warnings from the engine -- new feature

[stevencarlislewalker]

A discussion at a macpan tech meeting about when to round variables that should be integers (e.g. state variables in some models with process error) led to the idea that we should have a warning system for the engine. The standard Rf_warning mechanism is annoying in TMB because it will get called at each of the (5?) steps of the MakeADFun machinery and for each time step. Here is a design that addresses this issue and makes it straightforward to provide information about what time-step and expression was involved in the warning.

1. Define an enum of distinct warning messages.
2. When a warning is triggered it pushes back one int into each of the following vectors:
   • warn_msg_int : value of the enum for the triggered warning
   • warn_row : parse table row
   • warn_time : time step
3. These vectors are placed in the report.
4. On the r-side we can interpret every element of these vectors:
   • warn_msg_int -> character vector of warning messages
   • warn_row -> vector of expressions as characters (e.g. "x ~ round(0.5)")
   • warn_time -> unaltered vector of time-steps
5. Summarise this 'warning table' so that the time column is a string representing time-step ranges (e.g. "1, 5-10"), which would keep the table from being too long in models with many time-steps.
6. When this interpreted and summarised warning table has more than zero rows, throw a warning that prints the table.

[bbolker]

This seems like a good idea. It's too bad that we're working within C++ instead of R here; in R we could more easily write functions that modify their environments to warn only once in a session (as discussed by you in a recent lab meeting I think?)

I'm trying to think of alternative ways to achieve this goal without writing our own warning system ...

[stevencarlislewalker]

This is a great point. Can we get any inspiration from glmmTMB here?