Hi all ODE Stan modelers!
We are currently in the process to add a new adjoint ODE solver to Stan. The benefit of this approach is it’s much better scalability which makes it well suited for large ODE problems. In particular problems with many parameters and a modest number of ODE states (there can still be plenty states, no worries).
The crux of the method is that it’s no free lunch in that the method is rather complex from a numerical perspective. As a result one has to deal with a lot of control parameters for the integration process and we are struggling a bit with these. That is, we are not entirely sure on which control parameter to expose to users and which ones we do not need to expose. Therefore, we would appreciate if Stan users dealing with large ODE problems to try out this new feature and give us feedback on it.
Please head to the wiki page “Adjoint ODE Prototype” which should get you started. It’s as easy as installing an experimental version of cmdstan! To emphasize - this is an experimental cmdstan version and not an official Stan release.
The experimental evaluation phase is planned to last until Easter, April 4th. With the feedback collected we hope to wrap up coding of the feature during April such that we can include the new solver in the next mid May Stan version.
Please let me know if there is anything unclear here or on the wiki instructions page.