Thanks for the table
My runtimes are similar. It seems that there are problems with transition with interpolation: with no interpolation it converges but with interpilation it doesn’t (for the Guerrieri-Lorenzoni model).
should I take this as “better to set gridinterplayer=0 for transitions”?
I think so: with no interpolation it converges nicely and the solution looks reasonable. This suggests that there is something unstable/incorrect with the interpolation command.
GL2017 now converges with vfoptions.gridinterplayer=1.
But it is currently very very slow due to the change in convergence criteria for transition paths which implicitly massively increases the required accuracy; if you want to run GL2017 with grid interpolation layer I suggest setting transpathoptions.toleranceGEcondns=1e-3, else you’ll be waiting a while.
You can speed it up with the new ‘additional factors’.
I am hoping to update the GL2017 example soon with some better settings so it solves in reasonable times.
The reworking of the convergence criteria was the necessary first step in overhauling how the transition path updates are done, so hopefully some new faster and better algorithm choices are coming soon. I have completed the large amount of ground work that was required, the actual implementation should go fairly smoothly now.
I also did a lot of work on CoreInfHorzTests to improve the defaults for Howards, and on CoreInfHorzTPathTests to ensure the infinite horizon transition path codes are all working smoothly. Should make them much more reliable going forward.