Closed perrydv closed 3 months ago
Updates to this PR:
updateSettings
method.runLaplace
and runAGHQ
Remaining to-dos:
summaryAGHQ
to call `summaryLaplace to match name pairing system (AGHQ and Laplace, equivalent but both there).I cleaned up tests. Two notes:
test-ADaghq
emit lots of Warnings about non-cached values. I haven't investigated. It may be a numerical tolerance issue where this is checked. But the tests still run ok.tetst-ADaghq
, I had to relax some tolerances (but still reasonable) and also set some outer methods to "BFGS" (which they would have been using when the tests were written, before the switch to "nlminb" as default.test-ADlaplace
, I had to put a tolerance of 0.2 (quite large) for one of the standard error results. This is not good.@perrydv I wrote some RTMB code that does Laplace or AGHQ in a very basic way as a reality check. The beta-binomial example in the test takes a lot of nodes to be accurate. It might be helpful as a check that we didn't break anything in this update for some of the other examples (code is commented out on line 271 of test ADaghq) and very easy to setup for other examples.
I like the simplification of the setup code that you've done moving it into its own function. Reading through updateSettings
that looks really good. Much better to have it all in one place as it was getting messy with the number new methods added for such small controls.
To add is a test on 2 dimension AGHQ in test-ADaghq.
merging so can proceed with other steps. We can check testing on devel.
@paul-vdb @weizhangstats @paciorek @danielturek
This is the long-awaited PR with AGHQ from @paul-vdb along with more flexible
nimOptim
and many other changes.A few quick notes:
nlminb
.nlminb
and works withBFGS
.runLaplace
or such idea is not drafted