Gadget3 likelihood bounds_penalty action

Add a liklihood penalty for parameters leaving the bounds set in parameter_template

g3l_bounds_penalty(
        actions_or_parameter_template,
        weight = 1,
        run_at = g3_action_order$likelihood)

Arguments

actions_or_parameter_template

Either:

A list of actions, to extract parameters from and to add bounds to.

A parameter template generated by g3_to_tmb, with optimise, lower, upper populated, bounds for the parameters will be hard-coded.

weight

Weighting applied to this likelihood component.

run_at

Integer order that actions will be run within model, see g3_action_order.

Details

Whilst lower/upper can be passed to optim, not all methods can use them. Adding g3l_bounds_penalty OTOH can be used with any method.

Value

g3l_bounds_penalty

An action (i.e. list of formula objects) that will... If a actions list is supplied, add a large number to likelihood when any parameter is outside bounds. Bounds are updated whenever g3_tmb_adfun is run.

If a parameter_template is supplied, add a large number to likelihood when outside the bounds in the template. The bounds are baked into the model at this point.

Examples

library(magrittr)
anch <- g3_stock('anch', seq(20, 156, 4)) %>% g3s_age(3, 10)
actions <- list(
  g3a_time(1990, 1994),
  g3a_growmature(anch, g3a_grow_impl_bbinom(
    maxlengthgroupgrowth = 38L)),
  g3a_naturalmortality(anch),
  g3a_initialconditions_normalparam(anch),
  g3a_renewal_normalparam(anch,
    run_step = NULL),
  g3a_age(anch),
  NULL)

# Generate code with bounds added
model_code <- g3_to_tmb(c(actions, list(g3l_bounds_penalty(actions))))

attr(model_code, "parameter_template") %>%
  # Set lower / upper bounds for initial conditions
  g3_init_val("*.init.#", 10, lower = 0.001, upper = 200) %>%
  identity() -> params.in

# The objective function produced by g3_tmb_adfun() will honour the bounds
# above, without having to pass them to stats::optim()