enhance handling of optimization for Laplace

[paciorek]

WIP: do not merge without discussion.

In particular @perrydv see my changes to custom_optim_inner.

Flagging @weizhangstats @paul-vdb @kenkellner in case anyone has thoughts on item 3 below or other thoughts on readiness to merge this into devel.

This makes a few enhancements to the "laplaceRpieces" work:

1. Perry added capability for inner optimization to use inner Hessian to allow use of Newton's methods
2. Wei fixed a small bug in buildOneAGHQuad
3. Chris modified custom_optim_inner to return NaN rather than stopping if inner optimization fails. This allows outer optimization to proceed (e.g., in cases where BFGS tries an outer parameter value that causes inner gradients of NaN/Inf/huge values). He saw that for the CDFW carnivores example, such inner gradients occur (including when using TMB via unmarked), so we need robust handling of such inner optimization failures.

One open item. We have innerOptimWarning (default of FALSE) in our Laplace system to avoid overwhelming user with warnings about problems in the inner optimization. This needs to be extended into Chris' changes to custom_optim_inner presumably. Right now all warnings are being printed.

Also we'll presumably need to change the default inner optimization to nlminb per NCT issue discussion.

[weizhangstats]

The changes look good to me. :) I fixed a wee bug in setupMargNodes in nimbleQuad. Not sure when Laplace will be reallocated to nimbleQuad, so I just fixed it here as well.

[paciorek]

The test failure is because mcmcse is not being installed on the runner, causing test-mcem.R failures. It fails in this case because Rcpp fails because of a compilation issue. Strangely on devel, mcmcse is installed but no mention of installing Rcpp is given, and then there is a later message (after successful mcmcse installation) saying that mcmcse is not available. But then test-mcem.R runs fine.

I'm retriggering the failed test batch in hopes this was a stochastic issue for the runner.

Update: Rerunning resulted in test-mcmcse.R being successful.

[paciorek]

The most recent commit tries to finalize our Laplace interface and documentation.

@perrydv @weizhangstats @paul-vdb please let me know of any thoughts you have.

These include:

• Make nlminb the default for inner/outer optim for Laplace.
• Make last.best default starting point for inner Laplace.
• Silence R error messages in custom_optim_inner so that messages about inner convergence failure only occur if innerOptimWarning = TRUE.
• Update manual on Laplace to highlight runLaplace (mentioning individual Laplace methods second, and then discuss user provided optimizers.
• Clean up messaging in Laplace.R to better follow nimble standard.
• Return logLik and df in summaryLaplace (to allow for model selection)

[weizhangstats]

All looks good to me @paciorek. I just edited the last.best description a bit - to me, it means that the optimised random effects values that correspond to the currently best outer optimisation (i.e. largest marginal log likelihood value) will be used for inner optimisation. The previous wording "best inner optimisation so far" is a bit vague.

[weizhangstats]

@paciorek testADaghq failed here due to a wee tolerance issue but it seems to work ok on my mac.

[paciorek]

@weizhangstats thanks for your recent improvements. I'll figure out the tolerance issue on that last failing test.

[perrydv]

This all looks good to me. Thanks @paciorek