Trace value of variables in a G3 model

Trace value of variables in a G3 model, warning if conditions are met

g3a_trace_var(
        actions,
        check_finite = TRUE,
        check_positive = FALSE,
        check_strictly_positive = FALSE,
        on_error = c("continue", "browser", "stop"),
        print_var = FALSE,
        var_re = c("__num$", "__wgt$"))

Arguments

actions

A list of model actions to add tracing to

check_finite

Boolean, notify if variable is not finite (i.e. Inf, NA, NaN)

check_positive

Boolean, notify if variable is < 0

check_strictly_positive

Boolean, notify if variable is <= 0

on_error

What to do when a variable fails one of the checks? NB: "browser" will not work in a TMB-compiled model.

print_var

Boolean, if true print the value of the variable at the point the test fails. NB: This will not work in a TMB-compiled model.

var_re

Regular expression(s), variable whose name matches will be traced.

Details

The main reason to use g3a_trace_var is to find out why a model is producing NaN in reports / likelihood. Adding this to your model will help pinpoint the action this originally occurs in, so you can inspect closer for incorrect settings and/or bugs.

Suggested var_re settings

The var_re parameter chooses which variables are traced, and should be tweaked to further pinpoint the problem. Generally, once an error has been found, dig into the code (e.g. by doing edit(g3_to_r(actions))), and see what other variables are available for tracing. Some pre-canned suggestions follow:

c("__num$", "__wgt$") (i.e. default)

This will trace abundance/weight for all stocks, and a good starting point.

^[stock_name]__(num|wgt|cons|suit|totalpredate|consratio|feedinglevel)$

This will, once [stock_name] is replaced with the name of your stock, dig deeper into the predation mechanisms.

See also

g3a_predate_catchability_project

Value

g3_trace_var

A list of actions that will report when variables stop being finite (e.g.)

Examples

stocks <- list(
    st = g3_stock("st", 1:10 * 10) |> g3s_age(1, 5) )

actions <- list(
    g3a_time(1990, 1995, c(3,3,3,3)),
    g3a_initialconditions_normalcv(stocks$st),
    g3a_growmature(stocks$st, impl_f = gadget3::g3a_grow_impl_bbinom(
        maxlengthgroupgrowth = 2L) ),

    NULL )
model_fn <- g3_to_r(c(actions,
    list(g3a_trace_var(actions)),
    g3a_report_detail(actions) ))

# Configure set of working parameters
attr(model_fn, "parameter_template") |>
    g3_init_val("*.K", 0.3) |>
    g3_init_val("*.t0", 0.2) |>
    g3_init_val("*.Linf", 80) |>
    g3_init_val("*.lencv", 0.1) |>
    g3_init_val("*.walpha", 0.01) |>
    g3_init_val("*.wbeta", 3) |>
    g3_init_val("*.M.#", 0.01) |>
    identity() -> params.in
nll <- model_fn(params.in) ; r <- attributes(nll) ; nll <- as.vector(nll)

# Try setting parameters to NaN and see what fails:
r <- model_fn(params.in |> g3_init_val("*.t0", NaN))
#> st__num failed test at 1990-1, after 'g3a_initialconditions for st'
#> st__wgt failed test at 1990-1, after 'g3a_grow for st'
r <- model_fn(params.in |> g3_init_val("*.bbin", NaN))
#> st__num failed test at 1990-1, after 'g3a_grow for st'
#> st__wgt failed test at 1990-1, after 'g3a_grow for st'