obj_fn_plot debug helper worthwhile?

[lentinj]

I've been monkeying with random effects this week, and started using this debugging helper with my old friend curve():

obj_fn_plot <- function (model_code, params, par_name, from = params[par_name[[1]], "lower"], to = params[par_name[[1]], "upper"], ...) {
  stopifnot(par_name %in% rownames(params))
  params$optimise <- FALSE
  params$random <- FALSE
  params[par_name, "optimise"] <- TRUE

  obj.fn <- g3_tmb_adfun(model_code, params, type = c("ADFun", "Fun"))  # TODO: Just "Fun" is failing, maybe because there's only one param?
  fn <- Vectorize(function (x) { params[par_name, "value"] <- list(x) ; obj.fn$fn(g3_tmb_par(params)) })
  curve(fn(x), from = from, to = to, ...)
}

So I can do things like: obj_fn_plot(model_cpp, params.in, "fish_imm.rec.2010").

I'd been suggesting @vbartolino try something similar at one point, but didn't have any code to hand. Do we think something along these lines is worth adding into gadget3? Or at least parked in the debugging vignette?

• I can't help feel that as a debugging technique it's not particularly subtle. Maybe that's a good thing.
• Handing in C++ code rather than an objective function seems a bit ugly.
• What else should it do?
  • Multiple lines for multiple parameters could be interesting, where scales match (I could plot "fish_imm.rec.2010--2020" in the above).
  • A 3d surface for 2 parameters is another obvious thing to do, but I'm not convinced it's worth the faff.
  • Plot the maximum gradient instead of the double value

[vbartolino]

I'd been suggesting @vbartolino try something similar at one point, but didn't have any code to hand. Do we think something along these lines is worth adding into gadget3? Or at least parked in the debugging vignette?

maybe it's not subtle debugging but I think in some cases is useful to profile the likelihood.
Not sure where it belongs if gadget3 or vignette. I've the feeling that it's often easier to find functionalities when part of the package

[bthe]

This is akin to the old plot routine we had for gadget2, see (https://github.com/gadget-framework/rgadget/blob/53eef9e10a64fef96a5eab05460073da7a5aaaa4/R/gadgetfunctions.R#L690). It was mostly used to indicate convergence, but it probably not a bad idea to use this sort of thing to see if your model's maximum likelihood is attained at the boundary, and that there is a strong gradient at that point, suggesting that the boundary and/or model is improperly specified.

And if your likelihood is properly specified then the uncertainty bounds are defined by selecting the range where -2*logLik(x) < 1.96 - 2*logLik(opt).

[lentinj]

I've the feeling that it's often easier to find functionalities when part of the package

Yeah, I agree. I might dump it in g3experiments for the time being though.

the uncertainty bounds are defined by selecting the range where -2*logLik(x) < 1.96 - 2*logLik(opt)

The current "use lower/upper bounds" default is largely useless since the bounds penalty destroys anything of interest, so will try doing something along these lines and see how it goes.