.

Bob: Mainly answering tons of email, work on language parsing with Matthijs, Sean.

Yi: Working on updating Torsten with new Stan release. Two weeks to new Torsten release. Working on some new ODE solvers. Working on manual, user guide.

Aki: Taught class in Marseilles, working on new Rhat, updating new case studies. Coming to new your 27 Nov. Dan Simpson will be there.

Matthijs: Working on OCaml parser for Stan3.

Sebastian: Improvements on ODE system, 20% speedup.

============
Sebastian: Considering alternate math lib. Conclusion is that we should just extended. Can we have an optional coupled_ode_system in stan-math which enables an easy way to provide analytic Jacobians for ODE solving (leads to >2x speedups for cases I have seen)? We could allow for this using compile time flags, for example. Consensus is that we are open to specialized math functions as long as they are clearly called out as not being general.

Sebastian: Does it make sense to have ODE integrators provide the solutions of the states with higher relative accuracy vs the gradients? Does it buy sampling performance something (knowing lp more precisely vs the gradients is worthwhile, I think)? Are we up to yet another ODE call?

Bob: We should be running experiments on this a reporting back. Easy project, hugely useful should be a good intro project. He will write it up.

Andrew: Where does research happen? In inferences or C++ directly.