Tagging @charlesm93 , @wds15 , @yizhang , and @bbbales2 to start for this question, since I know you guys are working on various things related to the topic. I was wondering about the general state of priorities right now for ODEs in Stan and what present areas of development focus were. Over the past…

For a quick initial response, I’m thinking likely not and that going between libraries might be pretty troublesome. It might be more feasible to write something switching between solvers in Boost? Going to have to look more into this and read up some more.

[image] xiehw: I’m thinking likely not and that going between libraries might be pretty troublesome Right, but we only have to go Boost → CVODES, so it only has to work one way which might make it a simpler proposition.

[image] bbbales2: Right, but we only have to go Boost → CVODES, so it only has to work one way which might make it a simpler proposition. Good point. Based on the algorithms I’ve read about in the Mathematica documentation page and a discussion with Chris, it seems like once the one-way funct…

I think an RK45 -> BDF switch covers more things than an Adams -> BDF switch.

Definitely agree there. That said at this point, seeing more descriptions of RK --> Rosenbrock and Adams --> BDF, so I’m wondering whether those flows canonically work more easily somehow and whether the two-way stiffness detection works more easily for those families. Dumb question – would you hap…

[image] xiehw: how to estimate Jacobian eigenvalues at each step for the Boost algorithms I do not, and it doesn’t seem like a dumb question cuz it seems hard.

[image] bbbales2: I do not, and it doesn’t seem like a dumb question cuz it seems hard. Haha, yeah, felt nervous asking just in case it was an obvious thing, but glad it was not. Yeah, I think that’s the next step to figure out. A lot more to be done, but it looks like things are much more st…

I don’t want to stop anyone from putting work into this… but I really think that adjoint sensitivity analysis is the much more important thing to tackle if we ever want to solve larger systems. Speed problems in smaller ODE systems can be dealt with whenever we talk about a hierarchical problem (usi…

For sure, this is definitely very secondary to the adjoint sensitivity priority, a side project if you will. I can iteratively plug at the switching algorithm with minimal bothering of the devs so that y’all can focus on adjoint sensitivity.

@bbbales2 Having done a little bit more digging, it looks like the Boost-to-CVODE plan won’t be possible because stiffness detection is not included within Boost’s odeint. Also, going between libraries will also be tricky if there are a few of arguments being passed between (for example, in the case…

Stan differential equation functionality development priorities

Developers

Topic		Replies	Views
Adams Moulton ODE solver in Stan useful? General	24	1590	June 16, 2020
Coupled ode system in cmdstan 2.18 Developers	54	2768	February 12, 2019
ODE solver behavior change from Math v3.2.0 to v3.3.0 Developers	40	2171	March 26, 2022
Adjoint ODE Prototype - RFC - Please Test General ode	39	3189	April 7, 2021
Forced ODEs, a start for a case study? Developers	23	3978	May 25, 2017

Stan differential equation functionality development priorities

Related topics